input stringlengths 0 19.2k ⌀ | context stringlengths 688 257k | answers listlengths 1 182 ⌀ | length int32 258 40.5k ⌀ | dataset stringclasses 16
values | language stringclasses 4
values | all_classes listlengths 50 50 ⌀ | _id stringlengths 48 48 ⌀ | question stringlengths 19 19.3k | answer listlengths 1 182 | answer_prefix stringlengths 0 158 | task stringclasses 29
values | max_new_tokens int64 30 512 | __index_level_0__ int64 0 499 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
How is the ground truth for fake news established? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: 10pt
1.10pt
[ Characterizing Political Fake News in Twitter by its Meta-DataJulio Amador Díaz LópezAxel Oehmichen Miguel Molina-Solana( j.amador, axelfrancois.oehmichen11, mmolinas@imperial.ac.uk ) Imperial College London This article presents a preliminary approach towards characterizing political fake news on Twitter through the analysis of their meta-data. In particular, we focus on more than 1.5M tweets collected on the day of the election of Donald Trump as 45th president of the United States of America. We use the meta-data embedded within those tweets in order to look for differences between tweets containing fake news and tweets not containing them. Specifically, we perform our analysis only on tweets that went viral, by studying proxies for users' exposure to the tweets, by characterizing accounts spreading fake news, and by looking at their polarization. We found significant differences on the distribution of followers, the number of URLs on tweets, and the verification of the users.
]
Introduction
While fake news, understood as deliberately misleading pieces of information, have existed since long ago (e.g. it is not unusual to receive news falsely claiming the death of a celebrity), the term reached the mainstream, particularly so in politics, during the 2016 presidential election in the United States BIBREF0 . Since then, governments and corporations alike (e.g. Google BIBREF1 and Facebook BIBREF2 ) have begun efforts to tackle fake news as they can affect political decisions BIBREF3 . Yet, the ability to define, identify and stop fake news from spreading is limited.
Since the Obama campaign in 2008, social media has been pervasive in the political arena in the United States. Studies report that up to 62% of American adults receive their news from social media BIBREF4 . The wide use of platforms such as Twitter and Facebook has facilitated the diffusion of fake news by simplifying the process of receiving content with no significant third party filtering, fact-checking or editorial judgement. Such characteristics make these platforms suitable means for sharing news that, disguised as legit ones, try to confuse readers.
Such use and their prominent rise has been confirmed by Craig Silverman, a Canadian journalist who is a prominent figure on fake news BIBREF5 : “In the final three months of the US presidential campaign, the top-performing fake election news stories on Facebook generated more engagement than the top stories from major news outlet”.
Our current research hence departs from the assumption that social media is a conduit for fake news and asks the question of whether fake news (as spam was some years ago) can be identified, modelled and eventually blocked. In order to do so, we use a sample of more that 1.5M tweets collected on November 8th 2016 —election day in the United States— with the goal of identifying features that tweets containing fake news are likely to have. As such, our paper aims to provide a preliminary characterization of fake news in Twitter by looking into meta-data embedded in tweets. Considering meta-data as a relevant factor of analysis is in line with findings reported by Morris et al. BIBREF6 . We argue that understanding differences between tweets containing fake news and regular tweets will allow researchers to design mechanisms to block fake news in Twitter.
Specifically, our goals are: 1) compare the characteristics of tweets labelled as containing fake news to tweets labelled as not containing them, 2) characterize, through their meta-data, viral tweets containing fake news and the accounts from which they originated, and 3) determine the extent to which tweets containing fake news expressed polarized political views.
For our study, we used the number of retweets to single-out those that went viral within our sample. Tweets within that subset (viral tweets hereafter) are varied and relate to different topics. We consider that a tweet contains fake news if its text falls within any of the following categories described by Rubin et al. BIBREF7 (see next section for the details of such categories): serious fabrication, large-scale hoaxes, jokes taken at face value, slanted reporting of real facts and stories where the truth is contentious. The dataset BIBREF8 , manually labelled by an expert, has been publicly released and is available to researchers and interested parties.
From our results, the following main observations can be made:
Our findings resonate with similar work done on fake news such as the one from Allcot and Gentzkow BIBREF9 . Therefore, even if our study is a preliminary attempt at characterizing fake news on Twitter using only their meta-data, our results provide external validity to previous research. Moreover, our work not only stresses the importance of using meta-data, but also underscores which parameters may be useful to identify fake news on Twitter.
The rest of the paper is organized as follows. The next section briefly discusses where this work is located within the literature on fake news and contextualizes the type of fake news we are studying. Then, we present our hypotheses, the data, and the methodology we follow. Finally, we present our findings, conclusions of this study, and future lines of work.
Defining Fake news
Our research is connected to different strands of academic knowledge related to the phenomenon of fake news. In relation to Computer Science, a recent survey by Conroy and colleagues BIBREF10 identifies two popular approaches to single-out fake news. On the one hand, the authors pointed to linguistic approaches consisting in using text, its linguistic characteristics and machine learning techniques to automatically flag fake news. On the other, these researchers underscored the use of network approaches, which make use of network characteristics and meta-data, to identify fake news.
With respect to social sciences, efforts from psychology, political science and sociology, have been dedicated to understand why people consume and/or believe misinformation BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 . Most of these studies consistently reported that psychological biases such as priming effects and confirmation bias play an important role in people ability to discern misinformation.
In relation to the production and distribution of fake news, a recent paper in the field of Economics BIBREF9 found that most fake news sites use names that resemble those of legitimate organizations, and that sites supplying fake news tend to be short-lived. These authors also noticed that fake news items are more likely shared than legitimate articles coming from trusted sources, and they tend to exhibit a larger level of polarization.
The conceptual issue of how to define fake news is a serious and unresolved issue. As the focus of our work is not attempting to offer light on this, we will rely on work by other authors to describe what we consider as fake news. In particular, we use the categorization provided by Rubin et al. BIBREF7 . The five categories they described, together with illustrative examples from our dataset, are as follows:
Research Hypotheses
Previous works on the area (presented in the section above) suggest that there may be important determinants for the adoption and diffusion of fake news. Our hypotheses builds on them and identifies three important dimensions that may help distinguishing fake news from legit information:
Taking those three dimensions into account, we propose the following hypotheses about the features that we believe can help to identify tweets containing fake news from those not containing them. They will be later tested over our collected dataset.
Exposure.
Characterization.
Polarization.
Data and Methodology
For this study, we collected publicly available tweets using Twitter's public API. Given the nature of the data, it is important to emphasize that such tweets are subject to Twitter's terms and conditions which indicate that users consent to the collection, transfer, manipulation, storage, and disclosure of data. Therefore, we do not expect ethical, legal, or social implications from the usage of the tweets. Our data was collected using search terms related to the presidential election held in the United States on November 8th 2016. Particularly, we queried Twitter's streaming API, more precisely the filter endpoint of the streaming API, using the following hashtags and user handles: #MyVote2016, #ElectionDay, #electionnight, @realDonaldTrump and @HillaryClinton. The data collection ran for just one day (Nov 8th 2016).
One straightforward way of sharing information on Twitter is by using the retweet functionality, which enables a user to share a exact copy of a tweet with his followers. Among the reasons for retweeting, Body et al. BIBREF15 reported the will to: 1) spread tweets to a new audience, 2) to show one’s role as a listener, and 3) to agree with someone or validate the thoughts of others. As indicated, our initial interest is to characterize tweets containing fake news that went viral (as they are the most harmful ones, as they reach a wider audience), and understand how it differs from other viral tweets (that do not contain fake news). For our study, we consider that a tweet went viral if it was retweeted more than 1000 times.
Once we have the dataset of viral tweets, we eliminated duplicates (some of the tweets were collected several times because they had several handles) and an expert manually inspected the text field within the tweets to label them as containing fake news, or not containing them (according to the characterization presented before). This annotated dataset BIBREF8 is publicly available and can be freely reused.
Finally, we use the following fields within tweets (from the ones returned by Twitter's API) to compare their distributions and look for differences between viral tweets containing fake news and viral tweets not containing fake news:
In the following section, we provide graphical descriptions of the distribution of each of the identified attributes for the two sets of tweets (those labelled as containing fake news and those labelled as not containing them). Where appropriate, we normalized and/or took logarithms of the data for better representation. To gain a better understanding of the significance of those differences, we use the Kolmogorov-Smirnov test with the null hypothesis that both distributions are equal.
Results
The sample collected consisted on 1 785 855 tweets published by 848 196 different users. Within our sample, we identified 1327 tweets that went viral (retweeted more than 1000 times by the 8th of November 2016) produced by 643 users. Such small subset of viral tweets were retweeted on 290 841 occasions in the observed time-window.
The 1327 `viral' tweets were manually annotated as containing fake news or not. The annotation was carried out by a single person in order to obtain a consistent annotation throughout the dataset. Out of those 1327 tweets, we identified 136 as potentially containing fake news (according to the categories previously described), and the rest were classified as `non containing fake news'. Note that the categorization is far from being perfect given the ambiguity of fake news themselves and human judgement involved in the process of categorization. Because of this, we do not claim that this dataset can be considered a ground truth.
The following results detail characteristics of these tweets along the previously mentioned dimensions. Table TABREF23 reports the actual differences (together with their associated p-values) of the distributions of viral tweets containing fake news and viral tweets not containing them for every variable considered.
Exposure
Figure FIGREF24 shows that, in contrast to other kinds of viral tweets, those containing fake news were created more recently. As such, Twitter users were exposed to fake news related to the election for a shorter period of time.
However, in terms of retweets, Figure FIGREF25 shows no apparent difference between containing fake news or not containing them. That is confirmed by the Kolmogorov-Smirnoff test, which does not discard the hypothesis that the associated distributions are equal.
In relation to the number of favourites, users that generated at least a viral tweet containing fake news appear to have, on average, less favourites than users that do not generate them. Figure FIGREF26 shows the distribution of favourites. Despite the apparent visual differences, the difference are not statistically significant.
Finally, the number of hashtags used in viral fake news appears to be larger than those in other viral tweets. Figure FIGREF27 shows the density distribution of the number of hashtags used. However, once again, we were not able to find any statistical difference between the average number of hashtags in a viral tweet and the average number of hashtags in viral fake news.
Characterization
We found that 82 users within our sample were spreading fake news (i.e. they produced at least one tweet which was labelled as fake news). Out of those, 34 had verified accounts, and the rest were unverified. From the 48 unverified accounts, 6 have been suspended by Twitter at the date of writing, 3 tried to imitate legitimate accounts of others, and 4 accounts have been already deleted. Figure FIGREF28 shows the proportion of verified accounts to unverified accounts for viral tweets (containing fake news vs. not containing fake news). From the chart, it is clear that there is a higher chance of fake news coming from unverified accounts.
Turning to friends, accounts distributing fake news appear to have, on average, the same number of friends than those distributing tweets with no fake news. However, the density distribution of friends from the accounts (Figure FIGREF29 ) shows that there is indeed a statistically significant difference in their distributions.
If we take into consideration the number of followers, accounts generating viral tweets with fake news do have a very different distribution on this dimension, compared to those accounts generating viral tweets with no fake news (see Figure FIGREF30 ). In fact, such differences are statistically significant.
A useful representation for friends and followers is the ratio between friends/followers. Figures FIGREF31 and FIGREF32 show this representation. Notice that accounts spreading viral tweets with fake news have, on average, a larger ratio of friends/followers. The distribution of those accounts not generating fake news is more evenly distributed.
With respect to the number of mentions, Figure FIGREF33 shows that viral tweets labelled as containing fake news appear to use mentions to other users less frequently than viral tweets not containing fake news. In other words, tweets containing fake news mostly contain 1 mention, whereas other tweets tend to have two). Such differences are statistically significant.
The analysis (Figure FIGREF34 ) of the presence of media in the tweets in our dataset shows that tweets labelled as not containing fake news appear to present more media elements than those labelled as fake news. However, the difference is not statistically significant.
On the other hand, Figure FIGREF35 shows that viral tweets containing fake news appear to include more URLs to other sites than viral tweets that do not contain fake news. In fact, the difference between the two distributions is statistically significant (assuming INLINEFORM0 ).
Polarization
Finally, manual inspection of the text field of those viral tweets labelled as containing fake news shows that 117 of such tweets expressed support for Donald Trump, while only 8 supported Hillary Clinton. The remaining tweets contained fake news related to other topics, not expressing support for any of the candidates.
Discussion
As a summary, and constrained by our existing dataset, we made the following observations regarding differences between viral tweets labelled as containing fake news and viral tweets labelled as not containing them:
These findings (related to our initial hypothesis in Table TABREF44 ) clearly suggest that there are specific pieces of meta-data about tweets that may allow the identification of fake news. One such parameter is the time of exposure. Viral tweets containing fake news are shorter-lived than those containing other type of content. This notion seems to resonate with our findings showing that a number of accounts spreading fake news have already been deleted or suspended by Twitter by the time of writing. If one considers that researchers using different data have found similar results BIBREF9 , it appears that the lifetime of accounts, together with the age of the questioned viral content could be useful to identify fake news. In the light of this finding, accounts newly created should probably put under higher scrutiny than older ones. This in fact, would be a nice a-priori bias for a Bayesian classifier.
Accounts spreading fake news appear to have a larger proportion of friends/followers (i.e. they have, on average, the same number of friends but a smaller number of followers) than those spreading viral content only. Together with the fact that, on average, tweets containing fake news have more URLs than those spreading viral content, it is possible to hypothesize that, both, the ratio of friends/followers of the account producing a viral tweet and number of URLs contained in such a tweet could be useful to single-out fake news in Twitter. Not only that, but our finding related to the number of URLs is in line with intuitions behind the incentives to create fake news commonly found in the literature BIBREF9 (in particular that of obtaining revenue through click-through advertising).
Finally, it is interesting to notice that the content of viral fake news was highly polarized. This finding is also in line with those of Alcott et al. BIBREF9 . This feature suggests that textual sentiment analysis of the content of tweets (as most researchers do), together with the above mentioned parameters from meta-data, may prove useful for identifying fake news.
Conclusions
With the election of Donald Trump as President of the United States, the concept of fake news has become a broadly-known phenomenon that is getting tremendous attention from governments and media companies. We have presented a preliminary study on the meta-data of a publicly available dataset of tweets that became viral during the day of the 2016 US presidential election. Our aim is to advance the understanding of which features might be characteristic of viral tweets containing fake news in comparison with viral tweets without fake news.
We believe that the only way to automatically identify those deceitful tweets (i.e. containing fake news) is by actually understanding and modelling them. Only then, the automation of the processes of tagging and blocking these tweets can be successfully performed. In the same way that spam was fought, we anticipate fake news will suffer a similar evolution, with social platforms implementing tools to deal with them. With most works so far focusing on the actual content of the tweets, ours is a novel attempt from a different, but also complementary, angle.
Within the used dataset, we found there are differences around exposure, characteristics of accounts spreading fake news and the tone of the content. Those findings suggest that it is indeed possible to model and automatically detect fake news. We plan to replicate and validate our experiments in an extended sample of tweets (until 4 months after the US election), and tests the predictive power of the features we found relevant within our sample.
Author Disclosure Statement
No competing financial interest exist.
| [
"Ground truth is not established in the paper"
] | 3,141 | qasper | en | null | 3ac3eef636db11635a21a61804cb28e92c546a5686dd1e12 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How is the ground truth for fake news established?
Answer: | [
"Ground truth is not established in the paper"
] | qasper | 128 | 0 | |
What is the GhostVLAD approach? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: INTRODUCTION
The idea of language identification is to classify a given audio signal into a particular class using a classification algorithm. Commonly language identification task was done using i-vector systems [1]. A very well known approach for language identification proposed by N. Dahek et al. [1] uses the GMM-UBM model to obtain utterance level features called i-vectors. Recent advances in deep learning [15,16] have helped to improve the language identification task using many different neural network architectures which can be trained efficiently using GPUs for large scale datasets. These neural networks can be configured in various ways to obtain better accuracy for language identification task. Early work on using Deep learning for language Identification was published by Pavel Matejka et al. [2], where they used stacked bottleneck features extracted from deep neural networks for language identification task and showed that the bottleneck features learned by Deep neural networks are better than simple MFCC or PLP features. Later the work by I. Lopez-Moreno et al. [3] from Google showed how to use Deep neural networks to directly map the sequence of MFCC frames into its language class so that we can apply language identification at the frame level. Speech signals will have both spatial and temporal information, but simple DNNs are not able to capture temporal information. Work done by J. Gonzalez-Dominguez et al. [4] by Google developed an LSTM based language identification model which improves the accuracy over the DNN based models. Work done by Alicia et al. [5] used CNNs to improve upon i-vector [1] and other previously developed systems. The work done by Daniel Garcia-Romero et al. [6] has used a combination of Acoustic model trained for speech recognition with Time-delay neural networks where they train the TDNN model by feeding the stacked bottleneck features from acoustic model to predict the language labels at the frame level. Recently X-vectors [7] is proposed for speaker identification task and are shown to outperform all the previous state of the art speaker identification algorithms and are also used for language identification by David Snyder et al. [8].
In this paper, we explore multiple pooling strategies for language identification task. Mainly we propose Ghost-VLAD based pooling method for language identification. Inspired by the recent work by W. Xie et al. [9] and Y. Zhong et al. [10], we use Ghost-VLAD to improve the accuracy of language identification task for Indian languages. We explore multiple pooling strategies including NetVLAD pooling [11], Average pooling and Statistics pooling( as proposed in X-vectors [7]) and show that Ghost-VLAD pooling is the best pooling strategy for language identification. Our model obtains the best accuracy of 98.24%, and it outperforms all the other previously proposed pooling methods. We conduct all our experiments on 635hrs of audio data for 7 Indian languages collected from $\textbf {All India Radio}$ news channel. The paper is organized as follows. In section 2, we explain the proposed pooling method for language identification. In section 3, we explain our dataset. In section 4, we describe the experiments, and in section 5, we describe the results.
POOLING STRATEGIES
In any language identification model, we want to obtain utterance level representation which has very good language discriminative features. These representations should be compact and should be easily separable by a linear classifier. The idea of any pooling strategy is to pool the frame-level representations into a single utterance level representation. Previous works by [7] have used simple mean and standard deviation aggregation to pool the frame-level features from the top layer of the neural network to obtain the utterance level features. Recently [9] used VLAD based pooling strategy for speaker identification which is inspired from [10] proposed for face recognition. The NetVLAD [11] and Ghost-VLAD [10] methods are proposed for Place recognition and face recognition, respectively, and in both cases, they try to aggregate the local descriptors into global features. In our case, the local descriptors are features extracted from ResNet [15], and the global utterance level feature is obtained by using GhostVLAD pooling. In this section, we explain different pooling methods, including NetVLAD, Ghost-VLAD, Statistic pooling, and Average pooling.
POOLING STRATEGIES ::: NetVLAD pooling
The NetVLAD pooling strategy was initially developed for place recognition by R. Arandjelovic et al. [11]. The NetVLAD is an extension to VLAD [18] approach where they were able to replace the hard assignment based clustering with soft assignment based clustering so that it can be trained with neural network in an end to end fashion. In our case, we use the NetVLAD layer to map N local features of dimension D into a fixed dimensional vector, as shown in Figure 1 (Left side).
The model takes spectrogram as an input and feeds into CNN based ResNet architecture. The ResNet is used to map the spectrogram into 3D feature map of dimension HxWxD. We convert this 3D feature map into 2D by unfolding H and W dimensions, creating a NxD dimensional feature map, where N=HxW. The NetVLAD layer is kept on top of the feature extraction layer of ResNet, as shown in Figure 1. The NetVLAD now takes N features vectors of dimension D and computes a matrix V of dimension KxD, where K is the number clusters in the NetVLAD layer, and D is the dimension of the feature vector. The matrix V is computed as follows.
Where $w_k$,$b_k$ and $c_k$ are trainable parameters for the cluster $k$ and V(j,k) represents a point in the V matrix for (j,k)th location. The matrix is constructed using the equation (1) where the first term corresponds to the soft assignment of the input $x_i$ to the cluster $c_k$, whereas the second term corresponds to the residual term which tells how far the input descriptor $x_i$ is from the cluster center $c_k$.
POOLING STRATEGIES ::: GhostVLAD pooling
GhostVLAD is an extension of the NetVLAD approach, which we discussed in the previous section. The GhostVLAD model was proposed for face recognition by Y. Zhong [10]. GhostVLAD works exactly similar to NetVLAD except it adds Ghost clusters along with the NetVLAD clusters. So, now we will have a K+G number of clusters instead of K clusters. Where G is the number of ghost clusters, we want to add (typically 2-4). The Ghost clusters are added to map any noisy or irrelevant content into ghost clusters and are not included during the feature aggregation stage, as shown in Figure 1 (Right side). Which means that we compute the matrix V for both normal cluster K and ghost clusters G, but we will not include the vectors belongs to ghost cluster from V during concatenation of the features. Due to which, during feature aggregation stage the contribution of the noisy and unwanted features to normal VLAD clusters are assigned less weights while Ghost clusters absorb most of the weight. We illustrate this in Figure 1(Right Side), where the ghost clusters are shown in red color. We use Ghost clusters when we are computing the V matrix, but they are excluded during the concatenation stage. These concatenated features are fed into the projection layer, followed by softmax to predict the language label.
POOLING STRATEGIES ::: Statistic and average pooling
In statistic pooling, we compute the first and second order statistics of the local features from the top layer of the ResNet model. The 3-D feature map is unfolded to create N features of D dimensions, and then we compute the mean and standard deviation of all these N vectors and get two D dimensional vectors, one for mean and the other for standard deviation. We then concatenate these 2 features and feed it to the projection layer for predicting the language label.
In the Average pooling layer, we compute only the first-order statistics (mean) of the local features from the top layer of the CNN model. The feature map from the top layer of CNN is unfolded to create N features of D dimensions, and then we compute the mean of all these N vectors and get D dimensional representation. We then feed this feature to the projection layer followed by softmax for predicting the language label.
DATASET
In this section, we describe our dataset collection process. We collected and curated around 635Hrs of audio data for 7 Indian languages, namely Kannada, Hindi, Telugu, Malayalam, Bengali, and English. We collected the data from the All India Radio news channel where an actor will be reading news for about 5-10 mins. To cover many speakers for the dataset, we crawled data from 2010 to 2019. Since the audio is very long to train any deep neural network directly, we segment the audio clips into smaller chunks using Voice activity detector. Since the audio clips will have music embedded during the news, we use Inhouse music detection model to remove the music segments from the dataset to make the dataset clean and our dataset contains 635Hrs of clean audio which is divided into 520Hrs of training data containing 165K utterances and 115Hrs of testing data containing 35K utterances. The amount of audio data for training and testing for each of the language is shown in the table bellow.
EXPERIMENTS
In this section, we describe the feature extraction process and network architecture in detail. We use spectral features of 256 dimensions computed using 512 point FFT for every frame, and we add an energy feature for every frame giving us total 257 features for every frame. We use a window size of 25ms and frame shift of 10ms during feature computation. We crop random 5sec audio data from each utterance during training which results in a spectrogram of size 257x500 (features x number of features). We use these spectrograms as input to our CNN model during training. During testing, we compute the prediction score irrespective of the audio length.
For the network architecture, we use ResNet-34 architecture, as described in [9]. The model uses convolution layers with Relu activations to map the spectrogram of size 257x500 input into 3D feature map of size 1x32x512. This feature cube is converted into 2D feature map of dimension 32x512 and fed into Ghost-VLAD/NetVLAD layer to generate a representation that has more language discrimination capacity. We use Adam optimizer with an initial learning rate of 0.01 and a final learning rate of 0.00001 for training. Each model is trained for 15 epochs with early stopping criteria.
For the baseline, we train an i-vector model using GMM-UBM. We fit a small classifier on top of the generated i-vectors to measure the accuracy. This model is referred as i-vector+svm . To compare our model with the previous state of the art system, we set up the x-vector language identification system [8]. The x-vector model used time-delay neural networks (TDNN) along with statistic-pooling. We use 7 layer TDNN architecture similar to [8] for training. We refer to this model as tdnn+stat-pool . Finally, we set up a Deep LSTM based language identification system similar to [4] but with little modification where we add statistics pooling for the last layers hidden activities before classification. We use 3 layer Bi-LSTM with 256 hidden units at each layer. We refer to this model as LSTM+stat-pool. We train our i-vector+svm and TDNN+stat-pool using Kaldi toolkit. We train our NetVLAD and GhostVLAD experiments using Keras by modifying the code given by [9] for language identification. We train the LSTM+stat-pool and the remaining experiments using Pytorch [14] toolkit, and we will opensource all the codes and data soon.
RESULTS
In this section, we compare the performance of our system with the recent state of the art language identification approaches. We also compare different pooling strategies and finally, compare the robustness of our system to the length of the input spectrogram during training. We visualize the embeddings generated by the GhostVLAD method and conclude that the GhostVLAD embeddings shows very good feature discrimination capabilities.
RESULTS ::: Comparison with different approaches
We compare our system performance with the previous state of the art language identification approaches, as shown in Table 2. The i-vector+svm system is trained using GMM-UBM models to generate i-vectors as proposed in [1]. Once the i-vectors are extracted, we fit SVM classifier to classify the audio. The TDNN+stat-pool system is trained with a statistics pooling layer and is called the x-vector system as proposed by David Snyder et al. [11] and is currently the state of the art language identification approach as far as our knowledge. Our methods outperform the state of the art x-vector system by absolute 1.88% improvement in F1-score, as shown in Table 2.
RESULTS ::: Comparison with different pooling techniques
We compare our approach with different pooling strategies in Table 3. We use ResNet as our base feature extraction network. We keep the base network the same and change only the pooling layers to see which pooling approach performs better for language identification task. Our experiments show that GhostVLAD pooling outperforms all the other pooling methods by achieving 98.43% F1-Score.
RESULTS ::: Duration analysis
To observe the performance of our method with different input durations, we conducted an experiment where we train our model on different input durations. Since our model uses ResNet as the base feature extractor, we need to feed fixed-length spectrogram. We conducted 4 different experiments where we trained the model using 2sec, 3sec, 4sec and 5sec spectrograms containing 200,300,400 and 500 frames respectively. We observed that the model trained with a 5sec spectrogram is the best model, as shown in Table 4.
RESULTS ::: Visualization of embeddings
We visualize the embeddings generated by our approach to see the effectiveness. We extracted 512-dimensional embeddings for our testing data and reduced the dimensionality using t-sne projection. The t-sne plot of the embeddings space is shown in Figure 3. The plot shows that the embeddings learned by our approach has very good discriminative properties
Conclusion
In this work, we use Ghost-VLAD pooling approach that was originally proposed for face recognition to improve language identification performance for Indian languages. We collected and curated 630 hrs audio data from news All India Radio news channel for 7 Indian languages. Our experimental results shows that our approach outperforms the previous state of the art methods by an absolute 1.88% F1-score. We have also conducted experiments with different pooling strategies proposed in the past, and the GhostVLAD pooling approach turns out to be the best approach for aggregating frame-level features into a single utterance level feature. Our experiments also prove that our approach works much better even if the input during training contains smaller durations. Finally, we see that the embeddings generated by our method has very good language discriminative features and helps to improve the performance of language identification.
| [
"extension of the NetVLAD, adds Ghost clusters along with the NetVLAD clusters",
"An extension of NetVLAD which replaces hard assignment-based clustering with soft assignment-based clustering with the additon o fusing Ghost clusters to deal with noisy content."
] | 2,454 | qasper | en | null | db6c4d611b78c840972b57ee0b245003566b982feaff1084 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What is the GhostVLAD approach?
Answer: | [
"extension of the NetVLAD, adds Ghost clusters along with the NetVLAD clusters",
"An extension of NetVLAD which replaces hard assignment-based clustering with soft assignment-based clustering with the additon o fusing Ghost clusters to deal with noisy content."
] | qasper | 128 | 1 | |
By how much does their model outperform the state of the art results? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Recently, deep learning algorithms have successfully addressed problems in various fields, such as image classification, machine translation, speech recognition, text-to-speech generation and other machine learning related areas BIBREF0 , BIBREF1 , BIBREF2 . Similarly, substantial improvements in performance have been obtained when deep learning algorithms have been applied to statistical speech processing BIBREF3 . These fundamental improvements have led researchers to investigate additional topics related to human nature, which have long been objects of study. One such topic involves understanding human emotions and reflecting it through machine intelligence, such as emotional dialogue models BIBREF4 , BIBREF5 .
In developing emotionally aware intelligence, the very first step is building robust emotion classifiers that display good performance regardless of the application; this outcome is considered to be one of the fundamental research goals in affective computing BIBREF6 . In particular, the speech emotion recognition task is one of the most important problems in the field of paralinguistics. This field has recently broadened its applications, as it is a crucial factor in optimal human-computer interactions, including dialog systems. The goal of speech emotion recognition is to predict the emotional content of speech and to classify speech according to one of several labels (i.e., happy, sad, neutral, and angry). Various types of deep learning methods have been applied to increase the performance of emotion classifiers; however, this task is still considered to be challenging for several reasons. First, insufficient data for training complex neural network-based models are available, due to the costs associated with human involvement. Second, the characteristics of emotions must be learned from low-level speech signals. Feature-based models display limited skills when applied to this problem.
To overcome these limitations, we propose a model that uses high-level text transcription, as well as low-level audio signals, to utilize the information contained within low-resource datasets to a greater degree. Given recent improvements in automatic speech recognition (ASR) technology BIBREF7 , BIBREF2 , BIBREF8 , BIBREF9 , speech transcription can be carried out using audio signals with considerable skill. The emotional content of speech is clearly indicated by the emotion words contained in a sentence BIBREF10 , such as “lovely” and “awesome,” which carry strong emotions compared to generic (non-emotion) words, such as “person” and “day.” Thus, we hypothesize that the speech emotion recognition model will be benefit from the incorporation of high-level textual input.
In this paper, we propose a novel deep dual recurrent encoder model that simultaneously utilizes audio and text data in recognizing emotions from speech. Extensive experiments are conducted to investigate the efficacy and properties of the proposed model. Our proposed model outperforms previous state-of-the-art methods by 68.8% to 71.8% when applied to the IEMOCAP dataset, which is one of the most well-studied datasets. Based on an error analysis of the models, we show that our proposed model accurately identifies emotion classes. Moreover, the neutral class misclassification bias frequently exhibited by previous models, which focus on audio features, is less pronounced in our model.
Related work
Classical machine learning algorithms, such as hidden Markov models (HMMs), support vector machines (SVMs), and decision tree-based methods, have been employed in speech emotion recognition problems BIBREF11 , BIBREF12 , BIBREF13 . Recently, researchers have proposed various neural network-based architectures to improve the performance of speech emotion recognition. An initial study utilized deep neural networks (DNNs) to extract high-level features from raw audio data and demonstrated its effectiveness in speech emotion recognition BIBREF14 . With the advancement of deep learning methods, more complex neural-based architectures have been proposed. Convolutional neural network (CNN)-based models have been trained on information derived from raw audio signals using spectrograms or audio features such as Mel-frequency cepstral coefficients (MFCCs) and low-level descriptors (LLDs) BIBREF15 , BIBREF16 , BIBREF17 . These neural network-based models are combined to produce higher-complexity models BIBREF18 , BIBREF19 , and these models achieved the best-recorded performance when applied to the IEMOCAP dataset.
Another line of research has focused on adopting variant machine learning techniques combined with neural network-based models. One researcher utilized the multiobject learning approach and used gender and naturalness as auxiliary tasks so that the neural network-based model learned more features from a given dataset BIBREF20 . Another researcher investigated transfer learning methods, leveraging external data from related domains BIBREF21 .
As emotional dialogue is composed of sound and spoken content, researchers have also investigated the combination of acoustic features and language information, built belief network-based methods of identifying emotional key phrases, and assessed the emotional salience of verbal cues from both phoneme sequences and words BIBREF22 , BIBREF23 . However, none of these studies have utilized information from speech signals and text sequences simultaneously in an end-to-end learning neural network-based model to classify emotions.
Model
This section describes the methodologies that are applied to the speech emotion recognition task. We start by introducing the recurrent encoder model for the audio and text modalities individually. We then propose a multimodal approach that encodes both audio and textual information simultaneously via a dual recurrent encoder.
Audio Recurrent Encoder (ARE)
Motivated by the architecture used in BIBREF24 , BIBREF25 , we build an audio recurrent encoder (ARE) to predict the class of a given audio signal. Once MFCC features have been extracted from an audio signal, a subset of the sequential features is fed into the RNN (i.e., gated recurrent units (GRUs)), which leads to the formation of the network's internal hidden state INLINEFORM0 to model the time series patterns. This internal hidden state is updated at each time step with the input data INLINEFORM1 and the hidden state of the previous time step INLINEFORM2 as follows: DISPLAYFORM0
where INLINEFORM0 is the RNN function with weight parameter INLINEFORM1 , INLINEFORM2 represents the hidden state at t- INLINEFORM3 time step, and INLINEFORM4 represents the t- INLINEFORM5 MFCC features in INLINEFORM6 . After encoding the audio signal INLINEFORM7 with the RNN, the last hidden state of the RNN, INLINEFORM8 , is considered to be the representative vector that contains all of the sequential audio data. This vector is then concatenated with another prosodic feature vector, INLINEFORM9 , to generate a more informative vector representation of the signal, INLINEFORM10 . The MFCC and the prosodic features are extracted from the audio signal using the openSMILE toolkit BIBREF26 , INLINEFORM11 , respectively. Finally, the emotion class is predicted by applying the softmax function to the vector INLINEFORM12 . For a given audio sample INLINEFORM13 , we assume that INLINEFORM14 is the true label vector, which contains all zeros but contains a one at the correct class, and INLINEFORM15 is the predicted probability distribution from the softmax layer. The training objective then takes the following form: DISPLAYFORM0
where INLINEFORM0 is the calculated representative vector of the audio signal with dimensionality INLINEFORM1 . The INLINEFORM2 and the bias INLINEFORM3 are learned model parameters. C is the total number of classes, and N is the total number of samples used in training. The upper part of Figure shows the architecture of the ARE model.
Text Recurrent Encoder (TRE)
We assume that speech transcripts can be extracted from audio signals with high accuracy, given the advancement of ASR technologies BIBREF7 . We attempt to use the processed textual information as another modality in predicting the emotion class of a given signal. To use textual information, a speech transcript is tokenized and indexed into a sequence of tokens using the Natural Language Toolkit (NLTK) BIBREF27 . Each token is then passed through a word-embedding layer that converts a word index to a corresponding 300-dimensional vector that contains additional contextual meaning between words. The sequence of embedded tokens is fed into a text recurrent encoder (TRE) in such a way that the audio MFCC features are encoded using the ARE represented by equation EQREF2 . In this case, INLINEFORM0 is the t- INLINEFORM1 embedded token from the text input. Finally, the emotion class is predicted from the last hidden state of the text-RNN using the softmax function.
We use the same training objective as the ARE model, and the predicted probability distribution for the target class is as follows: DISPLAYFORM0
where INLINEFORM0 is last hidden state of the text-RNN, INLINEFORM1 , and the INLINEFORM2 and bias INLINEFORM3 are learned model parameters. The lower part of Figure indicates the architecture of the TRE model.
Multimodal Dual Recurrent Encoder (MDRE)
We present a novel architecture called the multimodal dual recurrent encoder (MDRE) to overcome the limitations of existing approaches. In this study, we consider multiple modalities, such as MFCC features, prosodic features and transcripts, which contain sequential audio information, statistical audio information and textual information, respectively. These types of data are the same as those used in the ARE and TRE cases. The MDRE model employs two RNNs to encode data from the audio signal and textual inputs independently. The audio-RNN encodes MFCC features from the audio signal using equation EQREF2 . The last hidden state of the audio-RNN is concatenated with the prosodic features to form the final vector representation INLINEFORM0 , and this vector is then passed through a fully connected neural network layer to form the audio encoding vector A. On the other hand, the text-RNN encodes the word sequence of the transcript using equation EQREF2 . The final hidden states of the text-RNN are also passed through another fully connected neural network layer to form a textual encoding vector T. Finally, the emotion class is predicted by applying the softmax function to the concatenation of the vectors A and T. We use the same training objective as the ARE model, and the predicted probability distribution for the target class is as follows: DISPLAYFORM0
where INLINEFORM0 is the feed-forward neural network with weight parameter INLINEFORM1 , and INLINEFORM2 , INLINEFORM3 are final encoding vectors from the audio-RNN and text-RNN, respectively. INLINEFORM4 and the bias INLINEFORM5 are learned model parameters.
Multimodal Dual Recurrent Encoder with Attention (MDREA)
Inspired by the concept of the attention mechanism used in neural machine translation BIBREF28 , we propose a novel multimodal attention method to focus on the specific parts of a transcript that contain strong emotional information, conditioning on the audio information. Figure shows the architecture of the MDREA model. First, the audio data and text data are encoded with the audio-RNN and text-RNN using equation EQREF2 . We then consider the final audio encoding vector INLINEFORM0 as a context vector. As seen in equation EQREF9 , during each time step t, the dot product between the context vector e and the hidden state of the text-RNN at each t-th sequence INLINEFORM1 is evaluated to calculate a similarity score INLINEFORM2 . Using this score INLINEFORM3 as a weight parameter, the weighted sum of the sequences of the hidden state of the text-RNN, INLINEFORM4 , is calculated to generate an attention-application vector Z. This attention-application vector is concatenated with the final encoding vector of the audio-RNN INLINEFORM5 (equation EQREF7 ), which will be passed through the softmax function to predict the emotion class. We use the same training objective as the ARE model, and the predicted probability distribution for the target class is as follows: DISPLAYFORM0
where INLINEFORM0 and the bias INLINEFORM1 are learned model parameters.
Dataset
We evaluate our model using the Interactive Emotional Dyadic Motion Capture (IEMOCAP) BIBREF18 dataset. This dataset was collected following theatrical theory in order to simulate natural dyadic interactions between actors. We use categorical evaluations with majority agreement. We use only four emotional categories happy, sad, angry, and neutral to compare the performance of our model with other research using the same categories. The IEMOCAP dataset includes five sessions, and each session contains utterances from two speakers (one male and one female). This data collection process resulted in 10 unique speakers. For consistent comparison with previous work, we merge the excitement dataset with the happiness dataset. The final dataset contains a total of 5531 utterances (1636 happy, 1084 sad, 1103 angry, 1708 neutral).
Feature extraction
To extract speech information from audio signals, we use MFCC values, which are widely used in analyzing audio signals. The MFCC feature set contains a total of 39 features, which include 12 MFCC parameters (1-12) from the 26 Mel-frequency bands and log-energy parameters, 13 delta and 13 acceleration coefficients The frame size is set to 25 ms at a rate of 10 ms with the Hamming function. According to the length of each wave file, the sequential step of the MFCC features is varied. To extract additional information from the data, we also use prosodic features, which show effectiveness in affective computing. The prosodic features are composed of 35 features, which include the F0 frequency, the voicing probability, and the loudness contours. All of these MFCC and prosodic features are extracted from the data using the OpenSMILE toolkit BIBREF26 .
Implementation details
Among the variants of the RNN function, we use GRUs as they yield comparable performance to that of the LSTM and include a smaller number of weight parameters BIBREF29 . We use a max encoder step of 750 for the audio input, based on the implementation choices presented in BIBREF30 and 128 for the text input because it covers the maximum length of the transcripts. The vocabulary size of the dataset is 3,747, including the “_UNK_" token, which represents unknown words, and the “_PAD_" token, which is used to indicate padding information added while preparing mini-batch data. The number of hidden units and the number of layers in the RNN for each model (ARE, TRE, MDRE and MDREA) are selected based on extensive hyperparameter search experiments. The weights of the hidden units are initialized using orthogonal weights BIBREF31 ], and the text embedding layer is initialized from pretrained word-embedding vectors BIBREF32 .
In preparing the textual dataset, we first use the released transcripts of the IEMOCAP dataset for simplicity. To investigate the practical performance, we then process all of the IEMOCAP audio data using an ASR system (the Google Cloud Speech API) and retrieve the transcripts. The performance of the Google ASR system is reflected by its word error rate (WER) of 5.53%.
Performance evaluation
As the dataset is not explicitly split beforehand into training, development, and testing sets, we perform 5-fold cross validation to determine the overall performance of the model. The data in each fold are split into training, development, and testing datasets (8:0.5:1.5, respectively). After training the model, we measure the weighted average precision (WAP) over the 5-fold dataset. We train and evaluate the model 10 times per fold, and the model performance is assessed in terms of the mean score and standard deviation.
We examine the WAP values, which are shown in Table 1. First, our ARE model shows the baseline performance because we use minimal audio features, such as the MFCC and prosodic features with simple architectures. On the other hand, the TRE model shows higher performance gain compared to the ARE. From this result, we note that textual data are informative in emotion prediction tasks, and the recurrent encoder model is effective in understanding these types of sequential data. Second, the newly proposed model, MDRE, shows a substantial performance gain. It thus achieves the state-of-the-art performance with a WAP value of 0.718. This result shows that multimodal information is a key factor in affective computing. Lastly, the attention model, MDREA, also outperforms the best existing research results (WAP 0.690 to 0.688) BIBREF19 . However, the MDREA model does not match the performance of the MDRE model, even though it utilizes a more complex architecture. We believe that this result arises because insufficient data are available to properly determine the complex model parameters in the MDREA model. Moreover, we presume that this model will show better performance when the audio signals are aligned with the textual sequence while applying the attention mechanism. We leave the implementation of this point as a future research direction.
To investigate the practical performance of the proposed models, we conduct further experiments with the ASR-processed transcript data (see “-ASR” models in Table ). The label accuracy of the processed transcripts is 5.53% WER. The TRE-ASR, MDRE-ASR and MDREA-ASR models reflect degraded performance compared to that of the TRE, MDRE and MDREA models. However, the performance of these models is still competitive; in particular, the MDRE-ASR model outperforms the previous best-performing model, 3CNN-LSTM10H (WAP 0.691 to 0.688).
Error analysis
We analyze the predictions of the ARE, TRE, and MDRE models. Figure shows the confusion matrix of each model. The ARE model (Fig. ) incorrectly classifies most instances of happy as neutral (43.51%); thus, it shows reduced accuracy (35.15%) in predicting the the happy class. Overall, most of the emotion classes are frequently confused with the neutral class. This observation is in line with the findings of BIBREF30 , who noted that the neutral class is located in the center of the activation-valence space, complicating its discrimination from the other classes.
Interestingly, the TRE model (Fig. ) shows greater prediction gains in predicting the happy class when compared to the ARE model (35.15% to 75.73%). This result seems plausible because the model can benefit from the differences among the distributions of words in happy and neutral expressions, which gives more emotional information to the model than that of the audio signal data. On the other hand, it is striking that the TRE model incorrectly predicts instances of the sad class as the happy class 16.20% of the time, even though these emotional states are opposites of one another.
The MDRE model (Fig. ) compensates for the weaknesses of the previous two models (ARE and TRE) and benefits from their strengths to a surprising degree. The values arranged along the diagonal axis show that all of the accuracies of the correctly predicted class have increased. Furthermore, the occurrence of the incorrect “sad-to-happy" cases in the TRE model is reduced from 16.20% to 9.15%.
Conclusions
In this paper, we propose a novel multimodal dual recurrent encoder model that simultaneously utilizes text data, as well as audio signals, to permit the better understanding of speech data. Our model encodes the information from audio and text sequences using dual RNNs and then combines the information from these sources using a feed-forward neural model to predict the emotion class. Extensive experiments show that our proposed model outperforms other state-of-the-art methods in classifying the four emotion categories, and accuracies ranging from 68.8% to 71.8% are obtained when the model is applied to the IEMOCAP dataset. In particular, it resolves the issue in which predictions frequently incorrectly yield the neutral class, as occurs in previous models that focus on audio features.
In the future work, we aim to extend the modalities to audio, text and video inputs. Furthermore, we plan to investigate the application of the attention mechanism to data derived from multiple modalities. This approach seems likely to uncover enhanced learning schemes that will increase performance in both speech emotion recognition and other multimodal classification tasks.
Acknowledgments
K. Jung is with the Department of Electrical and Computer Engineering, ASRI, Seoul National University, Seoul, Korea. This work was supported by the Ministry of Trade, Industry & Energy (MOTIE, Korea) under Industrial Technology Innovation Program (No.10073144).
| [
"the attention model, MDREA, also outperforms the best existing research results (WAP 0.690 to 0.688)"
] | 3,207 | qasper | en | null | e419f2bff9d2ab7c3b60b3250caccd2d9ae1285ec3e8e818 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: By how much does their model outperform the state of the art results?
Answer: | [
"the attention model, MDREA, also outperforms the best existing research results (WAP 0.690 to 0.688)"
] | qasper | 128 | 2 | |
What additional features and context are proposed? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Abusive language refers to any type of insult, vulgarity, or profanity that debases the target; it also can be anything that causes aggravation BIBREF0 , BIBREF1 . Abusive language is often reframed as, but not limited to, offensive language BIBREF2 , cyberbullying BIBREF3 , othering language BIBREF4 , and hate speech BIBREF5 .
Recently, an increasing number of users have been subjected to harassment, or have witnessed offensive behaviors online BIBREF6 . Major social media companies (i.e. Facebook, Twitter) have utilized multiple resources—artificial intelligence, human reviewers, user reporting processes, etc.—in effort to censor offensive language, yet it seems nearly impossible to successfully resolve the issue BIBREF7 , BIBREF8 .
The major reason of the failure in abusive language detection comes from its subjectivity and context-dependent characteristics BIBREF9 . For instance, a message can be regarded as harmless on its own, but when taking previous threads into account it may be seen as abusive, and vice versa. This aspect makes detecting abusive language extremely laborious even for human annotators; therefore it is difficult to build a large and reliable dataset BIBREF10 .
Previously, datasets openly available in abusive language detection research on Twitter ranged from 10K to 35K in size BIBREF9 , BIBREF11 . This quantity is not sufficient to train the significant number of parameters in deep learning models. Due to this reason, these datasets have been mainly studied by traditional machine learning methods. Most recently, Founta et al. founta2018large introduced Hate and Abusive Speech on Twitter, a dataset containing 100K tweets with cross-validated labels. Although this corpus has great potential in training deep models with its significant size, there are no baseline reports to date.
This paper investigates the efficacy of different learning models in detecting abusive language. We compare accuracy using the most frequently studied machine learning classifiers as well as recent neural network models. Reliable baseline results are presented with the first comparative study on this dataset. Additionally, we demonstrate the effect of different features and variants, and describe the possibility for further improvements with the use of ensemble models.
Related Work
The research community introduced various approaches on abusive language detection. Razavi et al. razavi2010offensive applied Naïve Bayes, and Warner and Hirschberg warner2012detecting used Support Vector Machine (SVM), both with word-level features to classify offensive language. Xiang et al. xiang2012detecting generated topic distributions with Latent Dirichlet Allocation BIBREF12 , also using word-level features in order to classify offensive tweets.
More recently, distributed word representations and neural network models have been widely applied for abusive language detection. Djuric et al. djuric2015hate used the Continuous Bag Of Words model with paragraph2vec algorithm BIBREF13 to more accurately detect hate speech than that of the plain Bag Of Words models. Badjatiya et al. badjatiya2017deep implemented Gradient Boosted Decision Trees classifiers using word representations trained by deep learning models. Other researchers have investigated character-level representations and their effectiveness compared to word-level representations BIBREF14 , BIBREF15 .
As traditional machine learning methods have relied on feature engineering, (i.e. n-grams, POS tags, user information) BIBREF1 , researchers have proposed neural-based models with the advent of larger datasets. Convolutional Neural Networks and Recurrent Neural Networks have been applied to detect abusive language, and they have outperformed traditional machine learning classifiers such as Logistic Regression and SVM BIBREF15 , BIBREF16 . However, there are no studies investigating the efficiency of neural models with large-scale datasets over 100K.
Methodology
This section illustrates our implementations on traditional machine learning classifiers and neural network based models in detail. Furthermore, we describe additional features and variant models investigated.
Traditional Machine Learning Models
We implement five feature engineering based machine learning classifiers that are most often used for abusive language detection. In data preprocessing, text sequences are converted into Bag Of Words (BOW) representations, and normalized with Term Frequency-Inverse Document Frequency (TF-IDF) values. We experiment with word-level features using n-grams ranging from 1 to 3, and character-level features from 3 to 8-grams. Each classifier is implemented with the following specifications:
Naïve Bayes (NB): Multinomial NB with additive smoothing constant 1
Logistic Regression (LR): Linear LR with L2 regularization constant 1 and limited-memory BFGS optimization
Support Vector Machine (SVM): Linear SVM with L2 regularization constant 1 and logistic loss function
Random Forests (RF): Averaging probabilistic predictions of 10 randomized decision trees
Gradient Boosted Trees (GBT): Tree boosting with learning rate 1 and logistic loss function
Neural Network based Models
Along with traditional machine learning approaches, we investigate neural network based models to evaluate their efficacy within a larger dataset. In particular, we explore Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and their variant models. A pre-trained GloVe BIBREF17 representation is used for word-level features.
CNN: We adopt Kim's kim2014convolutional implementation as the baseline. The word-level CNN models have 3 convolutional filters of different sizes [1,2,3] with ReLU activation, and a max-pooling layer. For the character-level CNN, we use 6 convolutional filters of various sizes [3,4,5,6,7,8], then add max-pooling layers followed by 1 fully-connected layer with a dimension of 1024.
Park and Fung park2017one proposed a HybridCNN model which outperformed both word-level and character-level CNNs in abusive language detection. In order to evaluate the HybridCNN for this dataset, we concatenate the output of max-pooled layers from word-level and character-level CNN, and feed this vector to a fully-connected layer in order to predict the output.
All three CNN models (word-level, character-level, and hybrid) use cross entropy with softmax as their loss function and Adam BIBREF18 as the optimizer.
RNN: We use bidirectional RNN BIBREF19 as the baseline, implementing a GRU BIBREF20 cell for each recurrent unit. From extensive parameter-search experiments, we chose 1 encoding layer with 50 dimensional hidden states and an input dropout probability of 0.3. The RNN models use cross entropy with sigmoid as their loss function and Adam as the optimizer.
For a possible improvement, we apply a self-matching attention mechanism on RNN baseline models BIBREF21 so that they may better understand the data by retrieving text sequences twice. We also investigate a recently introduced method, Latent Topic Clustering (LTC) BIBREF22 . The LTC method extracts latent topic information from the hidden states of RNN, and uses it for additional information in classifying the text data.
Feature Extension
While manually analyzing the raw dataset, we noticed that looking at the tweet one has replied to or has quoted, provides significant contextual information. We call these, “context tweets". As humans can better understand a tweet with the reference of its context, our assumption is that computers also benefit from taking context tweets into account in detecting abusive language.
As shown in the examples below, (2) is labeled abusive due to the use of vulgar language. However, the intention of the user can be better understood with its context tweet (1).
(1) I hate when I'm sitting in front of the bus and somebody with a wheelchair get on.
INLINEFORM0 (2) I hate it when I'm trying to board a bus and there's already an as**ole on it.
Similarly, context tweet (3) is important in understanding the abusive tweet (4), especially in identifying the target of the malice.
(3) Survivors of #Syria Gas Attack Recount `a Cruel Scene'.
INLINEFORM0 (4) Who the HELL is “LIKE" ING this post? Sick people....
Huang et al. huang2016modeling used several attributes of context tweets for sentiment analysis in order to improve the baseline LSTM model. However, their approach was limited because the meta-information they focused on—author information, conversation type, use of the same hashtags or emojis—are all highly dependent on data.
In order to avoid data dependency, text sequences of context tweets are directly used as an additional feature of neural network models. We use the same baseline model to convert context tweets to vectors, then concatenate these vectors with outputs of their corresponding labeled tweets. More specifically, we concatenate max-pooled layers of context and labeled tweets for the CNN baseline model. As for RNN, the last hidden states of context and labeled tweets are concatenated.
Dataset
Hate and Abusive Speech on Twitter BIBREF10 classifies tweets into 4 labels, “normal", “spam", “hateful" and “abusive". We were only able to crawl 70,904 tweets out of 99,996 tweet IDs, mainly because the tweet was deleted or the user account had been suspended. Table shows the distribution of labels of the crawled data.
Data Preprocessing
In the data preprocessing steps, user IDs, URLs, and frequently used emojis are replaced as special tokens. Since hashtags tend to have a high correlation with the content of the tweet BIBREF23 , we use a segmentation library BIBREF24 for hashtags to extract more information.
For character-level representations, we apply the method Zhang et al. zhang2015character proposed. Tweets are transformed into one-hot encoded vectors using 70 character dimensions—26 lower-cased alphabets, 10 digits, and 34 special characters including whitespace.
Training and Evaluation
In training the feature engineering based machine learning classifiers, we truncate vector representations according to the TF-IDF values (the top 14,000 and 53,000 for word-level and character-level representations, respectively) to avoid overfitting. For neural network models, words that appear only once are replaced as unknown tokens.
Since the dataset used is not split into train, development, and test sets, we perform 10-fold cross validation, obtaining the average of 5 tries; we divide the dataset randomly by a ratio of 85:5:10, respectively. In order to evaluate the overall performance, we calculate the weighted average of precision, recall, and F1 scores of all four labels, “normal”, “spam”, “hateful”, and “abusive”.
Empirical Results
As shown in Table , neural network models are more accurate than feature engineering based models (i.e. NB, SVM, etc.) except for the LR model—the best LR model has the same F1 score as the best CNN model.
Among traditional machine learning models, the most accurate in classifying abusive language is the LR model followed by ensemble models such as GBT and RF. Character-level representations improve F1 scores of SVM and RF classifiers, but they have no positive effect on other models.
For neural network models, RNN with LTC modules have the highest accuracy score, but there are no significant improvements from its baseline model and its attention-added model. Similarly, HybridCNN does not improve the baseline CNN model. For both CNN and RNN models, character-level features significantly decrease the accuracy of classification.
The use of context tweets generally have little effect on baseline models, however they noticeably improve the scores of several metrics. For instance, CNN with context tweets score the highest recall and F1 for “hateful" labels, and RNN models with context tweets have the highest recall for “abusive" tweets.
Discussion and Conclusion
While character-level features are known to improve the accuracy of neural network models BIBREF16 , they reduce classification accuracy for Hate and Abusive Speech on Twitter. We conclude this is because of the lack of labeled data as well as the significant imbalance among the different labels. Unlike neural network models, character-level features in traditional machine learning classifiers have positive results because we have trained the models only with the most significant character elements using TF-IDF values.
Variants of neural network models also suffer from data insufficiency. However, these models show positive performances on “spam" (14%) and “hateful" (4%) tweets—the lower distributed labels. The highest F1 score for “spam" is from the RNN-LTC model (0.551), and the highest for “hateful" is CNN with context tweets (0.309). Since each variant model excels in different metrics, we expect to see additional improvements with the use of ensemble models of these variants in future works.
In this paper, we report the baseline accuracy of different learning models as well as their variants on the recently introduced dataset, Hate and Abusive Speech on Twitter. Experimental results show that bidirectional GRU networks with LTC provide the most accurate results in detecting abusive language. Additionally, we present the possibility of using ensemble models of variant models and features for further improvements.
Acknowledgments
K. Jung is with the Department of Electrical and Computer Engineering, ASRI, Seoul National University, Seoul, Korea. This work was supported by the National Research Foundation of Korea (NRF) funded by the Korea government (MSIT) (No. 2016M3C4A7952632), the Technology Innovation Program (10073144) funded by the Ministry of Trade, Industry & Energy (MOTIE, Korea).
We would also like to thank Yongkeun Hwang and Ji Ho Park for helpful discussions and their valuable insights.
| [
"using tweets that one has replied or quoted to as contextual information",
"text sequences of context tweets"
] | 2,060 | qasper | en | null | 9fb085a1f47673d1907f2378c90843b4b6e8622a14fe1fa9 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What additional features and context are proposed?
Answer: | [
"using tweets that one has replied or quoted to as contextual information",
"text sequences of context tweets"
] | qasper | 128 | 3 | |
Which Facebook pages did they look at? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
This work is licenced under a Creative Commons Attribution 4.0 International Licence. Licence details: http://creativecommons.org/licenses/by/4.0/
In the spirit of the brevity of social media's messages and reactions, people have got used to express feelings minimally and symbolically, as with hashtags on Twitter and Instagram. On Facebook, people tend to be more wordy, but posts normally receive more simple “likes” than longer comments. Since February 2016, Facebook users can express specific emotions in response to a post thanks to the newly introduced reaction feature (see Section SECREF2 ), so that now a post can be wordlessly marked with an expression of say “joy" or “surprise" rather than a generic “like”.
It has been observed that this new feature helps Facebook to know much more about their users and exploit this information for targeted advertising BIBREF0 , but interest in people's opinions and how they feel isn't limited to commercial reasons, as it invests social monitoring, too, including health care and education BIBREF1 . However, emotions and opinions are not always expressed this explicitly, so that there is high interest in developing systems towards their automatic detection. Creating manually annotated datasets large enough to train supervised models is not only costly, but also—especially in the case of opinions and emotions—difficult, due to the intrinsic subjectivity of the task BIBREF2 , BIBREF3 . Therefore, research has focused on unsupervised methods enriched with information derived from lexica, which are manually created BIBREF3 , BIBREF4 . Since go2009twitter have shown that happy and sad emoticons can be successfully used as signals for sentiment labels, distant supervision, i.e. using some reasonably safe signals as proxies for automatically labelling training data BIBREF5 , has been used also for emotion recognition, for example exploiting both emoticons and Twitter hashtags BIBREF6 , but mainly towards creating emotion lexica. mohammad2015using use hashtags, experimenting also with highly fine-grained emotion sets (up to almost 600 emotion labels), to create the large Hashtag Emotion Lexicon. Emoticons are used as proxies also by hallsmarmulti, who use distributed vector representations to find which words are interchangeable with emoticons but also which emoticons are used in a similar context.
We take advantage of distant supervision by using Facebook reactions as proxies for emotion labels, which to the best of our knowledge hasn't been done yet, and we train a set of Support Vector Machine models for emotion recognition. Our models, differently from existing ones, exploit information which is acquired entirely automatically, and achieve competitive or even state-of-the-art results for some of the emotion labels on existing, standard evaluation datasets. For explanatory purposes, related work is discussed further and more in detail when we describe the benchmarks for evaluation (Section SECREF3 ) and when we compare our models to existing ones (Section SECREF5 ). We also explore and discuss how choosing different sets of Facebook pages as training data provides an intrinsic domain-adaptation method.
Facebook reactions as labels
For years, on Facebook people could leave comments to posts, and also “like” them, by using a thumbs-up feature to explicitly express a generic, rather underspecified, approval. A “like” could thus mean “I like what you said", but also “I like that you bring up such topic (though I find the content of the article you linked annoying)".
In February 2016, after a short trial, Facebook made a more explicit reaction feature available world-wide. Rather than allowing for the underspecified “like” as the only wordless response to a post, a set of six more specific reactions was introduced, as shown in Figure FIGREF1 : Like, Love, Haha, Wow, Sad and Angry. We use such reactions as proxies for emotion labels associated to posts.
We collected Facebook posts and their corresponding reactions from public pages using the Facebook API, which we accessed via the Facebook-sdk python library. We chose different pages (and therefore domains and stances), aiming at a balanced and varied dataset, but we did so mainly based on intuition (see Section SECREF4 ) and with an eye to the nature of the datasets available for evaluation (see Section SECREF5 ). The choice of which pages to select posts from is far from trivial, and we believe this is actually an interesting aspect of our approach, as by using different Facebook pages one can intrinsically tackle the domain-adaptation problem (See Section SECREF6 for further discussion on this). The final collection of Facebook pages for the experiments described in this paper is as follows: FoxNews, CNN, ESPN, New York Times, Time magazine, Huffington Post Weird News, The Guardian, Cartoon Network, Cooking Light, Home Cooking Adventure, Justin Bieber, Nickelodeon, Spongebob, Disney.
Note that thankful was only available during specific time spans related to certain events, as Mother's Day in May 2016.
For each page, we downloaded the latest 1000 posts, or the maximum available if there are fewer, from February 2016, retrieving the counts of reactions for each post. The output is a JSON file containing a list of dictionaries with a timestamp, the post and a reaction vector with frequency values, which indicate how many users used that reaction in response to the post (Figure FIGREF3 ). The resulting emotion vectors must then be turned into an emotion label.
In the context of this experiment, we made the simple decision of associating to each post the emotion with the highest count, ignoring like as it is the default and most generic reaction people tend to use. Therefore, for example, to the first post in Figure FIGREF3 , we would associate the label sad, as it has the highest score (284) among the meaningful emotions we consider, though it also has non-zero scores for other emotions. At this stage, we didn't perform any other entropy-based selection of posts, to be investigated in future work.
Emotion datasets
Three datasets annotated with emotions are commonly used for the development and evaluation of emotion detection systems, namely the Affective Text dataset, the Fairy Tales dataset, and the ISEAR dataset. In order to compare our performance to state-of-the-art results, we have used them as well. In this Section, in addition to a description of each dataset, we provide an overview of the emotions used, their distribution, and how we mapped them to those we obtained from Facebook posts in Section SECREF7 . A summary is provided in Table TABREF8 , which also shows, in the bottom row, what role each dataset has in our experiments: apart from the development portion of the Affective Text, which we used to develop our models (Section SECREF4 ), all three have been used as benchmarks for our evaluation.
Affective Text dataset
Task 14 at SemEval 2007 BIBREF7 was concerned with the classification of emotions and valence in news headlines. The headlines where collected from several news websites including Google news, The New York Times, BBC News and CNN. The used emotion labels were Anger, Disgust, Fear, Joy, Sadness, Surprise, in line with the six basic emotions of Ekman's standard model BIBREF8 . Valence was to be determined as positive or negative. Classification of emotion and valence were treated as separate tasks. Emotion labels were not considered as mututally exclusive, and each emotion was assigned a score from 0 to 100. Training/developing data amounted to 250 annotated headlines (Affective development), while systems were evaluated on another 1000 (Affective test). Evaluation was done using two different methods: a fine-grained evaluation using Pearson's r to measure the correlation between the system scores and the gold standard; and a coarse-grained method where each emotion score was converted to a binary label, and precision, recall, and f-score were computed to assess performance. As it is done in most works that use this dataset BIBREF3 , BIBREF4 , BIBREF9 , we also treat this as a classification problem (coarse-grained). This dataset has been extensively used for the evaluation of various unsupervised methods BIBREF2 , but also for testing different supervised learning techniques and feature portability BIBREF10 .
Fairy Tales dataset
This is a dataset collected by alm2008affect, where about 1,000 sentences from fairy tales (by B. Potter, H.C. Andersen and Grimm) were annotated with the same six emotions of the Affective Text dataset, though with different names: Angry, Disgusted, Fearful, Happy, Sad, and Surprised. In most works that use this dataset BIBREF3 , BIBREF4 , BIBREF9 , only sentences where all annotators agreed are used, and the labels angry and disgusted are merged. We adopt the same choices.
ISEAR
The ISEAR (International Survey on Emotion Antecedents and Reactions BIBREF11 , BIBREF12 ) is a dataset created in the context of a psychology project of the 1990s, by collecting questionnaires answered by people with different cultural backgrounds. The main aim of this project was to gather insights in cross-cultural aspects of emotional reactions. Student respondents, both psychologists and non-psychologists, were asked to report situations in which they had experienced all of seven major emotions (joy, fear, anger, sadness, disgust, shame and guilt). In each case, the questions covered the way they had appraised a given situation and how they reacted. The final dataset contains reports by approximately 3000 respondents from all over the world, for a total of 7665 sentences labelled with an emotion, making this the largest dataset out of the three we use.
Overview of datasets and emotions
We summarise datasets and emotion distribution from two viewpoints. First, because there are different sets of emotions labels in the datasets and Facebook data, we need to provide a mapping and derive a subset of emotions that we are going to use for the experiments. This is shown in Table TABREF8 , where in the “Mapped” column we report the final emotions we use in this paper: anger, joy, sadness, surprise. All labels in each dataset are mapped to these final emotions, which are therefore the labels we use for training and testing our models.
Second, the distribution of the emotions for each dataset is different, as can be seen in Figure FIGREF9 .
In Figure FIGREF9 we also provide the distribution of the emotions anger, joy, sadness, surprise per Facebook page, in terms of number of posts (recall that we assign to a post the label corresponding to the majority emotion associated to it, see Section SECREF2 ). We can observe that for example pages about news tend to have more sadness and anger posts, while pages about cooking and tv-shows have a high percentage of joy posts. We will use this information to find the best set of pages for a given target domain (see Section SECREF5 ).
Model
There are two main decisions to be taken in developing our model: (i) which Facebook pages to select as training data, and (ii) which features to use to train the model, which we discuss below. Specifically, we first set on a subset of pages and then experiment with features. Further exploration of the interaction between choice of pages and choice of features is left to future work, and partly discussed in Section SECREF6 . For development, we use a small portion of the Affective data set described in Section SECREF4 , that is the portion that had been released as development set for SemEval's 2007 Task 14 BIBREF7 , which contains 250 annotated sentences (Affective development, Section SECREF4 ). All results reported in this section are on this dataset. The test set of Task 14 as well as the other two datasets described in Section SECREF3 will be used to evaluate the final models (Section SECREF4 ).
Selecting Facebook pages
Although page selection is a crucial ingredient of this approach, which we believe calls for further and deeper, dedicated investigation, for the experiments described here we took a rather simple approach. First, we selected the pages that would provide training data based on intuition and availability, then chose different combinations according to results of a basic model run on development data, and eventually tested feature combinations, still on the development set.
For the sake of simplicity and transparency, we first trained an SVM with a simple bag-of-words model and default parameters as per the Scikit-learn implementation BIBREF13 on different combinations of pages. Based on results of the attempted combinations as well as on the distribution of emotions in the development dataset (Figure FIGREF9 ), we selected a best model (B-M), namely the combined set of Time, The Guardian and Disney, which yields the highest results on development data. Time and The Guardian perform well on most emotions but Disney helps to boost the performance for the Joy class.
Features
In selecting appropriate features, we mainly relied on previous work and intuition. We experimented with different combinations, and all tests were still done on Affective development, using the pages for the best model (B-M) described above as training data. Results are in Table TABREF20 . Future work will further explore the simultaneous selection of features and page combinations.
We use a set of basic text-based features to capture the emotion class. These include a tf-idf bag-of-words feature, word (2-3) and character (2-5) ngrams, and features related to the presence of negation words, and to the usage of punctuation.
This feature is used in all unsupervised models as a source of information, and we mainly include it to assess its contribution, but eventually do not use it in our final model.
We used the NRC10 Lexicon because it performed best in the experiments by BIBREF10 , which is built around the emotions anger, anticipation, disgust, fear, joy, sadness, and surprise, and the valence values positive and negative. For each word in the lexicon, a boolean value indicating presence or absence is associated to each emotion. For a whole sentence, a global score per emotion can be obtained by summing the vectors for all content words of that sentence included in the lexicon, and used as feature.
As additional feature, we also included Word Embeddings, namely distributed representations of words in a vector space, which have been exceptionally successful in boosting performance in a plethora of NLP tasks. We use three different embeddings:
Google embeddings: pre-trained embeddings trained on Google News and obtained with the skip-gram architecture described in BIBREF14 . This model contains 300-dimensional vectors for 3 million words and phrases.
Facebook embeddings: embeddings that we trained on our scraped Facebook pages for a total of 20,000 sentences. Using the gensim library BIBREF15 , we trained the embeddings with the following parameters: window size of 5, learning rate of 0.01 and dimensionality of 100. We filtered out words with frequency lower than 2 occurrences.
Retrofitted embeddings: Retrofitting BIBREF16 has been shown as a simple but efficient way of informing trained embeddings with additional information derived from some lexical resource, rather than including it directly at the training stage, as it's done for example to create sense-aware BIBREF17 or sentiment-aware BIBREF18 embeddings. In this work, we retrofit general embeddings to include information about emotions, so that emotion-similar words can get closer in space. Both the Google as well as our Facebook embeddings were retrofitted with lexical information obtained from the NRC10 Lexicon mentioned above, which provides emotion-similarity for each token. Note that differently from the previous two types of embeddings, the retrofitted ones do rely on handcrafted information in the form of a lexical resource.
Results on development set
We report precision, recall, and f-score on the development set. The average f-score is reported as micro-average, to better account for the skewed distribution of the classes as well as in accordance to what is usually reported for this task BIBREF19 .
From Table TABREF20 we draw three main observations. First, a simple tf-idf bag-of-word mode works already very well, to the point that the other textual and lexicon-based features don't seem to contribute to the overall f-score (0.368), although there is a rather substantial variation of scores per class. Second, Google embeddings perform a lot better than Facebook embeddings, and this is likely due to the size of the corpus used for training. Retrofitting doesn't seem to help at all for the Google embeddings, but it does boost the Facebook embeddings, leading to think that with little data, more accurate task-related information is helping, but corpus size matters most. Third, in combination with embeddings, all features work better than just using tf-idf, but removing the Lexicon feature, which is the only one based on hand-crafted resources, yields even better results. Then our best model (B-M) on development data relies entirely on automatically obtained information, both in terms of training data as well as features.
Results
In Table TABREF26 we report the results of our model on the three datasets standardly used for the evaluation of emotion classification, which we have described in Section SECREF3 .
Our B-M model relies on subsets of Facebook pages for training, which were chosen according to their performance on the development set as well as on the observation of emotions distribution on different pages and in the different datasets, as described in Section SECREF4 . The feature set we use is our best on the development set, namely all the features plus Google-based embeddings, but excluding the lexicon. This makes our approach completely independent of any manual annotation or handcrafted resource. Our model's performance is compared to the following systems, for which results are reported in the referred literature. Please note that no other existing model was re-implemented, and results are those reported in the respective papers.
Discussion, conclusions and future work
We have explored the potential of using Facebook reactions in a distant supervised setting to perform emotion classification. The evaluation on standard benchmarks shows that models trained as such, especially when enhanced with continuous vector representations, can achieve competitive results without relying on any handcrafted resource. An interesting aspect of our approach is the view to domain adaptation via the selection of Facebook pages to be used as training data.
We believe that this approach has a lot of potential, and we see the following directions for improvement. Feature-wise, we want to train emotion-aware embeddings, in the vein of work by tang:14, and iacobacci2015sensembed. Retrofitting FB-embeddings trained on a larger corpus might also be successful, but would rely on an external lexicon.
The largest room for yielding not only better results but also interesting insights on extensions of this approach lies in the choice of training instances, both in terms of Facebook pages to get posts from, as well as in which posts to select from the given pages. For the latter, one could for example only select posts that have a certain length, ignore posts that are only quotes or captions to images, or expand posts by including content from linked html pages, which might provide larger and better contexts BIBREF23 . Additionally, and most importantly, one could use an entropy-based measure to select only posts that have a strong emotion rather than just considering the majority emotion as training label. For the former, namely the choice of Facebook pages, which we believe deserves the most investigation, one could explore several avenues, especially in relation to stance-based issues BIBREF24 . In our dataset, for example, a post about Chile beating Colombia in a football match during the Copa America had very contradictory reactions, depending on which side readers would cheer for. Similarly, the very same political event, for example, would get very different reactions from readers if it was posted on Fox News or The Late Night Show, as the target audience is likely to feel very differently about the same issue. This also brings up theoretical issues related more generally to the definition of the emotion detection task, as it's strongly dependent on personal traits of the audience. Also, in this work, pages initially selected on availability and intuition were further grouped into sets to make training data according to performance on development data, and label distribution. Another criterion to be exploited would be vocabulary overlap between the pages and the datasets.
Lastly, we could develop single models for each emotion, treating the problem as a multi-label task. This would even better reflect the ambiguity and subjectivity intrinsic to assigning emotions to text, where content could be at same time joyful or sad, depending on the reader.
Acknowledgements
In addition to the anonymous reviewers, we want to thank Lucia Passaro and Barbara Plank for insightful discussions, and for providing comments on draft versions of this paper.
| [
"FoxNews, CNN, ESPN, New York Times, Time magazine, Huffington Post Weird News, The Guardian, Cartoon Network, Cooking Light, Home Cooking Adventure, Justin Bieber, Nickelodeon, Spongebob, Disney",
"FoxNews, CNN, ESPN, New York Times, Time magazine, Huffington Post Weird News, The Guardian, Cartoon Network, Cooki... | 3,411 | qasper | en | null | 14e88916523197e066dbb743339685621db47c4beb503abb | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Which Facebook pages did they look at?
Answer: | [
"FoxNews, CNN, ESPN, New York Times, Time magazine, Huffington Post Weird News, The Guardian, Cartoon Network, Cooking Light, Home Cooking Adventure, Justin Bieber, Nickelodeon, Spongebob, Disney",
"FoxNews, CNN, ESPN, New York Times, Time magazine, Huffington Post Weird News, The Guardian, Cartoon Network, Cooki... | qasper | 128 | 4 | |
Do the hashtag and SemEval datasets contain only English data? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
A hashtag is a keyphrase represented as a sequence of alphanumeric characters plus underscore, preceded by the # symbol. Hashtags play a central role in online communication by providing a tool to categorize the millions of posts generated daily on Twitter, Instagram, etc. They are useful in search, tracking content about a certain topic BIBREF0 , BIBREF1 , or discovering emerging trends BIBREF2 .
Hashtags often carry very important information, such as emotion BIBREF3 , sentiment BIBREF4 , sarcasm BIBREF5 , and named entities BIBREF6 , BIBREF7 . However, inferring the semantics of hashtags is non-trivial since many hashtags contain multiple tokens joined together, which frequently leads to multiple potential interpretations (e.g., lion head vs. lionhead). Table TABREF3 shows several examples of single- and multi-token hashtags. While most hashtags represent a mix of standard tokens, named entities and event names are prevalent and pose challenges to both human and automatic comprehension, as these are more likely to be rare tokens. Hashtags also tend to be shorter to allow fast typing, to attract attention or to satisfy length limitations imposed by some social media platforms. Thus, they tend to contain a large number of abbreviations or non-standard spelling variations (e.g., #iloveu4eva) BIBREF8 , BIBREF9 , which hinders their understanding.
The goal of our study is to build efficient methods for automatically splitting a hashtag into a meaningful word sequence. Our contributions are:
Our new dataset includes segmentation for 12,594 unique hashtags and their associated tweets annotated in a multi-step process for higher quality than the previous dataset of 1,108 hashtags BIBREF10 . We frame the segmentation task as a pairwise ranking problem, given a set of candidate segmentations. We build several neural architectures using this problem formulation which use corpus-based, linguistic and thesaurus based features. We further propose a multi-task learning approach which jointly learns segment ranking and single- vs. multi-token hashtag classification. The latter leads to an error reduction of 24.6% over the current state-of-the-art. Finally, we demonstrate the utility of our method by using hashtag segmentation in the downstream task of sentiment analysis. Feeding the automatically segmented hashtags to a state-of-the-art sentiment analysis method on the SemEval 2017 benchmark dataset results in a 2.6% increase in the official metric for the task.
Background and Preliminaries
Current approaches for hashtag segmentation can be broadly divided into three categories: (a) gazeteer and rule based BIBREF11 , BIBREF12 , BIBREF13 , (b) word boundary detection BIBREF14 , BIBREF15 , and (c) ranking with language model and other features BIBREF16 , BIBREF10 , BIBREF0 , BIBREF17 , BIBREF18 . Hashtag segmentation approaches draw upon work on compound splitting for languages such as German or Finnish BIBREF19 and word segmentation BIBREF20 for languages with no spaces between words such as Chinese BIBREF21 , BIBREF22 . Similar to our work, BIBREF10 BansalBV15 extract an initial set of candidate segmentations using a sliding window, then rerank them using a linear regression model trained on lexical, bigram and other corpus-based features. The current state-of-the-art approach BIBREF14 , BIBREF15 uses maximum entropy and CRF models with a combination of language model and hand-crafted features to predict if each character in the hashtag is the beginning of a new word.
Generating Candidate Segmentations. Microsoft Word Breaker BIBREF16 is, among the existing methods, a strong baseline for hashtag segmentation, as reported in BIBREF14 and BIBREF10 . It employs a beam search algorithm to extract INLINEFORM0 best segmentations as ranked by the n-gram language model probability: INLINEFORM1
where INLINEFORM0 is the word sequence of segmentation INLINEFORM1 and INLINEFORM2 is the window size. More sophisticated ranking strategies, such as Binomial and word length distribution based ranking, did not lead to a further improvement in performance BIBREF16 . The original Word Breaker was designed for segmenting URLs using language models trained on web data. In this paper, we reimplemented and tailored this approach to segmenting hashtags by using a language model specifically trained on Twitter data (implementation details in § SECREF26 ). The performance of this method itself is competitive with state-of-the-art methods (evaluation results in § SECREF46 ). Our proposed pairwise ranking method will effectively take the top INLINEFORM3 segmentations generated by this baseline as candidates for reranking.
However, in prior work, the ranking scores of each segmentation were calculated independently, ignoring the relative order among the top INLINEFORM0 candidate segmentations. To address this limitation, we utilize a pairwise ranking strategy for the first time for this task and propose neural architectures to model this.
Multi-task Pairwise Neural Ranking
We propose a multi-task pairwise neural ranking approach to better incorporate and distinguish the relative order between the candidate segmentations of a given hashtag. Our model adapts to address single- and multi-token hashtags differently via a multi-task learning strategy without requiring additional annotations. In this section, we describe the task setup and three variants of pairwise neural ranking models (Figure FIGREF11 ).
Segmentation as Pairwise Ranking
The goal of hashtag segmentation is to divide a given hashtag INLINEFORM0 into a sequence of meaningful words INLINEFORM1 . For a hashtag of INLINEFORM2 characters, there are a total of INLINEFORM3 possible segmentations but only one, or occasionally two, of them ( INLINEFORM4 ) are considered correct (Table TABREF9 ).
We transform this task into a pairwise ranking problem: given INLINEFORM0 candidate segmentations { INLINEFORM1 }, we rank them by comparing each with the rest in a pairwise manner. More specifically, we train a model to predict a real number INLINEFORM2 for any two candidate segmentations INLINEFORM3 and INLINEFORM4 of hashtag INLINEFORM5 , which indicates INLINEFORM6 is a better segmentation than INLINEFORM7 if positive, and vice versa. To quantify the quality of a segmentation in training, we define a gold scoring function INLINEFORM8 based on the similarities with the ground-truth segmentation INLINEFORM9 : INLINEFORM10
We use the Levenshtein distance (minimum number of single-character edits) in this paper, although it is possible to use other similarity measurements as alternatives. We use the top INLINEFORM0 segmentations generated by Microsoft Word Breaker (§ SECREF2 ) as initial candidates.
Pairwise Neural Ranking Model
For an input candidate segmentation pair INLINEFORM0 , we concatenate their feature vectors INLINEFORM1 and INLINEFORM2 , and feed them into a feedforward network which emits a comparison score INLINEFORM3 . The feature vector INLINEFORM4 or INLINEFORM5 consists of language model probabilities using Good-Turing BIBREF23 and modified Kneser-Ney smoothing BIBREF24 , BIBREF25 , lexical and linguistic features (more details in § SECREF23 ). For training, we use all the possible pairs INLINEFORM6 of the INLINEFORM7 candidates as the input and their gold scores INLINEFORM8 as the target. The training objective is to minimize the Mean Squared Error (MSE): DISPLAYFORM0
where INLINEFORM0 is the number of training examples.
To aggregate the pairwise comparisons, we follow a greedy algorithm proposed by BIBREF26 cohen1998learning and used for preference ranking BIBREF27 . For each segmentation INLINEFORM0 in the candidate set INLINEFORM1 , we calculate a single score INLINEFORM2 , and find the segmentation INLINEFORM3 corresponding to the highest score. We repeat the same procedure after removing INLINEFORM4 from INLINEFORM5 , and continue until INLINEFORM6 reduces to an empty set. Figure FIGREF11 (a) shows the architecture of this model.
Margin Ranking (MR) Loss
As an alternative to the pairwise ranker (§ SECREF15 ), we propose a pairwise model which learns from candidate pairs INLINEFORM0 but ranks each individual candidate directly rather than relatively. We define a new scoring function INLINEFORM1 which assigns a higher score to the better candidate, i.e., INLINEFORM2 , if INLINEFORM3 is a better candidate than INLINEFORM4 and vice-versa. Instead of concatenating the features vectors INLINEFORM5 and INLINEFORM6 , we feed them separately into two identical feedforward networks with shared parameters. During testing, we use only one of the networks to rank the candidates based on the INLINEFORM7 scores. For training, we add a ranking layer on top of the networks to measure the violations in the ranking order and minimize the Margin Ranking Loss (MR): DISPLAYFORM0
where INLINEFORM0 is the number of training samples. The architecture of this model is presented in Figure FIGREF11 (b).
Adaptive Multi-task Learning
Both models in § SECREF15 and § SECREF17 treat all the hashtags uniformly. However, different features address different types of hashtags. By design, the linguistic features capture named entities and multi-word hashtags that exhibit word shape patterns, such as camel case. The ngram probabilities with Good-Turing smoothing gravitate towards multi-word segmentations with known words, as its estimate for unseen ngrams depends on the fraction of ngrams seen once which can be very low BIBREF28 . The modified Kneser-Ney smoothing is more likely to favor segmentations that contain rare words, and single-word segmentations in particular. Please refer to § SECREF46 for a more detailed quantitative and qualitative analysis.
To leverage this intuition, we introduce a binary classification task to help the model differentiate single-word from multi-word hashtags. The binary classifier takes hashtag features INLINEFORM0 as the input and outputs INLINEFORM1 , which represents the probability of INLINEFORM2 being a multi-word hashtag. INLINEFORM3 is used as an adaptive gating value in our multi-task learning setup. The gold labels for this task are obtained at no extra cost by simply verifying whether the ground-truth segmentation has multiple words. We train the pairwise segmentation ranker and the binary single- vs. multi-token hashtag classifier jointly, by minimizing INLINEFORM4 for the pairwise ranker and the Binary Cross Entropy Error ( INLINEFORM5 ) for the classifier: DISPLAYFORM0
where INLINEFORM0 is the adaptive gating value, INLINEFORM1 indicates if INLINEFORM2 is actually a multi-word hashtag and INLINEFORM3 is the number of training examples. INLINEFORM4 and INLINEFORM5 are the weights for each loss. For our experiments, we apply equal weights.
More specifically, we divide the segmentation feature vector INLINEFORM0 into two subsets: (a) INLINEFORM1 with modified Kneser-Ney smoothing features, and (b) INLINEFORM2 with Good-Turing smoothing and linguistic features. For an input candidate segmentation pair INLINEFORM3 , we construct two pairwise vectors INLINEFORM4 and INLINEFORM5 by concatenation, then combine them based on the adaptive gating value INLINEFORM6 before feeding them into the feedforward network INLINEFORM7 for pairwise ranking: DISPLAYFORM0
We use summation with padding, as we find this simple ensemble method achieves similar performance in our experiments as the more complex multi-column networks BIBREF29 . Figure FIGREF11 (c) shows the architecture of this model. An analogue multi-task formulation can also be used for the Margin Ranking loss as: DISPLAYFORM0
Features
We use a combination of corpus-based and linguistic features to rank the segmentations. For a candidate segmentation INLINEFORM0 , its feature vector INLINEFORM1 includes the number of words in the candidate, the length of each word, the proportion of words in an English dictionary or Urban Dictionary BIBREF30 , ngram counts from Google Web 1TB corpus BIBREF31 , and ngram probabilities from trigram language models trained on the Gigaword corpus BIBREF32 and 1.1 billion English tweets from 2010, respectively. We train two language models on each corpus: one with Good-Turing smoothing using SRILM BIBREF33 and the other with modified Kneser-Ney smoothing using KenLM BIBREF34 . We also add boolean features, such as if the candidate is a named-entity present in the list of Wikipedia titles, and if the candidate segmentation INLINEFORM2 and its corresponding hashtag INLINEFORM3 satisfy certain word-shapes (more details in appendix SECREF61 ).
Similarly, for hashtag INLINEFORM0 , we extract the feature vector INLINEFORM1 consisting of hashtag length, ngram count of the hashtag in Google 1TB corpus BIBREF31 , and boolean features indicating if the hashtag is in an English dictionary or Urban Dictionary, is a named-entity, is in camel case, ends with a number, and has all the letters as consonants. We also include features of the best-ranked candidate by the Word Breaker model.
Implementation Details
We use the PyTorch framework to implement our multi-task pairwise ranking model. The pairwise ranker consists of an input layer, three hidden layers with eight nodes in each layer and hyperbolic tangent ( INLINEFORM0 ) activation, and a single linear output node. The auxiliary classifier consists of an input layer, one hidden layer with eight nodes and one output node with sigmoid activation. We use the Adam algorithm BIBREF35 for optimization and apply a dropout of 0.5 to prevent overfitting. We set the learning rate to 0.01 and 0.05 for the pairwise ranker and auxiliary classifier respectively. For each experiment, we report results obtained after 100 epochs.
For the baseline model used to extract the INLINEFORM0 initial candidates, we reimplementated the Word Breaker BIBREF16 as described in § SECREF2 and adapted it to use a language model trained on 1.1 billion tweets with Good-Turing smoothing using SRILM BIBREF33 to give a better performance in segmenting hashtags (§ SECREF46 ). For all our experiments, we set INLINEFORM1 .
Hashtag Segmentation Data
We use two datasets for experiments (Table TABREF29 ): (a) STAN INLINEFORM0 , created by BIBREF10 BansalBV15, which consists of 1,108 unique English hashtags from 1,268 randomly selected tweets in the Stanford Sentiment Analysis Dataset BIBREF36 along with their crowdsourced segmentations and our additional corrections; and (b) STAN INLINEFORM1 , our new expert curated dataset, which includes all 12,594 unique English hashtags and their associated tweets from the same Stanford dataset.
Experiments
In this section, we present experimental results that compare our proposed method with the other state-of-the-art approaches on hashtag segmentation datasets. The next section will show experiments of applying hashtag segmentation to the popular task of sentiment analysis.
Existing Methods
We compare our pairwise neural ranker with the following baseline and state-of-the-art approaches:
The original hashtag as a single token;
A rule-based segmenter, which employs a set of word-shape rules with an English dictionary BIBREF13 ;
A Viterbi model which uses word frequencies from a book corpus BIBREF0 ;
The specially developed GATE Hashtag Tokenizer from the open source toolkit, which combines dictionaries and gazetteers in a Viterbi-like algorithm BIBREF11 ;
A maximum entropy classifier (MaxEnt) trained on the STAN INLINEFORM0 training dataset. It predicts whether a space should be inserted at each position in the hashtag and is the current state-of-the-art BIBREF14 ;
Our reimplementation of the Word Breaker algorithm which uses beam search and a Twitter ngram language model BIBREF16 ;
A pairwise linear ranker which we implemented for comparison purposes with the same features as our neural model, but using perceptron as the underlying classifier BIBREF38 and minimizing the hinge loss between INLINEFORM0 and a scoring function similar to INLINEFORM1 . It is trained on the STAN INLINEFORM2 dataset.
Evaluation Metrics
We evaluate the performance by the top INLINEFORM0 ( INLINEFORM1 ) accuracy (A@1, A@2), average token-level F INLINEFORM2 score (F INLINEFORM3 @1), and mean reciprocal rank (MRR). In particular, the accuracy and MRR are calculated at the segmentation-level, which means that an output segmentation is considered correct if and only if it fully matches the human segmentation. Average token-level F INLINEFORM4 score accounts for partially correct segmentation in the multi-token hashtag cases.
Results
Tables TABREF32 and TABREF33 show the results on the STAN INLINEFORM0 and STAN INLINEFORM1 datasets, respectively. All of our pairwise neural rankers are trained on the 2,518 manually segmented hashtags in the training set of STAN INLINEFORM2 and perform favorably against other state-of-the-art approaches. Our best model (MSE+multitask) that utilizes different features adaptively via a multi-task learning procedure is shown to perform better than simply combining all the features together (MR and MSE). We highlight the 24.6% error reduction on STAN INLINEFORM3 and 16.5% on STAN INLINEFORM4 of our approach over the previous SOTA BIBREF14 on the Multi-token hashtags, and the importance of having a separate evaluation of multi-word cases as it is trivial to obtain 100% accuracy for Single-token hashtags. While our hashtag segmentation model is achieving a very high accuracy@2, to be practically useful, it remains a challenge to get the top one predication exactly correct. Some hashtags are very difficult to interpret, e.g., #BTVbrownSMB refers to the Social Media Breakfast (SMB) in Burlington, Vermont (BTV).
The improved Word Breaker with our addition of a Twitter-specific language model is a very strong baseline, which echos the findings of the original Word Breaker paper BIBREF16 that having a large in-domain language model is extremely helpful for word segmentation tasks. It is worth noting that the other state-of-the-art system BIBREF14 also utilized a 4-gram language model trained on 476 million tweets from 2009.
Analysis and Discussion
To empirically illustrate the effectiveness of different features on different types of hashtags, we show the results for models using individual feature sets in pairwise ranking models (MSE) in Table TABREF45 . Language models with modified Kneser-Ney smoothing perform best on single-token hashtags, while Good-Turing and Linguistic features work best on multi-token hashtags, confirming our intuition about their usefulness in a multi-task learning approach. Table TABREF47 shows a qualitative analysis with the first column ( INLINEFORM0 INLINEFORM1 INLINEFORM2 ) indicating which features lead to correct or wrong segmentations, their count in our data and illustrative examples with human segmentation.
As expected, longer hashtags with more than three tokens pose greater challenges and the segmentation-level accuracy of our best model (MSE+multitask) drops to 82.1%. For many error cases, our model predicts a close-to-correct segmentation, e.g., #youbrownknowyoubrownupttoobrownearly, #iseebrownlondoniseebrownfrance, which is also reflected by the higher token-level F INLINEFORM0 scores across hashtags with different lengths (Figure FIGREF51 ).
Since our approach heavily relies on building a Twitter language model, we experimented with its sizes and show the results in Figure FIGREF52 . Our approach can perform well even with access to a smaller amount of tweets. The drop in F INLINEFORM0 score for our pairwise neural ranker is only 1.4% and 3.9% when using the language models trained on 10% and 1% of the total 1.1 billion tweets, respectively.
Language use in Twitter changes with time BIBREF9 . Our pairwise ranker uses language models trained on the tweets from the year 2010. We tested our approach on a set of 500 random English hashtags posted in tweets from the year 2019 and show the results in Table TABREF55 . With a segmentation-level accuracy of 94.6% and average token-level F INLINEFORM0 score of 95.6%, our approach performs favorably on 2019 hashtags.
Extrinsic Evaluation: Twitter Sentiment Analysis
We attempt to demonstrate the effectiveness of our hashtag segmentation system by studying its impact on the task of sentiment analysis in Twitter BIBREF39 , BIBREF40 , BIBREF41 . We use our best model (MSE+multitask), under the name HashtagMaster, in the following experiments.
Experimental Setup
We compare the performance of the BiLSTM+Lex BIBREF42 sentiment analysis model under three configurations: (a) tweets with hashtags removed, (b) tweets with hashtags as single tokens excluding the # symbol, and (c) tweets with hashtags as segmented by our system, HashtagMaster. BiLSTM+Lex is a state-of-the-art open source system for predicting tweet-level sentiment BIBREF43 . It learns a context-sensitive sentiment intensity score by leveraging a Twitter-based sentiment lexicon BIBREF44 . We use the same settings as described by BIBREF42 teng-vo-zhang:2016:EMNLP2016 to train the model.
We use the dataset from the Sentiment Analysis in Twitter shared task (subtask A) at SemEval 2017 BIBREF41 . Given a tweet, the goal is to predict whether it expresses POSITIVE, NEGATIVE or NEUTRAL sentiment. The training and development sets consist of 49,669 tweets and we use 40,000 for training and the rest for development. There are a total of 12,284 tweets containing 12,128 hashtags in the SemEval 2017 test set, and our hashtag segmenter ended up splitting 6,975 of those hashtags present in 3,384 tweets.
Results and Analysis
In Table TABREF59 , we report the results based on the 3,384 tweets where HashtagMaster predicted a split, as for the rest of tweets in the test set, the hashtag segmenter would neither improve nor worsen the sentiment prediction. Our hashtag segmenter successfully improved the sentiment analysis performance by 2% on average recall and F INLINEFORM0 comparing to having hashtags unsegmented. This improvement is seemingly small but decidedly important for tweets where sentiment-related information is embedded in multi-word hashtags and sentiment prediction would be incorrect based only on the text (see Table TABREF60 for examples). In fact, 2,605 out of the 3,384 tweets have multi-word hashtags that contain words in the Twitter-based sentiment lexicon BIBREF44 and 125 tweets contain sentiment words only in the hashtags but not in the rest of the tweet. On the entire test set of 12,284 tweets, the increase in the average recall is 0.5%.
Other Related Work
Automatic hashtag segmentation can improve the performance of many applications besides sentiment analysis, such as text classification BIBREF13 , named entity linking BIBREF10 and modeling user interests for recommendations BIBREF45 . It can also help in collecting data of higher volume and quality by providing a more nuanced interpretation of its content, as shown for emotion analysis BIBREF46 , sarcasm and irony detection BIBREF11 , BIBREF47 . Better semantic analysis of hashtags can also potentially be applied to hashtag annotation BIBREF48 , to improve distant supervision labels in training classifiers for tasks such as sarcasm BIBREF5 , sentiment BIBREF4 , emotions BIBREF3 ; and, more generally, as labels for pre-training representations of words BIBREF49 , sentences BIBREF50 , and images BIBREF51 .
Conclusion
We proposed a new pairwise neural ranking model for hashtag segmention and showed significant performance improvements over the state-of-the-art. We also constructed a larger and more curated dataset for analyzing and benchmarking hashtag segmentation methods. We demonstrated that hashtag segmentation helps with downstream tasks such as sentiment analysis. Although we focused on English hashtags, our pairwise ranking approach is language-independent and we intend to extend our toolkit to languages other than English as future work.
Acknowledgments
We thank Ohio Supercomputer Center BIBREF52 for computing resources and the NVIDIA for providing GPU hardware. We thank Alan Ritter, Quanze Chen, Wang Ling, Pravar Mahajan, and Dushyanta Dhyani for valuable discussions. We also thank the annotators: Sarah Flanagan, Kaushik Mani, and Aswathnarayan Radhakrishnan. This material is based in part on research sponsored by the NSF under grants IIS-1822754 and IIS-1755898, DARPA through the ARO under agreement number W911NF-17-C-0095, through a Figure-Eight (CrowdFlower) AI for Everyone Award and a Criteo Faculty Research Award to Wei Xu. The views and conclusions contained in this publication are those of the authors and should not be interpreted as representing official policies or endorsements of the U.S. Government.
Word-shape rules
Our model uses the following word shape rules as boolean features. If the candidate segmentation INLINEFORM0 and its corresponding hashtag INLINEFORM1 satisfies a word shape rule, then the boolean feature is set to True.
| [
"Yes",
"Yes"
] | 3,735 | qasper | en | null | c109a25d23a8c5c95af8697ef1ac51cd3bee93c67521af6d | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Do the hashtag and SemEval datasets contain only English data?
Answer: | [
"Yes",
"Yes"
] | qasper | 128 | 5 | |
What type of evaluation is proposed for this task? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Multi-document summarization (MDS), the transformation of a set of documents into a short text containing their most important aspects, is a long-studied problem in NLP. Generated summaries have been shown to support humans dealing with large document collections in information seeking tasks BIBREF0 , BIBREF1 , BIBREF2 . However, when exploring a set of documents manually, humans rarely write a fully-formulated summary for themselves. Instead, user studies BIBREF3 , BIBREF4 show that they note down important keywords and phrases, try to identify relationships between them and organize them accordingly. Therefore, we believe that the study of summarization with similarly structured outputs is an important extension of the traditional task.
A representation that is more in line with observed user behavior is a concept map BIBREF5 , a labeled graph showing concepts as nodes and relationships between them as edges (Figure FIGREF2 ). Introduced in 1972 as a teaching tool BIBREF6 , concept maps have found many applications in education BIBREF7 , BIBREF8 , for writing assistance BIBREF9 or to structure information repositories BIBREF10 , BIBREF11 . For summarization, concept maps make it possible to represent a summary concisely and clearly reveal relationships. Moreover, we see a second interesting use case that goes beyond the capabilities of textual summaries: When concepts and relations are linked to corresponding locations in the documents they have been extracted from, the graph can be used to navigate in a document collection, similar to a table of contents. An implementation of this idea has been recently described by BIBREF12 .
The corresponding task that we propose is concept-map-based MDS, the summarization of a document cluster in the form of a concept map. In order to develop and evaluate methods for the task, gold-standard corpora are necessary, but no suitable corpus is available. The manual creation of such a dataset is very time-consuming, as the annotation includes many subtasks. In particular, an annotator would need to manually identify all concepts in the documents, while only a few of them will eventually end up in the summary.
To overcome these issues, we present a corpus creation method that effectively combines automatic preprocessing, scalable crowdsourcing and high-quality expert annotations. Using it, we can avoid the high effort for single annotators, allowing us to scale to document clusters that are 15 times larger than in traditional summarization corpora. We created a new corpus of 30 topics, each with around 40 source documents on educational topics and a summarizing concept map that is the consensus of many crowdworkers (see Figure FIGREF3 ).
As a crucial step of the corpus creation, we developed a new crowdsourcing scheme called low-context importance annotation. In contrast to traditional approaches, it allows us to determine important elements in a document cluster without requiring annotators to read all documents, making it feasible to crowdsource the task and overcome quality issues observed in previous work BIBREF13 . We show that the approach creates reliable data for our focused summarization scenario and, when tested on traditional summarization corpora, creates annotations that are similar to those obtained by earlier efforts.
To summarize, we make the following contributions: (1) We propose a novel task, concept-map-based MDS (§ SECREF2 ), (2) present a new crowdsourcing scheme to create reference summaries (§ SECREF4 ), (3) publish a new dataset for the proposed task (§ SECREF5 ) and (4) provide an evaluation protocol and baseline (§ SECREF7 ). We make these resources publicly available under a permissive license.
Task
Concept-map-based MDS is defined as follows: Given a set of related documents, create a concept map that represents its most important content, satisfies a specified size limit and is connected.
We define a concept map as a labeled graph showing concepts as nodes and relationships between them as edges. Labels are arbitrary sequences of tokens taken from the documents, making the summarization task extractive. A concept can be an entity, abstract idea, event or activity, designated by its unique label. Good maps should be propositionally coherent, meaning that every relation together with the two connected concepts form a meaningful proposition.
The task is complex, consisting of several interdependent subtasks. One has to extract appropriate labels for concepts and relations and recognize different expressions that refer to the same concept across multiple documents. Further, one has to select the most important concepts and relations for the summary and finally organize them in a graph satisfying the connectedness and size constraints.
Related Work
Some attempts have been made to automatically construct concept maps from text, working with either single documents BIBREF14 , BIBREF9 , BIBREF15 , BIBREF16 or document clusters BIBREF17 , BIBREF18 , BIBREF19 . These approaches extract concept and relation labels from syntactic structures and connect them to build a concept map. However, common task definitions and comparable evaluations are missing. In addition, only a few of them, namely Villalon.2012 and Valerio.2006, define summarization as their goal and try to compress the input to a substantially smaller size. Our newly proposed task and the created large-cluster dataset fill these gaps as they emphasize the summarization aspect of the task.
For the subtask of selecting summary-worthy concepts and relations, techniques developed for traditional summarization BIBREF20 and keyphrase extraction BIBREF21 are related and applicable. Approaches that build graphs of propositions to create a summary BIBREF22 , BIBREF23 , BIBREF24 , BIBREF25 seem to be particularly related, however, there is one important difference: While they use graphs as an intermediate representation from which a textual summary is then generated, the goal of the proposed task is to create a graph that is directly interpretable and useful for a user. In contrast, these intermediate graphs, e.g. AMR, are hardly useful for a typical, non-linguist user.
For traditional summarization, the most well-known datasets emerged out of the DUC and TAC competitions. They provide clusters of news articles with gold-standard summaries. Extending these efforts, several more specialized corpora have been created: With regard to size, Nakano.2010 present a corpus of summaries for large-scale collections of web pages. Recently, corpora with more heterogeneous documents have been suggested, e.g. BIBREF26 and BIBREF27 . The corpus we present combines these aspects, as it has large clusters of heterogeneous documents, and provides a necessary benchmark to evaluate the proposed task.
For concept map generation, one corpus with human-created summary concept maps for student essays has been created BIBREF28 . In contrast to our corpus, it only deals with single documents, requires a two orders of magnitude smaller amount of compression of the input and is not publicly available .
Other types of information representation that also model concepts and their relationships are knowledge bases, such as Freebase BIBREF29 , and ontologies. However, they both differ in important aspects: Whereas concept maps follow an open label paradigm and are meant to be interpretable by humans, knowledge bases and ontologies are usually more strictly typed and made to be machine-readable. Moreover, approaches to automatically construct them from text typically try to extract as much information as possible, while we want to summarize a document.
Low-Context Importance Annotation
Lloret.2013 describe several experiments to crowdsource reference summaries. Workers are asked to read 10 documents and then select 10 summary sentences from them for a reward of $0.05. They discovered several challenges, including poor work quality and the subjectiveness of the annotation task, indicating that crowdsourcing is not useful for this purpose.
To overcome these issues, we introduce a new task design, low-context importance annotation, to determine summary-worthy parts of documents. Compared to Lloret et al.'s approach, it is more in line with crowdsourcing best practices, as the tasks are simple, intuitive and small BIBREF30 and workers receive reasonable payment BIBREF31 . Most importantly, it is also much more efficient and scalable, as it does not require workers to read all documents in a cluster.
Task Design
We break down the task of importance annotation to the level of single propositions. The goal of our crowdsourcing scheme is to obtain a score for each proposition indicating its importance in a document cluster, such that a ranking according to the score would reveal what is most important and should be included in a summary. In contrast to other work, we do not show the documents to the workers at all, but provide only a description of the document cluster's topic along with the propositions. This ensures that tasks are small, simple and can be done quickly (see Figure FIGREF4 ).
In preliminary tests, we found that this design, despite the minimal context, works reasonably on our focused clusters on common educational topics. For instance, consider Figure FIGREF4 : One can easily say that P1 is more important than P2 without reading the documents.
We distinguish two task variants:
Instead of enforcing binary importance decisions, we use a 5-point Likert-scale to allow more fine-grained annotations. The obtained labels are translated into scores (5..1) and the average of all scores for a proposition is used as an estimate for its importance. This follows the idea that while single workers might find the task subjective, the consensus of multiple workers, represented in the average score, tends to be less subjective due to the “wisdom of the crowd”. We randomly group five propositions into a task.
As an alternative, we use a second task design based on pairwise comparisons. Comparisons are known to be easier to make and more consistent BIBREF32 , but also more expensive, as the number of pairs grows quadratically with the number of objects. To reduce the cost, we group five propositions into a task and ask workers to order them by importance per drag-and-drop. From the results, we derive pairwise comparisons and use TrueSkill BIBREF35 , a powerful Bayesian rank induction model BIBREF34 , to obtain importance estimates for each proposition.
Pilot Study
To verify the proposed approach, we conducted a pilot study on Amazon Mechanical Turk using data from TAC2008 BIBREF36 . We collected importance estimates for 474 propositions extracted from the first three clusters using both task designs. Each Likert-scale task was assigned to 5 different workers and awarded $0.06. For comparison tasks, we also collected 5 labels each, paid $0.05 and sampled around 7% of all possible pairs. We submitted them in batches of 100 pairs and selected pairs for subsequent batches based on the confidence of the TrueSkill model.
Following the observations of Lloret.2013, we established several measures for quality control. First, we restricted our tasks to workers from the US with an approval rate of at least 95%. Second, we identified low quality workers by measuring the correlation of each worker's Likert-scores with the average of the other four scores. The worst workers (at most 5% of all labels) were removed.
In addition, we included trap sentences, similar as in BIBREF13 , in around 80 of the tasks. In contrast to Lloret et al.'s findings, both an obvious trap sentence (This sentence is not important) and a less obvious but unimportant one (Barack Obama graduated from Harvard Law) were consistently labeled as unimportant (1.08 and 1.14), indicating that the workers did the task properly.
For Likert-scale tasks, we follow Snow.2008 and calculate agreement as the average Pearson correlation of a worker's Likert-score with the average score of the remaining workers. This measure is less strict than exact label agreement and can account for close labels and high- or low-scoring workers. We observe a correlation of 0.81, indicating substantial agreement. For comparisons, the majority agreement is 0.73. To further examine the reliability of the collected data, we followed the approach of Kiritchenko.2016 and simply repeated the crowdsourcing for one of the three topics. Between the importance estimates calculated from the first and second run, we found a Pearson correlation of 0.82 (Spearman 0.78) for Likert-scale tasks and 0.69 (Spearman 0.66) for comparison tasks. This shows that the approach, despite the subjectiveness of the task, allows us to collect reliable annotations.
In addition to the reliability studies, we extrinsically evaluated the annotations in the task of summary evaluation. For each of the 58 peer summaries in TAC2008, we calculated a score as the sum of the importance estimates of the propositions it contains. Table TABREF13 shows how these peer scores, averaged over the three topics, correlate with the manual responsiveness scores assigned during TAC in comparison to ROUGE-2 and Pyramid scores. The results demonstrate that with both task designs, we obtain importance annotations that are similarly useful for summary evaluation as pyramid annotations or gold-standard summaries (used for ROUGE).
Based on the pilot study, we conclude that the proposed crowdsourcing scheme allows us to obtain proper importance annotations for propositions. As workers are not required to read all documents, the annotation is much more efficient and scalable as with traditional methods.
Corpus Creation
This section presents the corpus construction process, as outlined in Figure FIGREF16 , combining automatic preprocessing, scalable crowdsourcing and high-quality expert annotations to be able to scale to the size of our document clusters. For every topic, we spent about $150 on crowdsourcing and 1.5h of expert annotations, while just a single annotator would already need over 8 hours (at 200 words per minute) to read all documents of a topic.
Source Data
As a starting point, we used the DIP corpus BIBREF37 , a collection of 49 clusters of 100 web pages on educational topics (e.g. bullying, homeschooling, drugs) with a short description of each topic. It was created from a large web crawl using state-of-the-art information retrieval. We selected 30 of the topics for which we created the necessary concept map annotations.
Proposition Extraction
As concept maps consist of propositions expressing the relation between concepts (see Figure FIGREF2 ), we need to impose such a structure upon the plain text in the document clusters. This could be done by manually annotating spans representing concepts and relations, however, the size of our clusters makes this a huge effort: 2288 sentences per topic (69k in total) need to be processed. Therefore, we resort to an automatic approach.
The Open Information Extraction paradigm BIBREF38 offers a representation very similar to the desired one. For instance, from
Students with bad credit history should not lose hope and apply for federal loans with the FAFSA.
Open IE systems extract tuples of two arguments and a relation phrase representing propositions:
(s. with bad credit history, should not lose, hope)
(s. with bad credit history, apply for, federal loans with the FAFSA)
While the relation phrase is similar to a relation in a concept map, many arguments in these tuples represent useful concepts. We used Open IE 4, a state-of-the-art system BIBREF39 to process all sentences. After removing duplicates, we obtained 4137 tuples per topic.
Since we want to create a gold-standard corpus, we have to ensure that we produce high-quality data. We therefore made use of the confidence assigned to every extracted tuple to filter out low quality ones. To ensure that we do not filter too aggressively (and miss important aspects in the final summary), we manually annotated 500 tuples sampled from all topics for correctness. On the first 250 of them, we tuned the filter threshold to 0.5, which keeps 98.7% of the correct extractions in the unseen second half. After filtering, a topic had on average 2850 propositions (85k in total).
Proposition Filtering
Despite the similarity of the Open IE paradigm, not every extracted tuple is a suitable proposition for a concept map. To reduce the effort in the subsequent steps, we therefore want to filter out unsuitable ones. A tuple is suitable if it (1) is a correct extraction, (2) is meaningful without any context and (3) has arguments that represent proper concepts. We created a guideline explaining when to label a tuple as suitable for a concept map and performed a small annotation study. Three annotators independently labeled 500 randomly sampled tuples. The agreement was 82% ( INLINEFORM0 ). We found tuples to be unsuitable mostly because they had unresolvable pronouns, conflicting with (2), or arguments that were full clauses or propositions, conflicting with (3), while (1) was mostly taken care of by the confidence filtering in § SECREF21 .
Due to the high number of tuples we decided to automate the filtering step. We trained a linear SVM on the majority voted annotations. As features, we used the extraction confidence, length of arguments and relations as well as part-of-speech tags, among others. To ensure that the automatic classification does not remove suitable propositions, we tuned the classifier to avoid false negatives. In particular, we introduced class weights, improving precision on the negative class at the cost of a higher fraction of positive classifications. Additionally, we manually verified a certain number of the most uncertain negative classifications to further improve performance. When 20% of the classifications are manually verified and corrected, we found that our model trained on 350 labeled instances achieves 93% precision on negative classifications on the unseen 150 instances. We found this to be a reasonable trade-off of automation and data quality and applied the model to the full dataset.
The classifier filtered out 43% of the propositions, leaving 1622 per topic. We manually examined the 17k least confident negative classifications and corrected 955 of them. We also corrected positive classifications for certain types of tuples for which we knew the classifier to be imprecise. Finally, each topic was left with an average of 1554 propositions (47k in total).
Importance Annotation
Given the propositions identified in the previous step, we now applied our crowdsourcing scheme as described in § SECREF4 to determine their importance. To cope with the large number of propositions, we combine the two task designs: First, we collect Likert-scores from 5 workers for each proposition, clean the data and calculate average scores. Then, using only the top 100 propositions according to these scores, we crowdsource 10% of all possible pairwise comparisons among them. Using TrueSkill, we obtain a fine-grained ranking of the 100 most important propositions.
For Likert-scores, the average agreement over all topics is 0.80, while the majority agreement for comparisons is 0.78. We repeated the data collection for three randomly selected topics and found the Pearson correlation between both runs to be 0.73 (Spearman 0.73) for Likert-scores and 0.72 (Spearman 0.71) for comparisons. These figures show that the crowdsourcing approach works on this dataset as reliably as on the TAC documents.
In total, we uploaded 53k scoring and 12k comparison tasks to Mechanical Turk, spending $4425.45 including fees. From the fine-grained ranking of the 100 most important propositions, we select the top 50 per topic to construct a summary concept map in the subsequent steps.
Proposition Revision
Having a manageable number of propositions, an annotator then applied a few straightforward transformations that correct common errors of the Open IE system. First, we break down propositions with conjunctions in either of the arguments into separate propositions per conjunct, which the Open IE system sometimes fails to do. And second, we correct span errors that might occur in the argument or relation phrases, especially when sentences were not properly segmented. As a result, we have a set of high quality propositions for our concept map, consisting of, due to the first transformation, 56.1 propositions per topic on average.
Concept Map Construction
In this final step, we connect the set of important propositions to form a graph. For instance, given the following two propositions
(student, may borrow, Stafford Loan)
(the student, does not have, a credit history)
one can easily see, although the first arguments differ slightly, that both labels describe the concept student, allowing us to build a concept map with the concepts student, Stafford Loan and credit history. The annotation task thus involves deciding which of the available propositions to include in the map, which of their concepts to merge and, when merging, which of the available labels to use. As these decisions highly depend upon each other and require context, we decided to use expert annotators rather than crowdsource the subtasks.
Annotators were given the topic description and the most important, ranked propositions. Using a simple annotation tool providing a visualization of the graph, they could connect the propositions step by step. They were instructed to reach a size of 25 concepts, the recommended maximum size for a concept map BIBREF6 . Further, they should prefer more important propositions and ensure connectedness. When connecting two propositions, they were asked to keep the concept label that was appropriate for both propositions. To support the annotators, the tool used ADW BIBREF40 , a state-of-the-art approach for semantic similarity, to suggest possible connections. The annotation was carried out by graduate students with a background in NLP after receiving an introduction into the guidelines and tool and annotating a first example.
If an annotator was not able to connect 25 concepts, she was allowed to create up to three synthetic relations with freely defined labels, making the maps slightly abstractive. On average, the constructed maps have 0.77 synthetic relations, mostly connecting concepts whose relation is too obvious to be explicitly stated in text (e.g. between Montessori teacher and Montessori education).
To assess the reliability of this annotation step, we had the first three maps created by two annotators. We casted the task of selecting propositions to be included in the map as a binary decision task and observed an agreement of 84% ( INLINEFORM0 ). Second, we modeled the decision which concepts to join as a binary decision on all pairs of common concepts, observing an agreement of 95% ( INLINEFORM1 ). And finally, we compared which concept labels the annotators decided to include in the final map, observing 85% ( INLINEFORM2 ) agreement. Hence, the annotation shows substantial agreement BIBREF41 .
Corpus Analysis
In this section, we describe our newly created corpus, which, in addition to having summaries in the form of concept maps, differs from traditional summarization corpora in several aspects.
Document Clusters
The corpus consists of document clusters for 30 different topics. Each of them contains around 40 documents with on average 2413 tokens, which leads to an average cluster size of 97,880 token. With these characteristics, the document clusters are 15 times larger than typical DUC clusters of ten documents and five times larger than the 25-document-clusters (Table TABREF26 ). In addition, the documents are also more variable in terms of length, as the (length-adjusted) standard deviation is twice as high as in the other corpora. With these properties, the corpus represents an interesting challenge towards real-world application scenarios, in which users typically have to deal with much more than ten documents.
Because we used a large web crawl as the source for our corpus, it contains documents from a variety of genres. To further analyze this property, we categorized a sample of 50 documents from the corpus. Among them, we found professionally written articles and blog posts (28%), educational material for parents and kids (26%), personal blog posts (16%), forum discussions and comments (12%), commented link collections (12%) and scientific articles (6%).
In addition to the variety of genres, the documents also differ in terms of language use. To capture this property, we follow Zopf.2016 and compute, for every topic, the average Jensen-Shannon divergence between the word distribution of one document and the word distribution in the remaining documents. The higher this value is, the more the language differs between documents. We found the average divergence over all topics to be 0.3490, whereas it is 0.3019 in DUC 2004 and 0.3188 in TAC 2008A.
Concept Maps
As Table TABREF33 shows, each of the 30 reference concept maps has exactly 25 concepts and between 24 and 28 relations. Labels for both concepts and relations consist on average of 3.2 tokens, whereas the latter are a bit shorter in characters.
To obtain a better picture of what kind of text spans have been used as labels, we automatically tagged them with their part-of-speech and determined their head with a dependency parser. Concept labels tend to be headed by nouns (82%) or verbs (15%), while they also contain adjectives, prepositions and determiners. Relation labels, on the other hand, are almost always headed by a verb (94%) and contain prepositions, nouns and particles in addition. These distributions are very similar to those reported by Villalon.2010 for their (single-document) concept map corpus.
Analyzing the graph structure of the maps, we found that all of them are connected. They have on average 7.2 central concepts with more than one relation, while the remaining ones occur in only one proposition. We found that achieving a higher number of connections would mean compromising importance, i.e. including less important propositions, and decided against it.
Baseline Experiments
In this section, we briefly describe a baseline and evaluation scripts that we release, with a detailed documentation, along with the corpus.
Conclusion
In this work, we presented low-context importance annotation, a novel crowdsourcing scheme that we used to create a new benchmark corpus for concept-map-based MDS. The corpus has large-scale document clusters of heterogeneous web documents, posing a challenging summarization task. Together with the corpus, we provide implementations of a baseline method and evaluation scripts and hope that our efforts facilitate future research on this variant of summarization.
Acknowledgments
We would like to thank Teresa Botschen, Andreas Hanselowski and Markus Zopf for their help with the annotation work and Christian Meyer for his valuable feedback. This work has been supported by the German Research Foundation as part of the Research Training Group “Adaptive Preparation of Information from Heterogeneous Sources” (AIPHES) under grant No. GRK 1994/1.
| [
"Answer with content missing: (Evaluation Metrics section) Precision, Recall, F1-scores, Strict match, METEOR, ROUGE-2"
] | 4,263 | qasper | en | null | 072d3de1a7122730a13a31db3eede4113af2d920814f0aaa | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What type of evaluation is proposed for this task?
Answer: | [
"Answer with content missing: (Evaluation Metrics section) Precision, Recall, F1-scores, Strict match, METEOR, ROUGE-2"
] | qasper | 128 | 6 | |
What are the datasets used for evaluation? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Language model pretraining has advanced the state of the art in many NLP tasks ranging from sentiment analysis, to question answering, natural language inference, named entity recognition, and textual similarity. State-of-the-art pretrained models include ELMo BIBREF1, GPT BIBREF2, and more recently Bidirectional Encoder Representations from Transformers (Bert; BIBREF0). Bert combines both word and sentence representations in a single very large Transformer BIBREF3; it is pretrained on vast amounts of text, with an unsupervised objective of masked language modeling and next-sentence prediction and can be fine-tuned with various task-specific objectives.
In most cases, pretrained language models have been employed as encoders for sentence- and paragraph-level natural language understanding problems BIBREF0 involving various classification tasks (e.g., predicting whether any two sentences are in an entailment relationship; or determining the completion of a sentence among four alternative sentences). In this paper, we examine the influence of language model pretraining on text summarization. Different from previous tasks, summarization requires wide-coverage natural language understanding going beyond the meaning of individual words and sentences. The aim is to condense a document into a shorter version while preserving most of its meaning. Furthermore, under abstractive modeling formulations, the task requires language generation capabilities in order to create summaries containing novel words and phrases not featured in the source text, while extractive summarization is often defined as a binary classification task with labels indicating whether a text span (typically a sentence) should be included in the summary.
We explore the potential of Bert for text summarization under a general framework encompassing both extractive and abstractive modeling paradigms. We propose a novel document-level encoder based on Bert which is able to encode a document and obtain representations for its sentences. Our extractive model is built on top of this encoder by stacking several inter-sentence Transformer layers to capture document-level features for extracting sentences. Our abstractive model adopts an encoder-decoder architecture, combining the same pretrained Bert encoder with a randomly-initialized Transformer decoder BIBREF3. We design a new training schedule which separates the optimizers of the encoder and the decoder in order to accommodate the fact that the former is pretrained while the latter must be trained from scratch. Finally, motivated by previous work showing that the combination of extractive and abstractive objectives can help generate better summaries BIBREF4, we present a two-stage approach where the encoder is fine-tuned twice, first with an extractive objective and subsequently on the abstractive summarization task.
We evaluate the proposed approach on three single-document news summarization datasets representative of different writing conventions (e.g., important information is concentrated at the beginning of the document or distributed more evenly throughout) and summary styles (e.g., verbose vs. more telegraphic; extractive vs. abstractive). Across datasets, we experimentally show that the proposed models achieve state-of-the-art results under both extractive and abstractive settings. Our contributions in this work are three-fold: a) we highlight the importance of document encoding for the summarization task; a variety of recently proposed techniques aim to enhance summarization performance via copying mechanisms BIBREF5, BIBREF6, BIBREF7, reinforcement learning BIBREF8, BIBREF9, BIBREF10, and multiple communicating encoders BIBREF11. We achieve better results with a minimum-requirement model without using any of these mechanisms; b) we showcase ways to effectively employ pretrained language models in summarization under both extractive and abstractive settings; we would expect any improvements in model pretraining to translate in better summarization in the future; and c) the proposed models can be used as a stepping stone to further improve summarization performance as well as baselines against which new proposals are tested.
Background ::: Pretrained Language Models
Pretrained language models BIBREF1, BIBREF2, BIBREF0, BIBREF12, BIBREF13 have recently emerged as a key technology for achieving impressive gains in a wide variety of natural language tasks. These models extend the idea of word embeddings by learning contextual representations from large-scale corpora using a language modeling objective. Bidirectional Encoder Representations from Transformers (Bert; BIBREF0) is a new language representation model which is trained with a masked language modeling and a “next sentence prediction” task on a corpus of 3,300M words.
The general architecture of Bert is shown in the left part of Figure FIGREF2. Input text is first preprocessed by inserting two special tokens. [cls] is appended to the beginning of the text; the output representation of this token is used to aggregate information from the whole sequence (e.g., for classification tasks). And token [sep] is inserted after each sentence as an indicator of sentence boundaries. The modified text is then represented as a sequence of tokens $X=[w_1,w_2,\cdots ,w_n]$. Each token $w_i$ is assigned three kinds of embeddings: token embeddings indicate the meaning of each token, segmentation embeddings are used to discriminate between two sentences (e.g., during a sentence-pair classification task) and position embeddings indicate the position of each token within the text sequence. These three embeddings are summed to a single input vector $x_i$ and fed to a bidirectional Transformer with multiple layers:
where $h^0=x$ are the input vectors; $\mathrm {LN}$ is the layer normalization operation BIBREF14; $\mathrm {MHAtt}$ is the multi-head attention operation BIBREF3; superscript $l$ indicates the depth of the stacked layer. On the top layer, Bert will generate an output vector $t_i$ for each token with rich contextual information.
Pretrained language models are usually used to enhance performance in language understanding tasks. Very recently, there have been attempts to apply pretrained models to various generation problems BIBREF15, BIBREF16. When fine-tuning for a specific task, unlike ELMo whose parameters are usually fixed, parameters in Bert are jointly fine-tuned with additional task-specific parameters.
Background ::: Extractive Summarization
Extractive summarization systems create a summary by identifying (and subsequently concatenating) the most important sentences in a document. Neural models consider extractive summarization as a sentence classification problem: a neural encoder creates sentence representations and a classifier predicts which sentences should be selected as summaries. SummaRuNNer BIBREF7 is one of the earliest neural approaches adopting an encoder based on Recurrent Neural Networks. Refresh BIBREF8 is a reinforcement learning-based system trained by globally optimizing the ROUGE metric. More recent work achieves higher performance with more sophisticated model structures. Latent BIBREF17 frames extractive summarization as a latent variable inference problem; instead of maximizing the likelihood of “gold” standard labels, their latent model directly maximizes the likelihood of human summaries given selected sentences. Sumo BIBREF18 capitalizes on the notion of structured attention to induce a multi-root dependency tree representation of the document while predicting the output summary. NeuSum BIBREF19 scores and selects sentences jointly and represents the state of the art in extractive summarization.
Background ::: Abstractive Summarization
Neural approaches to abstractive summarization conceptualize the task as a sequence-to-sequence problem, where an encoder maps a sequence of tokens in the source document $\mathbf {x} = [x_1, ..., x_n]$ to a sequence of continuous representations $\mathbf {z} = [z_1, ..., z_n]$, and a decoder then generates the target summary $\mathbf {y} = [y_1, ..., y_m]$ token-by-token, in an auto-regressive manner, hence modeling the conditional probability: $p(y_1, ..., y_m|x_1, ..., x_n)$.
BIBREF20 and BIBREF21 were among the first to apply the neural encoder-decoder architecture to text summarization. BIBREF6 enhance this model with a pointer-generator network (PTgen) which allows it to copy words from the source text, and a coverage mechanism (Cov) which keeps track of words that have been summarized. BIBREF11 propose an abstractive system where multiple agents (encoders) represent the document together with a hierarchical attention mechanism (over the agents) for decoding. Their Deep Communicating Agents (DCA) model is trained end-to-end with reinforcement learning. BIBREF9 also present a deep reinforced model (DRM) for abstractive summarization which handles the coverage problem with an intra-attention mechanism where the decoder attends over previously generated words. BIBREF4 follow a bottom-up approach (BottomUp); a content selector first determines which phrases in the source document should be part of the summary, and a copy mechanism is applied only to preselected phrases during decoding. BIBREF22 propose an abstractive model which is particularly suited to extreme summarization (i.e., single sentence summaries), based on convolutional neural networks and additionally conditioned on topic distributions (TConvS2S).
Fine-tuning Bert for Summarization ::: Summarization Encoder
Although Bert has been used to fine-tune various NLP tasks, its application to summarization is not as straightforward. Since Bert is trained as a masked-language model, the output vectors are grounded to tokens instead of sentences, while in extractive summarization, most models manipulate sentence-level representations. Although segmentation embeddings represent different sentences in Bert, they only apply to sentence-pair inputs, while in summarization we must encode and manipulate multi-sentential inputs. Figure FIGREF2 illustrates our proposed Bert architecture for Summarization (which we call BertSum).
In order to represent individual sentences, we insert external [cls] tokens at the start of each sentence, and each [cls] symbol collects features for the sentence preceding it. We also use interval segment embeddings to distinguish multiple sentences within a document. For $sent_i$ we assign segment embedding $E_A$ or $E_B$ depending on whether $i$ is odd or even. For example, for document $[sent_1, sent_2, sent_3, sent_4, sent_5]$, we would assign embeddings $[E_A, E_B, E_A,E_B, E_A]$. This way, document representations are learned hierarchically where lower Transformer layers represent adjacent sentences, while higher layers, in combination with self-attention, represent multi-sentence discourse.
Position embeddings in the original Bert model have a maximum length of 512; we overcome this limitation by adding more position embeddings that are initialized randomly and fine-tuned with other parameters in the encoder.
Fine-tuning Bert for Summarization ::: Extractive Summarization
Let $d$ denote a document containing sentences $[sent_1, sent_2, \cdots , sent_m]$, where $sent_i$ is the $i$-th sentence in the document. Extractive summarization can be defined as the task of assigning a label $y_i \in \lbrace 0, 1\rbrace $ to each $sent_i$, indicating whether the sentence should be included in the summary. It is assumed that summary sentences represent the most important content of the document.
With BertSum, vector $t_i$ which is the vector of the $i$-th [cls] symbol from the top layer can be used as the representation for $sent_i$. Several inter-sentence Transformer layers are then stacked on top of Bert outputs, to capture document-level features for extracting summaries:
where $h^0=\mathrm {PosEmb}(T)$; $T$ denotes the sentence vectors output by BertSum, and function $\mathrm {PosEmb}$ adds sinusoid positional embeddings BIBREF3 to $T$, indicating the position of each sentence.
The final output layer is a sigmoid classifier:
where $h^L_i$ is the vector for $sent_i$ from the top layer (the $L$-th layer ) of the Transformer. In experiments, we implemented Transformers with $L=1, 2, 3$ and found that a Transformer with $L=2$ performed best. We name this model BertSumExt.
The loss of the model is the binary classification entropy of prediction $\hat{y}_i$ against gold label $y_i$. Inter-sentence Transformer layers are jointly fine-tuned with BertSum. We use the Adam optimizer with $\beta _1=0.9$, and $\beta _2=0.999$). Our learning rate schedule follows BIBREF3 with warming-up ($ \operatorname{\operatorname{warmup}}=10,000$):
Fine-tuning Bert for Summarization ::: Abstractive Summarization
We use a standard encoder-decoder framework for abstractive summarization BIBREF6. The encoder is the pretrained BertSum and the decoder is a 6-layered Transformer initialized randomly. It is conceivable that there is a mismatch between the encoder and the decoder, since the former is pretrained while the latter must be trained from scratch. This can make fine-tuning unstable; for example, the encoder might overfit the data while the decoder underfits, or vice versa. To circumvent this, we design a new fine-tuning schedule which separates the optimizers of the encoder and the decoder.
We use two Adam optimizers with $\beta _1=0.9$ and $\beta _2=0.999$ for the encoder and the decoder, respectively, each with different warmup-steps and learning rates:
where $\tilde{lr}_{\mathcal {E}}=2e^{-3}$, and $\operatorname{\operatorname{warmup}}_{\mathcal {E}}=20,000$ for the encoder and $\tilde{lr}_{\mathcal {D}}=0.1$, and $\operatorname{\operatorname{warmup}}_{\mathcal {D}}=10,000$ for the decoder. This is based on the assumption that the pretrained encoder should be fine-tuned with a smaller learning rate and smoother decay (so that the encoder can be trained with more accurate gradients when the decoder is becoming stable).
In addition, we propose a two-stage fine-tuning approach, where we first fine-tune the encoder on the extractive summarization task (Section SECREF8) and then fine-tune it on the abstractive summarization task (Section SECREF13). Previous work BIBREF4, BIBREF23 suggests that using extractive objectives can boost the performance of abstractive summarization. Also notice that this two-stage approach is conceptually very simple, the model can take advantage of information shared between these two tasks, without fundamentally changing its architecture. We name the default abstractive model BertSumAbs and the two-stage fine-tuned model BertSumExtAbs.
Experimental Setup
In this section, we describe the summarization datasets used in our experiments and discuss various implementation details.
Experimental Setup ::: Summarization Datasets
We evaluated our model on three benchmark datasets, namely the CNN/DailyMail news highlights dataset BIBREF24, the New York Times Annotated Corpus (NYT; BIBREF25), and XSum BIBREF22. These datasets represent different summary styles ranging from highlights to very brief one sentence summaries. The summaries also vary with respect to the type of rewriting operations they exemplify (e.g., some showcase more cut and paste operations while others are genuinely abstractive). Table TABREF12 presents statistics on these datasets (test set); example (gold-standard) summaries are provided in the supplementary material.
Experimental Setup ::: Summarization Datasets ::: CNN/DailyMail
contains news articles and associated highlights, i.e., a few bullet points giving a brief overview of the article. We used the standard splits of BIBREF24 for training, validation, and testing (90,266/1,220/1,093 CNN documents and 196,961/12,148/10,397 DailyMail documents). We did not anonymize entities. We first split sentences with the Stanford CoreNLP toolkit BIBREF26 and pre-processed the dataset following BIBREF6. Input documents were truncated to 512 tokens.
Experimental Setup ::: Summarization Datasets ::: NYT
contains 110,540 articles with abstractive summaries. Following BIBREF27, we split these into 100,834/9,706 training/test examples, based on the date of publication (the test set contains all articles published from January 1, 2007 onward). We used 4,000 examples from the training as validation set. We also followed their filtering procedure, documents with summaries less than 50 words were removed from the dataset. The filtered test set (NYT50) includes 3,452 examples. Sentences were split with the Stanford CoreNLP toolkit BIBREF26 and pre-processed following BIBREF27. Input documents were truncated to 800 tokens.
Experimental Setup ::: Summarization Datasets ::: XSum
contains 226,711 news articles accompanied with a one-sentence summary, answering the question “What is this article about?”. We used the splits of BIBREF22 for training, validation, and testing (204,045/11,332/11,334) and followed the pre-processing introduced in their work. Input documents were truncated to 512 tokens.
Aside from various statistics on the three datasets, Table TABREF12 also reports the proportion of novel bi-grams in gold summaries as a measure of their abstractiveness. We would expect models with extractive biases to perform better on datasets with (mostly) extractive summaries, and abstractive models to perform more rewrite operations on datasets with abstractive summaries. CNN/DailyMail and NYT are somewhat abstractive, while XSum is highly abstractive.
Experimental Setup ::: Implementation Details
For both extractive and abstractive settings, we used PyTorch, OpenNMT BIBREF28 and the `bert-base-uncased' version of Bert to implement BertSum. Both source and target texts were tokenized with Bert's subwords tokenizer.
Experimental Setup ::: Implementation Details ::: Extractive Summarization
All extractive models were trained for 50,000 steps on 3 GPUs (GTX 1080 Ti) with gradient accumulation every two steps. Model checkpoints were saved and evaluated on the validation set every 1,000 steps. We selected the top-3 checkpoints based on the evaluation loss on the validation set, and report the averaged results on the test set. We used a greedy algorithm similar to BIBREF7 to obtain an oracle summary for each document to train extractive models. The algorithm generates an oracle consisting of multiple sentences which maximize the ROUGE-2 score against the gold summary.
When predicting summaries for a new document, we first use the model to obtain the score for each sentence. We then rank these sentences by their scores from highest to lowest, and select the top-3 sentences as the summary.
During sentence selection we use Trigram Blocking to reduce redundancy BIBREF9. Given summary $S$ and candidate sentence $c$, we skip $c$ if there exists a trigram overlapping between $c$ and $S$. The intuition is similar to Maximal Marginal Relevance (MMR; BIBREF29); we wish to minimize the similarity between the sentence being considered and sentences which have been already selected as part of the summary.
Experimental Setup ::: Implementation Details ::: Abstractive Summarization
In all abstractive models, we applied dropout (with probability $0.1$) before all linear layers; label smoothing BIBREF30 with smoothing factor $0.1$ was also used. Our Transformer decoder has 768 hidden units and the hidden size for all feed-forward layers is 2,048. All models were trained for 200,000 steps on 4 GPUs (GTX 1080 Ti) with gradient accumulation every five steps. Model checkpoints were saved and evaluated on the validation set every 2,500 steps. We selected the top-3 checkpoints based on their evaluation loss on the validation set, and report the averaged results on the test set.
During decoding we used beam search (size 5), and tuned the $\alpha $ for the length penalty BIBREF31 between $0.6$ and 1 on the validation set; we decode until an end-of-sequence token is emitted and repeated trigrams are blocked BIBREF9. It is worth noting that our decoder applies neither a copy nor a coverage mechanism BIBREF6, despite their popularity in abstractive summarization. This is mainly because we focus on building a minimum-requirements model and these mechanisms may introduce additional hyper-parameters to tune. Thanks to the subwords tokenizer, we also rarely observe issues with out-of-vocabulary words in the output; moreover, trigram-blocking produces diverse summaries managing to reduce repetitions.
Results ::: Automatic Evaluation
We evaluated summarization quality automatically using ROUGE BIBREF32. We report unigram and bigram overlap (ROUGE-1 and ROUGE-2) as a means of assessing informativeness and the longest common subsequence (ROUGE-L) as a means of assessing fluency. Table TABREF23 summarizes our results on the CNN/DailyMail dataset. The first block in the table includes the results of an extractive Oracle system as an upper bound. We also present the Lead-3 baseline (which simply selects the first three sentences in a document). The second block in the table includes various extractive models trained on the CNN/DailyMail dataset (see Section SECREF5 for an overview). For comparison to our own model, we also implemented a non-pretrained Transformer baseline (TransformerExt) which uses the same architecture as BertSumExt, but with fewer parameters. It is randomly initialized and only trained on the summarization task. TransformerExt has 6 layers, the hidden size is 512, and the feed-forward filter size is 2,048. The model was trained with same settings as in BIBREF3. The third block in Table TABREF23 highlights the performance of several abstractive models on the CNN/DailyMail dataset (see Section SECREF6 for an overview). We also include an abstractive Transformer baseline (TransformerAbs) which has the same decoder as our abstractive BertSum models; the encoder is a 6-layer Transformer with 768 hidden size and 2,048 feed-forward filter size. The fourth block reports results with fine-tuned Bert models: BertSumExt and its two variants (one without interval embeddings, and one with the large version of Bert), BertSumAbs, and BertSumExtAbs. Bert-based models outperform the Lead-3 baseline which is not a strawman; on the CNN/DailyMail corpus it is indeed superior to several extractive BIBREF7, BIBREF8, BIBREF19 and abstractive models BIBREF6. Bert models collectively outperform all previously proposed extractive and abstractive systems, only falling behind the Oracle upper bound. Among Bert variants, BertSumExt performs best which is not entirely surprising; CNN/DailyMail summaries are somewhat extractive and even abstractive models are prone to copying sentences from the source document when trained on this dataset BIBREF6. Perhaps unsurprisingly we observe that larger versions of Bert lead to performance improvements and that interval embeddings bring only slight gains. Table TABREF24 presents results on the NYT dataset. Following the evaluation protocol in BIBREF27, we use limited-length ROUGE Recall, where predicted summaries are truncated to the length of the gold summaries. Again, we report the performance of the Oracle upper bound and Lead-3 baseline. The second block in the table contains previously proposed extractive models as well as our own Transformer baseline. Compress BIBREF27 is an ILP-based model which combines compression and anaphoricity constraints. The third block includes abstractive models from the literature, and our Transformer baseline. Bert-based models are shown in the fourth block. Again, we observe that they outperform previously proposed approaches. On this dataset, abstractive Bert models generally perform better compared to BertSumExt, almost approaching Oracle performance.
Table TABREF26 summarizes our results on the XSum dataset. Recall that summaries in this dataset are highly abstractive (see Table TABREF12) consisting of a single sentence conveying the gist of the document. Extractive models here perform poorly as corroborated by the low performance of the Lead baseline (which simply selects the leading sentence from the document), and the Oracle (which selects a single-best sentence in each document) in Table TABREF26. As a result, we do not report results for extractive models on this dataset. The second block in Table TABREF26 presents the results of various abstractive models taken from BIBREF22 and also includes our own abstractive Transformer baseline. In the third block we show the results of our Bert summarizers which again are superior to all previously reported models (by a wide margin).
Results ::: Model Analysis ::: Learning Rates
Recall that our abstractive model uses separate optimizers for the encoder and decoder. In Table TABREF27 we examine whether the combination of different learning rates ($\tilde{lr}_{\mathcal {E}}$ and $\tilde{lr}_{\mathcal {D}}$) is indeed beneficial. Specifically, we report model perplexity on the CNN/DailyMail validation set for varying encoder/decoder learning rates. We can see that the model performs best with $\tilde{lr}_{\mathcal {E}}=2e-3$ and $\tilde{lr}_{\mathcal {D}}=0.1$.
Results ::: Model Analysis ::: Position of Extracted Sentences
In addition to the evaluation based on ROUGE, we also analyzed in more detail the summaries produced by our model. For the extractive setting, we looked at the position (in the source document) of the sentences which were selected to appear in the summary. Figure FIGREF31 shows the proportion of selected summary sentences which appear in the source document at positions 1, 2, and so on. The analysis was conducted on the CNN/DailyMail dataset for Oracle summaries, and those produced by BertSumExt and the TransformerExt. We can see that Oracle summary sentences are fairly smoothly distributed across documents, while summaries created by TransformerExt mostly concentrate on the first document sentences. BertSumExt outputs are more similar to Oracle summaries, indicating that with the pretrained encoder, the model relies less on shallow position features, and learns deeper document representations.
Results ::: Model Analysis ::: Novel N-grams
We also analyzed the output of abstractive systems by calculating the proportion of novel n-grams that appear in the summaries but not in the source texts. The results are shown in Figure FIGREF33. In the CNN/DailyMail dataset, the proportion of novel n-grams in automatically generated summaries is much lower compared to reference summaries, but in XSum, this gap is much smaller. We also observe that on CNN/DailyMail, BertExtAbs produces less novel n-ngrams than BertAbs, which is not surprising. BertExtAbs is more biased towards selecting sentences from the source document since it is initially trained as an extractive model. The supplementary material includes examples of system output and additional ablation studies.
Results ::: Human Evaluation
In addition to automatic evaluation, we also evaluated system output by eliciting human judgments. We report experiments following a question-answering (QA) paradigm BIBREF33, BIBREF8 which quantifies the degree to which summarization models retain key information from the document. Under this paradigm, a set of questions is created based on the gold summary under the assumption that it highlights the most important document content. Participants are then asked to answer these questions by reading system summaries alone without access to the article. The more questions a system can answer, the better it is at summarizing the document as a whole. Moreover, we also assessed the overall quality of the summaries produced by abstractive systems which due to their ability to rewrite content may produce disfluent or ungrammatical output. Specifically, we followed the Best-Worst Scaling BIBREF34 method where participants were presented with the output of two systems (and the original document) and asked to decide which one was better according to the criteria of Informativeness, Fluency, and Succinctness.
Both types of evaluation were conducted on the Amazon Mechanical Turk platform. For the CNN/DailyMail and NYT datasets we used the same documents (20 in total) and questions from previous work BIBREF8, BIBREF18. For XSum, we randomly selected 20 documents (and their questions) from the release of BIBREF22. We elicited 3 responses per HIT. With regard to QA evaluation, we adopted the scoring mechanism from BIBREF33; correct answers were marked with a score of one, partially correct answers with 0.5, and zero otherwise. For quality-based evaluation, the rating of each system was computed as the percentage of times it was chosen as better minus the times it was selected as worse. Ratings thus range from -1 (worst) to 1 (best).
Results for extractive and abstractive systems are shown in Tables TABREF37 and TABREF38, respectively. We compared the best performing BertSum model in each setting (extractive or abstractive) against various state-of-the-art systems (whose output is publicly available), the Lead baseline, and the Gold standard as an upper bound. As shown in both tables participants overwhelmingly prefer the output of our model against comparison systems across datasets and evaluation paradigms. All differences between BertSum and comparison models are statistically significant ($p<0.05$), with the exception of TConvS2S (see Table TABREF38; XSum) in the QA evaluation setting.
Conclusions
In this paper, we showcased how pretrained Bert can be usefully applied in text summarization. We introduced a novel document-level encoder and proposed a general framework for both abstractive and extractive summarization. Experimental results across three datasets show that our model achieves state-of-the-art results across the board under automatic and human-based evaluation protocols. Although we mainly focused on document encoding for summarization, in the future, we would like to take advantage the capabilities of Bert for language generation.
Acknowledgments
This research is supported by a Google PhD Fellowship to the first author. We gratefully acknowledge the support of the European Research Council (Lapata, award number 681760, “Translating Multiple Modalities into Text”). We would also like to thank Shashi Narayan for providing us with the XSum dataset.
| [
"CNN/DailyMail news highlights, New York Times Annotated Corpus, XSum",
"the CNN/DailyMail news highlights dataset BIBREF24, the New York Times Annotated Corpus (NYT; BIBREF25), XSum BIBREF22"
] | 4,369 | qasper | en | null | 8fa5af6a36dd0b6b73900b2ec6f6e43a652a3e7d2b827a58 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What are the datasets used for evaluation?
Answer: | [
"CNN/DailyMail news highlights, New York Times Annotated Corpus, XSum",
"the CNN/DailyMail news highlights dataset BIBREF24, the New York Times Annotated Corpus (NYT; BIBREF25), XSum BIBREF22"
] | qasper | 128 | 7 | |
How does this approach compare to other WSD approaches employing word embeddings? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Language modelling in its inception had one-hot vector encoding of words. However, it captures only alphabetic ordering but not the word semantic similarity. Vector space models helps to learn word representations in a lower dimensional space and also captures semantic similarity. Learning word embedding aids in natural language processing tasks such as question answering and reasoning BIBREF0, stance detection BIBREF1, claim verification BIBREF2.
Recent models BIBREF3, BIBREF4 work on the basis that words with similar context share semantic similarity. BIBREF4 proposes a neural probabilistic model which models the target word probability conditioned on the previous words using a recurrent neural network. Word2Vec models BIBREF3 such as continuous bag-of-words (CBOW) predict the target word given the context, and skip-gram model works in reverse of predicting the context given the target word. While, GloVe embeddings were based on a Global matrix factorization on local contexts BIBREF5. However, the aforementioned models do not handle words with multiple meanings (polysemies).
BIBREF6 proposes a neural network approach considering both local and global contexts in learning word embeddings (point estimates). Their multiple prototype model handles polysemous words by providing apriori heuristics about word senses in the dataset. BIBREF7 proposes an alternative to handle polysemous words by a modified skip-gram model and EM algorithm. BIBREF8 presents a non-parametric based alternative to handle polysemies. However, these approaches fail to consider entailment relations among the words. BIBREF9 learn a Gaussian distribution per word using the expected likelihood kernel. However, for polysemous words, this may lead to word distributions with larger variances as it may have to cover various senses.
BIBREF10 proposes multimodal word distribution approach. It captures polysemy. However, the energy based objective function fails to consider asymmetry and hence entailment. Textual entailment recognition is necessary to capture lexical inference relations such as causality (for example, mosquito $\rightarrow $ malaria), hypernymy (for example, dog $\models $ animal) etc.
In this paper, we propose to obtain multi-sense word embedding distributions by using a variant of max margin objective based on the asymmetric KL divergence energy function to capture textual entailment. Multi-sense distributions are advantageous in capturing polysemous nature of words and in reducing the uncertainty per word by distributing it across senses. However, computing KL divergence between mixtures of Gaussians is intractable, and we use a KL divergence approximation based on stricter upper and lower bounds. While capturing textual entailment (asymmetry), we have also not compromised on capturing symmetrical similarity between words (for example, funny and hilarious) which will be elucidated in Section $3.1$. We also show the effectiveness of the proposed approach on the benchmark word similarity and entailment datasets in the experimental section.
Methodology ::: Word Representation
Probabilistic representation of words helps one model uncertainty in word representation, and polysemy. Given a corpus $V$, containing a list of words each represented as $w$, the probability density for a word $w$ can be represented as a mixture of Gaussians with $C$ components BIBREF10.
Here, $p_{w,j}$ represents the probability of word $w$ belonging to the component $j$, $\operatorname{\mathbf {\mu }}_{w,j}$ represents $D$ dimensional word representation corresponding to the $j^{th}$ component sense of the word $w$, and $\Sigma _{w,j}$ represents the uncertainty in representation for word $w$ belonging to component $j$.
Objective function
The model parameters (means, covariances and mixture weights) $\theta $ can be learnt using a variant of max-margin objective BIBREF11.
Here $E_\theta (\cdot , \cdot )$ represents an energy function which assigns a score to the pair of words, $w$ is a particular word under consideration, $c$ its positive context (same context), and $c^{\prime }$ the negative context. The objective aims to push the margin of the difference between the energy function of a word $w$ to its positive context $c$ higher than its negative context $c$ by a threshold of $m$. Thus, word pairs in the same context gets a higher energy than the word pairs in the dissimilar context. BIBREF10 consider the energy function to be an expected likelihood kernel which is defined as follows.
This is similar to the cosine similarity metric over vectors and the energy between two words is maximum when they have similar distributions. But, the expected likelihood kernel is a symmetric metric which will not be suitable for capturing ordering among words and hence entailment.
Objective function ::: Proposed Energy function
As each word is represented by a mixture of Gaussian distributions, KL divergence is a better choice of energy function to capture distance between distributions. Since, KL divergence is minimum when the distributions are similar and maximum when they are dissimilar, energy function is taken as exponentiated negative KL divergence.
However, computing KL divergence between Gaussian mixtures is intractable and obtaining exact KL value is not possible. One way of approximating the KL is by Monte-Carlo approximation but it requires large number of samples to get a good approximation and is computationally expensive on high dimensional embedding space.
Alternatively, BIBREF12 presents a KL approximation between Gaussian mixtures where they obtain an upper bound through product of Gaussian approximation method and a lower bound through variational approximation method. In BIBREF13, the authors combine the lower and upper bounds from approximation methods of BIBREF12 to provide a stricter bound on KL between Gaussian mixtures. Lets consider Gaussian mixtures for the words $w$ and $v$ as follows.
The approximate KL divergence between the Gaussian mixture representations over the words $w$ and $v$ is shown in equation DISPLAY_FORM8. More details on approximation is included in the Supplementary Material.
where $EL_{ik}(w,w) = \int f_{w,i} (\operatorname{\mathbf {x}}) f_{w,k} (\operatorname{\mathbf {x}}) d\operatorname{\mathbf {x}}$ and $EL_{ij}(w,v) = \int f_{w,i} (\operatorname{\mathbf {x}}) f_{v,k} (\operatorname{\mathbf {x}}) d\operatorname{\mathbf {x}}$. Note that the expected likelihood kernel appears component wise inside the approximate KL divergence derivation.
One advantage of using KL as energy function is that it enables to capture asymmetry in entailment datasets. For eg., let us consider the words 'chair' with two senses as 'bench' and 'sling', and 'wood' with two senses as 'trees' and 'furniture'. The word chair ($w$) is entailed within wood ($v$), i.e. chair $\models $ wood. Now, minimizing the KL divergence necessitates maximizing $\log {\sum _j p_{v,j} \exp ({-KL(f_{w,i} (\operatorname{\mathbf {x}})||f_{v,j}(\operatorname{\mathbf {x}}))})}$ which in turn minimizes $KL(f_{w,i}(\operatorname{\mathbf {x}})||f_{v,j}(\operatorname{\mathbf {x}}))$. This will result in the support of the $i^{th}$ component of $w$ to be within the $j^{th}$ component of $v$, and holds for all component pairs leading to the entailment of $w$ within $v$. Consequently, we can see that bench $\models $ trees, bench $\models $ furniture, sling $\models $ trees, and sling $\models $ furniture. Thus, it introduces lexical relationship between the senses of child word and that of the parent word. Minimizing the KL also necessitates maximizing $\log {\sum _j {p_{v,j}} EL_{ij}(w,v)}$ term for all component pairs among $w$ and $v$. This is similar to maximizing expected likelihood kernel, which brings the means of $f_{w,i}(\operatorname{\mathbf {x}})$ and $f_{v,j}(\operatorname{\mathbf {x}})$ closer (weighted by their co-variances) as discussed in BIBREF10. Hence, the proposed approach captures the best of both worlds, thereby catering to both word similarity and entailment.
We also note that minimizing the KL divergence necessitates minimizing $\log {\sum _k p_{w,k} \exp ({-KL(f_{w,i}||f_{w,k})})}$ which in turn maximizes $KL(f_{w,i}||f_{w,k})$. This prevents the different mixture components of a word converging to single Gaussian and encourages capturing different possible senses of the word. The same is also achieved by minimizing $\sum _k {p_{w,k}} EL_{ik}(w,w)$ term and act as a regularization term which promotes diversity in learning senses of a word.
Experimentation and Results
We train our proposed model GM$\_$KL (Gaussian Mixture using KL Divergence) on the Text8 dataset BIBREF14 which is a pre-processed data of $17M$ words from wikipedia. Of which, 71290 unique and frequent words are chosen using the subsampling trick in BIBREF15. We compare GM$\_$KL with the previous approaches w2g BIBREF9 ( single Gaussian model) and w2gm BIBREF10 (mixture of Gaussian model with expected likelihood kernel). For all the models used for experimentation, the embedding size ($D$) was set to 50, number of mixtures to 2, context window length to 10, batch size to 128. The word embeddings were initialized using a uniform distribution in the range of $[-\sqrt{\frac{3}{D}}$, $\sqrt{\frac{3}{D}}]$ such that the expectation of variance is 1 and mean 0 BIBREF16. One could also consider initializing the word embeddings using other contextual representations such as BERT BIBREF17 and ELMo BIBREF18 in the proposed approach. In order to purely analyze the performance of $\emph {GM\_KL}$ over the other models, we have chosen initialization using uniform distribution for experiments. For computational benefits, diagonal covariance is used similar to BIBREF10. Each mixture probability is constrained in the range $[0,1]$, summing to 1 by optimizing over unconstrained scores in the range $(-\infty ,\infty )$ and converting scores to probability using softmax function. The mixture scores are initialized to 0 to ensure fairness among all the components. The threshold for negative sampling was set to $10^{-5}$, as recommended in BIBREF3. Mini-batch gradient descent with Adagrad optimizer BIBREF19 was used with initial learning rate set to $0.05$.
Table TABREF9 shows the qualitative results of GM$\_$KL. Given a query word and component id, the set of nearest neighbours along with their respective component ids are listed. For eg., the word `plane' in its 0th component captures the `geometry' sense and so are its neighbours, and its 1st component captures `vehicle' sense and so are its corresponding neighbours. Other words such as `rock' captures both the `metal' and `music' senses, `star' captures `celebrity' and `astronomical' senses, and `phone' captures `telephony' and `internet' senses.
We quantitatively compare the performance of the GM$\_$KL, w2g, and w2gm approaches on the SCWS dataset BIBREF6. The dataset consists of 2003 word pairs of polysemous and homonymous words with labels obtained by an average of 10 human scores. The Spearman correlation between the human scores and the model scores are computed. To obtain the model score, the following metrics are used:
MaxCos: Maximum cosine similarity among all component pairs of words $w$ and $v$:
AvgCos: Average component-wise cosine similarity between the words $w$ and $v$.
KL$\_$approx: Formulated as shown in (DISPLAY_FORM8) between the words $w$ and $v$.
KL$\_$comp: Maximum component-wise negative KL between words $w$ and $v$:
Table TABREF17 compares the performance of the approaches on the SCWS dataset. It is evident from Table TABREF17 that GM$\_$KL achieves better correlation than existing approaches for various metrics on SCWS dataset.
Table TABREF18 shows the Spearman correlation values of GM$\_$KL model evaluated on the benchmark word similarity datasets: SL BIBREF20, WS, WS-R, WS-S BIBREF21, MEN BIBREF22, MC BIBREF23, RG BIBREF24, YP BIBREF25, MTurk-287 and MTurk-771 BIBREF26, BIBREF27, and RW BIBREF28. The metric used for comparison is 'AvgCos'. It can be seen that for most of the datasets, GM$\_$KL achieves significantly better correlation score than w2g and w2gm approaches. Other datasets such as MC and RW consist of only a single sense, and hence w2g model performs better and GM$\_$KL achieves next better performance. The YP dataset have multiple senses but does not contain entailed data and hence could not make use of entailment benefits of GM$\_$KL.
Table TABREF19 shows the evaluation results of GM$\_$KL model on the entailment datasets such as entailment pairs dataset BIBREF29 created from WordNet with both positive and negative labels, a crowdsourced dataset BIBREF30 of 79 semantic relations labelled as entailed or not and annotated distributionally similar nouns dataset BIBREF31. The 'MaxCos' similarity metric is used for evaluation and the best precision and best F1-score is shown, by picking the optimal threshold. Overall, GM$\_$KL performs better than both w2g and w2gm approaches.
Conclusion
We proposed a KL divergence based energy function for learning multi-sense word embedding distributions modelled as Gaussian mixtures. Due to the intractability of the Gaussian mixtures for the KL divergence measure, we use an approximate KL divergence function. We also demonstrated that the proposed GM$\_$KL approaches performed better than other approaches on the benchmark word similarity and entailment datasets.
tocsectionAppendices
Approximation for KL divergence between mixtures of gaussians
KL between gaussian mixtures $f_{w}(\operatorname{\mathbf {x}})$ and $f_{v}(\operatorname{\mathbf {x}})$ can be decomposed as:
BIBREF12 presents KL approximation between gaussian mixtures using
product of gaussian approximation method where KL is approximated using product of component gaussians and
variational approximation method where KL is approximated by introducing some variational parameters.
The product of component gaussian approximation method using Jensen's inequality provides upper bounds as shown in equations DISPLAY_FORM23 and .
The variational approximation method provides lower bounds as shown in equations DISPLAY_FORM24 and DISPLAY_FORM25.
where $H$ represents the entropy term and the entropy of $i^{th}$ component of word $w$ with dimension $D$ is given as
In BIBREF13, the authors combine the lower and upper bounds from approximation methods of BIBREF12 to formulate a stricter bound on KL between gaussian mixtures.
From equations DISPLAY_FORM23 and DISPLAY_FORM25, a stricter lower bound for KL between gaussian mixtures is obtained as shown in equation DISPLAY_FORM26
From equations and DISPLAY_FORM24, a stricter upper bound for KL between gaussian mixtures is obtained as shown in equation DISPLAY_FORM27
Finally, the KL between gaussian mixtures is taken as the mean of KL upper and lower bounds as shown in equation DISPLAY_FORM28.
| [
"GM$\\_$KL achieves better correlation than existing approaches for various metrics on SCWS dataset."
] | 2,189 | qasper | en | null | 5f00d4f6e62f4b99484eb78491f803f8143cc1b13ad33816 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How does this approach compare to other WSD approaches employing word embeddings?
Answer: | [
"GM$\\_$KL achieves better correlation than existing approaches for various metrics on SCWS dataset."
] | qasper | 128 | 8 | |
How does their ensemble method work? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Since humans amass more and more generally available data in the form of unstructured text it would be very useful to teach machines to read and comprehend such data and then use this understanding to answer our questions. A significant amount of research has recently focused on answering one particular kind of questions the answer to which depends on understanding a context document. These are cloze-style questions BIBREF0 which require the reader to fill in a missing word in a sentence. An important advantage of such questions is that they can be generated automatically from a suitable text corpus which allows us to produce a practically unlimited amount of them. That opens the task to notoriously data-hungry deep-learning techniques which now seem to outperform all alternative approaches.
Two such large-scale datasets have recently been proposed by researchers from Google DeepMind and Facebook AI: the CNN/Daily Mail dataset BIBREF1 and the Children's Book Test (CBT) BIBREF2 respectively. These have attracted a lot of attention from the research community BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 with a new state-of-the-art model coming out every few weeks.
However if our goal is a production-level system actually capable of helping humans, we want the model to use all available resources as efficiently as possible. Given that
we believe that if the community is striving to bring the performance as far as possible, it should move its work to larger data.
This thinking goes in line with recent developments in the area of language modelling. For a long time models were being compared on several "standard" datasets with publications often presenting minuscule improvements in performance. Then the large-scale One Billion Word corpus dataset appeared BIBREF15 and it allowed Jozefowicz et al. to train much larger LSTM models BIBREF16 that almost halved the state-of-the-art perplexity on this dataset.
We think it is time to make a similar step in the area of text comprehension. Hence we are introducing the BookTest, a new dataset very similar to the Children's Book test but more than 60 times larger to enable training larger models even in the domain of text comprehension. Furthermore the methodology used to create our data can later be used to create even larger datasets when the need arises thanks to further technological progress.
We show that if we evaluate a model trained on the new dataset on the now standard Children's Book Test dataset, we see an improvement in accuracy much larger than other research groups achieved by enhancing the model architecture itself (while still using the original CBT training data). By training on the new dataset, we reduce the prediction error by almost one third. On the named-entity version of CBT this brings the ensemble of our models to the level of human baseline as reported by Facebook BIBREF2 . However in the final section we show in our own human study that there is still room for improvement on the CBT beyond the performance of our model.
Task Description
A natural way of testing a reader's comprehension of a text is to ask her a question the answer to which can be deduced from the text. Hence the task we are trying to solve consists of answering a cloze-style question, the answer to which depends on the understanding of a context document provided with the question. The model is also provided with a set of possible answers from which the correct one is to be selected. This can be formalized as follows:
The training data consist of tuples INLINEFORM0 , where INLINEFORM1 is a question, INLINEFORM2 is a document that contains the answer to question INLINEFORM3 , INLINEFORM4 is a set of possible answers and INLINEFORM5 is the ground-truth answer. Both INLINEFORM6 and INLINEFORM7 are sequences of words from vocabulary INLINEFORM8 . We also assume that all possible answers are words from the vocabulary, that is INLINEFORM9 . In the CBT and CNN/Daily Mail datasets it is also true that the ground-truth answer INLINEFORM10 appears in the document. This is exploited by many machine learning models BIBREF2 , BIBREF4 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF10 , BIBREF11 , BIBREF12 , however some do not explicitly depend on this property BIBREF1 , BIBREF3 , BIBREF5 , BIBREF9
Current Landscape
We will now briefly review what datasets for text comprehension have been published up to date and look at models which have been recently applied to solving the task we have just described.
Datasets
A crucial condition for applying deep-learning techniques is to have a huge amount of data available for training. For question answering this specifically means having a large number of document-question-answer triples available. While there is an unlimited amount of text available, coming up with relevant questions and the corresponding answers can be extremely labour-intensive if done by human annotators. There were efforts to provide such human-generated datasets, e.g. Microsoft's MCTest BIBREF17 , however their scale is not suitable for deep learning without pre-training on other data BIBREF18 (such as using pre-trained word embedding vectors).
Google DeepMind managed to avoid this scale issue with their way of generating document-question-answer triples automatically, closely followed by Facebook with a similar method. Let us now briefly introduce the two resulting datasets whose properties are summarized in Table TABREF8 .
These two datasets BIBREF1 exploit a useful feature of online news articles – many articles include a short summarizing sentence near the top of the page. Since all information in the summary sentence is also presented in the article body, we get a nice cloze-style question about the article contents by removing a word from the short summary.
The dataset's authors also replaced all named entities in the dataset by anonymous tokens which are further shuffled for each new batch. This forces the model to rely solely on information from the context document, not being able to transfer any meaning of the named entities between documents.
This restricts the task to one specific aspect of context-dependent question answering which may be useful however it moves the task further from the real application scenario, where we would like the model to use all information available to answer questions. Furthermore Chen et al. BIBREF5 have suggested that this can make about 17% of the questions unanswerable even by humans. They also claim that more than a half of the question sentences are mere paraphrases or exact matches of a single sentence from the context document. This raises a question to what extent the dataset can test deeper understanding of the articles.
The Children's Book Test BIBREF2 uses a different source - books freely available thanks to Project Gutenberg. Since no summary is available, each example consists of a context document formed from 20 consecutive sentences from the story together with a question formed from the subsequent sentence.
The dataset comes in four flavours depending on what type of word is omitted from the question sentence. Based on human evaluation done in BIBREF2 it seems that NE and CN are more context dependent than the other two types – prepositions and verbs. Therefore we (and all of the recent publications) focus only on these two word types.
Several new datasets related to the (now almost standard) ones above emerged recently. We will now briefly present them and explain how the dataset we are introducing in this article differs from them.
The LAMBADA dataset BIBREF19 is designed to measure progress in understanding common-sense questions about short stories that can be easily answered by humans but cannot be answered by current standard machine-learning models (e.g. plain LSTM language models). This dataset is useful for measuring the gap between humans and machine learning algorithms. However, by contrast to our BookTest dataset, it will not allow us to track progress towards the performance of the baseline systems or on examples where machine learning may show super-human performance. Also LAMBADA is just a diagnostic dataset and does not provide ready-to-use question-answering training data, just a plain-text corpus which may moreover include copyrighted books making its use potentially problematic for some purposes. We are providing ready training data consisting of copyright-free books only.
The SQuAD dataset BIBREF20 based on Wikipedia and the Who-did-What dataset BIBREF21 based on Gigaword news articles are factoid question-answering datasets where a multi-word answer should be extracted from a context document. This is in contrast to the previous datasets, including CNN/DM, CBT, LAMBADA and our new dataset, which require only single-word answers. Both these datasets however provide less than 130,000 training questions, two orders of magnitude less than our dataset does.
The Story Cloze Test BIBREF22 provides a crowd-sourced corpus of 49,255 commonsense stories for training and 3,744 testing stories with right and wrong endings. Hence the dataset is again rather small. Similarly to LAMBADA, the Story Cloze Test was designed to be easily answerable by humans.
In the WikiReading BIBREF23 dataset the context document is formed from a Wikipedia article and the question-answer pair is taken from the corresponding WikiData page. For each entity (e.g. Hillary Clinton), WikiData contain a number of property-value pairs (e.g. place of birth: Chicago) which form the datasets's question-answer pairs. The dataset is certainly relevant to the community, however the questions are of very limited variety with only 20 properties (and hence unique questions) covering INLINEFORM0 of the dataset. Furthermore many of the frequent properties are mentioned at a set spot within the article (e.g. the date of birth is almost always in brackets behind the name of a person) which may make the task easier for machines. We are trying to provide a more varied dataset.
Although there are several datasets related to task we are aiming to solve, they differ sufficiently for our dataset to bring new value to the community. Its biggest advantage is its size which can furthermore be easily upscaled without expensive human annotation. Finally while we are emphasizing the differences, models could certainly benefit from as diverse a collection of datasets as possible.
Machine Learning Models
A first major work applying deep-learning techniques to text comprehension was Hermann et al. BIBREF1 . This work was followed by the application of Memory Networks to the same task BIBREF2 . Later three models emerged around the same time BIBREF3 , BIBREF4 , BIBREF5 including our psr model BIBREF4 . The AS Reader inspired several subsequent models that use it as a sub-component in a diverse ensemble BIBREF8 ; extend it with a hierarchical structure BIBREF6 , BIBREF24 , BIBREF7 ; compute attention over the context document for every word in the query BIBREF10 or use two-way context-query attention mechanism for every word in the context and the query BIBREF11 that is similar in its spirit to models recently proposed in different domains, e.g. BIBREF25 in information retrieval. Other neural approaches to text comprehension are explored in BIBREF9 , BIBREF12 .
Possible Directions for Improvements
Accuracy in any machine learning tasks can be enhanced either by improving a machine learning model or by using more in-domain training data. Current state of the art models BIBREF6 , BIBREF7 , BIBREF8 , BIBREF11 improve over AS Reader's accuracy on CBT NE and CN datasets by 1-2 percent absolute. This suggests that with current techniques there is only limited room for improvement on the algorithmic side.
The other possibility to improve performance is simply to use more training data. The importance of training data was highlighted by the frequently quoted Mercer's statement that “There is no data like more data.” The observation that having more data is often more important than having better algorithms has been frequently stressed since then BIBREF13 , BIBREF14 .
As a step in the direction of exploiting the potential of more data in the domain of text comprehension, we created a new dataset called BookTest similar to, but much larger than the widely used CBT and CNN/DM datasets.
BookTest
Similarly to the CBT, our BookTest dataset is derived from books available through project Gutenberg. We used 3555 copyright-free books to extract CN examples and 10507 books for NE examples, for comparison the CBT dataset was extracted from just 108 books.
When creating our dataset we follow the same procedure as was used to create the CBT dataset BIBREF2 . That is, we detect whether each sentence contains either a named entity or a common noun that already appeared in one of the preceding twenty sentences. This word is then replaced by a gap tag (XXXXX) in this sentence which is hence turned into a cloze-style question. The preceding 20 sentences are used as the context document. For common noun and named entity detection we use the Stanford POS tagger BIBREF27 and Stanford NER BIBREF28 .
The training dataset consists of the original CBT NE and CN data extended with new NE and CN examples. The new BookTest dataset hence contains INLINEFORM0 training examples and INLINEFORM1 tokens.
The validation dataset consists of INLINEFORM0 NE and INLINEFORM1 CN questions. We have one test set for NE and one for CN, each containing INLINEFORM2 examples. The training, validation and test sets were generated from non-overlapping sets of books.
When generating the dataset we removed all editions of books used to create CBT validation and test sets from our training dataset. Therefore the models trained on the BookTest corpus can be evaluated on the original CBT data and they can be compared with recent text-comprehension models utilizing this dataset BIBREF2 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 .
Baselines
We will now use our psr model to evaluate the performance gain from increasing the dataset size.
AS Reader
In BIBREF4 we introduced the psr , which at the time of publication significantly outperformed all other architectures on the CNN, DM and CBT datasets. This model is built to leverage the fact that the answer is a single word from the context document. Similarly to many other models it uses attention over the document – intuitively a measure of how relevant each word is to answering the question. However while most previous models used this attention as weights to calculate a blended representation of the answer word, we simply sum the attention across all occurrences of each unique words and then simply select the word with the highest sum as the final answer. While simple, this trick seems both to improve accuracy and to speed-up training. It was adopted by many subsequent models BIBREF8 , BIBREF6 , BIBREF7 , BIBREF10 , BIBREF11 , BIBREF24 .
Let us now describe the model in more detail. Figure FIGREF21 may help you in understanding the following paragraphs.
The words from the document and the question are first converted into vector embeddings using a look-up matrix INLINEFORM0 . The document is then read by a bidirectional GRU network BIBREF29 . A concatenation of the hidden states of the forward and backward GRUs at each word is then used as a contextual embedding of this word, intuitively representing the context in which the word is appearing. We can also understand it as representing the set of questions to which this word may be an answer.
Similarly the question is read by a bidirectional GRU but in this case only the final hidden states are concatenated to form the question embedding.
The attention over each word in the context is then calculated as the dot product of its contextual embedding with the question embedding. This attention is then normalized by the softmax function and summed across all occurrences of each answer candidate. The candidate with most accumulated attention is selected as the final answer.
For a more detailed description of the model including equations check BIBREF4 . More details about the training setup and model hyperparameters can be found in the Appendix.
During our past experiments on the CNN, DM and CBT datasets BIBREF4 each unique word from the training, validation and test datasets had its row in the look-up matrix INLINEFORM0 . However as we radically increased the dataset size, this would result in an extremely large number of model parameters so we decided to limit the vocabulary size to INLINEFORM1 most frequent words. For each example, each unique out-of-vocabulary word is now mapped on one of 1000 anonymous tokens which are randomly initialized and untrained. Fixing the embeddings of these anonymous tags proved to significantly improve the performance.
While mostly using the original AS Reader model, we have also tried introducing a minor tweak in some instances of the model. We tried initializing the context encoder GRU's hidden state by letting the encoder read the question first before proceeding to read the context document. Intuitively this allows the encoder to know in advance what to look for when reading over the context document.
Including models of this kind in the ensemble helped to improve the performance.
Results
Table TABREF25 shows the accuracy of the psr and other architectures on the CBT validation and test data. The last two rows show the performance of the psr trained on the BookTest dataset; all the other models have been trained on the original CBT training data.
If we take the best psr ensemble trained on CBT as a baseline, improving the model architecture as in BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , continuing to use the original CBT training data, lead to improvements of INLINEFORM0 and INLINEFORM1 absolute on named entities and common nouns respectively. By contrast, inflating the training dataset provided a boost of INLINEFORM2 while using the same model. The ensemble of our models even exceeded the human baseline provided by Facebook BIBREF2 on the Common Noun dataset.
Our model takes approximately two weeks to converge when trained on the BookTest dataset on a single Nvidia Tesla K40 GPU.
Discussion
Embracing the abundance of data may mean focusing on other aspects of system design than with smaller data. Here are some of the challenges that we need to face in this situation.
Firstly, since the amount of data is practically unlimited – we could even generate them on the fly resulting in continuous learning similar to the Never-Ending Language Learning by Carnegie Mellon University BIBREF30 – it is now the speed of training that determines how much data the model is able to see. Since more training data significantly help the model performance, focusing on speeding up the algorithm may be more important than ever before. This may for instance influence the decision whether to use regularization such as dropout which does seem to somewhat improve the model performance, however usually at a cost of slowing down training.
Thanks to its simplicity, the psr seems to be training fast - for example around seven times faster than the models proposed by Chen et al. BIBREF5 . Hence the psr may be particularly suitable for training on large datasets.
The second challenge is how to generalize the performance gains from large data to a specific target domain. While there are huge amounts of natural language data in general, it may not be the case in the domain where we may want to ultimately apply our model.
Hence we are usually not facing a scenario of simply using a larger amount of the same training data, but rather extending training to a related domain of data, hoping that some of what the model learns on the new data will still help it on the original task.
This is highlighted by our observations from applying a model trained on the BookTest to Children's Book Test test data. If we move model training from joint CBT NE+CN training data to a subset of the BookTest of the same size (230k examples), we see a drop in accuracy of around 10% on the CBT test datasets.
Hence even though the Children's Book Test and BookTest datasets are almost as close as two disjoint datasets can get, the transfer is still very imperfect . Rightly choosing data to augment the in-domain training data is certainly a problem worth exploring in future work.
Our results show that given enough data the AS Reader was able to exceed the human performance on CBT CN reported by Facebook. However we hypothesized that the system is still not achieving its full potential so we decided to examine the room for improvement in our own small human study.
Human Study
After adding more data we have the performance on the CBT validation and test datasets soaring. However is there still potential for much further growth beyond the results we have observed?
We decided to explore the remaining space for improvement on the CBT by testing humans on a random subset of 50 named entity and 50 common noun validation questions that the psr ensemble could not answer correctly. These questions were answered by 10 non-native English speakers from our research laboratory, each on a disjoint subset of questions.. Participants had unlimited time to answer the questions and were told that these questions were not correctly answered by a machine, providing additional motivation to prove they are better than computers. The results of the human study are summarized in Table TABREF28 . They show that a majority of questions that our system could not answer so far are in fact answerable. This suggests that 1) the original human baselines might have been underestimated, however, it might also be the case that there are some examples that can be answered by machines and not by humans; 2) there is still space for improvement.
A system that would answer correctly every time when either our ensemble or human answered correctly would achieve accuracy over 92% percent on both validation and test NE datasets and over 96% on both CN datasets. Hence it still makes sense to use CBT dataset to study further improvements of text-comprehension systems.
Conclusion
Few ways of improving model performance are as solidly established as using more training data. Yet we believe this principle has been somewhat neglected by recent research in text comprehension. While there is a practically unlimited amount of data available in this field, most research was performed on unnecessarily small datasets.
As a gentle reminder to the community we have shown that simply infusing a model with more data can yield performance improvements of up to INLINEFORM0 where several attempts to improve the model architecture on the same training data have given gains of at most INLINEFORM1 compared to our best ensemble result. Yes, experiments on small datasets certainly can bring useful insights. However we believe that the community should also embrace the real-world scenario of data abundance.
The BookTest dataset we are proposing gives the reading-comprehension community an opportunity to make a step in that direction.
Training Details
The training details are similar to those in BIBREF4 however we are including them here for completeness.
To train the model we used stochastic gradient descent with the ADAM update rule BIBREF32 and learning rates of INLINEFORM0 , INLINEFORM1 and INLINEFORM2 . The best learning rate in our experiments was INLINEFORM3 . We minimized negative log-likelihood as the training objective.
The initial weights in the word-embedding matrix were drawn randomly uniformly from the interval INLINEFORM0 . Weights in the GRU networks were initialized by random orthogonal matrices BIBREF34 and biases were initialized to zero. We also used a gradient clipping BIBREF33 threshold of 10 and batches of sizes between 32 or 256. Increasing the batch from 32 to 128 seems to significantly improve performance on the large dataset - something we did not observe on the original CBT data. Increasing the batch size much above 128 is currently difficult due to memory constraints of the GPU.
During training we randomly shuffled all examples at the beginning of each epoch. To speed up training, we always pre-fetched 10 batches worth of examples and sorted them according to document length. Hence each batch contained documents of roughly the same length.
We also did not use pre-trained word embeddings.
We did not perform any text pre-processing since the datasets were already tokenized.
During training we evaluated the model performance every 12 hours and at the end of each epoch and stopped training when the error on the 20k BookTest validation set started increasing. We explored the hyperparameter space by training 67 different models The region of the parameter space that we explored together with the parameters of the model with best validation accuracy are summarized in Table TABREF29 .
Our model was implemented using Theano BIBREF31 and Blocks BIBREF35 .
The ensembles were formed by simply averaging the predictions from the constituent single models. These single models were selected using the following algorithm.
We started with the best performing model according to validation performance. Then in each step we tried adding the best performing model that had not been previously tried. We kept it in the ensemble if it did improve its validation performance and discarded it otherwise. This way we gradually tried each model once. We call the resulting model a greedy ensemble. We used the INLINEFORM0 BookTest validation dataset for this procedure.
The algorithm was offered 10 models and selected 5 of them for the final ensemble.
| [
"simply averaging the predictions from the constituent single models"
] | 4,212 | qasper | en | null | 91dd7b7a6ead4025763812d70dc51c6674b0acf31bd5a5f0 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How does their ensemble method work?
Answer: | [
"simply averaging the predictions from the constituent single models"
] | qasper | 128 | 9 | |
What are the sources of the datasets? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Emotion detection has long been a topic of interest to scholars in natural language processing (NLP) domain. Researchers aim to recognize the emotion behind the text and distribute similar ones into the same group. Establishing an emotion classifier can not only understand each user's feeling but also be extended to various application, for example, the motivation behind a user's interests BIBREF0. Based on releasing of large text corpus on social media and the emotion categories proposed by BIBREF1, BIBREF2, numerous models have provided and achieved fabulous precision so far. For example, DeepMoji BIBREF3 which utilized transfer learning concept to enhance emotions and sarcasm understanding behind the target sentence. CARER BIBREF4 learned contextualized affect representations to make itself more sensitive to rare words and the scenario behind the texts.
As methods become mature, text-based emotion detecting applications can be extended from a single utterance to a dialogue contributed by a series of utterances. Table TABREF2 illustrates the difference between single utterance and dialogue emotion recognition. The same utterances in Table TABREF2, even the same person said the same sentence, the emotion it convey may be various, which may depend on different background of the conversation, tone of speaking or personality. Therefore, for emotion detection, the information from preceding utterances in a conversation is relatively critical.
In SocialNLP 2019 EmotionX, the challenge is to recognize emotions for all utterances in EmotionLines dataset, a dataset consists of dialogues. According to the needs for considering context at the same time, we develop two classification models, inspired by bidirectional encoder representations from transformers (BERT) BIBREF5, FriendsBERT and ChatBERT. In this paper, we introduce our approaches including causal utterance modeling, model pre-training, and fine-turning.
Dataset
EmotionLines BIBREF6 is a dialogue dataset composed of two subsets, Friends and EmotionPush, according to the source of the dialogues. The former comes from the scripts of the Friends TV sitcom. The other is made up of Facebook messenger chats. Each subset includes $1,000$ English dialogues, and each dialogue can be further divided into a few consecutive utterances. All the utterances are annotated by five annotators on a crowd-sourcing platform (Amazon Mechanical Turk), and the labeling work is only based on the textual content. Annotator votes for one of the seven emotions, namely Ekman’s six basic emotions BIBREF1, plus the neutral. If none of the emotion gets more than three votes, the utterance will be marked as “non-neutral”.
For the datasets, there are properties worth additional mentioning. Although Friends and EmotionPush share the same data format, they are quite different in nature. Friends is a speech-based dataset which is annotated dialogues from the TV sitcom. It means most of the utterances are generated by the a few main characters. The personality of a character often affects the way of speaking, and therefore “who is the speaker" might provide extra clues for emotion prediction. In contrast, EmotionPush does not have this trait due to the anonymous mechanism. In addition, features such as typo, hyperlink, and emoji that only appear in chat-based data will need some domain-specific techniques to process.
Incidentally, the objective of the challenge is to predict the emotion for each utterance. Just, according to EmotionX 2019 specification, there are only four emotions be selected as our label candidates, which are Joy, Sadness, Anger, and Neutral. These emotions will be considered during performance evaluation. The technical detail will also be introduced and discussed in following Section SECREF13 and Section SECREF26.
Model Description
For this challenge, we adapt BERT which is proposed by BIBREF5 to help understand the context at the same time. Technically, BERT, designed on end-to-end architecture, is a deep pre-trained transformer encoder that dynamically provides language representation and BERT already achieved multiple state-of-the-art results on GLUE benchmark BIBREF7 and many tasks. A quick recap for BERT's architecture and its pre-training tasks will be illustrated in the following subsections.
Model Description ::: Model Architecture
BERT, the Bidirectional Encoder Representations from Transformers, consists of several transformer encoder layers that enable the model to extract very deep language features on both token-level and sentence-level. Each transformer encoder contains multi-head self-attention layers that provide ability to learn multiple attention feature of each word from their bidirectional context. The transformer and its self-attention mechanism are proposed by BIBREF8. This self-attention mechanism can be interpreted as a key-value mapping given query. By given the embedding vector for token input, the query ($Q$), key ($K$) and value ($V$) are produced by the projection from each three parameter matrices where $W^Q \in \mathbb {R}^{d_{{\rm model}} \times d_{k}}, W^K \in \mathbb {R}^{d_{\rm model} \times d_{k}}$ and $W^V \in \mathbb {R}^{d_{\rm model} \times d_{v}}$. The self-attention BIBREF8 is formally represented as:
The $ d_k = d_v = d_{\rm model} = 1024$ in BERT large version and 768 in BERT base version. Once model can extract attention feature, we can extend one self-attention into multi-head self-attention, this extension makes sub-space features can be extracted in same time by this multi-head configuration. Overall, the multi-attention mechanism is adopt for each transformer encoder, and several of encoder layer will be stacked together to form a deep transformer encoder.
For the model input, BERT allow us take one sentence as input sequence or two sentences together as one input sequence, and the maximum length of input sequence is 512. The way that BERT was designed is for giving model the sentence-level and token-level understanding. In two sentences case, a special token ([SEP]) will be inserted between two sentences. In addition, the first input token is also a special token ([CLS]), and its corresponding ouput will be vector place for classification during fine-tuning. The outputs of the last encoder layer corresponding to each input token can be treated as word representations for each token, and the word representation of the first token ([CLS]) will be consider as classification (output) representation for further fine-tuning tasks. In BERT, this vector is denoted as $C \in \mathbb {R}^{d_{\rm model}} $, and a classification layer is denoted as $ W \in \mathbb {R}^{K \times d_{\rm model}}$, where $K$ is number of classification labels. Finally, the prediction $P$ of BERT is represented as $P = {\rm softmax}(CW^T)$.
Model Description ::: Pre-training Tasks
In pre-training, intead of using unidirectional language models, BERT developed two pre-training tasks: (1) Masked LM (cloze test) and (2) Next Sentence Prediction. At the first pre-training task, bidirectional language modeling can be done at this cloze-like pre-training. In detail, 15% tokens of input sequence will be masked at random and model need to predict those masked tokens. The encoder will try to learn contextual representations from every given tokens due to masking tokens at random. Model will not know which part of the input is going to be masked, so that the information of each masked tokens should be inferred by remaining tokens. At Next Sentence Prediction, two sentences concatenated together will be considered as model input. In order to give model a good nature language understanding, knowing relationship between sentence is one of important abilities. When generating input sequences, 50% of time the sentence B is actually followed by sentence A, and rest 50% of the time the sentence B will be picked randomly from dataset, and model need to predict if the sentence B is next sentence of sentence A. That is, the attention information will be shared between sentences. Such sentence-level understanding may have difficulties to be learned at first pre-training task (Masked LM), therefore, the pre-training task (NSP) is developed as second training goal to capture the cross sentence relationship.
In this competition, limited by the size of dataset and the challenge in contextual emotion recognition, we consider BERT with both two pre-training tasks can give a good starting point to extract emotion changing during dialogue-like conversation. Especially the second pre-training task, it might be more important for dialogue-like conversation where the emotion may various by the context of continuous utterances. That is, given a set of continues conversations, the emotion of current utterance might be influenced by previous utterance. By this assumption and with supporting from the experiment results of BERT, we can take sentence A as one-sentence context and consider sentence B as the target sentence for emotion prediction. The detail will be described in Section SECREF4.
Methodology
The main goal of the present work is to predict the emotion of utterance within the dialogue. Following are four major difficulties we concern about:
The emotion of the utterances depends not only on the text but also on the interaction happened earlier.
The source of the two datasets are different. Friends is speech-based dialogues and EmotionPush is chat-based dialogues. It makes datasets possess different characteristics.
There are only $1,000$ dialogues in both training datasets which are not large enough for the stability of training a complex neural-based model.
The prediction targets (emotion labels) are highly unbalanced.
The proposed approach is summarized in Figure FIGREF3, which aims to overcome these challenges. The framework could be separated into three steps and described as follow:
Methodology ::: Causal Utterance Modeling
Given a dialogue $D^{(i)}$ which includes sequence of utterances denoted as $D^{(i)}=(u^{(i)}_{1}, u^{(i)}_{2}, ..., u^{(i)}_{n})$, where $i$ is the index in dataset and $n$ is the number of utterances in the given dialogue. In order to conserve the emotional information of both utterance and conversation, we rearrange each two consecutive utterances $u_{t}, u_{t-1}$ into a single sentence representation $x_{t}$ as
The corresponding sentence representation corpus $X^{(i)}$ are denoted as $X^{(i)}=(x^{(i)}_{1}, x^{(i)}_{2}, ..., x^{(i)}_{n})$. Note that the first utterance within a conversation does not have its causal utterance (previous sentence), therefore, the causal utterance will be set as [None]. A practical example of sentence representation is shown in Table TABREF11.
Since the characteristics of two datasets are not identical, we customize different causal utterance modeling strategies to refine the information in text.
For Friends, there are two specific properties. The first one is that most dialogues are surrounding with the six main characters, including Rachel, Monica, Phoebe, Joey, Chandler, and Ross. The utterance ratio of given by the six roles is up to $83.4\%$. Second, the personal characteristics of the six characters are very clear. Each leading role has its own emotion undulated rule. To make use of these features, we introduce the personality tokenization which help learning the personality of the six characters. Personality tokenization concatenate the speaker and says tokens before the input utterance if the speaker is one of the six characters. The example is shown in Table TABREF12.
For EmotionPush, the text are informal chats which including like slang, acronym, typo, hyperlink, and emoji. Another characteristic is that the specific name entities are tokenized with random index. (e.g. “organization_80”, “person_01”, and “time_12”). We consider some of these informal text are related to expressing emotion such as repeated typing, purposed capitalization, and emoji (e.g. “:D”, “:(”, and “<3”)). Therefore, we keep most informal expressions but only process hyperlinks, empty utterance, and name entities by unifying the tokens.
Methodology ::: Model Pre-training
Since the size of both datasets are not large enough for complex neural-based model training as well as BERT model is only pre-train on formal text datasets, the issues of overfitting and domain bias are important considerations for design the pre-training process.
To avoid our model overfitting on the training data and increase the understanding of informal text, we adapted BERT and derived two models, namely FriendsBERT and ChatBERT, with different pre-training tasks before the formal training process for Friends and EmotionPush dataset, respectively. The pre-training strategies are described below.
For pre-training FriendsBERT, we collect the completed scripts of all ten seasons of Friends TV shows from emorynlp which includes 3,107 scenes within 61,309 utterances. All the utterances are followed the preprocessing methods mentions above to compose the corpus for Masked language model pre-training task. The consequent utterances in the same scenes are considered as the consequent sentences to pre-train the Next Sentence Prediction task. In the pre-training process, the training loss is the sum of the mean likelihood of two pre-train tasks.
For pre-training ChatBERT, we pre-train our model on the Twitter dataset, since the text and writing style on Twitter are close to the chat text where both may involved with many informal words or emoticons as well. The Twitter emotion dataset, 8 basic emotions from emotion wheel BIBREF1, was collected by twitter streaming API with specific emotion-related hashtags, such as #anger, #joy, #cry, #sad and etc. The hashtags in tweets are treated as emotion label for model fine-tuning. The tweets were fine-grined processing followed the rules in BIBREF9, BIBREF4, including duplicate tweets removing, the emotion hashtags must appearing in the last position of a tweet, and etc. The statis of tweets were summarized in Table TABREF17. Each tweet and corresponding emotion label composes an emotion classification dataset for pre-training.
Methodology ::: Fine-tuning
Since our emotion recognition task is treated as a sequence-level classification task, the model would be fine-tuned on the processed training data. Following the BERT construction, we take the first embedding vector which corresponds to the special token [CLS] from the final hidden state of the Transformer encoder. This vector represents the embedding vector of the corresponding conversation utterances which is denoted as $\mathbf {C} \in \mathbb {R}^{H}$, where $H$ is the embedding size. A dense neural layer is treated as a classification layer which consists of parameters $\mathbf {W} \in \mathbb {R}^{K\times H}$ and $\mathbf {b} \in \mathbb {R}^{K}$, where $K$ is the number of emotion class. The emotion prediction probabilities $\mathbf {P} \in \mathbb {R}^{K}$ are computed by a softmax activation function as
All the parameters in BERT and the classification layer would be fine-turned together to minimize the Negative Log Likelihood (NLL) loss function, as Equation (DISPLAY_FORM22), based on the ground truth emotion label $c$.
In order to tackle the problem of highly unbalanced emotion labels, we apply weighted balanced warming on NLL loss function, as Equation (DISPLAY_FORM23), in the first epoch of fine-tuning procedure.
where $\mathbf {w}$ are the weights of corresponding emotion label $c$ which are computed and normalize by the frequency as
By adding the weighted balanced warming on NLL loss, the model could learn to predict the minor emotions (e.g. anger and sadness) earlier and make the training process more stable. Since the major evaluation metrics micro F1-score is effect by the number of each label, we only apply the weighted balanced warming in first epoch to optimize the performance.
Experiments
Since the EmotionX challenge only provided the gold labels in training data, we pick the best performance model (weights) to predict the testing data. In this section, we present the experiment and evaluation results.
Experiments ::: Experimental Setup
The EmotionX challenge consists of $1,000$ dialogues for both Friends and EmotionPush. In all of our experiments, each dataset is separated into top 800 dialogues for training and last 200 dialogues for validation. Since the EmotionX challenge considers only the four emotions (anger, joy, neutral, and sadness) in the evaluation stage, we ignore all the data point corresponding to other emotions directly. The details of emotions distribution are shown in Table TABREF18.
The hyperparameters and training setup of our models (FriendsBERT and ChatBERT) are shown in Table TABREF25. Some common and easily implemented methods are selected as the baselines embedding methods and classification models. The baseline embedding methods are including bag-of-words (BOW), term frequency–inverse document frequency (TFIDF), and neural-based word embedding. The classification models are including Logistic Regression (LR), Random Forest (RF), TextCNN BIBREF10 with initial word embedding as GloVe BIBREF11, and our proposed model. All the experiment results are based on the best performances of validation results.
Experiments ::: Performance
The experiment results of validation on Friends are shown in Table TABREF19. The proposed model and baselines are evaluated based on the Precision (P.), Recall (R.), and F1-measure (F1).
For the traditional baselines, namely BOW and TFIDF, we observe that they achieve surprising high F1 scores around $0.81$, however, the scores for Anger and Sadness are lower. This explains that traditional approaches tend to predict the labels with large sample size, such as Joy and Neutral, but fail to take of scarce samples even when an ensemble random forest classifier is adopted. In order to prevent the unbalanced learning, we choose the weighted loss mechanism for both TextCNN and causal modeling TextCNN (C-TextCNN), these models suffer less than the traditional baselines and achieve a slightly balance performance, where there are around 15% and 7% improvement on Anger and Sadness, respectively. We following adopt the casual utterance modeling to original TextCNN, mapping previous utterance as well as target utterance into model. The causal utterance modeling improve the C-TextCNN over TextCNN for 6%, 2% and 1% on Anger, Joy and overall F1 score. Motivated from these preliminary experiments, the proposed FriendsBERT also adopt the ideas of both weighted loss and causal utterance modeling. As compared to the original BERT, single sentence BERT (FriendsBERT-base-s), the proposed FriendsBERT-base improve 1% for Joy and overall F1, and 2% for Sadness. For the final validation performance, our proposed approach achieves the highest scores, which are $0.85$ and $0.86$ for FriendsBERT-base and FriendsBERT-large, respectively.
Overall, the proposed FriendsBERT successfully captures the sentence-level context-awarded information and outperforms all the baselines, which not only achieves high performance on large sample labels, but also on small sample labels. The similar settings are also adapted to EmotionPush dataset for the final evaluation.
Experiments ::: Evaluation Results
The testing dataset consists of 240 dialogues including $3,296$ and $3,536$ utterances in Friends and EmotionPush respectively. We re-train our FriendsBERT and ChatBERT with top 920 training dialogues and predict the evaluation results using the model performing the best validation results. The results are shown in Table TABREF29 and Table TABREF30. The present method achieves $81.5\%$ and $88.5\%$ micro F1-score on the testing dataset of Friends and EmotionPush, respectively.
Conclusion and Future work
In the present work, we propose FriendsBERT and ChatBERT for the multi-utterance emotion recognition task on EmotionLines dataset. The proposed models are adapted from BERT BIBREF5 with three main improvement during the model training procedure, which are the causal utterance modeling mechanism, specific model pre-training, and adapt weighted loss. The causal utterance modeling takes the advantages of the sentence-level context information during model inference. The specific model pre-training helps to against the bias in different text domain. The weighted loss avoids our model to only predict on large size sample. The effectiveness and generalizability of the proposed methods are demonstrated from the experiments.
In future work, we consider to include the conditional probabilistic constraint $P ({\rm Emo}_{B} | \hat{\rm Emo}_{A})$. Model should predict the emotion based on a certain understanding about context emotions. This might be more reasonable for guiding model than just predicting emotion of ${\rm Sentence}_B$ directly. In addition, due to the limitation of BERT input format, ambiguous number of input sentences is now becoming an important design requirement for our future work. Also, personality embedding development will be another future work of the emotion recognition. The personality embedding will be considered as sentence embedding injected into word embedding, and it seems this additional information can contribute some improvement potentially.
| [
"Friends TV sitcom, Facebook messenger chats"
] | 3,185 | qasper | en | null | fe6e5087f91071369646d9d868d784a290d0d13013292b61 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What are the sources of the datasets?
Answer: | [
"Friends TV sitcom, Facebook messenger chats"
] | qasper | 128 | 10 | |
what language does this paper focus on? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Text simplification aims to reduce the lexical and structural complexity of a text, while still retaining the semantic meaning, which can help children, non-native speakers, and people with cognitive disabilities, to understand text better. One of the methods of automatic text simplification can be generally divided into three categories: lexical simplification (LS) BIBREF0 , BIBREF1 , rule-based BIBREF2 , and machine translation (MT) BIBREF3 , BIBREF4 . LS is mainly used to simplify text by substituting infrequent and difficult words with frequent and easier words. However, there are several challenges for the LS approach: a great number of transformation rules are required for reasonable coverage and should be applied based on the specific context; third, the syntax and semantic meaning of the sentence is hard to retain. Rule-based approaches use hand-crafted rules for lexical and syntactic simplification, for example, substituting difficult words in a predefined vocabulary. However, such approaches need a lot of human-involvement to manually define these rules, and it is impossible to give all possible simplification rules. MT-based approach has attracted great attention in the last several years, which addresses text simplification as a monolingual machine translation problem translating from 'ordinary' and 'simplified' sentences.
In recent years, neural Machine Translation (NMT) is a newly-proposed deep learning approach and achieves very impressive results BIBREF5 , BIBREF6 , BIBREF7 . Unlike the traditional phrased-based machine translation system which operates on small components separately, NMT system is being trained end-to-end, without the need to have external decoders, language models or phrase tables. Therefore, the existing architectures in NMT are used for text simplification BIBREF8 , BIBREF4 . However, most recent work using NMT is limited to the training data that are scarce and expensive to build. Language models trained on simplified corpora have played a central role in statistical text simplification BIBREF9 , BIBREF10 . One main reason is the amount of available simplified corpora typically far exceeds the amount of parallel data. The performance of models can be typically improved when trained on more data. Therefore, we expect simplified corpora to be especially helpful for NMT models.
In contrast to previous work, which uses the existing NMT models, we explore strategy to include simplified training corpora in the training process without changing the neural network architecture. We first propose to pair simplified training sentences with synthetic ordinary sentences during training, and treat this synthetic data as additional training data. We obtain synthetic ordinary sentences through back-translation, i.e. an automatic translation of the simplified sentence into the ordinary sentence BIBREF11 . Then, we mix the synthetic data into the original (simplified-ordinary) data to train NMT model. Experimental results on two publicly available datasets show that we can improve the text simplification quality of NMT models by mixing simplified sentences into the training set over NMT model only using the original training data.
Related Work
Automatic TS is a complicated natural language processing (NLP) task, which consists of lexical and syntactic simplification levels BIBREF12 . It has attracted much attention recently as it could make texts more accessible to wider audiences, and used as a pre-processing step, improve performances of various NLP tasks and systems BIBREF13 , BIBREF14 , BIBREF15 . Usually, hand-crafted, supervised, and unsupervised methods based on resources like English Wikipedia and Simple English Wikipedia (EW-SEW) BIBREF10 are utilized for extracting simplification rules. It is very easy to mix up the automatic TS task and the automatic summarization task BIBREF3 , BIBREF16 , BIBREF6 . TS is different from text summarization as the focus of text summarization is to reduce the length and redundant content.
At the lexical level, lexical simplification systems often substitute difficult words using more common words, which only require a large corpus of regular text to obtain word embeddings to get words similar to the complex word BIBREF1 , BIBREF9 . Biran et al. BIBREF0 adopted an unsupervised method for learning pairs of complex and simpler synonyms from a corpus consisting of Wikipedia and Simple Wikipedia. At the sentence level, a sentence simplification model was proposed by tree transformation based on statistical machine translation (SMT) BIBREF3 . Woodsend and Lapata BIBREF17 presented a data-driven model based on a quasi-synchronous grammar, a formalism that can naturally capture structural mismatches and complex rewrite operations. Wubben et al. BIBREF18 proposed a phrase-based machine translation (PBMT) model that is trained on ordinary-simplified sentence pairs. Xu et al. BIBREF19 proposed a syntax-based machine translation model using simplification-specific objective functions and features to encourage simpler output.
Compared with SMT, neural machine translation (NMT) has shown to produce state-of-the-art results BIBREF5 , BIBREF7 . The central approach of NMT is an encoder-decoder architecture implemented by recurrent neural networks, which can represent the input sequence as a vector, and then decode that vector into an output sequence. Therefore, NMT models were used for text simplification task, and achieved good results BIBREF8 , BIBREF4 , BIBREF20 . The main limitation of the aforementioned NMT models for text simplification depended on the parallel ordinary-simplified sentence pairs. Because ordinary-simplified sentence pairs are expensive and time-consuming to build, the available largest data is EW-SEW that only have 296,402 sentence pairs. The dataset is insufficiency for NMT model if we want to NMT model can obtain the best parameters. Considering simplified data plays an important role in boosting fluency for phrase-based text simplification, and we investigate the use of simplified data for text simplification. We are the first to show that we can effectively adapt neural translation models for text simplifiation with simplified corpora.
Simplified Corpora
We collected a simplified dataset from Simple English Wikipedia that are freely available, which has been previously used for many text simplification methods BIBREF0 , BIBREF10 , BIBREF3 . The simple English Wikipedia is pretty easy to understand than normal English Wikipedia. We downloaded all articles from Simple English Wikipedia. For these articles, we removed stubs, navigation pages and any article that consisted of a single sentence. We then split them into sentences with the Stanford CorNLP BIBREF21 , and deleted these sentences whose number of words are smaller than 10 or large than 40. After removing repeated sentences, we chose 600K sentences as the simplified data with 11.6M words, and the size of vocabulary is 82K.
Text Simplification using Neural Machine Translation
Our work is built on attention-based NMT BIBREF5 as an encoder-decoder network with recurrent neural networks (RNN), which simultaneously conducts dynamic alignment and generation of the target simplified sentence.
The encoder uses a bidirectional RNN that consists of forward and backward RNN. Given a source sentence INLINEFORM0 , the forward RNN and backward RNN calculate forward hidden states INLINEFORM1 and backward hidden states INLINEFORM2 , respectively. The annotation vector INLINEFORM3 is obtained by concatenating INLINEFORM4 and INLINEFORM5 .
The decoder is a RNN that predicts a target simplificated sentence with Gated Recurrent Unit (GRU) BIBREF22 . Given the previously generated target (simplified) sentence INLINEFORM0 , the probability of next target word INLINEFORM1 is DISPLAYFORM0
where INLINEFORM0 is a non-linear function, INLINEFORM1 is the embedding of INLINEFORM2 , and INLINEFORM3 is a decoding state for time step INLINEFORM4 .
State INLINEFORM0 is calculated by DISPLAYFORM0
where INLINEFORM0 is the activation function GRU.
The INLINEFORM0 is the context vector computed as a weighted annotation INLINEFORM1 , computed by DISPLAYFORM0
where the weight INLINEFORM0 is computed by DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 , INLINEFORM1 and INLINEFORM2 are weight matrices. The training objective is to maximize the likelihood of the training data. Beam search is employed for decoding.
Synthetic Simplified Sentences
We train an auxiliary system using NMT model from the simplified sentence to the ordinary sentence, which is first trained on the available parallel data. For leveraging simplified sentences to improve the quality of NMT model for text simplification, we propose to adapt the back-translation approach proposed by Sennrich et al. BIBREF11 to our scenario. More concretely, Given one sentence in simplified sentences, we use the simplified-ordinary system in translate mode with greedy decoding to translate it to the ordinary sentences, which is denoted as back-translation. This way, we obtain a synthetic parallel simplified-ordinary sentences. Both the synthetic sentences and the available parallel data are used as training data for the original NMT system.
Evaluation
We evaluate the performance of text simplification using neural machine translation on available parallel sentences and additional simplified sentences.
Dataset. We use two simplification datasets (WikiSmall and WikiLarge). WikiSmall consists of ordinary and simplified sentences from the ordinary and simple English Wikipedias, which has been used as benchmark for evaluating text simplification BIBREF17 , BIBREF18 , BIBREF8 . The training set has 89,042 sentence pairs, and the test set has 100 pairs. WikiLarge is also from Wikipedia corpus whose training set contains 296,402 sentence pairs BIBREF19 , BIBREF20 . WikiLarge includes 8 (reference) simplifications for 2,359 sentences split into 2,000 for development and 359 for testing.
Metrics. Three metrics in text simplification are chosen in this paper. BLEU BIBREF5 is one traditional machine translation metric to assess the degree to which translated simplifications differed from reference simplifications. FKGL measures the readability of the output BIBREF23 . A small FKGL represents simpler output. SARI is a recent text-simplification metric by comparing the output against the source and reference simplifications BIBREF20 .
We evaluate the output of all systems using human evaluation. The metric is denoted as Simplicity BIBREF8 . The three non-native fluent English speakers are shown reference sentences and output sentences. They are asked whether the output sentence is much simpler (+2), somewhat simpler (+1), equally (0), somewhat more difficult (-1), and much more difficult (-2) than the reference sentence.
Methods. We use OpenNMT BIBREF24 as the implementation of the NMT system for all experiments BIBREF5 . We generally follow the default settings and training procedure described by Klein et al.(2017). We replace out-of-vocabulary words with a special UNK symbol. At prediction time, we replace UNK words with the highest probability score from the attention layer. OpenNMT system used on parallel data is the baseline system. To obtain a synthetic parallel training set, we back-translate a random sample of 100K sentences from the collected simplified corpora. OpenNMT used on parallel data and synthetic data is our model. The benchmarks are run on a Intel(R) Core(TM) i7-5930K CPU@3.50GHz, 32GB Mem, trained on 1 GPU GeForce GTX 1080 (Pascal) with CUDA v. 8.0.
We choose three statistical text simplification systems. PBMT-R is a phrase-based method with a reranking post-processing step BIBREF18 . Hybrid performs sentence splitting and deletion operations based on discourse representation structures, and then simplifies sentences with PBMT-R BIBREF25 . SBMT-SARI BIBREF19 is syntax-based translation model using PPDB paraphrase database BIBREF26 and modifies tuning function (using SARI). We choose two neural text simplification systems. NMT is a basic attention-based encoder-decoder model which uses OpenNMT framework to train with two LSTM layers, hidden states of size 500 and 500 hidden units, SGD optimizer, and a dropout rate of 0.3 BIBREF8 . Dress is an encoder-decoder model coupled with a deep reinforcement learning framework, and the parameters are chosen according to the original paper BIBREF20 . For the experiments with synthetic parallel data, we back-translate a random sample of 60 000 sentences from the collected simplified sentences into ordinary sentences. Our model is trained on synthetic data and the available parallel data, denoted as NMT+synthetic.
Results. Table 1 shows the results of all models on WikiLarge dataset. We can see that our method (NMT+synthetic) can obtain higher BLEU, lower FKGL and high SARI compared with other models, except Dress on FKGL and SBMT-SARI on SARI. It verified that including synthetic data during training is very effective, and yields an improvement over our baseline NMF by 2.11 BLEU, 1.7 FKGL and 1.07 SARI. We also substantially outperform Dress, who previously reported SOTA result. The results of our human evaluation using Simplicity are also presented in Table 1. NMT on synthetic data is significantly better than PBMT-R, Dress, and SBMT-SARI on Simplicity. It indicates that our method with simplified data is effective at creating simpler output.
Results on WikiSmall dataset are shown in Table 2. We see substantial improvements (6.37 BLEU) than NMT from adding simplified training data with synthetic ordinary sentences. Compared with statistical machine translation models (PBMT-R, Hybrid, SBMT-SARI), our method (NMT+synthetic) still have better results, but slightly worse FKGL and SARI. Similar to the results in WikiLarge, the results of our human evaluation using Simplicity outperforms the other models. In conclusion, Our method produces better results comparing with the baselines, which demonstrates the effectiveness of adding simplified training data.
Conclusion
In this paper, we propose one simple method to use simplified corpora during training of NMT systems, with no changes to the network architecture. In the experiments on two datasets, we achieve substantial gains in all tasks, and new SOTA results, via back-translation of simplified sentences into the ordinary sentences, and treating this synthetic data as additional training data. Because we do not change the neural network architecture to integrate simplified corpora, our method can be easily applied to other Neural Text Simplification (NTS) systems. We expect that the effectiveness of our method not only varies with the quality of the NTS system used for back-translation, but also depends on the amount of available parallel and simplified corpora. In the paper, we have only utilized data from Wikipedia for simplified sentences. In the future, many other text sources are available and the impact of not only size, but also of domain should be investigated.
| [
"English",
"Simple English"
] | 2,243 | qasper | en | null | 51b9066a5f2845e2fdf0d1dcde6833f70ae49ed01aa306db | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: what language does this paper focus on?
Answer: | [
"English",
"Simple English"
] | qasper | 128 | 11 | |
What sentiment analysis dataset is used? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
There have been many implementations of the word2vec model in either of the two architectures it provides: continuous skipgram and CBoW (BIBREF0). Similar distributed models of word or subword embeddings (or vector representations) find usage in sota, deep neural networks like BERT and its successors (BIBREF1, BIBREF2, BIBREF3). These deep networks generate contextual representations of words after been trained for extended periods on large corpora, unsupervised, using the attention mechanisms (BIBREF4).
It has been observed that various hyper-parameter combinations have been used in different research involving word2vec with the possibility of many of them being sub-optimal (BIBREF5, BIBREF6, BIBREF7). Therefore, the authors seek to address the research question: what is the optimal combination of word2vec hyper-parameters for intrinsic and extrinsic NLP purposes? There are astronomically high numbers of combinations of hyper-parameters possible for neural networks, even with just a few layers. Hence, the scope of our extensive work over three corpora is on dimension size, training epochs, window size and vocabulary size for the training algorithms (hierarchical softmax and negative sampling) of both skipgram and CBoW. The corpora used for word embeddings are English Wiki News Abstract by BIBREF8 of about 15MB, English Wiki Simple (SW) Articles by BIBREF9 of about 711MB and the Billion Word (BW) of 3.9GB by BIBREF10. The corpus used for sentiment analysis is the IMDb dataset of movie reviews by BIBREF11 while that for NER is Groningen Meaning Bank (GMB) by BIBREF12, containing 47,959 sentence samples. The IMDb dataset used has a total of 25,000 sentences with half being positive sentiments and the other half being negative sentiments. The GMB dataset has 17 labels, with 9 main labels and 2 context tags. It is however unbalanced due to the high percentage of tokens with the label 'O'. This skew in the GMB dataset is typical with NER datasets.
The objective of this work is to determine the optimal combinations of word2vec hyper-parameters for intrinsic evaluation (semantic and syntactic analogies) and extrinsic evaluation tasks (BIBREF13, BIBREF14), like SA and NER. It is not our objective in this work to record sota results. Some of the main contributions of this research are the empirical establishment of optimal combinations of word2vec hyper-parameters for NLP tasks, discovering the behaviour of quality of vectors viz-a-viz increasing dimensions and the confirmation of embeddings being task-specific for the downstream. The rest of this paper is organised as follows: the literature review that briefly surveys distributed representation of words, particularly word2vec; the methodology employed in this research work; the results obtained and the conclusion.
Literature Review
Breaking away from the non-distributed (high-dimensional, sparse) representations of words, typical of traditional bag-of-words or one-hot-encoding (BIBREF15), BIBREF0 created word2vec. Word2Vec consists of two shallow neural network architectures: continuous skipgram and CBoW. It uses distributed (low-dimensional, dense) representations of words that group similar words. This new model traded the complexity of deep neural network architectures, by other researchers, for more efficient training over large corpora. Its architectures have two training algorithms: negative sampling and hierarchical softmax (BIBREF16). The released model was trained on Google news dataset of 100 billion words. Implementations of the model have been undertaken by researchers in the programming languages Python and C++, though the original was done in C (BIBREF17).
Continuous skipgram predicts (by maximizing classification of) words before and after the center word, for a given range. Since distant words are less connected to a center word in a sentence, less weight is assigned to such distant words in training. CBoW, on the other hand, uses words from the history and future in a sequence, with the objective of correctly classifying the target word in the middle. It works by projecting all history or future words within a chosen window into the same position, averaging their vectors. Hence, the order of words in the history or future does not influence the averaged vector. This is similar to the traditional bag-of-words, which is oblivious of the order of words in its sequence. A log-linear classifier is used in both architectures (BIBREF0). In further work, they extended the model to be able to do phrase representations and subsample frequent words (BIBREF16). Being a NNLM, word2vec assigns probabilities to words in a sequence, like other NNLMs such as feedforward networks or recurrent neural networks (BIBREF15). Earlier models like latent dirichlet allocation (LDA) and latent semantic analysis (LSA) exist and effectively achieve low dimensional vectors by matrix factorization (BIBREF18, BIBREF19).
It's been shown that word vectors are beneficial for NLP tasks (BIBREF15), such as sentiment analysis and named entity recognition. Besides, BIBREF0 showed with vector space algebra that relationships among words can be evaluated, expressing the quality of vectors produced from the model. The famous, semantic example: vector("King") - vector("Man") + vector("Woman") $\approx $ vector("Queen") can be verified using cosine distance. Another type of semantic meaning is the relationship between a capital city and its corresponding country. Syntactic relationship examples include plural verbs and past tense, among others. Combination of both syntactic and semantic analyses is possible and provided (totaling over 19,000 questions) as Google analogy test set by BIBREF0. WordSimilarity-353 test set is another analysis tool for word vectors (BIBREF20). Unlike Google analogy score, which is based on vector space algebra, WordSimilarity is based on human expert-assigned semantic similarity on two sets of English word pairs. Both tools rank from 0 (totally dissimilar) to 1 (very much similar or exact, in Google analogy case).
A typical artificial neural network (ANN) has very many hyper-parameters which may be tuned. Hyper-parameters are values which may be manually adjusted and include vector dimension size, type of algorithm and learning rate (BIBREF19). BIBREF0 tried various hyper-parameters with both architectures of their model, ranging from 50 to 1,000 dimensions, 30,000 to 3,000,000 vocabulary sizes, 1 to 3 epochs, among others. In our work, we extended research to 3,000 dimensions. Different observations were noted from the many trials. They observed diminishing returns after a certain point, despite additional dimensions or larger, unstructured training data. However, quality increased when both dimensions and data size were increased together. Although BIBREF16 pointed out that choice of optimal hyper-parameter configurations depends on the NLP problem at hand, they identified the most important factors are architecture, dimension size, subsampling rate, and the window size. In addition, it has been observed that variables like size of datasets improve the quality of word vectors and, potentially, performance on downstream tasks (BIBREF21, BIBREF0).
Methodology
The models were generated in a shared cluster running Ubuntu 16 with 32 CPUs of 32x Intel Xeon 4110 at 2.1GHz. Gensim (BIBREF17) python library implementation of word2vec was used with parallelization to utilize all 32 CPUs. The downstream experiments were run on a Tesla GPU on a shared DGX cluster running Ubuntu 18. Pytorch deep learning framework was used. Gensim was chosen because of its relative stability, popular support and to minimize the time required in writing and testing a new implementation in python from scratch.
To form the vocabulary, words occurring less than 5 times in the corpora were dropped, stop words removed using the natural language toolkit (NLTK) (BIBREF22) and data pre-processing carried out. Table TABREF2 describes most hyper-parameters explored for each dataset. In all, 80 runs (of about 160 minutes) were conducted for the 15MB Wiki Abstract dataset with 80 serialized models totaling 15.136GB while 80 runs (for over 320 hours) were conducted for the 711MB SW dataset, with 80 serialized models totaling over 145GB. Experiments for all combinations for 300 dimensions were conducted on the 3.9GB training set of the BW corpus and additional runs for other dimensions for the window 8 + skipgram + heirarchical softmax combination to verify the trend of quality of word vectors as dimensions are increased.
Google (semantic and syntactic) analogy tests and WordSimilarity-353 (with Spearman correlation) by BIBREF20 were chosen for intrinsic evaluations. They measure the quality of word vectors. The analogy scores are averages of both semantic and syntactic tests. NER and SA were chosen for extrinsic evaluations. The GMB dataset for NER was trained in an LSTM network, which had an embedding layer for input. The network diagram is shown in fig. FIGREF4. The IMDb dataset for SA was trained in a BiLSTM network, which also used an embedding layer for input. Its network diagram is given in fig. FIGREF4. It includes an additional hidden linear layer. Hyper-parameter details of the two networks for the downstream tasks are given in table TABREF3. The metrics for extrinsic evaluation include F1, precision, recall and accuracy scores. In both tasks, the default pytorch embedding was tested before being replaced by pre-trained embeddings released by BIBREF0 and ours. In each case, the dataset was shuffled before training and split in the ratio 70:15:15 for training, validation (dev) and test sets. Batch size of 64 was used. For each task, experiments for each embedding was conducted four times and an average value calculated and reported in the next section
Results and Discussion
Table TABREF5 summarizes key results from the intrinsic evaluations for 300 dimensions. Table TABREF6 reveals the training time (in hours) and average embedding loading time (in seconds) representative of the various models used. Tables TABREF11 and TABREF12 summarize key results for the extrinsic evaluations. Figures FIGREF7, FIGREF9, FIGREF10, FIGREF13 and FIGREF14 present line graph of the eight combinations for different dimension sizes for Simple Wiki, trend of Simple Wiki and Billion Word corpora over several dimension sizes, analogy score comparison for models across datasets, NER mean F1 scores on the GMB dataset and SA mean F1 scores on the IMDb dataset, respectively. Combination of the skipgram using hierarchical softmax and window size of 8 for 300 dimensions outperformed others in analogy scores for the Wiki Abstract. However, its results are so poor, because of the tiny file size, they're not worth reporting here. Hence, we'll focus on results from the Simple Wiki and Billion Word corpora.
Best combination changes when corpus size increases, as will be noticed from table TABREF5. In terms of analogy score, for 10 epochs, w8s0h0 performs best while w8s1h0 performs best in terms of WordSim and corresponding Spearman correlation. Meanwhile, increasing the corpus size to BW, w4s1h0 performs best in terms of analogy score while w8s1h0 maintains its position as the best in terms of WordSim and Spearman correlation. Besides considering quality metrics, it can be observed from table TABREF6 that comparative ratio of values between the models is not commensurate with the results in intrinsic or extrinsic values, especially when we consider the amount of time and energy spent, since more training time results in more energy consumption (BIBREF23).
Information on the length of training time for the released Mikolov model is not readily available. However, it's interesting to note that their presumed best model, which was released is also s1h0. Its analogy score, which we tested and report, is confirmed in their paper. It beats our best models in only analogy score (even for Simple Wiki), performing worse in others. This is inspite of using a much bigger corpus of 3,000,000 vocabulary size and 100 billion words while Simple Wiki had vocabulary size of 367,811 and is 711MB. It is very likely our analogy scores will improve when we use a much larger corpus, as can be observed from table TABREF5, which involves just one billion words.
Although the two best combinations in analogy (w8s0h0 & w4s0h0) for SW, as shown in fig. FIGREF7, decreased only slightly compared to others with increasing dimensions, the increased training time and much larger serialized model size render any possible minimal score advantage over higher dimensions undesirable. As can be observed in fig. FIGREF9, from 100 dimensions, scores improve but start to drop after over 300 dimensions for SW and after over 400 dimensions for BW. More becomes worse! This trend is true for all combinations for all tests. Polynomial interpolation may be used to determine the optimal dimension in both corpora. Our models are available for confirmation and source codes are available on github.
With regards to NER, most pretrained embeddings outperformed the default pytorch embedding, with our BW w4s1h0 model (which is best in BW analogy score) performing best in F1 score and closely followed by BIBREF0 model. On the other hand, with regards to SA, pytorch embedding outperformed the pretrained embeddings but was closely followed by our SW w8s0h0 model (which also had the best SW analogy score). BIBREF0 performed second worst of all, despite originating from a very huge corpus. The combinations w8s0h0 & w4s0h0 of SW performed reasonably well in both extrinsic tasks, just as the default pytorch embedding did.
Conclusion
This work analyses, empirically, optimal combinations of hyper-parameters for embeddings, specifically for word2vec. It further shows that for downstream tasks, like NER and SA, there's no silver bullet! However, some combinations show strong performance across tasks. Performance of embeddings is task-specific and high analogy scores do not necessarily correlate positively with performance on downstream tasks. This point on correlation is somewhat similar to results by BIBREF24 and BIBREF14. It was discovered that increasing dimension size depreciates performance after a point. If strong considerations of saving time, energy and the environment are made, then reasonably smaller corpora may suffice or even be better in some cases. The on-going drive by many researchers to use ever-growing data to train deep neural networks can benefit from the findings of this work. Indeed, hyper-parameter choices are very important in neural network systems (BIBREF19).
Future work that may be investigated are performance of other architectures of word or sub-word embeddings, the performance and comparison of embeddings applied to languages other than English and how embeddings perform in other downstream tasks. In addition, since the actual reason for the changes in best model as corpus size increases is not clear, this will also be suitable for further research.
The work on this project is partially funded by Vinnova under the project number 2019-02996 "Språkmodeller för svenska myndigheter"
Acronyms
| [
"IMDb dataset of movie reviews",
"IMDb"
] | 2,327 | qasper | en | null | bae15e10e0f414a92fb0e943871ed25c3fc16183a3028012 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What sentiment analysis dataset is used?
Answer: | [
"IMDb dataset of movie reviews",
"IMDb"
] | qasper | 128 | 12 | |
What accuracy does the proposed system achieve? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
There has been significant progress on Named Entity Recognition (NER) in recent years using models based on machine learning algorithms BIBREF0 , BIBREF1 , BIBREF2 . As with other Natural Language Processing (NLP) tasks, building NER systems typically requires a massive amount of labeled training data which are annotated by experts. In real applications, we often need to consider new types of entities in new domains where we do not have existing annotated data. For such new types of entities, however, it is very hard to find experts to annotate the data within short time limits and hiring experts is costly and non-scalable, both in terms of time and money.
In order to quickly obtain new training data, we can use crowdsourcing as one alternative way at lower cost in a short time. But as an exchange, crowd annotations from non-experts may be of lower quality than those from experts. It is one biggest challenge to build a powerful NER system on such a low quality annotated data. Although we can obtain high quality annotations for each input sentence by majority voting, it can be a waste of human labors to achieve such a goal, especially for some ambiguous sentences which may require a number of annotations to reach an agreement. Thus majority work directly build models on crowd annotations, trying to model the differences among annotators, for example, some of the annotators may be more trustful BIBREF3 , BIBREF4 .
Here we focus mainly on the Chinese NER, which is more difficult than NER for other languages such as English for the lack of morphological variations such as capitalization and in particular the uncertainty in word segmentation. The Chinese NE taggers trained on news domain often perform poor in other domains. Although we can alleviate the problem by using character-level tagging to resolve the problem of poor word segmentation performances BIBREF5 , still there exists a large gap when the target domain changes, especially for the texts of social media. Thus, in order to get a good tagger for new domains and also for the conditions of new entity types, we require large amounts of labeled data. Therefore, crowdsourcing is a reasonable solution for these situations.
In this paper, we propose an approach to training a Chinese NER system on the crowd-annotated data. Our goal is to extract additional annotator independent features by adversarial training, alleviating the annotation noises of non-experts. The idea of adversarial training in neural networks has been used successfully in several NLP tasks, such as cross-lingual POS tagging BIBREF6 and cross-domain POS tagging BIBREF7 . They use it to reduce the negative influences of the input divergences among different domains or languages, while we use adversarial training to reduce the negative influences brought by different crowd annotators. To our best knowledge, we are the first to apply adversarial training for crowd annotation learning.
In the learning framework, we perform adversarial training between the basic NER and an additional worker discriminator. We have a common Bi-LSTM for representing annotator-generic information and a private Bi-LSTM for representing annotator-specific information. We build another label Bi-LSTM by the crowd-annotated NE label sequence which reflects the mind of the crowd annotators who learn entity definitions by reading the annotation guidebook. The common and private Bi-LSTMs are used for NER, while the common and label Bi-LSTMs are used as inputs for the worker discriminator. The parameters of the common Bi-LSTM are learned by adversarial training, maximizing the worker discriminator loss and meanwhile minimizing the NER loss. Thus the resulting features of the common Bi-LSTM are worker invariant and NER sensitive.
For evaluation, we create two Chinese NER datasets in two domains: dialog and e-commerce. We require the crowd annotators to label the types of entities, including person, song, brand, product, and so on. Identifying these entities is useful for chatbot and e-commerce platforms BIBREF8 . Then we conduct experiments on the newly created datasets to verify the effectiveness of the proposed adversarial neural network model. The results show that our system outperforms very strong baseline systems. In summary, we make the following contributions:
Related Work
Our work is related to three lines of research: Sequence labeling, Adversarial training, and Crowdsourcing.
Sequence labeling. NER is widely treated as a sequence labeling problem, by assigning a unique label over each sentential word BIBREF9 . Early studies on sequence labeling often use the models of HMM, MEMM, and CRF BIBREF10 based on manually-crafted discrete features, which can suffer the feature sparsity problem and require heavy feature engineering. Recently, neural network models have been successfully applied to sequence labeling BIBREF1 , BIBREF11 , BIBREF2 . Among these work, the model which uses Bi-LSTM for feature extraction and CRF for decoding has achieved state-of-the-art performances BIBREF11 , BIBREF2 , which is exploited as the baseline model in our work.
Adversarial Training. Adversarial Networks have achieved great success in computer vision such as image generation BIBREF12 , BIBREF13 . In the NLP community, the method is mainly exploited under the settings of domain adaption BIBREF14 , BIBREF7 , cross-lingual BIBREF15 , BIBREF6 and multi-task learning BIBREF16 , BIBREF17 . All these settings involve the feature divergences between the training and test examples, and aim to learn invariant features across the divergences by an additional adversarial discriminator, such as domain discriminator. Our work is similar to these work but is applies on crowdsourcing learning, aiming to find invariant features among different crowdsourcing workers.
Crowdsourcing. Most NLP tasks require a massive amount of labeled training data which are annotated by experts. However, hiring experts is costly and non-scalable, both in terms of time and money. Instead, crowdsourcing is another solution to obtain labeled data at a lower cost but with relative lower quality than those from experts. BIBREF18 snow2008cheap collected labeled results for several NLP tasks from Amazon Mechanical Turk and demonstrated that non-experts annotations were quite useful for training new systems. In recent years, a series of work have focused on how to use crowdsourcing data efficiently in tasks such as classification BIBREF19 , BIBREF20 , and compare quality of crowd and expert labels BIBREF21 .
In sequence labeling tasks, BIBREF22 dredze2009sequence viewed this task as a multi-label problem while BIBREF3 rodrigues2014sequence took workers identities into account by assuming that each sentential word was tagged correctly by one of the crowdsourcing workers and proposed a CRF-based model with multiple annotators. BIBREF4 nguyen2017aggregating introduced a crowd representation in which the crowd vectors were added into the LSTM-CRF model at train time, but ignored them at test time. In this paper, we apply adversarial training on crowd annotations on Chinese NER in new domains, and achieve better performances than previous studies on crowdsourcing learning.
Baseline: LSTM-CRF
We use a neural CRF model as the baseline system BIBREF9 , treating NER as a sequence labeling problem over Chinese characters, which has achieved state-of-the-art performances BIBREF5 . To this end, we explore the BIEO schema to convert NER into sequence labeling, following BIBREF2 lample-EtAl:2016:N16-1, where sentential character is assigned with one unique tag. Concretely, we tag the non-entity character by label “O”, the beginning character of an entity by “B-XX”, the ending character of an entity by “E-XX” and the other character of an entity by “I-XX”, where “XX” denotes the entity type.
We build high-level neural features from the input character sequence by a bi-directional LSTM BIBREF2 . The resulting features are combined and then are fed into an output CRF layer for decoding. In summary, the baseline model has three main components. First, we make vector representations for sentential characters $\mathbf {x}_1\mathbf {x}_2\cdots \mathbf {x}_n$ , transforming the discrete inputs into low-dimensional neural inputs. Second, feature extraction is performed to obtain high-level features $\mathbf {h}_1^{\text{ner}}\mathbf {h}_2^{\text{ner}}\cdots \mathbf {h}_n^{\text{ner}}$ , by using a bi-directional LSTM (Bi-LSTM) structure together with a linear transformation over $\mathbf {x}_1\mathbf {x}_2\cdots \mathbf {x}_n$ . Third, we apply a CRF tagging module over $\mathbf {h}_1^{\text{ner}}\mathbf {h}_2^{\text{ner}}\cdots \mathbf {h}_n^{\text{ner}}$ , obtaining the final output NE labels. The overall framework of the baseline model is shown by the right part of Figure 1 .
Vector Representation of Characters
To represent Chinese characters, we simply exploit a neural embedding layer to map discrete characters into the low-dimensional vector representations. The goal is achieved by a looking-up table $\mathbf {E}^W$ , which is a model parameter and will be fine-tuned during training. The looking-up table can be initialized either by random or by using a pretrained embeddings from large scale raw corpus. For a given Chinese character sequence $c_1c_2\cdots c_n$ , we obtain the vector representation of each sentential character by: $ \mathbf {x}_t = \text{look-up}(c_t, \mathbf {E}^W), \text{~~~} t \in [1, n]$ .
Feature Extraction
Based on the vector sequence $\mathbf {x}_1\mathbf {x}_2\cdots \mathbf {x}_n$ , we extract higher-level features $\mathbf {h}_1^{\text{ner}}\mathbf {h}_2^{\text{ner}}\cdots \mathbf {h}_n^{\text{ner}}$ by using a bidirectional LSTM module and a simple feed-forward neural layer, which are then used for CRF tagging at the next step.
LSTM is a type of recurrent neural network (RNN), which is designed for solving the exploding and diminishing gradients of basic RNNs BIBREF23 . It has been widely used in a number of NLP tasks, including POS-tagging BIBREF11 , BIBREF24 , parsing BIBREF25 and machine translation BIBREF26 , because of its strong capabilities of modeling natural language sentences.
By traversing $\mathbf {x}_1\mathbf {x}_2\cdots \mathbf {x}_n$ by order and reversely, we obtain the output features $\mathbf {h}_1^{\text{private}}\mathbf {h}_2^{\text{private}}\cdots \mathbf {h}_n^{\text{private}}$ of the bi-LSTM, where $\mathbf {h}_t^{\text{private}} = \overrightarrow{\mathbf {h}}_t \oplus \overleftarrow{\mathbf {h}}_t $ . Here we refer this Bi-LSTM as private in order to differentiate it with the common Bi-LSTM over the same character inputs which will be introduced in the next section.
Further we make an integration of the output vectors of bi-directional LSTM by a linear feed-forward neural layer, resulting in the features $\mathbf {h}_1^{\text{ner}}\mathbf {h}_2^{\text{ner}}\cdots \mathbf {h}_n^{\text{ner}}$ by equation:
$$\mathbf {h}_t^{\text{ner}} = \mathbf {W} \mathbf {h}_t^{\text{private}} + \mathbf {b},$$ (Eq. 6)
where $\mathbf {W}$ and $\mathbf {b}$ are both model parameters.
CRF Tagging
Finally we feed the resulting features $\mathbf {h}_t^{\text{ner}}, t\in [1, n]$ into a CRF layer directly for NER decoding. CRF tagging is one globally normalized model, aiming to find the best output sequence considering the dependencies between successive labels. In the sequence labeling setting for NER, the output label of one position has a strong dependency on the label of the previous position. For example, the label before “I-XX” must be either “B-XX” or “I-XX”, where “XX” should be exactly the same.
CRF involves two parts for prediction. First we should compute the scores for each label based $\mathbf {h}_t^{\text{ner}}$ , resulting in $\mathbf {o}_t^{\text{ner}}$ , whose dimension is the number of output labels. The other part is a transition matrix $\mathbf {T}$ which defines the scores of two successive labels. $\mathbf {T}$ is also a model parameter. Based on $\mathbf {o}_t^{\text{ner}}$ and $\mathbf {T}$ , we use the Viterbi algorithm to find the best-scoring label sequence.
We can formalize the CRF tagging process as follows:
$$\begin{split} & \mathbf {o}_t^{\text{ner}} = \mathbf {W}^{\text{ner}} \mathbf {h}_t^{\text{ner}}, \text{~~~~} t \in [1,n] \\ & \text{score}(\mathbf {X}, \mathbf {y}) = \sum _{t = 1}^{n}(\mathbf {o}_{t,y_t} + T_{y_{t-1},y_t}) \\ & \mathbf {y}^{\text{ner}} = \mathop {arg~max}_{\mathbf {y}}\big (\text{score}(\mathbf {X}, \mathbf {y}))\big ), \\ \end{split}$$ (Eq. 8)
where $\text{score}(\cdot )$ is the scoring function for a given output label sequence $\mathbf {y} = y_1y_2 \cdots y_n$ based on input $\mathbf {X}$ , $\mathbf {y}^{\text{ner}}$ is the resulting label sequence, $\mathbf {W}^{\text{ner}}$ is a model parameter.
Training
To train model parameters, we exploit a negative log-likelihood objective as the loss function. We apply softmax over all candidate output label sequences, thus the probability of the crowd-annotated label sequence is computed by:
$$p(\mathbf {\bar{y}}|\mathbf {X}) = \frac{\exp \big (\text{score}(\mathbf {X}, \mathbf {\bar{y}})\big )}{\sum _{\mathbf {y} \in \mathbf {Y}_{\mathbf {X}}} \exp \big (\text{score}(\mathbf {X}, \mathbf {y})\big )},$$ (Eq. 10)
where $\mathbf {\bar{y}}$ is the crowd-annotated label sequences and $\mathbf {Y}_{\mathbf {X}}$ is all candidate label sequence of input $\mathbf {X}$ .
Based on the above formula, the loss function of our baseline model is:
$$\text{loss}(\Theta , \mathbf {X}, \mathbf {\bar{y}}) = -\log p(\mathbf {\bar{y}}|\mathbf {X}),$$ (Eq. 11)
where $\Theta $ is the set of all model parameters. We use standard back-propagation method to minimize the loss function of the baseline CRF model.
Worker Adversarial
Adversarial learning has been an effective mechanism to resolve the problem of the input features between the training and test examples having large divergences BIBREF27 , BIBREF13 . It has been successfully applied on domain adaption BIBREF7 , cross-lingual learning BIBREF15 and multi-task learning BIBREF17 . All settings involve feature shifting between the training and testing.
In this paper, our setting is different. We are using the annotations from non-experts, which are noise and can influence the final performances if they are not properly processed. Directly learning based on the resulting corpus may adapt the neural feature extraction into the biased annotations. In this work, we assume that individual workers have their own guidelines in mind after short training. For example, a perfect worker can annotate highly consistently with an expert, while common crowdsourcing workers may be confused and have different understandings on certain contexts. Based on the assumption, we make an adaption for the original adversarial neural network to our setting.
Our adaption is very simple. Briefly speaking, the original adversarial learning adds an additional discriminator to classify the type of source inputs, for example, the domain category in the domain adaption setting, while we add a discriminator to classify the annotation workers. Solely the features from the input sentence is not enough for worker classification. The annotation result of the worker is also required. Thus the inputs of our discriminator are different. Here we exploit both the source sentences and the crowd-annotated NE labels as basic inputs for the worker discrimination.
In the following, we describe the proposed adversarial learning module, including both the submodels and the training method. As shown by the left part of Figure 1 , the submodel consists of four parts: (1) a common Bi-LSTM over input characters; (2) an additional Bi-LSTM to encode crowd-annotated NE label sequence; (3) a convolutional neural network (CNN) to extract features for worker discriminator; (4) output and prediction.
Common Bi-LSTM over Characters
To build the adversarial part, first we create a new bi-directional LSTM, named by the common Bi-LSTM:
$$\mathbf {h}_1^{\text{\tiny common}} \mathbf {h}_2^{\text{\tiny common}} \cdots \mathbf {h}_n^{\text{\tiny common}} = \text{Bi-LSTM}(\mathbf {x}_1\mathbf {x}_2\cdots \mathbf {x}_n).$$ (Eq. 13)
As shown in Figure 1 , this Bi-LSTM is constructed over the same input character representations of the private Bi-LSTM, in order to extract worker independent features.
The resulting features of the common Bi-LSTM are used for both NER and the worker discriminator, different with the features of private Bi-LSTM which are used for NER only. As shown in Figure 1 , we concatenate the outputs of the common and private Bi-LSTMs together, and then feed the results into the feed-forward combination layer of the NER part. Thus Formula 6 can be rewritten as:
$$\mathbf {h}_t^{\text{ner}} = \mathbf {W} (\mathbf {h}_t^{\text{common}} \oplus \mathbf {h}_t^{\text{private}}) + \mathbf {b},$$ (Eq. 14)
where $\mathbf {W}$ is wider than the original combination because the newly-added $\mathbf {h}_t^{\text{common}}$ .
Noticeably, although the resulting common features are used for the worker discriminator, they actually have no capability to distinguish the workers. Because this part is exploited to maximize the loss of the worker discriminator, it will be interpreted in the later training subsection. These features are invariant among different workers, thus they can have less noises for NER. This is the goal of adversarial learning, and we hope the NER being able to find useful features from these worker independent features.
Additional Bi-LSTM over Annotated NER Labels
In order to incorporate the annotated NE labels to predict the exact worker, we build another bi-directional LSTM (named by label Bi-LSTM) based on the crowd-annotated NE label sequence. This Bi-LSTM is used for worker discriminator only. During the decoding of the testing phase, we will never have this Bi-LSTM, because the worker discriminator is no longer required.
Assuming the crowd-annotated NE label sequence annotated by one worker is $\mathbf {\bar{y}} = \bar{y}_1\bar{y}_2 \cdots \bar{y}_n$ , we exploit a looking-up table $\mathbf {E}^{L}$ to obtain the corresponding sequence of their vector representations $\mathbf {x^{\prime }}_1\mathbf {x^{\prime }}_2\cdots \mathbf {x^{\prime }}_n$ , similar to the method that maps characters into their neural representations. Concretely, for one NE label $\bar{y}_t$ ( $t \in [1, n]$ ), we obtain its neural vector by: $\mathbf {x^{\prime }}_t = \text{look-up}(\bar{y}_t, \mathbf {E}^L)$ .
Next step we apply bi-directional LSTM over the sequence $\mathbf {x^{\prime }}_1\mathbf {x^{\prime }}_2\cdots \mathbf {x^{\prime }}_n$ , which can be formalized as:
$$\mathbf {h}_1^{\text{label}} \mathbf {h}_2^{\text{label}} \cdots \mathbf {h}_n^{\text{label}} = \text{Bi-LSTM}(\mathbf {x^{\prime }}_1\mathbf {x^{\prime }}_2\cdots \mathbf {x^{\prime }}_n).$$ (Eq. 16)
The resulting feature sequence is concatenated with the outputs of the common Bi-LSTM, and further be used for worker classification.
CNN
Following, we add a convolutional neural network (CNN) module based on the concatenated outputs of the common Bi-LSTM and the label Bi-LSTM, to produce the final features for worker discriminator. A convolutional operator with window size 5 is used, and then max pooling strategy is applied over the convolution sequence to obtain the final fixed-dimensional feature vector. The whole process can be described by the following equations:
$$\begin{split} &\mathbf {h}_t^{\text{worker}} = \mathbf {h}_t^{\text{common}} \oplus \mathbf {h}_t^{\text{label}} \\ &\mathbf {\tilde{h}}_t^{\text{worker}} = \tanh (\mathbf {W}^{\text{cnn}}[\mathbf {h}_{t-2}^{\text{worker}}, \mathbf {h}_{t-1}^{\text{worker}}, \cdots , \mathbf {h}_{t+2}^{\text{worker}}]) \\ &\mathbf {h}^{\text{worker}} = \text{max-pooling}(\mathbf {\tilde{h}}_1^{\text{worker}}\mathbf {\tilde{h}}_2^{\text{worker}} \cdots \mathbf {\tilde{h}}_n^{\text{worker}}) \\ \end{split}$$ (Eq. 18)
where $t \in [1,n]$ and $\mathbf {W}^{\text{cnn}}$ is one model parameter. We exploit zero vector to paddle the out-of-index vectors.
Output and Prediction
After obtaining the final feature vector for the worker discriminator, we use it to compute the output vector, which scores all the annotation workers. The score function is defined by:
$$\mathbf {o}^{\text{worker}} = \mathbf {W}^{\text{worker}} \mathbf {h}^{\text{worker}},$$ (Eq. 20)
where $\mathbf {W}^{\text{worker}}$ is one model parameter and the output dimension equals the number of total non-expert annotators. The prediction is to find the worker which is responsible for this annotation.
Adversarial Training
The training objective with adversarial neural network is different from the baseline model, as it includes the extra worker discriminator. Thus the new objective includes two parts, one being the negative log-likelihood from NER which is the same as the baseline, and the other being the negative the log-likelihood from the worker discriminator.
In order to obtain the negative log-likelihood of the worker discriminator, we use softmax to compute the probability of the actual worker $\bar{z}$ as well, which is defined by:
$$p(\bar{z}|\mathbf {X}, \mathbf {\bar{y}}) = \frac{\exp (\mathbf {o}^{\text{worker}}_{\bar{z}})}{\sum _{z} \exp (\mathbf {o}^{\text{worker}}_z)},$$ (Eq. 22)
where $z$ should enumerate all workers.
Based on the above definition of probability, our new objective is defined as follows:
$$\begin{split} \text{R}(\Theta , \Theta ^{\prime }, \mathbf {X}, \mathbf {\bar{y}}, \bar{z}) &= \text{loss}(\Theta , \mathbf {X}, \mathbf {\bar{y}}) - \text{loss}(\Theta , \Theta ^{\prime }, \mathbf {X}) \\ \text{~~~~~~} &= -\log p(\mathbf {\bar{y}}|\mathbf {X}) + \log p(\bar{z}|\mathbf {X}, \mathbf {\bar{y}}), \end{split}$$ (Eq. 23)
where $\Theta $ is the set of all model parameters related to NER, and $\Theta ^{\prime }$ is the set of the remaining parameters which are only related to the worker discriminator, $\mathbf {X}$ , $\mathbf {\bar{y}}$ and $\bar{z}$ are the input sentence, the crowd-annotated NE labels and the corresponding annotator for this annotation, respectively. It is worth noting that the parameters of the common Bi-LSTM are included in the set of $\Theta $ by definition.
In particular, our goal is not to simply minimize the new objective. Actually, we aim for a saddle point, finding the parameters $\Theta $ and $\Theta ^{\prime }$ satisfying the following conditions:
$$\begin{split} \hat{\Theta } &= \mathop {arg~min}_{\Theta }\text{R}(\Theta , \Theta ^{\prime }, \mathbf {X}, \mathbf {\bar{y}}, \bar{z}) \\ \hat{\Theta }^{\prime } &= \mathop {arg~max}_{\Theta ^{\prime }}\text{R}(\hat{\Theta }, \Theta ^{\prime }, \mathbf {X}, \mathbf {\bar{y}}, \bar{z}) \\ \end{split}$$ (Eq. 24)
where the first equation aims to find one $\Theta $ that minimizes our new objective $\text{R}(\cdot )$ , and the second equation aims to find one $\Theta ^{\prime }$ maximizing the same objective.
Intuitively, the first equation of Formula 24 tries to minimize the NER loss, but at the same time maximize the worker discriminator loss by the shared parameters of the common Bi-LSTM. Thus the resulting features of common Bi-LSTM actually attempt to hurt the worker discriminator, which makes these features worker independent since they are unable to distinguish different workers. The second equation tries to minimize the worker discriminator loss by its own parameter $\Theta ^{\prime }$ .
We use the standard back-propagation method to train the model parameters, the same as the baseline model. In order to incorporate the term of the argmax part of Formula 24 , we follow the previous work of adversarial training BIBREF13 , BIBREF15 , BIBREF17 , by introducing a gradient reverse layer between the common Bi-LSTM and the CNN module, whose forward does nothing but the backward simply negates the gradients.
Data Sets
With the purpose of obtaining evaluation datasets from crowd annotators, we collect the sentences from two domains: Dialog and E-commerce domain. We hire undergraduate students to annotate the sentences. They are required to identify the predefined types of entities in the sentences. Together with the guideline document, the annotators are educated some tips in fifteen minutes and also provided with 20 exemplifying sentences.
Labeled Data: DL-PS. In Dialog domain (DL), we collect raw sentences from a chatbot application. And then we randomly select 20K sentences as our pool and hire 43 students to annotate the sentences. We ask the annotators to label two types of entities: Person-Name and Song-Name. The annotators label the sentences independently. In particular, each sentence is assigned to three annotators for this data. Although the setting can be wasteful of labor, we can use the resulting dataset to test several well-known baselines such as majority voting.
After annotation, we remove some illegal sentences reported by the annotators. Finally, we have 16,948 sentences annotated by the students. Table 1 shows the information of annotated data. The average Kappa value among the annotators is 0.6033, indicating that the crowd annotators have moderate agreement on identifying entities on this data.
In order to evaluate the system performances, we create a set of corpus with gold annotations. Concretely, we randomly select 1,000 sentences from the final dataset and let two experts generate the gold annotations. Among them, we use 300 sentences as the development set and the remaining 700 as the test set. The rest sentences with only student annotations are used as the training set.
Labeled data: EC-MT and EC-UQ. In E-commerce domain (EC), we collect raw sentences from two types of texts: one is titles of merchandise entries (EC-MT) and another is user queries (EC-UQ). The annotators label five types of entities: Brand, Product, Model, Material, and Specification. These five types of entities are very important for E-commerce platform, for example building knowledge graph of merchandises. Five students participate the annotations for this domain since the number of sentences is small. We use the similar strategy as DL-PS to annotate the sentences, except that only two annotators are assigned for each sentence, because we aim to test the system performances under very small duplicated annotations.
Finally, we obtain 2,337 sentences for EC-MT and 2,300 for EC-UQ. Table 1 shows the information of annotated results. Similarly, we produce the development and test datasets for system evaluation, by randomly selecting 400 sentences and letting two experts to generate the groundtruth annotations. Among them, we use 100 sentences as the development set and the remaining 300 as the test set. The rest sentences with only crowdsourcing annotations are used as the training set.
Unlabeled data. The vector representations of characters are basic inputs of our baseline and proposed models, which are obtained by the looking-up table $\mathbf {E}^W$ . As introduced before, we can use pretrained embeddings from large-scale raw corpus to initialize the table. In order to pretrain the character embeddings, we use one large-scale unlabeled data from the user-generated content in Internet. Totally, we obtain a number of 5M sentences. Finally, we use the tool word2vec to pretrain the character embeddings based on the unlabeled dataset in our experiments.
Settings
For evaluation, we use the entity-level metrics of Precision (P), Recall (R), and their F1 value in our experiments, treating one tagged entity as correct only when it matches the gold entity exactly.
There are several hyper-parameters in the baseline LSTM-CRF and our final models. We set them empirically by the development performances. Concretely, we set the dimension size of the character embeddings by 100, the dimension size of the NE label embeddings by 50, and the dimension sizes of all the other hidden features by 200.
We exploit online training with a mini-batch size 128 to learn model parameters. The max-epoch iteration is set by 200, and the best-epoch model is chosen according to the development performances. We use RMSprop BIBREF28 with a learning rate $10^{-3}$ to update model parameters, and use $l_2$ -regularization by a parameter $10^{-5}$ . We adopt the dropout technique to avoid overfitting by a drop value of $0.2$ .
Comparison Systems
The proposed approach (henceforward referred to as “ALCrowd”) is compared with the following systems:
CRF: We use the Crfsuite tool to train a model on the crowdsourcing labeled data. As for the feature settings, we use the supervised version of BIBREF0 zhao2008unsupervised.
CRF-VT: We use the same settings of the CRF system, except that the training data is the voted version, whose groundtruths are produced by majority voting at the character level for each annotated sentence.
CRF-MA: The CRF model proposed by BIBREF3 rodrigues2014sequence, which uses a prior distributation to model multiple crowdsourcing annotators. We use the source code provided by the authors.
LSTM-CRF: Our baseline system trained on the crowdsourcing labeled data.
LSTM-CRF-VT: Our baseline system trained on the voted corpus, which is the same as CRF-VT.
LSTM-Crowd: The LSTM-CRF model with crowd annotation learning proposed by BIBREF4 nguyen2017aggregating. We use the source code provided by the authors.
The first three systems are based on the CRF model using traditional handcrafted features, and the last three systems are based on the neural LSTM-CRF model. Among them, CRF-MA, LSTM-Crowd and our system with adversarial learning (ALCrowd) are based on crowd annotation learning that directly trains the model on the crowd-annotations. Five systems, including CRF, CRF-MA, LSTM-CRF, LSTM-Crowd, and ALCrowd, are trained on the original version of labeled data, while CRF-VT and LSTM-CRF-VT are trained on the voted version. Since CRF-VT, CRF-MA and LSTM-CRF-VT all require ground-truth answers for each training sentence, which are difficult to be produced with only two annotations, we do not apply the three models on the two EC datasets.
Main Results
In this section, we show the model performances of our proposed crowdsourcing learning system (ALCrowd), and meanwhile compare it with the other systems mentioned above. Table 2 shows the experimental results on the DL-PS datasets and Table 3 shows the experiment results on the EC-MT and EC-UQ datasets, respectively.
The results of CRF and LSTM-CRF mean that the crowd annotation is an alternative solution with low cost for labeling data that could be used for training a NER system even there are some inconsistencies. Compared with CRF, LSTM-CRF achieves much better performances on all the three data, showing +6.12 F1 improvement on DL-PS, +4.51 on EC-MT, and +9.19 on EC-UQ. This indicates that LSTM-CRF is a very strong baseline system, demonstrating the effectiveness of neural network.
Interestingly, when compared with CRF and LSTM-CRF, CRF-VT and LSTM-CRF-VT trained on the voted version perform worse in the DL-PS dataset. This trend is also mentioned in BIBREF4 nguyen2017aggregating. This fact shows that the majority voting method might be unsuitable for our task. There are two possible reasons accounting for the observation. On the one hand, simple character-level voting based on three annotations for each sentence may be still not enough. In the DL-PS dataset, even with only two predefined entity types, one character can have nine NE labels. Thus the majority-voting may be incapable of handling some cases. While the cost by adding more annotations for each sentence would be greatly increased. On the other hand, the lost information produced by majority-voting may be important, at least the ambiguous annotations denote that the input sentence is difficult for NER. The normal CRF and LSTM-CRF models without discard any annotations can differentiate these difficult contexts through learning.
Three crowd-annotation learning systems provide better performances than their counterpart systems, (CRF-MA VS CRF) and (LSTM-Crowd/ALCrowd VS LSTM-CRF). Compared with the strong baseline LSTM-CRF, ALCrowd shows its advantage with +1.08 F1 improvements on DL-PS, +1.24 on EC-MT, and +2.38 on EC-UQ, respectively. This indicates that adding the crowd-annotation learning is quite useful for building NER systems. In addition, ALCrowd also outperforms LSTM-Crowd on all the datasets consistently, demonstrating the high effectiveness of ALCrowd in extracting worker independent features. Among all the systems, ALCrowd performs the best, and significantly better than all the other models (the p-value is below $10^{-5}$ by using t-test). The results indicate that with the help of adversarial training, our system can learn a better feature representation from crowd annotation.
Discussion
Impact of Character Embeddings. First, we investigate the effect of the pretrained character embeddings in our proposed crowdsourcing learning model. The comparison results are shown in Figure 2 , where Random refers to the random initialized character embeddings, and Pretrained refers to the embeddings pretrained on the unlabeled data. According to the results, we find that our model with the pretrained embeddings significantly outperforms that using the random embeddings, demonstrating that the pretrained embeddings successfully provide useful information.
Case Studies. Second, we present several case studies in order to study the differences between our baseline and the worker adversarial models. We conduct a closed test on the training set, the results of which can be regarded as modifications of the training corpus, since there exist inconsistent annotations for each training sentence among the different workers. Figure 3 shows the two examples from the DL-PS dataset, which compares the outputs of the baseline and our final models, as well as the majority-voting strategy.
In the first case, none of the annotations get the correct NER result, but our proposed model can capture it. The result of LSTM-CRF is the same as majority-voting. In the second example, the output of majority-voting is the worst, which can account for the reason why the same model trained on the voted corpus performs so badly, as shown in Table 2 . The model of LSTM-CRF fails to recognize the named entity “Xiexie” because of not trusting the second annotation, treating it as one noise annotation. Our proposed model is able to recognize it, because of its ability of extracting worker independent features.
Conclusions
In this paper, we presented an approach to performing crowd annotation learning based on the idea of adversarial training for Chinese Named Entity Recognition (NER). In our approach, we use a common and private Bi-LSTMs for representing annotator-generic and -specific information, and learn a label Bi-LSTM from the crowd-annotated NE label sequences. Finally, the proposed approach adopts a LSTM-CRF model to perform tagging. In our experiments, we create two data sets for Chinese NER tasks in the dialog and e-commerce domains. The experimental results show that the proposed approach outperforms strong baseline systems.
Acknowledgments
This work is supported by the National Natural Science Foundation of China (Grant No. 61572338, 61525205, and 61602160). This work is also partially supported by the joint research project of Alibaba and Soochow University. Wenliang is also partially supported by Collaborative Innovation Center of Novel Software Technology and Industrialization.
| [
"F1 scores of 85.99 on the DL-PS data, 75.15 on the EC-MT data and 71.53 on the EC-UQ data ",
"F1 of 85.99 on the DL-PS dataset (dialog domain); 75.15 on EC-MT and 71.53 on EC-UQ (e-commerce domain)"
] | 5,310 | qasper | en | null | d022dfe02fb2a55b4baa40fe436f616aecd3f3ced1a58d7c | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What accuracy does the proposed system achieve?
Answer: | [
"F1 scores of 85.99 on the DL-PS data, 75.15 on the EC-MT data and 71.53 on the EC-UQ data ",
"F1 of 85.99 on the DL-PS dataset (dialog domain); 75.15 on EC-MT and 71.53 on EC-UQ (e-commerce domain)"
] | qasper | 128 | 13 | |
Did they experiment with this new dataset? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
How humans process language has become increasingly relevant in natural language processing since physiological data during language understanding is more accessible and recorded with less effort. In this work, we focus on eye-tracking and electroencephalography (EEG) recordings to capture the reading process. On one hand, eye movement data provides millisecond-accurate records about where humans look when they are reading, and is highly correlated with the cognitive load associated with different stages of text processing. On the other hand, EEG records electrical brain activity across the scalp and is a direct measure of physiological processes, including language processing. The combination of both measurement methods enables us to study the language understanding process in a more natural setting, where participants read full sentences at a time, in their own speed. Eye-tracking then permits us to define exact word boundaries in the timeline of a subject reading a sentence, allowing the extraction of brain activity signals for each word.
Human cognitive language processing data is immensely useful for NLP: Not only can it be leveraged to improve NLP applications (e.g. barrett2016weakly for part-of-speech tagging or klerke2016improving for sentence compression), but also to evaluate state-of-the-art machine learning systems. For example, hollenstein2019cognival evaluate word embeddings, or schwartz2019inducing fine-tune language models with brain-relevant bias.
Additionally, the availability of labelled data plays a crucial role in all supervised machine learning applications. Physiological data can be used to understand and improve the labelling process (e.g. tokunaga2017eye), and, for instance, to build cost models for active learning scenarios BIBREF0. Is it possible to replace this expensive manual work with models trained on physiological activity data recorded from humans while reading? That is to say, can we find and extract relevant aspects of text understanding and annotation directly from the source, i.e. eye-tracking and brain activity signals during reading?
Motivated by these questions and our previously released dataset, ZuCo 1.0 BIBREF1, we developed this new corpus, where we specifically aim to collect recordings during natural reading as well as during annotation.
We provide the first dataset of simultaneous eye movement and brain activity recordings to analyze and compare normal reading to task-specific reading during annotation. The Zurich Cognitive Language Processing Corpus (ZuCo) 2.0, including raw and preprocessed eye-tracking and electroencephalography (EEG) data of 18 subjects, as well as the recording and preprocessing scripts, is publicly available at https://osf.io/2urht/. It contains physiological data of each subject reading 739 English sentences from Wikipedia (see example in Figure FIGREF1). We want to highlight the re-use potential of this data. In addition to the psycholinguistic motivation, this corpus is especially tailored for training and evaluating machine learning algorithms for NLP purposes. We conduct a detailed technical validation of the data as proof of the quality of the recordings.
Related Work
Some eye-tracking corpora of natural reading (e.g. the Dundee BIBREF2, Provo BIBREF3 and GECO corpus BIBREF4), and a few EEG corpora (for example, the UCL corpus BIBREF5) are available. It has been shown that this type of cognitive processing data is useful for improving and evaluating NLP methods (e.g. barrett2018sequence,hollenstein2019cognival, hale2018finding). However, before the Zurich Cognitive Language Processing Corpus (ZuCo 1.0), there was no available data for simultaneous eye-tracking and EEG recordings of natural reading. dimigen2011coregistration studied the linguistic effects of eye movements and EEG co-registration in natural reading and showed that they accurately represent lexical processing. Moreover, the simultaneous recordings are crucial to extract word-level brain activity signals.
While the above mentioned studies analyze and leverage natural reading, some NLP work has used eye-tracking during annotation (but, as of yet, not EEG data). mishra2016predicting and joshi2014measuring recorded eye-tracking during binary sentiment annotation (positive/negative). This data was used to determine the annotation complexity of the text passages based on eye movement metrics and for sarcasm detection BIBREF6. Moreover, eye-tracking has been used to analyze the word sense annotation process in Hindi BIBREF7, named entity annotation in Japanese BIBREF8, and to leverage annotator gaze behaviour for coreference resolution BIBREF9. Finally, tomanek2010cognitive used eye-tracking data during entity annotation to build a cost model for active learning. However, until now there is no available data or research that analyzes the differences in the human processing of normal reading versus annotation.
Related Work ::: ZuCo1.0
In previous work, we recorded a first dataset of simultaneous eye-tracking and EEG during natural reading BIBREF1. ZuCo 1.0 consists of three reading tasks, two of which contain very similar reading material and experiments as presented in the current work. However, the main difference and reason for recording ZuCo 2.0, consists in the experiment procedure. For ZuCo 1.0 the normal reading and task-specific reading paradigms were recorded in different sessions on different days. Therefore, the recorded data is not appropriate as a means of comparison between natural reading and annotation, since the differences in the brain activity data might result mostly from the different sessions due to the sensitivity of EEG. This, and extending the dataset with more sentences and more subjects, were the main factors for recording the current corpus. We purposefully maintained an overlap of some sentences between both datasets to allow additional analyses (details are described in Section SECREF7).
Corpus Construction
In this section we describe the contents and experimental design of the ZuCo 2.0 corpus.
Corpus Construction ::: Participants
We recorded data from 19 participants and discarded the data of one of them due to technical difficulties with the eye-tracking calibration. Hence, we share the data of 18 participants. All participants are healthy adults (mean age = 34 (SD=8.3), 10 females). Their native language is English, originating from Australia, Canada, UK, USA or South Africa. Two participants are left-handed and three participants wear glasses for reading. Details on subject demographics can be found in Table TABREF4. All participants gave written consent for their participation and the re-use of the data prior to the start of the experiments. The study was approved by the Ethics Commission of the University of Zurich.
Corpus Construction ::: Reading materials
During the recording session, the participants read 739 sentences that were selected from the Wikipedia corpus provided by culotta2006integrating. This corpus was chosen because it provides annotations of semantic relations. We included seven of the originally defined relation types: political_affiliation, education, founder, wife/husband, job_title, nationality, and employer. The sentences were chosen in the same length range as ZuCo 1.0, and with similar Flesch reading ease scores. The dataset statistics are shown in Table TABREF2.
Of the 739 sentences, the participants read 349 sentences in a normal reading paradigm, and 390 sentences in a task-specific reading paradigm, in which they had to determine whether a certain relation type occurred in the sentence or not. Table TABREF3 shows the distribution of the different relation types in the sentences of the task-specific annotation paradigm.
Purposefully, there are 63 duplicates between the normal reading and the task-specific sentences (8% of all sentences). The intention of these duplicate sentences is to provide a set of sentences read twice by all participants with a different task in mind. Hence, this enables the comparison of eye-tracking and brain activity data when reading normally and when annotating specific relations (see examples in Section SECREF4).
Furthermore, there is also an overlap in the sentences between ZuCo 1.0 and ZuCo 2.0. 100 normal reading and 85 task-specific sentences recorded for this dataset were already recorded in ZuCo 1.0. This allows for comparisons between the different recording procedures (i.e. session-specific effects) and between more participants (subject-specific effects).
Corpus Construction ::: Experimental design
As mentioned above, we recorded two different reading tasks for the ZuCo 2.0 dataset. During both tasks the participants were able to read in their own speed, using a control pad to move to the next sentence and to answer the control questions, which allowed for natural reading. Since each subject reads at their own personal pace, the reading speed between varies between subjects. Table TABREF4 shows the average reading speed for each task, i.e. the average number of seconds a subject spends per sentence before switching to the next one.
All 739 sentences were recorded in a single session for each participant. The duration of the recording sessions was between 100 and 180 minutes, depending on the time required to set up and calibrate the devices, and the personal reading speed of the participants.
We recorded 14 blocks of approx. 50 sentences, alternating between tasks: 50 sentences of normal reading, followed by 50 sentences of task-specific reading. The order of blocks and sentences within blocks was identical for all subjects. Each sentence block was preceded by a practice round of three sentences.
Corpus Construction ::: Experimental design ::: Normal reading (NR)
In the first task, participants were instructed to read the sentences naturally, without any specific task other than comprehension. Participants were told to read the sentences normally without any special instructions. Figure FIGREF8 (left) shows an example sentence as it was depicted on the screen during recording. As shown in Figure FIGREF8 (middle), the control condition for this task consisted of single-choice questions about the content of the previous sentence. 12% of randomly selected sentences were followed by such a comprehension question with three answer options on a separate screen.
Corpus Construction ::: Experimental design ::: Task-specific reading (TSR)
In the second task, the participants were instructed to search for a specific relation in each sentence they read. Instead of comprehension questions, the participants had to decide for each sentence whether it contains the relation or not, i.e. they were actively annotating each sentence. Figure FIGREF8 (right) shows an example screen for this task. 17% of the sentences did not include the relation type and were used as control conditions. All sentences within one block involved the same relation type. The blocks started with a practice round, which described the relation and was followed by three sample sentences, so that the participants would be familiar with the respective relation type.
Corpus Construction ::: Linguistic assessment
As a linguistic assessment, the vocabulary and language proficiency of the participants was tested with the LexTALE test (Lexical Test for Advanced Learners of English, lemhofer2012introducing). This is an unspeeded lexical decision task designed for intermediate to highly proficient language users. The average LexTALE score over all participants was 88.54%. Moreover, we also report the scores the participants achieved with their answers to the reading comprehension control questions and their relation annotations. The detailed scores for all participants are also presented in Table TABREF4.
Corpus Construction ::: Data acquisition
Data acquisition took place in a sound-attenuated and dark experiment room. Participants were seated at a distance of 68cm from a 24-inch monitor with a resolution of 800x600 pixels. A stable head position was ensured via a chin rest. Participants were instructed to stay as still as possible during the tasks to avoid motor EEG artifacts. Participants were also offered snacks and water during the breaks and were encouraged to rest. All sentences were presented at the same position on the screen and could span multiple lines. The sentences were presented in black on a light grey background with font size 20-point Arial, resulting in a letter height of 0.8 mm. The experiment was programmed in MATLAB 2016b BIBREF10, using PsychToolbox BIBREF11. Participants completed the tasks sitting alone in the room, while two research assistants were monitoring their progress in the adjoining room. All recording scripts including detailed participant instructions are available alongside the data.
Corpus Construction ::: Data acquisition ::: Eye-tracking acquisition
Eye position and pupil size were recorded with an infrared video-based eye tracker (EyeLink 1000 Plus, SR Research) at a sampling rate of 500 Hz. The eye tracker was calibrated with a 9-point grid at the beginning of the session and re-validated before each block of sentences.
Corpus Construction ::: Data acquisition ::: EEG acquisition
High-density EEG data were recorded at a sampling rate of 500 Hz with a bandpass of 0.1 to 100 Hz, using a 128-channel EEG Geodesic Hydrocel system (Electrical Geodesics). The recording reference was set at electrode Cz. The head circumference of each participant was measured to select an appropriately sized EEG net. To ensure good contact, the impedance of each electrode was checked prior to recording, and was kept below 40 kOhm. Electrode impedance levels were checked after every third block of 50 sentences (approx. every 30 mins) and reduced if necessary.
Corpus Construction ::: Preprocessing and feature extraction ::: Eye-tracking
The eye-tracking data consists of (x,y) gaze location entries for all individual fixations (Figure FIGREF1b). Coordinates were given in pixels with respect to the monitor coordinates (the upper left corner of the screen was (0,0) and down/right was positive). We provide this raw data as well as various engineered eye-tracking features. For this feature extraction only fixations within the boundaries of each displayed word were extracted. Data points distinctly not associated with reading (minimum distance of 50 pixels to the text) were excluded. Additionally, fixations shorter than 100 ms were excluded from the analyses, because these are unlikely to reflect fixations relevant for reading BIBREF12. On the basis of the GECO and ZuCo 1.0 corpora, we extracted the following features: (i) gaze duration (GD), the sum of all fixations on the current word in the first-pass reading before the eye moves out of the word; (ii) total reading time (TRT), the sum of all fixation durations on the current word, including regressions; (iii) first fixation duration (FFD), the duration of the first fixation on the prevailing word; (iv) single fixation duration (SFD), the duration of the first and only fixation on the current word; and (v) go-past time (GPT), the sum of all fixations prior to progressing to the right of the current word, including regressions to previous words that originated from the current word. For each of these eye-tracking features we additionally computed the pupil size. Furthermore, we extracted the number of fixations and mean pupil size for each word and sentence.
Corpus Construction ::: Preprocessing and feature extraction ::: EEG
The EEG data shared in this project are available as raw data, but also preprocessed with Automagic (version 1.4.6, pedroni2019automagic), a tool for automatic EEG data cleaning and validation. 105 EEG channels (i.e. electrodes) were used from the scalp recordings. 9 EOG channels were used for artifact removal and additional 14 channels lying mainly on the neck and face were discarded before data analysis. Bad channels were identified and interpolated. We used the Multiple Artifact Rejection Algorithm (MARA), a supervised machine learning algorithm that evaluates ICA components, for automatic artifact rejection. MARA has been trained on manual component classifications, and thus captures a wide range of artifacts. MARA is especially effective at detecting and removing eye and muscle artifact components. The effect of this preprocessing can be seen in Figure FIGREF1d.
After preprocessing, we synchronized the EEG and eye-tracking data to enable EEG analyses time-locked to the onsets of fixations. To compute oscillatory power measures, we band-pass filtered the continuous EEG signals across an entire reading task for five different frequency bands resulting in a time-series for each frequency band. The independent frequency bands were determined as follows: theta$_1$ (4–6 Hz), theta$_2$ (6.5–8 Hz), alpha$_1$ (8.5–10 Hz), alpha$_2$ (10.5–13 Hz), beta$_1$ (13.5–18 Hz), beta$_2$ (18.5–30 Hz), gamma$_1$ (30.5–40 Hz), and gamma$_2$ (40–49.5 Hz). We then applied a Hilbert transformation to each of these time-series. We specifically chose the Hilbert transformation to maintain the temporal information of the amplitude of the frequency bands, to enable the power of the different frequencies for time segments defined through the fixations from the eye-tracking recording. Thus, for each eye-tracking feature we computed the corresponding EEG feature in each frequency band. Furthermore, we extracted sentence-level EEG features by calculating the power in each frequency band, and additionally, the difference of the power spectra between frontal left and right homologue electrodes pairs. For each eye-tracking based EEG feature, all channels were subject to an artifact rejection criterion of $90\mu V$ to exclude trials with transient noise.
Data Validation
The aim of the technical validation of the data is to guarantee good recording quality and to replicate findings of previous studies investigating co-registration of EEG and eye movement data during natural reading tasks (e.g. dimigen2011coregistration). We also compare the results to ZuCo 1.0 BIBREF1, which allows a more direct comparison due to the analogous recording procedure.
Data Validation ::: Eye-tracking
We validated the recorded eye-tracking data by analyzing the fixations made by all subjects through their reading speed and omission rate on sentence level. The omission rate is defined as the percentage of words that is not fixated in a sentence. Figure FIGREF10 (middle) shows the mean reading speed over all subjects, measured in seconds per sentence and Figure FIGREF10 (right) shows the mean omission rates aggregated over all subjects for each task. Clearly, the participants made less fixations during the task-specific reading, which lead to faster reading speed.
Moreover, we corroborated these sentence-level metrics by visualizing the skipping proportion on word level (Figure FIGREF13). The skipping proportion is the average rate of words being skipped (i.e. not being fixated) in a sentence. As expected, this also increases in the task-specific reading.
Although the reading material is from the same source and of the same length range (see Figure FIGREF10 (left)), in the first task (NR) passive reading was recorded, while in the second task (TSR) the subjects had to annotate a specific relation type in each sentence. Thus, the task-specific annotation reading lead to shorter passes because the goal was merely to recognize a relation in the text, but not necessarily to process the every word in each sentence. This distinct reading behavior is shown in Figure FIGREF15, where fixations occur until the end of the sentence during normal reading, while during task-specific reading the fixations stop after the decisive words to detect a given relation type. Finally, we also analyzed the average reading times for each of the extracted eye-tracking features. The means and distributions for both tasks are shown in Figure FIGREF21. These results are in line with the recorded data in ZuCo 1.0, as well as with the features extracted in the GECO corpus BIBREF4.
Data Validation ::: EEG
As a first validation step, we extracted fixation-related potentials (FRPs), where the EEG signal during all fixations of one task are averaged. Figure FIGREF24 shows the time-series of the resulting FRPs for two electrodes (PO8 and Cz), as well as topographies of the voltage distributions across the scalp at selected points in time. The five components (for which the scalp topographies are plotted) are highly similar in the time-course of the chosen electrodes to dimigen2011coregistration as well as to ZuCo 1.0.
Moreover, these previous studies were able to show an effect of fixation duration on the resulting FRPs. To show this dependency we followed two approaches. First, for each reading task, all single-trial FRPs were ordered by fixation duration and a vertical sliding time-window was used to smooth the data BIBREF13. Figure FIGREF25 (bottom) shows the resulting plots. In line with this previous work, a first positivation can be identified at 100 ms post-fixation onset. A second positive peak is located dependent on the duration of the fixation, which can be explained by the time-jittered succeeding fixation. The second approach is based on henderson2013co in which single trial EEG segments are clustered by the duration of the current fixation. As shown in Figure FIGREF25 (top), we chose four clusters and averaged the data within each cluster to four distinct FRPs, depending on the fixation duration. Again, the same positivation peaks become apparent. Both findings are consistent with the previous work mentioned and with our findings from ZuCo 1.0.
Conclusion
We presented a new, freely available corpus of eye movement and electrical brain activity recordings during natural reading as well as during annotation. This is the first dataset that allows for the comparison between these two reading paradigms. We described the materials and experiment design in detail and conducted an extensive validation to ensure the quality of the recorded data. Since this corpus is tailored to cognitively-inspired NLP, the applications and re-use potentials of this data are extensive. The provided word-level and sentence-level eye-tracking and EEG features can be used to improve and evaluate NLP and machine learning methods, for instance, to evaluate linguistic phenomena in neural models via psycholinguistic data. In addition, because the sentences contains semantic relation labels and the annotations of all participants, it can also be widely used for relation extraction and classification. Finally, the two carefully constructed reading paradigms allow for the comparison between normal reading and reading during annotation, which can be relevant to improve the manual labelling process as well as the quality of the annotations for supervised machine learning.
| [
"No"
] | 3,445 | qasper | en | null | 43279ddf85ada1b163aa1b316a4df1418957058206501c26 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Did they experiment with this new dataset?
Answer: | [
"No"
] | qasper | 128 | 14 | |
What datasets are used? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Back to 42 BC, the philosopher Cicero has raised the issue that although there were many Oratory classes, there were none for Conversational skills BIBREF0 . He highlighted how important they were not only for politics, but also for educational purpose. Among other conversational norms, he claimed that people should be able to know when to talk in a conversation, what to talk depending on the subject of the conversation, and that they should not talk about themselves.
Norms such as these may become social conventions and are not learnt at home or at school. Social conventions are dynamic and may change according to context, culture and language. In online communication, new commonsense practices are evolved faster and accepted as a norm BIBREF1 , BIBREF2 . There is not a discipline for that on elementary or high schools and there are few linguistics researchers doing research on this field.
On the other hand, within the Artificial Intelligence area, some Conversational Systems have been created in the past decades since the test proposed by Alan Turing in 1950. The test consists of a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from that of a human BIBREF3 . Turing proposed that a human evaluator would judge natural language conversations between a human and a machine that is designed to generate human-like responses. Since then, many systems have been created to pass the Turing's test. Some of them have won prizes, some not BIBREF4 . Although in this paper we do not focus on creating a solution that is able to build conversational systems that pass the Turing's test, we focus on NDS. From BIBREF5 , "NDS are systems that try to improve usability and user satisfaction by imitating human behavior". We refer to Conversational Systems as NDS, where the dialogues are expressed as natural language texts, either from artificial intelligent agents (a.k.a. bots) or from humans.
That said, the current popular name to systems that have the ability to make a conversation with humans using natural language is Chatbot. Chatbots are typically used in conversational systems for various practical purposes, including customer service or information acquisition. Chatbots are becoming more widely used by social media software vendors. For example, Facebook recently announced that it would make Facebook Messenger (its 900-million-user messaging app by 2016), into a full-fledged platform that allows businesses to communicate with users via chatbots. Google is also building a new mobile-messaging service that uses artificial intelligence know-how and chatbot technology. In addition, according to the Wall Street Journal, there are more than 2 billion users of mobile apps. Still, people can be reluctant to install apps. So it is believed that social messaging can be a platform and chatbots may provide a new conversational interface for interacting with online services, as chatbots are easier to build and deploy than apps BIBREF6 .
China seems to be the place where chatbots adoption and use is most advanced today. For example, China's popular WeChat messaging platform can take payments, scan QR codes, and integrate chatbot systems. WeChat integrates e-mail, chat, videocalls and sharing of large multimedia files. Users can book flights or hotels using a mixed, multimedia interaction with active bots. WeChat was first released in 2011 by Tecent, a Chinese online-gaming and social-media firm, and today more than 700 million people use it, being one of the most popular messaging apps in the world (The Economist 2016). WeChat has a mixture of real-live customer service agents and automated replies (Olson 2016).
Still, current existing chatbot engines do not properly handle a group chat with many users and many chatbots. This makes the chatbots considerably less social, which is a problem since there is a strong demand of having social chatbots that are able to provide different kinds of services, from traveling packages to finance advisors. This happens because there is a lack of methods and tools to design and engineer the coordination and mediation among chatbots and humans, as we present in Sections 2 and 3. In this paper, we refer to conversational systems that are able to interact with one or more people or chatbots in a multi-party chat as MPCS. Altogether, this paper is not meant to advance the state of the art on the norms for MPCS. Instead, the main contributions of this paper are threefold:
We then present some discussion and future work in the last section.
Challenges on Chattering
There are plenty of challenges in conversation contexts, and even bigger ones when people and machines participate in those contexts. Conversation is a specialized form of interaction, which follows social conventions. Social interaction makes it possible to inform, context, create, ratify, refute, and ascribe, among other things, power, class, gender, ethnicity, and culture BIBREF2 . Social structures are the norms that emerge from the contact people have with others BIBREF7 , for example, the communicative norms of a negotiation, taking turns in a group, the cultural identity of a person, or power relationships in a work context.
Conventions, norms and patterns from everyday real conversations are applied when designing those systems to result in adoption and match user's expectations. BIBREF8 describes implicit interactions in a framework of interactions between humans and machines. The framework is based on the theory of implicit interactions which posits that people rely on conventions of interaction to communicate queries, offers, responses, and feedback to one another. Conventions and patterns drive our expectations about interactive behaviors. This framework helps designers and developers create interactions that are more socially appropriate. According to the author, we have interfaces which are based on explicit interaction and implicit ones. The explicit are the interactions or interfaces where people rely on explicit input and output, whereas implicit interactions are the ones that occur without user awareness of the computer behavior.
Social practices and actions are essential for a conversation to take place during the turn-by-turn moments of communication. BIBREF9 highlights that a distinguishing feature of ordinary conversation is "the local, moment-by-moment management of the distribution of turns, of their size, and what gets done in them, those things being accomplished in the course of each current speaker's turn." Management of turns and subject change in each course is a situation that occurs in real life conversations based on circumstances (internal and external) to speakers in a dialogue. Nowadays, machines are not prepared to fully understand context and change the course of conversations as humans. Managing dialogues with machines is challenging, which increases even more when more than one conversational agent is part of the same conversation. Some of those challenges in the dialogue flow were addressed by BIBREF10 . According to them, we have system-initiative, user-initiative, and mixed-initiative systems.
In the first case, system-initiative systems restrict user options, asking direct questions, such as (Table TABREF5 ): "What is the initial amount of investment?" Doing so, those types of systems are more successful and easier to answer to. On the other hand, user-initiative systems are the ones where users have freedom to ask what they wish. In this context, users may feel uncertain of the capabilities of the system and starting asking questions or requesting information or services which might be quite far from the system domain and understanding capacity, leading to user frustration. There is also a mixed-initiative approach, that is, a goal-oriented dialogue which users and computers participate interactively using a conversational paradigm. Challenges of this last classification are to understand interruptions, human utterances, and unclear sentences that were not always goal-oriented.
The dialog in Table TABREF5 has the system initiative in a question and answer mode, while the one in Table TABREF7 is a natural dialogue system where both the user and the system take the initiative. If we add another user in the chat, then we face other challenges.
In Table TABREF12 , line 4, the user U1 invites another person to the chat and the system does not reply to this utterance, nor to utterances on lines 6, 7 and 8 which are the ones when only the users (wife and husband) should reply to. On the other hand, when the couple agrees on the period and initial value of the investment (line 9), then the system S1 (at the time the only system in the chat) replies indicating that it will invite more systems (chatbots) that are experts on this kind of pair INLINEFORM0 period, initial value INLINEFORM1 . They then join the chat and start interacting with each other. At the end, on line 17, the user U2 interacts with U1 and they agree with the certificate option. Then, the chatbot responsible for that, S3, is the only one that replies indicating how to invest.
Table TABREF12 is one example of interactions on which the chatbots require knowledge of when to reply given the context of the dialog. In general, we acknowledge that exist four dimensions of understanding and replying to an utterance in MPCS which a chatbot that interacts in a multi-party chat group should fulfill:
In the next section we present the state of the art and how they fullfil some of these dimensions.
Conversational Systems
In this section we discuss the state of the art on conversational systems in three perspectives: types of interactions, types of architecture, and types of context reasoning. Then we present a table that consolidates and compares all of them.
ELIZA BIBREF11 was one of the first softwares created to understand natural language processing. Joseph Weizenbaum created it at the MIT in 1966 and it is well known for acting like a psychotherapist and it had only to reflect back onto patient's statements. ELIZA was created to tackle five "fundamental technical problems": the identification of critical words, the discovery of a minimal context, the choice of appropriate transformations, the generation of appropriate responses to the transformation or in the absence of critical words, and the provision of an ending capacity for ELIZA scripts.
Right after ELIZA came PARRY, developed by Kenneth Colby, who is psychiatrist at Stanford University in the early 1970s. The program was written using the MLISP language (meta-lisp) on the WAITS operating system running on a DEC PDP-10 and the code is non-portable. Parts of it were written in PDP-10 assembly code and others in MLISP. There may be other parts that require other language translators. PARRY was the first system to pass the Turing test - the psychiatrists were able to make the correct identification only 48 percent of the time, which is the same as a random guessing.
A.L.I.C.E. (Artificial Linguistic Internet Computer Entity) BIBREF12 appeared in 1995 but current version utilizes AIML, an XML language designed for creating stimulus-response chat robots BIBREF13 . A.L.I.C.E. bot has, at present, more than 40,000 categories of knowledge, whereas the original ELIZA had only about 200. The program is unable to pass the Turing test, as even the casual user will often expose its mechanistic aspects in short conversations.
Cleverbot (1997-2014) is a chatbot developed by the British AI scientist Rollo Carpenter. It passed the 2011 Turing Test at the Technique Techno-Management Festival held by the Indian Institute of Technology Guwahati. Volunteers participate in four-minute typed conversations with either Cleverbot or humans, with Cleverbot voted 59.3 per cent human, while the humans themselves were rated just 63.3 per cent human BIBREF14 .
Types of Interactions
Although most part of the research literature focuses on the dialogue of two persons, the reality of everyday life interactions shows a substantial part of multi-user conversations, such as in meetings, classes, family dinners, chats in bars and restaurants, and in almost every collaborative or competitive environment such as hospitals, schools, offices, sports teams, etc. The ability of human beings to organize, manage, and (mostly) make productive such complex interactive structures which are multi-user conversations is nothing less than remarkable. The advent of social media platforms and messaging systems such as WhatsApp in the first 15 years of the 21st century expanded our ability as a society to have asynchronous conversations in text form, from family and friends chatgroups to whole nations conversing in a highly distributed form in social media BIBREF15 .
In this context, many technological advances in the early 2010s in natural language processing (spearheaded by the IBM Watson's victory in Jeopardy BIBREF16 ) spurred the availability in the early 2010s of text-based chatbots in websites and apps (notably in China BIBREF17 ) and spoken speech interfaces such as Siri by Apple, Cortana by Microsoft, Alexa by Amazon, and Allo by Google. However, the absolute majority of those chatbot deployments were in contexts of dyadic dialog, that is, a conversation between a single chatbot with a single user. Most of the first toolkits for chatbot design and development of this initial period implicit assume that an utterance from the user is followed by an utterance of the chatbot, which greatly simplifies the management of the conversation as discussed in more details later. Therefore, from the interaction point of view, there are two types: 1) one in which the chatbot was designed to chat with one person or chatbot, and 2) other in which the chatbot can interact with more than two members in the chat.
Dyadic Chatbot
A Dyadic Chatbot is a chatbot that does know when to talk. If it receives an utterance, it will always handle and try to reply to the received utterance. For this chatbot to behave properly, either there are only two members in the chat, and the chatbot is one of them, or there are more, but the chatbot replies only when its name or nickname is mentioned. This means that a dyadic chatbot does not know how to coordinate with many members in a chat group. It lacks the social ability of knowing when it is more suitable to answer or not. Also, note that we are not considering here the ones that would use this social ability as an advantage in the conversation, because if the chatbot is doing with this intention, it means that the chatbot was designed to be aware of the social issues regarding a chat with multiple members, which is not the case of a dyadic chatbot. Most existing chatbots, from the first system, ELIZA BIBREF11 , until modern state-of-the-art ones fall into this category.
Multiparty Conversations
In multiparty conversations between people and computer systems, natural language becomes the communication protocol exchanged not only by the human users, but also among the bots themselves. When every actor, computer or user, understands human language and is able to engage effectively in a conversation, a new, universal computer protocol of communication is feasible, and one for which people are extremely good at.
There are many differences between dyadic and multiparty conversations, but chiefly among them is turn-taking, that is, how a participant determines when it is appropriate to make an utterance and how that is accomplished. There are many social settings, such as assemblies, debates, one-channel radio communications, and some formal meetings, where there are clear and explicit norms of who, when, and for long a participant can speak.
The state of the art for the creation of chatbots that can participate on multiparty conversations currently is a combination of the research on the creation of chatbots and research on the coordination or governance of multi-agents systems. A definition that mixes both concepts herein present is: A chatbot is an agent that interacts through natural language. Although these areas complement each other, there is a lack of solutions for creating multiparty-aware chatbots or governed chatbots, which can lead to higher degree of system trust.
Multi-Dyadic Chatbots
Turn-taking in generic, multiparty spoken conversation has been studied by, for example, Sacks et al. BIBREF18 . In broad terms, it was found that participants in general do not overlap their utterances and that the structure of the language and the norms of conversation create specific moments, called transition-relevance places, where turns can occur. In many cases, the last utterances make clear to the participants who should be the next speaker (selected-next-speaker), and he or she can take that moment to start to talk. Otherwise, any other participant can start speaking, with preference for the first starter to get the turn; or the current speaker can continue BIBREF18 .
A key part of the challenge is to determine whether the context of the conversation so far have or have not determined the next speaker. In its simplest form, a vocative such as the name of the next speaker is uttered. Also, there is a strong bias towards the speaker before the current being the most likely candidate to be the next speaker.
In general the detection of transition-relevance places and of the selected-next-speaker is still a challenge for speech-based machine conversational systems. However, in the case of text message chats, transition-relevance places are often determined by the acting of posting a message, so the main problem facing multiparty-enabled textual chatbots is in fact determining whether there is and who is the selected-next-speaker. In other words, chatbots have to know when to shut up. Bohus and Horowitz BIBREF19 have proposed a computational probabilistic model for speech-based systems, but we are not aware of any work dealing with modeling turn-taking in textual chats.
Coordination of Multi-Agent Systems
A multi-agent system (MAS) can be defined as a computational environment in which individual software agents interact with each other, in a cooperative manner, or in a competitive manner, and sometimes autonomously pursuing their individual goals. During this process, they access the environment's resources and services and occasionally produce results for the entities that initiated these software agents. As the agents interact in a concurrent, asynchronous and decentralized manner, this kind of system can be categorized as a complex system BIBREF20 .
Research in the coordination of multi-agent systems area does not address coordination using natural dialogue, as usually all messages are structured and formalized so the agents can reason and coordinate themselves. On the other hand, chatbots coordination have some relations with general coordination mechanisms of multi-agent systems in that they specify and control interactions between agents. However, chatbots coordination mechanisms is meant to regulate interactions and actions from a social perspective, whereas general coordination languages and mechanisms focus on means for expressing synchronization and coordination of activities and exchange of information, at a lower computational level.
In open multi-agent systems the development takes place without a centralized control, thus it is necessary to ensure the reliability of these systems in a way that all the interactions between agents will occur according to the specification and that these agents will obey the specified scenario. For this, these applications must be built upon a law-governed architecture.
Minsky published the first ideas about laws in 1987 BIBREF21 . Considering that a law is a set of norms that govern the interaction, afterwards, he published a seminal paper with the Law-Governed Interaction (LGI) conceptual model about the role of interaction laws on distributed systems BIBREF22 . Since then, he conducted further work and experimentation based on those ideas BIBREF23 . Although at the low level a multiparty conversation system is a distributed system and the LGI conceptual model can be used in a variety of application domains, it is composed of abstractions basically related to low level information about communication issues of distributed systems (like the primitives disconnected, reconnected, forward, and sending or receiving of messages), lacking the ability to express high level information of social systems.
Following the same approach, the Electronic Institution (EI) BIBREF24 solution also provides support for interaction norms. An EI has a set of high-level abstractions that allow for the specification of laws using concepts such as agent roles, norms and scenes.
Still at the agent level but more at the social level, the XMLaw description language and the M-Law framework BIBREF25 BIBREF26 were proposed and developed to support law-governed mechanism. They implement a law enforcement approach as an object-oriented framework and it allows normative behavior through the combination between norms and clocks. The M-Law framework BIBREF26 works by intercepting messages exchanged between agents, verifying the compliance of the messages with the laws and subsequently redirecting the message to the real addressee, if the laws allow it. If the message is not compliant, then the mediator blocks the message and applies the consequences specified in the law, if any. They are called laws in the sense that they enforce the norms, which represent what can be done (permissions), what cannot be done (prohibitions) and what must be done (obligations).
Coordinated Aware Chatbots in a Multiparty Conversation
With regard to chatbot engines, there is a lack of research directed to building coordination laws integrated with natural language. To the best of our knowledge, the architecture proposed in this paper is the first one in the state of the art designed to support the design and development of coordinated aware chatbots in a multiparty conversation.
Types of Architectures
There are mainly three types of architectures when building conversational systems: totally rule-oriented, totally data-oriented, and a mix of rules and data-oriented.
Rule-oriented
A rule-oriented architecture provides a manually coded reply for each recognized utterance. Classical examples of rule-based chatbots include Eliza and Parry. Eliza could also extract some words from sentences and then create another sentence with these words based on their syntatic functions. It was a rule-based solution with no reasoning. Eliza could not "understand" what she was parsing. More sophisticated rule-oriented architectures contain grammars and mappings for converting sentences to appropriate sentences using some sort of knowledge. They can be implemented with propositional logic or first-order logic (FOL). Propositional logic assumes the world contains facts (which refer to events, phenomena, symptoms or activities). Usually, a set of facts (statements) is not sufficient to describe a domain in a complete manner. On the other hand, FOL assumes the world contains Objects (e.g., people, houses, numbers, etc.), Relations (e.g. red, prime, brother of, part of, comes between, etc.), and Functions (e.g. father of, best friend, etc.), not only facts as in propositional logic. Moreover, FOL contains predicates, quantifiers and variables, which range over individuals (which are domain of discourse).
Prolog (from French: Programmation en Logique) was one of the first logic programming languages (created in the 1970s), and it is one of the most important languages for expressing phrases, rules and facts. A Prolog program consists of logical formulas and running a program means proving a theorem. Knowledge bases, which include rules in addition to facts, are the basis for most rule-oriented chatbots created so far.
In general, a rule is presented as follows: DISPLAYFORM0
Prolog made it possible to perform the language of Horn clauses (implications with only one conclusion). The concept of Prolog is based on predicate logic, and proving theorems involves a resolute system of denials. Prolog can be distinguished from classic programming languages due to its possibility of interpreting the code in both a procedural and declarative way. Although Prolog is a set of specifications in FOL, it adopts the closed-world assumption, i.e. all knowledge of the world is present in the database. If a term is not in the database, Prolog assumes it is false.
In case of Prolog, the FOL-based set of specifications (formulas) together with the facts compose the knowledge base to be used by a rule-oriented chatbot. However an Ontology could be used. For instance, OntBot BIBREF27 uses mapping technique to transform ontologies and knowledge into relational database and then use that knowledge to drive its chats. One of the main issues currently facing such a huge amount of ontologies stored in a database is the lack of easy to use interfaces for data retrieval, due to the need to use special query languages or applications.
In rule-oriented chatbots, the degree of intelligent behavior depends on the knowledge base size and quality (which represents the information that the chatbot knows), poor ones lead to weak chatbot responses while good ones do the opposite. However, good knowledge bases may require years to be created, depending on the domain.
Data-oriented
As opposed to rule-oriented architectures, where rules have to be explicitly defined, data-oriented architectures are based on learning models from samples of dialogues, in order to reproduce the behavior of the interaction that are observed in the data. Such kind of learning can be done by means of machine learning approach, or by simply extracting rules from data instead of manually coding them.
Among the different technologies on which these system can be based, we can highlight classical information retrieval algorithms, neural networks BIBREF28 , Hidden Markov Models (HMM) BIBREF29 , and Partially Observable Markov Decision Process (POMDP) BIBREF30 . Examples include Cleverbot and Tay BIBREF31 . Tay was a chatbot developed by Microsoft that after one day live learning from interaction with teenagers on Twitter, started replying impolite utterances. Microsoft has developed others similar chatbots in China (Xiaoice) and in Japan (Rinna). Microsoft has not associated its publications with these chatbots, but they have published a data-oriented approach BIBREF32 that proposes a unified multi-turn multi-task spoken language understanding (SLU) solution capable of handling multiple context sensitive classification (intent determination) and sequence labeling (slot filling) tasks simultaneously. The proposed architecture is based on recurrent convolutional neural networks (RCNN) with shared feature layers and globally normalized sequence modeling components.
A survey of public available corpora for can be found in BIBREF33 . A corpus can be classified into different categories, according to: the type of data, whether it is spoken dialogues, transcripts of spoken dialogues, or directly written; the type of interaction, if it is human-human or human-machine; and the domain, whether it is restricted or unconstrained. Two well-known corpora are the Switchboard dataset, which consists of transcripts of spoken, unconstrained, dialogues, and the set of tasks for the Dialog State Tracking Challenge (DSTC), which contain more constrained tasks, for instance the restaurant and travel information sets.
Rule and Data-oriented
The model of learning in current A.L.I.C.E. BIBREF13 is incremental or/and interactive learning because a person monitors the robot's conversations and creates new AIML content to make the responses more appropriate, accurate, believable, "human", or whatever he/she intends. There are algorithms for automatic detection of patterns in the dialogue data and this process provides the person with new input patterns that do not have specific replies yet, permitting a process of almost continuous supervised refinement of the bot.
As already mentioned, A.L.I.C.E. consists of roughly 41,000 elements called categories which is the basic unit of knowledge in AIML. Each category consists of an input question, an output answer, and an optional context. The question, or stimulus, is called the pattern. The answer, or response, is called the template. The two types of optional context are called that and topic. The keyword that refers to the robot's previous utterance. The AIML pattern language consists only of words, spaces, and the wildcard symbols "_" and "*". The words may consist only of letters and numerals. The pattern language is case invariant. Words are separated by a single space, and the wildcard characters function like words, similar to the initial pattern matching strategy of the Eliza system. More generally, AIML tags transform the reply into a mini computer program which can save data, activate other programs, give conditional responses, and recursively call the pattern matcher to insert the responses from other categories. Most AIML tags in fact belong to this template side sublanguage BIBREF13 .
AIML language allows:
Symbolic reduction: Reduce complex grammatical forms to simpler ones.
Divide and conquer: Split an input into two or more subparts, and combine the responses to each.
Synonyms: Map different ways of saying the same thing to the same reply.
Spelling or grammar corrections: the bot both corrects the client input and acts as a language tutor.
Detecting keywords anywhere in the input that act like triggers for a reply.
Conditionals: Certain forms of branching to produce a reply.
Any combination of (1)-(6).
When the bot chats with multiple clients, the predicates are stored relative to each client ID. For example, the markup INLINEFORM0 set name INLINEFORM1 "name" INLINEFORM2 Matthew INLINEFORM3 /set INLINEFORM4 stores the string Matthew under the predicate named "name". Subsequent activations of INLINEFORM5 get name="name" INLINEFORM6 return "Matthew". In addition, one of the simple tricks that makes ELIZA and A.L.I.C.E. so believable is a pronoun swapping substitution. For instance:
U: My husband would like to invest with me.
S: Who else in your family would like to invest with you?
Types of Intentions
According to the types of intentions, conversational systems can be classified into two categories: a) goal-driven or task oriented, and b) non-goal-driven or end-to-end systems.
In a goal-driven system, the main objective is to interact with the user so that back-end tasks, which are application specific, are executed by a supporting system. As an example of application we can cite technical support systems, for instance air ticket booking systems, where the conversation system must interact with the user until all the required information is known, such as origin, destination, departure date and return date, and the supporting system must book the ticket. The most widely used approaches for developing these systems are Partially-observed Decision Processes (POMDP) BIBREF30 , Hidden Markov Models (HMM) BIBREF29 , and more recently, Memory Networks BIBREF28 . Given that these approaches are data-oriented, a major issue is to collect a large corpora of annotated task-specific dialogs. For this reason, it is not trivial to transfer the knowledge from one to domain to another. In addition, it might be difficult to scale up to larger sets of tasks.
Non-goal-driven systems (also sometimes called reactive systems), on the other hand, generate utterances in accordance to user input, e.g. language learning tools or computer games characters. These systems have become more popular in recent years, mainly owning to the increase of popularity of Neural Networks, which is also a data-oriented approach. The most recent state of the art to develop such systems have employed Recurrent Neural Networs (RNN) BIBREF34 , Dynamic Context-Sensitive Generation BIBREF35 , and Memory Networks BIBREF36 , just to name a few. Nevertheless, probabilistic methods such as Hidden Topic Markov Models (HTMM) BIBREF37 have also been evaluated. Goal-driven approach can create both pro-active and reactive chatbots, while non-goal-driven approach creates reactive chatbots. In addition, they can serve as a tool to goal-driven systems as in BIBREF28 . That is, when trained on corpora of a goal-driven system, non-goal-driven systems can be used to simulate user interaction to then train goal-driven models.
Types of Context Reasoning
A dialogue system may support the context reasoning or not. Context reasoning is necessary in many occasions. For instance, when partial information is provided the chatbot needs to be able to interact one or more turns in order to get the complete information in order to be able to properly answer. In BIBREF38 , the authors present a taxonomy of errors in conversational systems. The ones regarding context-level errors are the ones that are perceived as the top-10 confusing and they are mainly divided into the following:
Excess/lack of proposition: the utterance does not provide any new proposition to the discourse context or provides excessive information than required.
Contradiction: the utterance contains propositions that contradict what has been said by the system or by the user.
Non-relevant topic: the topic of the utterance is irrelevant to the current context such as when the system suddenly jumps to some other topic triggered by some particular word in the previous user utterance.
Unclear relation: although the utterance might relate to the previous user utterance, its relation to the current topic is unclear.
Topic switch error: the utterance displays the fact that the system missed the switch in topic by the user, continuing with the previous topic.
Rule-oriented
In the state of the art most of the proposed approaches for context reasoning lies on rules using logics and knowledge bases as described in the Rule-oriented architecture sub-section. Given a set of facts extracted from the dialogue history and encoded in, for instance, FOL statements, a queries can be posed to the inference engine and produce answers. For instance, see the example in Table TABREF37 . The sentences were extracted from BIBREF36 (which does not use a rule-oriented approach), and the first five statements are their respective facts. The system then apply context reasoning for the query Q: Where is the apple.
If statements above are received on the order present in Table TABREF37 , if the query Q: Where is the apple is sent, the inference engine will produce the answer A: Bedroom (i.e., the statement INLINEFORM0 is found by the model and returned as True).
Nowadays, the most common way to store knowledge bases is on triple stores, or RDF (Resource Description Framework) stores. A triple store is a knowledge base for the storage and retrieval of triples through semantic queries. A triple is a data entity composed of subject-predicate-object, like "Sam is at the kitchen" or "The apple is with Sam", for instance. A query language is needed for storing and retrieving data from a triple store. While SPARQL is a RDF query language, Rya is an open source scalable RDF triple store built on top of Apache Accumulo. Originally developed by the Laboratory for Telecommunication Sciences and US Naval Academy, Rya is currently being used by a number of american government agencies for storing, inferencing, and querying large amounts of RDF data.
A SPARQL query has a SQL-like syntax for finding triples matching specific patterns. For instance, see the query below. It retrieves all the people that works at IBM and lives in New York:
SELECT ?people
WHERE {
?people <worksAt> <IBM> .
?people <livesIn> <New York>.
}
Since triple stores can become huge, Rya provides three triple table index BIBREF39 to help speeding up queries:
SPO: subject, predicate, object
POS: predicate, object, subject
OSP: object, subject, predicate
While Rya is an example of an optimized triple store, a rule-oriented chatbot can make use of Rya or any triple store and can call the semantic search engine in order to inference and generate proper answers.
Data-oriented
Recent papers have used neural networks to predict the next utterance on non-goal-driven systems considering the context, for instance with Memory Networks BIBREF40 . In this work BIBREF36 , for example the authors were able to generate answers for dialogue like below:
Sam walks into the kitchen.
Sam picks up an apple.
Sam walks into the bedroom.
Sam drops the apple.
Q: Where is the apple?
A: Bedroom
Sukhbaatar's model represents the sentence as a vector in a way that the order of the words matter, and the model encodes the temporal context enhancing the memory vector with a matrix that contains the temporal information. During the execution phase, Sukhbaatar's model takes a discrete set of inputs INLINEFORM0 that are to be stored in the memory, a query INLINEFORM1 , and outputs an answer INLINEFORM2 . Each of the INLINEFORM3 , INLINEFORM4 , and INLINEFORM5 contains symbols coming from a dictionary with INLINEFORM6 words. The model writes all INLINEFORM7 to the memory up to a fixed buffer size, and then finds a continuous representation for the INLINEFORM8 and INLINEFORM9 . The continuous representation is then processed via multiple computational steps to output INLINEFORM10 . This allows back propagation of the error signal through multiple memory accesses back to the input during training. Sukhbaatar's also presents the state of the art of recent efforts that have explored ways to capture dialogue context, treated as long-term structure within sequences, using RNNs or LSTM-based models. The problem of this approach is that it is has not been tested for goal-oriented systems. In addition, it works with a set of sentences but not necessary from multi-party bots.
Platforms
Regarding current platforms to support the development of conversational systems, we can categorize them into three types: platforms for plugging chatbots, for creating chatbots and for creating service chatbots. The platforms for plugging chatbots provide tools for integrating them another system, like Slack. The chatbots need to receive and send messages in a specific way, which depends on the API and there is no support for actually helping on building chatbots behavior with natural language understanding. The platforms for creating chatbots mainly provide tools for adding and training intentions together with dialogue flow specification and some entities extraction, with no reasoning support. Once the models are trained and the dialogue flow specified, the chatbots are able to reply to the received intention. The platforms for creating service chatbots provide the same functionalities as the last one and also provide support for defining actions to be executed by the chatbots when they are answering to an utterance. Table TABREF43 summarizes current platforms on the market accordingly to these categories. There is a lack on platforms that allow to create chatbots that can be coordinated in a multiparty chat with governance or mediation.
A Conceptual Architecture for Multiparty-Aware Chatbots
In this section the conceptual architecture for creating a hybrid rule and machine learning-based MPCS is presented. The MPCS is defined by the the entities and relationships illustrated in Fig. FIGREF44 which represents the chatbot's knowledge. A Chat Group contains several Members that join the group with a Role. The role may constrain the behavior of the member in the group. Chatbot is a type of Role, to differentiate from persons that may also join with different roles. For instance, a person may assume the role of the owner of the group, or someone that was invited by the owner, or a domain role like an expert, teacher or other.
When a Member joins the Chat Group, it/he/she can send Utterances. The Member then classifies each Utterance with an Intent which has a Speech Act. The Intent class, Speech Act class and the Intent Flow trigger the Action class to be executed by the Member that is a Chatbot. The Chatbots associated to the Intention are the only ones that know how to answer to it by executing Actions. The Action, which implements one Speech Act, produces answers which are Utterances, so, for instance, the Get_News action produces an Utterance for which Intention's speech act is Inform_News. The Intent Flow holds the intent's class conversation graph which maps the dialog state as a decision tree. The answer's intention class is mapped in the Intent Flow as a directed graph G defined as following: DISPLAYFORM0
From the graph definitions, INLINEFORM0 is for vertices and INLINEFORM1 is for relations, which are the arrows in the graph. And in Equation EQREF46 :
INLINEFORM0 is the set of intentions pairs,
INLINEFORM0 is the set of paths to navigate through the intentions,
INLINEFORM0 is the arrow's head, and
INLINEFORM0 is the arrow's tail.
This arrow represents a turn from an utterance with INLINEFORM0 intention class which is replying to an utterance with INLINEFORM1 intention class to the state which an utterance with INLINEFORM2 intention's class is sent.
INLINEFORM0 is the intention class of the answer to be provided to the received INLINEFORM1 intention class.
In addition, each intent's class may refer to many Entities which, in turn, may be associated to several Features. For instance, the utterance
"I would like to invest USD10,000 in Savings Account for 2 years"
contains one entity – the Savings Account's investment option – and two features – money (USD10,000) and period of time (2 years). The Intent Flow may need this information to choose the next node which will give the next answer. Therefore, if the example is changed a little, like
"I would like to invest in Savings Account",
INLINEFORM0 is constrained by the "Savings Account" entity which requires the two aforementioned features. Hence, a possible answer by one Member of the group would be
"Sure, I can simulate for you, what would be the initial amount and the period of time of the investment?"
With these conceptual model's elements, a MPCS system can be built with multiple chatbots. Next subsection further describes the components workflow.
Workflow
Figure FIGREF48 illustrates from the moment that an utterance is sent in a chat group to the moment a reply is generated in the same chat group, if the case. One or more person may be in the chat, while one or more chatbots too. There is a Hub that is responsible for broadcasting the messages to every Member in the group, if the case. The flow starts when a Member sends the utterance which goes to the Hub and, if allowed, is broadcasted. Many or none interactions norms can be enforced at this level depending on the application. Herein, a norm can be a prohibition, obligation or permission to send an utterance in the chat group.
Once the utterance is broadcasted, a chatbot needs to handle the utterance. In order to properly handle it, the chatbot parses the utterance with several parsers in the Parsing phase: a Topic Classifier, the Dependency Parsing, which includes Part-of-Speech tags and semantics tags, and any other that can extract metadata from the utterance useful for the reasoning. All these metadata, together with more criteria, may be used in the Frame parsing which is useful for context reasoning. All knowledge generated in this phase can be stored in the Context. Then, the Intent Classifier tries to detect the intent class of the utterance. If detected, the Speech Act is also retrieved. And an Event Detector can also check if there is any dialog inconsistency during this phase.
After that, the Filtering phase receives the object containing the utterance, the detected intent, and all metadata extracted so far and decides if an action should be performed to reply to the utterance. If yes, it is sent to the Acting phase which performs several steps. First the Action Classifier tries to detect the action to be performed. If detected, the action is executed. At this step, many substeps may be performed, like searching for an information, computing maths, or generating information to create the answer. All of this may require a search in the Context and also may activate the Error Detector component to check if the dialog did not run into a wrong state. After the answer is generated, the Filtering phase is activated again to check if the reply should be really sent. If so, it is sent to the Hub which, again may check if it can be broadcasted before actually doing it.
The topic classifier is domain-dependent and is not mandatory. However, the chatbot can better react when the intent or action is not detected, which means that it does not know how to answer. Many reasons might explain this situation: the set of intents might be incomplete, the action might not have produced the proper behavior, misunderstanding might happen, or the chatbot was not designed to reply to a particular topic. In all cases, it must be able to produce a proper reply, if needed. Because this might happen throughout the workflow, the sooner that information is available, the better the chatbot reacts. Therefore it is one of the first executions of the flow.
Dependency is the notion that linguistic units, e.g. words, are connected to each other by directed links. The (finite) verb is taken to be the structural center of clause structure. All other syntactic units (words) are either directly or indirectly connected to the verb in terms of the directed links, which are called dependencies. It is a one-to-one correspondence: for every element (e.g. word or morph) in the sentence, there is exactly one node in the structure of that sentence that corresponds to that element. The result of this one-to-one correspondence is that dependency grammars are word (or morph) grammars. All that exist are the elements and the dependencies that connect the elements into a structure. Dependency grammar (DG) is a class of modern syntactic theories that are all based on the dependency relation.
Semantic dependencies are understood in terms of predicates and their arguments. Morphological dependencies obtain between words or parts of words. To facilitate future research in unsupervised induction of syntactic structure and to standardize best-practices, a tagset that consists of twelve universal part-of-speech categories was proposed BIBREF41 .
Dependency parsers have to cope with a high degree of ambiguity and nondeterminism which let to different techniques than the ones used for parsing well-defined formal languages. Currently the mainstream approach uses algorithms that derive a potentially very large set of analyses in parallel and when disambiguation is required, this approach can be coupled with a statistical model for parse selection that ranks competing analyses with respect to plausibility BIBREF42 .
Below we present an example of a dependency tree for the utterance:
"I want to invest 10 thousands":
[s]""blue[l]:black
"tree": {
"want VERB ROOT": {
"I PRON nsubj": {},
"to ADP mark": {},
"invest VERB nmod": {
"thousands NOUN nmod": {
"10 NUM nummod": {}
}
}
}
The coarse-grained part-of-speech tags, or morphological dependencies (VERB, PRON, ADP, NOUN and NUM) encode basic grammatical categories and the grammatical relationships (nsubjs, nmod, nummod) are defined in the Universal Dependencies project BIBREF41 .
In this module, the dependency tree generated is used together with a set of rules to extract information that is saved in the context using the Frame-based approach. This approach fills the slots of the frame with the extracted values from the dialogue. Frames are like forms and slots are like fields. Using the knowledge's conceptual model, the fields are represented by the elements Entities and Features. In the dependency tree example, the entity would be the implicit concept: the investment option, and the feature is the implicit concept: initial amount – 10 thousands. Since the goal is to invest, and there are more entities needed for that (i.e., fields to be filled), the next node in the Intention Flow tree would return an utterance which asks the user the time of investment, if he/she has not provided yet.
This module could be implemented using different approaches according to the domain, but tree search algorithms will be necessary for doing the tree parsing.
The Intent Classifier component aims at recognizing not only the Intent but the goal of the utterance sent by a Member, so it can properly react. The development of an intent classifier needs to deal with the following steps:
i) the creation of dataset of intents, to train the classification algorithm;
ii) the design of a classification algorithm that provides a reasonable level of accuracy;
iii) the creation of dataset of trees of intents, the same as defined in i) and which maps the goals;
iv) the design of a plan-graph search algorithm that maps the goal's state to a node in the graph;
There are several approaches to create training sets for dialogues: from an incremental approach to crowdsourcing. In the incremental approach, the Wizard of Oz method can be applied to a set of potential users of the system, and from this study, a set of questions that the users asked posted to the `fake' system can be collected. These questions have to be manually classified into a set of intent classes, and used to train the first version of the system. Next, this set has to be increased both in terms of number of classes and samples per class.
The Speech Act Classifier can be implemented with many speech act classes as needed by the application. The more classes, the more flexible the chatbot is. It can be built based on dictionaries, or a machine learning-based classifier can be trained. In the table below we present the main and more general speech act classes BIBREF43 used in the Chatbots with examples to differentiate one from another:
There are at least as many Action classes as Speech Act classes, since the action is the realization of a speech act. The domain specific classes, like "Inform_News" or "Inform_Factoids", enhance the capabilities of answering of a chatbot.
The Action Classifier can be defined as a multi-class classifier with the tuple DISPLAYFORM0
where INLINEFORM0 is the intent of the answer defined in ( EQREF46 ), INLINEFORM1 is the speech act of the answer, INLINEFORM2 and INLINEFORM3 are the sets of entities and features needed to produce the answer, if needed, respectively.
This component is responsible for implementing the behavior of the Action class. Basic behaviors may exist and be shared among different chatbots, like the ones that implement the greetings, thanks or not understood. Although they can be generic, they can also be personalized to differentiate the bot from one another and also to make it more "real". Other cases like to inform, to send a query, to send a proposal, they are all domain-dependent and may require specific implementations.
Anyway, figure FIGREF59 shows at the high level the generic workflow. If action class detected is task-oriented, the system will implement the execution of the task, say to guide a car, to move a robot's arm, or to compute the return of investments. The execution might need to access an external service in the Internet in order to complete the task, like getting the inflation rate, or the interest rate, or to get information about the environment, or any external factor. During the execution or after it is finished, the utterance is generated as a reply and, if no more tasks are needed, the action execution is finished.
In the case of coordination of chatbots, one or more chatbots with the role of mediator may exist in the chat group and, at this step, it is able to invite one or more chatbots to the chat group and it is also able to redirect the utterances, if the case.
The proposed architecture addresses the challenges as the following:
What is the message/utterance about? solved by the Parsing phase;
Who should reply to the utterance? solved by the Filtering phase and may be enforced by the Hub;
How the reply should be built/generated? solved by the Acting phase;
When should the reply be sent? may be solved by the Acting phase or the Filtering phase, and may be enforced by the Hub;
And Context and Logging module is used throughout all phases.
Architecture Implementation and Evaluation
This section presents one implementation of the conceptual architecture presented in last section. After many refactorings, a framework called SABIA (Speech-Act-Based Intelligent Agents Framework) has been developed and CognIA (Cognitive Investment Advisor) application has been developed as an instantiation of SABIA framework. We present then the accuracy and some automated tests of this implementation.
Speech-Act-based Intelligent Agents Framework
SABIA was developed on top of Akka middleware. Akka is a toolkit and runtime that implements the Actor Model on the JVM. Akka's features, like concurrency, distributed computing, resilience, and message-passing were inspired by Erlang's actor model BIBREF44 BIBREF45 . The actor model is a mathematical model of concurrent computation that treats "actors" as the universal primitives of concurrent computation. In response to a message that it receives, an actor can: make local decisions, create more actors, send more messages, and determine how to respond to the next received message. Actors may modify private state, but can only affect each other through messages (avoiding the need for any locks). Akka middleware manages the actors life cycle and actors look up by theirs name, locally or remotely.
We implemented each Member of the Chat Group as an Actor by extending the UntypedActor class of Akka middleware. Yet, we created and implemented the SabiaActorSystem as a singleton (i.e., a single instance of it exists in the system) BIBREF46 that has a reference to Akka's ActorSystem. During SabiaActorSystem's initialization, all parsers that consume too much memory during their initialization to load models are instantiated as singletons. In this way, we save time on their calls during the runtime. Moreover, all chat group management, like to join or leave the group, or to broadcast or filter a message at the Hub level is implemented in SABIA through the Chat Group behavior.
This is implemented in SABIA as a singleton that is initialized during the SabiaActorSystem initialization with the URL of the service that implements the dependency parsing and is used on each utterance's arrival through the execution of the tagUtterance method. The service must retrieve a JSON Object with the dependency tree which is then parsed using depth-first search.
SABIA does not support invariants for frame parsing. We are leaving this task to the instantiated application.
There are two intent classifiers that can be loaded with trained models in order to be ready to be used at runtime: the 1-nearest-neighbor (1NN) and the SVM-based classifier.
SABIA implements the Action Classifier assuming that the application uses a relational database with a data schema that implements the conceptual model presented in Figure FIGREF44 . Then the invariants parts that use SQL are already present and the application only needs to implement the database connection and follow the required data schema.
SABIA provides partial implemented behavior for the Action through the Template method design pattern BIBREF46 , which implements the invariants parts of the action execution and leaves placeholders for customization.
CognIA: A Cognitive Investment Advisor
We developed CognIA, which is an instantiation of Sabia framework. A conversation is composed of a group chat that can contain multiple users and multiple chatbots. This example, in particular, has a mediator that can help users on financial matters, more specifically on investment options. For example, consider the following dialogue in the table below:
The Table TABREF71 shows an example that uses the mixed-initiative dialogue strategy, and a dialogue mediator to provide coordination control. In this example of an application, there are many types of intentions that should be answered: Q&A (question and answer) about definitions, investment options, and about the current finance indexes, simulation of investments, which is task-oriented and requires computation, and opinions, which can be highly subjective.
In Table SECREF72 , we present the interaction norms that were needed in Cognia. The Trigger column describes the event that triggers the Behavior specified in the third column. The Pre-Conditions column specifies what must happen in order to start the behavior execution. So, for instance, line 2, when the user sends an utterance in the chat group, an event is triggered and, if the utterance's topic is CDB (Certificate of Deposit which is a fixed rate investment) or if it is about the Savings Account investment option and the speech act is not Query_Calculation and the CDB and Savings Account members are not in the chat, then the behavior is activated. The bot members that implement these behaviors are called cdbguru and poupancaguru. Therefore these names are used when there is a mention.
Note that these interactions norms are not explicitly defined as obligations, permissions, and prohibitions. They are implict from the behavior described. During this implementation, we did not worry about explicitly defining the norms, because the goal was to evaluate the overall architecture, not to enhance the state of the art on norms specification for conversational systems. In addition, CognIA has only the presented interaction norms defined in Table SECREF72 , which is a very small set that that does not required model checking or verification of conflicts.
|p2cm|p5.0cm|p5.0cm|Cognia Interaction NormsCognia Interaction Norms
Trigger Pre-Conditions Behavior
On group chat creation Cognia chatbot is available Cognia chatbot joins the chat with the mediator role and user joins the chat with the owner_user role
On utterance sent by user Utterance's topic is CDB (cdbguru) or Savings Account (poupancaguru) and speech act is not Query_Calculation and they are not in the chat Cognia invites experts to the chat and repeats the utterance to them
On utterance sent by user Utterance's topic is CDB (cdbguru) or Savings Account (poupancaguru) and speech act is not Query_Calculation and they are in the chat Cognia waits for while and cdbguru or poupancaguru respectively handles the utterance. If they don't understand, they don't reply
On utterance sent by the experts If Cognia is waiting for them and has received both replies Cognia does not wait anymore
On utterance sent Utterance mentions cdbguru or poupancaguru cdbguru or poupancaguru respectively handles the utterance
On utterance sent Utterance mentions cdbguru or poupancaguru and they don't reply after a while and speech act is Query_Calculation Cognia sends I can only chat about investments...
On utterance sent Utterance mentions cdbguru or poupancaguru and they don't reply after while and speech act is not Query_Calculation Cognia sends I didn't understand
On utterance sent Utterance's speech act is Query_Calculation and period or initial amount of investment were not specified Cognia asks the user the missing information
On utterance sent Utterance's speech act is Query_Calculation and period and initial amount of investment were specified and the experts are not in the chat Cognia invites experts to the chat and repeats the utterance to them
On utterance sent Utterance's speech act is Query_Calculation and period and initial amount of investment were specified and the experts are in the chat Cognia repeats the utterance to experts
On utterance sent Utterance's speech act is Query_Calculation Cognia extracts variables and saves the context
On utterance sent Utterance's speech act is Query_Calculation and the experts are in the chat and the experts are mentioned Experts extract information, save in the context, compute calculation and send information
On utterance sent Utterance's speech act is Inform_Calculation and Cognia received all replies Cognia compares the results and inform comparison
On utterance sent Utterance mentions a chatbot but has no other text The chatbot replies How can I help you?
On utterance sent Utterance is not understood and speech act is Question The chatbot replies I don't know... I can only talk about topic X
On utterance sent Utterance is not understood and speech act is not Question The chatbot replies I didn't understand
On utterance sent Utterance's speech act is one of { Greetings, Thank, Bye } All chatbots reply to utterance
On group chat end All chatbots leave the chat, and the date and time of the end of chat is registered
We instantiated SABIA to develop CognIA as follows: the Mediator, Savings Account, CDB and User Actors are the Members of the Chat Group. The Hub was implemented using two servers: Socket.io and Node.JS which is a socket client of the Socket.io server. The CognIA system has also one Socket Client for receiving the broadcast and forwarding to the Group Chat Manager. The former will actually do the broadcast to every member after enforcing the norms that applies specified in Table SECREF72 . Each Member will behave according to this table too. For each user of the chat group, on a mobile or a desktop, there is its corresponding actor represented by the User Actor in the figure. Its main job is to receive Akka's broadcast and forward to the Socket.io server, so it can be finally propagated to the users.
All the intents, actions, factual answers, context and logging data are saved in DashDB (a relational Database-as-a-Service system). When an answer is not retrieved, a service which executes the module Search Finance on Social Media on a separate server is called. This service was implemented with the assumption that finance experts post relevant questions and answers on social media. Further details are explained in the Action execution sub-section.
We built a small dictionary-based topic classifier to identify if an utterance refers to finance or not, and if it refers to the two investment options (CDB or Savings Account) or not.
The dependency parsing is extremely important for computing the return of investment when the user sends an utterance with this intention. Our first implementation used regular expressions which led to a very fragile approach. Then we used a TensorFlow implementation BIBREF47 of a SyntaxNet model for Portuguese and used it to generate the dependency parse trees of the utterances. The SyntaxNet model is a feed-forward neural network that operates on a task-specific transition system and achieves the state-of-the-art on part-of-speech tagging, dependency parsing and sentence compression results BIBREF48 . Below we present output of the service for the utterance:
"I want to invest 10 thousands in 40 months":
[s]""blue[l]:black
{ "original": "I would like to invest 10 thousands in 40 months",
"start_pos": [
23,
32],
"end_pos": [
27,
33],
"digits": [
10000,
40],
"converted": "I would like to invest 10000 in 40 months",
"tree": {
"like VERB ROOT": {
"I PRON nsubj": {},
"would MD aux":{
"invest VERB xcomp":{
"to TO aux": {},
"10000 NUM dobj": {},
"in IN prep": {
"months NOUN pobj":{
"40 NUM num": {}}}}}}}
The service returns a JSON Object containing six fields: original, start_pos, end_pos, digits, converted and tree. The original field contains the original utterance sent to the service. The converted field contains the utterance replaced with decimal numbers, if the case (for instance, "10 thousands" was converted to "10000" and replaced in the utterance). The start_pos and end_pos are arrays that contain the start and end char positions of the numbers in the converted utterance. While the tree contains the dependency parse tree for the converted utterance.
Given the dependency tree, we implemented the frame parsing which first extracts the entities and features from the utterance and saves them in the context. Then, it replaces the extracted entities and features for reserved characters.
extract_period_of_investment (utteranceTree) [1] [t] numbersNodes INLINEFORM0 utteranceTree.getNumbersNodes(); [t] foreach(numberNode in numbersNodes) do [t] parentsOfNumbersNode INLINEFORM1 numbersNode.getParents() [t] foreach(parent in parentsOfNumbersNodes) do [t] if ( parent.name contains { "day", "month", "year"} ) then [t] parentOfParent INLINEFORM2 parent.getParent() [t] if ( parentOfParent is not null and
parentOfParent.getPosTag==Verb and
parentOfParent.name in investmentVerbsSet ) then [t] return numberNode
Therefore an utterance like "I would like to invest 10 thousands in 3 years" becomes "I would like to invest #v in #dt years". Or "10 in 3 years" becomes "#v in #dt years", and both intents have the same intent class.
For doing that we implemented a few rules using a depth-first search algorithm combined with the rules as described in Algorithm UID79 , Algorithm UID79 and Algorithm UID79 . Note that our parser works only for short texts on which the user's utterance mentions only one period of time and/ or initial amount of investment in the same utterance.
extract_initial_amount_of_investment (utteranceTree) [1] [t] numbersNodes INLINEFORM0 utteranceTree.getNumbersNodes(); [t] foreach(numberNode in numbersNodes) do [t] parentsOfNumbersNode INLINEFORM1 numbersNode.getParents() [t] foreach(parent in parentsOfNumbersNodes) do [t] if ( parent.name does not contain { "day", "month", "year"} ) then [t] return numberNode
frame_parsing(utterance, utteranceTree) [1] [t] period INLINEFORM0 extract_period_of_investment (utteranceTree) [t] save_period_of_investment(period) [t] value INLINEFORM1 extract_initial_amount_of_investment (utteranceTree) [t] save_initial_amount_of_investment(value) [t] new_intent INLINEFORM2 replace(new_intent, period, "#dt") [t] new_intent INLINEFORM3 replace(new_intent, value, "#v")
In CognIA we have complemented the speech act classes with the ones related to the execution of specific actions. Therefore, if the chatbot needed to compute the return of investment, then, once it is computed, the speech act of the reply will be Inform_Calculation and the one that represents the query for that is Query_Calculation. In table TABREF81 we list the specific ones.
Given that there is no public dataset available with financial intents in Portuguese, we have employed the incremental approach to create our own training set for the Intent Classifier. First, we applied the Wizard of Oz method and from this study, we have collected a set of 124 questions that the users asked. Next, after these questions have been manually classified into a set of intent classes, and used to train the first version of the system, this set has been increased both in terms of number of classes and samples per class, resulting in a training set with 37 classes of intents, and a total 415 samples, with samples per class ranging from 3 to 37.
We have defined our classification method based on features extracted from word vectors. Word vectors consist of a way to encode the semantic meaning of the words, based on their frequency of co-occurrence. To create domain-specific word vectors, a set of thousand documents are needed related to desired domain. Then each intent from the training set needs to be encoded with its corresponding mean word vector. The mean word vector is then used as feature vector for standard classifiers.
We have created domain-specific word vectors by considering a set 246,945 documents, corresponding to of 184,001 Twitter posts and and 62,949 news articles, all related to finance .
The set of tweets has been crawled from the feeds of blog users who are considered experts in the finance domain. The news article have been extracted from links included in these tweets. This set contained a total of 63,270,124 word occurrences, with a vocabulary of 97,616 distinct words. With the aforementioned word vectors, each intent from the training set has been encoded with its corresponding mean word vector. The mean word vector has been then used as feature vector for standard classifiers.
As the base classifier, we have pursued with a two-step approach. In the first step, the main goal was to make use of a classifier that could be easily retrained to include new classes and intents. For this reason, the first implementation of the system considered an 1-nearest-neighbor (1NN) classifier, which is simply a K-nearest-neighbor classifier with K set to 1. With 1NN, the developer of the system could simply add new intents and classes to the classifier, by means of inserting new lines into the database storing the training set. Once we have considered that the training set was stable enough for the system, we moved the focus to an approach that would be able to provide higher accuracy rates than 1NN. For this, we have employed Support Vector Machines (SVM) with a Gaussian kernel, the parameters of which are optimized by means of a grid search.
We manually mapped the intent classes used to train the intent classifier to action classes and the dependent entities and features, when the case. Table TABREF85 summarizes the number of intent classes per action class that we used in CognIA.
For the majority of action classes we used SABIA's default behavior. For instance, Greet and Bye actions classes are implemented using rapport, which means that if the user says "Hi" the chatbot will reply "Hi".
The Search News, Compute and Ask More classes are the ones that require specific implemention for CognIA as following:
Search News: search finance on social media service BIBREF49 , BIBREF50 receives the utterance as input, searches on previously indexed Twitter data for finance for Portuguese and return to the one which has the highest score, if found.
Ask More: If the user sends an utterance that has the intention class of simulating the return of investment, while not all variables to compute the return of investment are extracted from the dialogue, the mediator keeps asking the user these information before it actually redirects the query to the experts. This action then checks the state of the context given the specified intent flow as described in ( EQREF46 ) and ( EQREF57 ) in section SECREF4 to decide which variables are missing. For CognIA we manually added these dependencies on the database.
Compute: Each expert Chatbot implements this action according to its expertise. The savings account chatbot computes the formula ( EQREF90 ) and the certificate of deposit computes the formula ( EQREF92 ). Both are currently formulas for estimating in Brazil. DISPLAYFORM0
where INLINEFORM0 is the return of investment for the savings account, INLINEFORM1 is the initial value of investment, INLINEFORM2 is the savings account interest rate and INLINEFORM3 is the savings account rate base. DISPLAYFORM0
where INLINEFORM0 is the return of investment for certificate of deposit, INLINEFORM1 is the initial value of investment, INLINEFORM2 is the Interbank Deposit rate (DI in Portuguese), INLINEFORM3 is the ID's percentual payed by the bank (varies from 90% to 120%), INLINEFORM4 is the number of days the money is invested, and finally INLINEFORM5 is the income tax on the earnings.
Intention Classifier Accuracy
In Table TABREF95 we present the comparison of some distinct classification on the first version of the training set, i.e. the set used to deploy the first classifier into the system. Roughly speaking, the 1NN classifier has been able to achieve a level of accuracy that is higher than other well-known classifiers, such as Logistic Regression and Naïve Bayes, showing that 1NN is suitable as a development classifier. Nevertheless, a SVM can perform considerable better than 1NN, reaching accuracies of about 12 percentage points higher, which demonstrates that this type of base classifier is a better choice to be deployed once the system is stable enough. It is worth mentioning that these results consider the leave-one-out validation procedure, given the very low number of samples in some classes.
As we mentioned, the use of an 1NN classifier has allowed the developer of the system to easily add new intent classes and samples whenever they judged it necessary, so that the system could present new actions, or the understanding of the intents could be improved. As a consequence, the initial training set grew from 37 to 63 classes, and from 415 to 659 samples, with the number of samples per class varying from 2 to 63. For visualizing the impact on the accuracy of the system, in Table TABREF96 we present the accuracy of the same classifiers used in the previous evaluation, in the new set. In this case, we observe some drop in accuracy for 1NN, showing that this classifier suffers in dealing with scalability. On the other hand, SVM has shown to scale very well to more classes and samples, since its accuracy kept at a very similar level than that with the other set, with a difference of about only 1 percentage point.
Testing SABIA
In this section, we describe the validation framework that we created for integration tests. For this, we developed it as a new component of SABIA's system architecture and it provides a high level language which is able to specify interaction scenarios that simulate users interacting with the deployed chatbots. The system testers provide a set of utterances and their corresponding expected responses, and the framework automatically simulates users interacting with the bots and collect metrics, such as time taken to answer an utterance and other resource consumption metrics (e.g., memory, CPU, network bandwidth). Our goal was to: (i) provide a tool for integration tests, (ii) to validate CognIA's implementation, and (iii) to support the system developers in understanding the behavior of the system and which aspects can be improved. Thus, whenever developers modify the system's source code, the modifications must first pass the automatic test before actual deployment.
The test framework works as follows. The system testers provide a set INLINEFORM0 of dialogues as input. Each dialogue INLINEFORM1 INLINEFORM2 INLINEFORM3 is an ordered set whose elements are represented by INLINEFORM4 , where INLINEFORM5 is the user utterance and INLINEFORM6 is an ordered set of pairs INLINEFORM7 that lists each response INLINEFORM8 each chatbot INLINEFORM9 should respond when the user says INLINEFORM10 . For instance, Table UID98 shows a typical dialogue ( INLINEFORM11 ) between a user and the CognIA system. Note that we are omitting part of the expected answer with "..." just to better visualize the content of the table.
|p3.6cmp0.4cmp4.5cmp3.2cm|Content of dialogue INLINEFORM0 (example of dialogue in CognIA)Content of dialogue INLINEFORM1 (example of dialogue in CognIA
User utterance INLINEFORM0 rId Expected response INLINEFORM1 Chatbot INLINEFORM2
gray!25 hello 1 Hello Mediator
white what is cdb? 2 @CDBExpert what is cdb? Mediator
white 3 CDB is a type of investment that... CDB Expert
gray!25 which is better: cdb or savings account? 4 I found a post in the social media for.... Mediator
white i would like to invest R$ 50 in six months 5 @SavingsAccountExpert and @CDBExpert, could you do a simulation... Mediator
white 6 If you invest in Savings Account, ... Savings Account Exp.
white 7 If you invest in CDB,... CDB Expert
white 8 Thanks Mediator
white 9 @User, there is no significant difference.. Mediator
gray!25 so i want to invest R$ 10000 in 2 years 10 @SavingsAccountExpert and @CDBExpert, could you do a simulation... Mediator
gray!25 11 If you invest in Savings Account,... Savings Account Exp.
gray!25 12 If you invest in CDB,... CDB Expert
gray!25 13 Thanks Mediator
gray!25 14 @User, in that case, it is better... Mediator
white what if i invest R$10,000 in 5 years? 15 @SavingsAccountExpert and @CDBExpert, could you do a simulation... Mediator
white 16 If you invest in Saving Account,... Savings Account Exp.
white 17 If you invest in CDB,... CDB Expert
white 18 Thanks Mediator
white 19 @User, in that case, it is better... Mediator
gray!25 how about 15 years? 20 @SavingsAccountExpert and @CDBExpert, could you do a simulation... Mediator
gray!25 21 If you invest in Savings Account,... Savings Account Exp
gray!25 22 If you invest in CDB,... CDB Expert
gray!25 23 Thanks Mediator
gray!25 24 @User, in that case, it is better... Mediator
white and 50,0000? 25 @SavingsAccountExpert and @CDBExpert, could you do a simulation... Mediator
white 26 If you invest in Savings Account,... Savings Account Exp.
white 27 If you invest in CDB,... CDB Expert
white 28 Thanks Mediator
white 29 @User, in that case, it is better.. Mediator
gray!25 I want to invest in 50,000 for 15 years in CDB 30 Sure, follow this link to your bank... Mediator
white thanks 31 You are welcome. Mediator
The testers may also inform the number of simulated users that will concurrently use the platform. Then, for each simulated user, the test framework iterates over the dialogues in INLINEFORM0 and iterates over the elements in each dialogue to check if each utterance INLINEFORM1 was correctly responded with INLINEFORM2 by the chatbot INLINEFORM3 . There is a maximum time to wait. If a bot does not respond with the expected response in the maximum time (defined by the system developers), an error is raised and the test is stopped to inform the developers about the error. Otherwise, for each correct bot response, the test framework collects the time taken to respond that specific utterance by the bot for that specific user and continues for the next user utterance. Other consumption resource metrics (memory, CPU, network, disk). The framework is divided into two parts. One part is responsible to gather resource consumption metrics and it resides inside SABIA. The other part works as clients (users) interacting with the server. It collects information about time taken to answer utterances and checks if the utterances are answered correctly.
By doing this, we not only provide a sanity test for the domain application (CognIA) developed in SABIA framework, but also a performance analysis of the platform. That is, we can: validate if the bots are answering correctly given a pre-defined set of known dialogues, check if they are answering in a reasonable time, and verify the amount of computing resources that were consumed to answer a specific utterance. Given the complexity of CognIA, these tests enable debugging of specific features like: understanding the amount of network bandwidth to use external services, or analyzing CPU and memory consumption when responding a specific utterance. The later may happen when the system is performing more complex calculations to indicate the investment return, for instance.
CognIA was deployed on IBM Bluemix, a platform as a service, on a Liberty for Java Cloud Foundry app with 3 GB RAM memory and 1 GB disk. Each of the modules shown in Figure FIGREF74 are deployed on separate Bluemix servers. Node.JS and Socket.IO servers are both deployed as Node Cloud Foundry apps, with 256 MB RAM memory and 512 MB disk each. Search Finance on Social Media is on a Go build pack Cloud Foundry app with 128 MB RAM memory and 128 GB disk. For the framework part that simulates clients, we instantiated a virtual machine with 8 cores on IBM's SoftLayer that is able to communicate with Bluemix. Then, the system testers built two dialogues, i.e., INLINEFORM0 . The example shown in Table UID98 is the dialogue test INLINEFORM1 . For the dialogue INLINEFORM2 , although it also has 10 utterances, the testers varied some of them to check if other utterances in the finance domain (different from the ones in dialogue INLINEFORM3 ) are being responded as expected by the bots. Then, two tests are performed and the results are analyzed next. All tests were repeated until the standard deviation of the values was less than 1%. The results presented next are the average of these values within the 1% margin.
Test 1: The first test consists of running both dialogues INLINEFORM0 and INLINEFORM1 for only one user for sanity check. We set 30 seconds as the maximum time a simulated user should wait for a bot correct response before raising an error. The result is that all chatbots (Mediator, CDBExpert, and SavingsAccountExpert) responded all expected responses before the maximum time. Additionally, the framework collected how long each chatbot took to respond an expected answer.
In Figure FIGREF101 , we show the results for those time measurements for dialogue INLINEFORM0 , as for the dialogue INLINEFORM1 the results are approximately the same. The x-axis (Response Identifier) corresponds to the second column (Resp. Id) in Table UID98 . We can see, for example, that when the bot CDBExpert responds with the message 3 to the user utterance "what is cdb?", it is the only bot that takes time different than zero to answer, which is the expected behavior. We can also see that the Mediator bot is the one that takes the longest, as it is responsible to coordinate the other bots and the entire dialogue with the user. Moreover, when the expert bots (CDBExpert and SavingsAccountExpert) are called by the Mediator to respond to the simulation calculations (this happens in responses 6, 7, 11, 12, 16, 17, 21, 22, 26, 27), they take approximately the same to respond. Finally, we see that when the concluding responses to the simulation calculations are given by the Mediator (this happens in responses 9, 14, 19, 24, 29), the response times reaches the greatest values, being 20 seconds the greatest value in response 19. These results support the system developers to understand the behavior of the system when simulated users interact with it and then focus on specific messages that are taking longer.
Test 2: This test consists of running dialogue INLINEFORM0 , but now using eight concurrent simulated users. We set the maximum time to wait to 240 seconds, i.e., eight times the maximum set up for the single user in Test 1. The results are illustrated in Figure FIGREF102 , where we show the median time for the eight users. The maximum and minimum values are also presented with horizontal markers. Note that differently than what has been shown in Figure FIGREF101 , where each series represents one specific chatbot, in Figure FIGREF102 , the series represents the median response time for the responses in the order (x-axis) they are responded, regardless the chatbot.
Comparing the results in Figure FIGREF102 with the ones in Figure FIGREF101 , we can see that the bots take longer to respond when eight users are concurrently using the platform than when a single user uses it, as expected. For example, CDBExpert takes approximately 5 times longer to respond response 3 to eight users than to respond to one user. On average, the concluding responses to the simulation questions (i.e., responses 9, 14, 19, 24, 29) take approximately 7.3 times more to be responded with eight users than with one user, being the response 9 the one that presented greatest difference (11.4 times longer with eight users than with one). These results help the system developers to diagnose the scalability of the system architecture and to plan sizing and improvements.
Conclusions and Future Work
In this article, we explored the challenges of engineering MPCS and we have presented a hybrid conceptual architecture and its implementation with a finance advisory system.
We are currently evolving this architecture to be able to support decoupled interaction norms specification, and we are also developing a multi-party governance service that uses that specification to enforce exchange of compliant utterances.
In addition, we are exploring a micro-service implementation of SABIA in order to increase its scalability and performance, so thousands of members can join the system within thousands of conversations.
Acknowledgments
The authors would like to thank Maximilien de Bayser, Ana Paula Appel, Flavio Figueiredo and Marisa Vasconcellos, who contributed with discussions during SABIA and CognIA's implementation.
| [
"Custom dataset with user questions; set of documents, twitter posts and news articles, all related to finance.",
"a self-collected financial intents dataset in Portuguese"
] | 13,401 | qasper | en | null | f27a64d129091a6c8973c001ff789b8f68955b8ff0ae70af | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What datasets are used?
Answer: | [
"Custom dataset with user questions; set of documents, twitter posts and news articles, all related to finance.",
"a self-collected financial intents dataset in Portuguese"
] | qasper | 128 | 15 | |
Which stock market sector achieved the best performance? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Natural Language Processing (NLP) has increasingly attracted the attention of the financial community. This trend can be explained by at least three major factors. The first factor refers to the business perspective. It is the economics of gaining competitive advantage using alternative sources of data and going beyond historical stock prices, thus, trading by analyzing market news automatically. The second factor is the major advancements in the technologies to collect, store, and query massive amounts of user-generated data almost in real-time. The third factor refers to the progress made by the NLP community in understanding unstructured text. Over the last decades the number of studies using NLP for financial forecasting has experienced exponential growth. According to BIBREF0 , until 2008, less than five research articles were published per year mentioning both “stock market” and “text mining” or “sentiment analysis” keywords. In 2012, this number increased to slightly more than ten articles per year. The last numbers available for 2016 indicates this has increased to sixty articles per year.
The ability to mechanically harvest the sentiment from texts using NLP has shed light on conflicting theories of financial economics. Historically, there has been two differing views on whether disagreement among market participants induces more trades. The “non-trade theorem” BIBREF1 states that assuming all market participants have common knowledge about a market event, the level of disagreement among the participants does not increase the number of trades but only leads to a revision of the market quotes. In contrast, the theoretically framework proposed in BIBREF2 advocates that disagreement among market participants increases trading volume. Using textual data from Yahoo and RagingBull.com message boards to measure the dispersion of opinions (positive or negative) among traders, it was shown in BIBREF3 that disagreement among users' messages helps to predict subsequent trading volume and volatility. Similar relation between disagreement and increased trading volume was found in BIBREF4 using Twitter posts. Additionally, textual analysis is adding to the theories of medium-term/long-term momentum/reversal in stock markets BIBREF5 . The unified Hong and Stein model BIBREF6 on stock's momentum/reversal proposes that investors underreact to news, causing slow price drifts, and overreact to price shocks not accompanied by news, hence inducing reversals. This theoretical predicated behaviour between price and news was systematically estimated and supported in BIBREF7 , BIBREF8 using financial media headlines and in BIBREF9 using the Consumer Confidence Index® published by The Conference Board BIBREF10 . Similarly, BIBREF11 uses the Harvard IV-4 sentiment lexicon to count the occurrence of words with positive and negative connotation of the Wall Street Journal showing that negative sentiment is a good predictor of price returns and trading volumes.
Accurate models for forecasting both price returns and volatility are equally important in the financial domain. Volatility measures how wildly the asset is expected to oscillate in a given time period and is related to the second moment of the price return distribution. In general terms, forecasting price returns is relevant to take speculative positions. The volatility, on the other hand, measures the risk of these positions. On a daily basis, financial institutions need to assess the short-term risk of their portfolios. Measuring the risk is essential in many aspects. It is imperative for regulatory capital disclosures required by banking supervision bodies. Moreover, it is useful to dynamically adjust position sizing accordingly to market conditions, thus, maintaining the risk within reasonable levels.
Although, it is crucial to predict the short-term volatility from the financial markets application perspective, much of the current NLP research on volatility forecasting focus on the volatility prediction for very long-term horizons (see BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 ). Predominately, these works are built on extensions of the bag-of-words representation that has the main drawback of not capturing word order. Financial forecasting, however, requires the ability to capture semantics that is dependent upon word order. For example, the headline “Qualcomm sues Apple for contract breach” and “Apple sues Qualcomm for contract breach” trigger different responses for each stock and for the market aggregated index, however, they share the same bag-of-words representation. Additionally, these works use features from a pretrained sentiment analyis model to train the financial forecasting model. A key limitation of this process is that it requires a labelled sentiment dataset. Additionally, the error propagation is not end-to-end. In this work, we fill in the gaps of volatility prediction research in the following manner:
Related work
Previous work in BIBREF12 incorporates sections of the “Form 10-K” to predict the volatility twelve months after the report is released. They train a Support Vector Regression model on top of sparse representation (bag-of-words) with standard term weighting (e.g. Term-Frequency). This work was extended in BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 by employing the Loughran-McDonald Sentiment Word Lists BIBREF20 , which contain three lists where words are grouped by their sentiments (positive, negative and neutral). In all these works, the textual representation is engineered using the following steps: 1) For each sentiment group, the list is expanded by retrieving 20 most similar words for each word using Word2Vec word embeddings BIBREF21 . 2) Finally, each 10-K document is represented using the expanded lists of words. The weight of each word in this sparse representation is defined using Information Retrieval (IR) methods such as term-frequency (tf) and term-frequency with inverted document frequency (tfidf). Particularly, BIBREF16 shows that results can be improved using enhanced IR methods and projecting each sparse feature into a dense space using Principal Component Analysis (PCA).
The works described above ( BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 ) target long-horizon volatility predictions (one year or quarterly BIBREF16 ). In particular, BIBREF16 and BIBREF15 uses market data (price) features along with the textual representation of the 10-K reports. These existing works that employ multi-modal learning BIBREF22 are based on a late fusion approach. For example, stacking ensembles to take into account the price and text predictions BIBREF16 . In contrast, our end-to-end trained model can learn the joint distribution of both price and text.
Predicting the price direction rather than the volatility was the focus in BIBREF23 . They extracted sentiment words from Twitter posts to build a time series of collective Profile of Mood States (POMS). Their results show that collective mood accurately predicts the direction of Down Jones stock index (86.7% accuracy). In BIBREF24 handcrafted text representations including term count, noun-phrase tags and extracted named entities are employed for predicting stock market direction using Support Vector Machine (SVM). An extension of Latent Dirichlet Allocation (LDA) is proposed in BIBREF25 to learn a joint latent space of topics and sentiments.
Our deep learning models bear a close resemblance to works focused on directional price forecasting BIBREF26 , BIBREF27 . In BIBREF26 , headline news are processed using Stanford OpenIE to generate triples that are fed into a Neural Tensor Network to create the final headline representation. In BIBREF27 , a character-level embedding is pre-trained in an unsupervised manner. The character embedding is used as input to a sequence model to learn the headline representation. Particularly, both works average all headline representations in a given day, rather than attempting to weight the most relevant ones. In this work, we propose a neural attention mechanism to capture the News Relevance and provide experimental evidence that it is a key component of the end-to-end learning process. Our attention extends the previous deep learning methods from BIBREF26 , BIBREF27 .
Despite the fact that end-to-end deep learning models have attained state-of-the-art performance, the large number of parameters make them prone to overfitting. Additionally, end-to-end models are trained from scratch requiring large datasets and computational resources. Transfer learning (TL) alleviates this problem by adapting representations learnt from a different and potentially weakly related source domain to the new target domain. For example, in computer vision tasks the convolutional features learnt from ImageNet BIBREF28 dataset (source domain) have been successfully transferred to multiple domain target tasks with much smaller datasets such as object classification and scene recognition BIBREF29 . In this work, we consider TL in our experiments for two main reasons. First, it address the question whether our proposed dataset is suitable for end-to-end training since the performance of the transferred representations can be compared with end-to-end learning. Second, it is still to be investigated which dataset transfers better to the forecasting problem. Recently, the NLP community has focused on universal representations of sentences BIBREF17 , BIBREF19 , which are dense representations that carry the meaning of a full sentence. BIBREF17 found that transferring the sentence representation trained on the Stanford Natural Language Inference (SNLI) BIBREF30 dataset achieves state-of-the-art sentence representations to multiple NLP tasks (e.g. sentiment analysis, question-type and opinion polarity). Following BIBREF17 , in this work, we investigate the suitability of SNLI and Reuters RCV1 BIBREF31 datasets to transfer learning to the volatility forecasting task. To the best of our knowledge, the hierarchical attention mechanism at headline level, proposed in our work, has not being applied to volatility prediction so far; neither has been investigated the ability to transfer sentence encoders from source datasets to the target forecasting problem (Transfer Learning).
Our dataset
Our corpus covers a broad range of news including news around earnings dates and complements the 10-K reports content. As an illustration, the headlines “Walmart warns that strong U.S. dollar will cost $15B in sales” and “Procter & Gamble Co raises FY organic sales growth forecast after sales beat” describe the company financial conditions and performance from the management point of view – these are also typical content present in Section 7 of the 10-K reports.
In this section, we describe the steps involved in compiling our dataset of financial news at stock level, which comprises a broad range of business sectors.
Sectors and stocks
The first step in compiling our corpus was to choose the constituents stocks. Our goal was to consider stocks in a broad range of sectors, aiming a diversified financial domain corpus. We found that Exchange Traded Funds (ETF) provide a mechanical way to aggregate the most relevant stocks in a given industry/sector. An ETF is a fund that owns assets, e.g. stock shares or currencies, but, unlike mutual funds are traded in stock exchanges. These ETFs are extremely liquid and track different investment themes. We decided to use SPDR Setcor Funds constituents stocks in our work since the company is the largest provider of sector funds in the United States. We included in our analysis the top 5 (five) sector ETFs by financial trading volume (as in Jan/2018). Among the most traded sectors we also filtered out the sectors that were similar to each other. For example, the Consumer Staples and Consumer Discretionary sectors are both part of the parent Consumer category. For each of the top 5 sectors we selected the top 10 holdings, which are deemed the most relevant stocks. tbl:stockuniverse, details our dataset sectors and its respective stocks.
Stock specific data
We assume that an individual stock news as the one that explicitly mention the stock name or any of its surface forms in the headline. As an illustration, in order to collect all news for the stock code PG, Procter & Gamble company name, we search all the headlines with any of these words: Procter&Gamble OR Procter and Gamble OR P&G. In this example, the first word is just the company name and the remaining words are the company surface forms.
We automatically derived the surface forms for each stock by starting with a seed of surface forms extracted from the DBpedia Knowledge Base (KB). We then applied the following procedure:
Relate each company name with the KB entity unique identifier.
Retrieve all values of the wikiPageRedirects property. The property holds the names of different pages that points to the same entity/company name. This step sets the initial seed of surface forms.
Manually, filter out some noisy property values. For instance, from the Procter & Glamble entity page we were able to automatically extract dbr:Procter_and_gamble and dbr:P_&_G, but had to manually exclude the noisy associations dbr:Female_pads and dbr:California_Natural.
The result of the steps above is a dictionary of surface forms $wd_{sc}$ .
Stock headlines
Our corpus is built at stock code level by collecting headlines from the Reuters Archive. This archive groups the headlines by date, starting from 1 January 2007. Each headline is a html link (<a href> tag) to the full body of the news, where the anchor text is the headline content followed by the release time. For example, the page dated 16 Dec 2016 has the headline “Procter & Gamble appoints Nelson Peltz to board 5:26PM UTC”.
For each of the 50 stocks (5 sectors times 10 stocks per sector) selected using the criteria described in sub:corpussecstock, we retrieved all the headlines from the Reuters Archive raging from 01/01/2007 to 30/12/2017. This process takes the following steps:
For a given stock code ( $sc$ ) retrieve all surface forms $wd_{sc}$ .
For each day, store only the headlines content matching any word in $wd_{sc}$ . For each stored headline we also store the time and timezone.
Convert the news date and time to Eastern Daylight Time (EDT).
Categorize the news release time. We consider the following category set: {before market, during market , after market, holidays, weekends}. during market contains news between 9:30AM and 4:00PM. before market before 9:30AM and after market after 4:00PM.
The time categories prevents any misalignment between text and stock price data. Moreover, it prevents data leakage and, consequently, unrealistic predictive model performance. In general, news released after 4:00PM EDT can drastically change market expectations and the returns calculated using close to close prices as in the GARCH(1,1) model (see eq:closingreturn). Following BIBREF3 , to deal with news misalignment, news issued after 4:00PM (after market) are grouped with the pre-market (before market) on the following trading day.
tbl:stocktimecat shows the distribution of news per sector for each time category. We can see a high concentration of news released before the market opens (55% on average). In contrast, using a corpus compiled from message boards, a large occurrence of news during market hours was found BIBREF3 . This behaviour indicating day traders' activity. Our corpus comprise financial news agency headlines, a content more focused on corporate events (e.g. lawsuits, merges & acquisitions, research & development) and on economic news (see tbl:stockheadlinesexmaples for a sample of our dataset). These headlines are mostly factual. On the other hand, user-generated content such as Twitter and message boards (as in BIBREF3 , BIBREF4 ) tends to be more subjective.
U.S. macroeconomic indicators such as Retail Sales, Jobless Claims and GDP are mostly released around 8:30AM (one hour before the market opens). These numbers are key drivers of market activity and, as such, have a high media coverage. Specific sections of these economic reports impact several stocks and sectors. Another factor that contribute to the high activity of news outside regular trading hours are company earnings reports. These are rarely released during trading hours. Finally, before the market opens news agencies provide a summary of the international markets developments, e.g. the key facts during the Asian and Australian trading hours. All these factors contribute to the high concentration of pre-market news.
Background
We start this section by reviewing the GARCH(1,1) model, which is a strong benchmark used to evaluate our neural model. We then review the source datasets proposed in the literature that were trained independently and transfered to our volatility prediction model. Finally, we review the general architectures of sequence modelling and attention mechanisms.
GARCH model
Financial institutions use the concept of “Value at risk” to measure the expected volatility of their portfolios. The widespread econometric model for volatility forecasting is the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) BIBREF32 , BIBREF33 . Previous research shows that the GARCH(1,1) model is hard to beat. For example, BIBREF34 compared GARCH(1,1) with 330 different econometric volatility models showing that they are not significantly better than GARCH(1,1). Let $p_t$ be the price of an stock at the end of a trading period with closing returns $r_t$ given by
$$r_t = \frac{p_t}{p_{t-1}} - 1 $$ (Eq. 29)
The GARCH process explicitly models the time-varying volatility of asset returns. In the GARCH(1,1) specification the returns series $r_t$ follow the process:
$$r_t &= \mu + \epsilon _t \\ \epsilon _t &= \sigma _t z_t \\ \sigma ^2_t &= a_0 + a_1 \epsilon _{t-1}^2 + b_1 \sigma _{t-1}^2$$ (Eq. 30)
where $\mu $ is a constant (return drift) and $z_t$ is a sequence of i.i.d. random variables with mean zero and unit variance. It is worth noting that although the conditional mean return described in eq:garchcondmean has a constant value, the conditional volatility $\sigma _t$ is time-dependent and modeled by eq:att.
The one-step ahead expected volatility forecast can be computed directly from eq:garchcondvariance and is given by
$$E_T[\sigma _{T+1}^2] = a_0 + a_1 E_T[\epsilon ^2] + b_1 E_T[\sigma _{T}^2] $$ (Eq. 32)
In general, the $t^{\prime }$ -steps ahead expected volatility $E_T[\sigma _{T+t^{\prime }}^2]$ can be easily expressed in terms of the previous step expected volatility. It is easy to prove by induction that the forecast for any horizon can be represented in terms of the one-step ahead forecast and is given by
$$E_T[\sigma _{T+t^{\prime }}^2] - \sigma _u^2 = (a_1 + b_1)^{(t^{\prime } -1)} \left(E_T[\sigma _{T+1}^2] - \sigma _u^2\right)$$ (Eq. 33)
where $\sigma _u$ is the unconditional volatility:
$$\sigma _u = \sqrt{a_0 / (1 - a_1 - b_1)} $$ (Eq. 34)
From the equation above we can see that for long horizons, i.e. $t^\prime \rightarrow \infty $ , the volatility forecast in eq:forecastrecursive converges to the unconditional volatility in eq:unvar.
All the works reviewed in sec:introduction ( BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 ) consider GARCH(1,1) benchmark. However, given the long horizon of their predictions (e.g. quarterly or annual), the models are evaluated using the unconditional volatility $\sigma _u$ in eq:unvar. In this work, we focus on the short-term volatility prediction and use the GARCH(1,1) one-day ahead conditional volatility prediction in eq:forecastoneperiod to evaluate our models.
Let $\sigma _{t+1}$ denote the ex-post “true” daily volatility at a given time $t$ . The performance on a set with $N$ daily samples can be evaluated using the standard Mean Squared Error ( $MSE$ ) and Mean Absolute Error ( $MAE$ )
$$MSE &= \frac{1}{N} \sum _{t=1}^{N} \left( E_t[\sigma _{t+1}] - \sigma _{t+1}\right)^2 \\ MAE &= \frac{1}{N} \sum _{t=1}^{N}\left|E_t[\sigma _{t+1}] - \sigma _{t+1} \right|$$ (Eq. 36)
Additionally, following BIBREF35 , the models are also evaluated using the coefficient of determination $R^2$ of the regression
$$\sigma _{t+1} = a + b E_t[\sigma _{t+1}] + e_t$$ (Eq. 37)
where
$$R^2 = 1 - \frac{\sum _{t=1}^{N}e^{2}_{t}}{\sum _{t=1}^{N}\left(E_t[\sigma _{t+1}] - \frac{1}{N} \sum _{t=1}^{N}E_t[\sigma _{t+1}]\right)^{2}}$$ (Eq. 38)
One of the challenges in evaluating GARCH models is the fact that the ex-post volatility $\sigma _{t+1}$ is not directly observed. Apparently, the squared daily returns $r_{t+1}^{2}$ in eq:closingreturn could stand as a good proxy for the ex-post volatility. However, the squared returns yield very noisy measurements. This is a direct consequence of the term $z^t$ that connects the squared return to the latent volatility factor in eq:garchwhitenoise. The use of intraday prices to estimate the ex-post daily volayility was first proposed in BIBREF35 . They argue that volatility estimators using intraday prices is the proper way to evaluate the GARCH(1,1) model, as opposed to squared daily returns. For example, considering the Deutsche Mark the GARCH(1,1) model $R^2$ improves from $0.047$ (squared returns) to $0.33$ (intraday returns) BIBREF35 .
It is clear from the previous section that any volatility model evaluation using the noisy squared returns as the ex-post volatility proxy will lead to very poor performance. Therefore, high-frequency intraday data is fundamental to short-term volatility performance evaluation. However, intraday data is difficult to acquire and costly. Fortunately, there are statistically efficient daily volatility estimators that only depend on the open, high, low and close prices. These price “ranges” are widely available. In this section, we discuss these estimators.
Let $O_t$ , $H_t$ , $L_t$ , $C_t$ be the open, high, low and close prices of an asset in a given day $t$ . Assuming that the daily price follows a geometric Brownian motion with zero drift and constant daily volatility $\sigma $ , Parkinson (1980) derived the first daily volatility estimator
$$\widehat{\sigma _{PK,t}^2} = \frac{\ln \left(\frac{H_t}{L_t}\right)^2}{4\ln (2)} $$ (Eq. 41)
which represents the daily volatility in terms of its price range. Hence, it contains information about the price path. Given this property, it is expected that $\sigma _{PK}$ is less noisy than the volatility calculated using squared returns. The Parkinson's volatility estimator was extended by Garman-Klass (1980) which incorporates additional information about the opening ( $O_t$ ) and closing ( $C_t$ ) prices and is defined as
$$\widehat{\sigma _{GK,t}^{2}} = \frac{1}{2} \ln \left(\frac{H_t}{L_t}\right)^2 - (2\ln (2) - 1) \ln \left(\frac{C_t}{O_t}\right)^2 $$ (Eq. 42)
The relative noisy of different estimators $\hat{\sigma }$ can be measured in terms of its relative efficiency to the daily volatility $\sigma $ and is defined as
$$e\left(\widehat{\sigma ^{2}}, \sigma ^2\right) \equiv \frac{Var[\sigma ^2]}{Var[\widehat{\sigma ^{2}}]}$$ (Eq. 43)
where $Var[\cdot ]$ is the variance operator. It follows directly from eq:garchwhitenoise that the squared return has efficiency 1 and therefore, very noisy. BIBREF36 reports Parkinson ( $\widehat{\sigma _{PK,t}^2}$ ) volatility estimator has 4.9 relative efficiency and Garman-Klass ( $\widehat{\sigma _{GK,t}^2}$ ) 7.4. Additionally, all the described estimators are unbiased.
Many alternative estimators to daily volatility have been proposed in the literature. However, experiments in BIBREF36 rate the Garman-Klass volatility estimator as the best volatility estimator based only on open, high, low and close prices. In this work, we train our models to predict the state-of-the-art Garman-Klass estimator. Moreover, we evaluate our models and GARCH(1,1) using the metrics described in sub:evalution, but with the appropriate volatility proxies, i.e. Parkinson and Garman-Klass estimators.
Transfer Learning from other source domains
Vector representations of words, also known as Word embeddings BIBREF21 , BIBREF37 , that represent a word as a dense vector has become the standard building blocks of almost all NLP tasks. These embeddings are trained on large unlabeled corpus and are able to capture context and similarity among words.
Some attempts have been made to learn vector representations of a full sentence, rather than only a single word, using unsupervised approaches similar in nature to word embeddings. Recently, BIBREF17 showed state-of-the-art performance when a sentence encoder is trained end-to-end on a supervised source task and transferred to other target tasks. Inspired by this work, we investigate the performance of sentence encoders trained on the Text categorization and Natural Language Inference (NLI) tasks and use these encoders in our main short-term volatility prediction task.
A generic sentence encoder $S_e$ receives the sentence words as input and returns a vector representing the sentence. This can be expressed as a mapping
$$S_e \colon \mathbb {R}^{T^{S} \times d_w} \rightarrow \mathbb {R}^{d_S}$$ (Eq. 45)
from a variable size sequence of words to a sentence vector $S$ of fixed-size $d_S$ , where $T^{S}$ is the sentence number of words and $d_w$ is the pre-trained word embedding dimension.
In the following sections, we describe the datasets and architectures to train the sentence encoders of the auxiliary transfer learning tasks.
The Reuters Corpus Volume I (RCV1) is corpus containing 806,791 news articles in the English language collected from 20/08/1996 to 19/08/1997 BIBREF31 . The topic of each news was human-annotated using a hierarchical structure. At the top of the hierarchy, lies the coarse-grained categories: CCAT (Corporate), ECAT (Economics), GCAT (Government), and MCAT (Markets). A news article can be assigned to more than one category meaning that the text categorization task is mutilabel. Each news is stored in a separate XML file. lst:rcv1xmlexample shows the typical structure of an article.
<?xml version="1.0" encoding="iso-8859-1" ?>
<newsitem itemid="6159" id="root" date="1996-08-21" xml:lang="en">
<headline>Colombia raises internal coffee price.</headline>
<dateline>BOGOTA 1996-08-21</dateline>
<copyright>(c) Reuters Limited 1996</copyright>
<metadata>
<codes class="bip:topics:1.0">
<code code="C13">
<editdetail attribution="Reuters BIP Coding Group" action="confirmed" date="1996-08-21"/>
</code>
<code code="C31">
<editdetail attribution="Reuters BIP Coding Group" action="confirmed" date="1996-08-21"/>
</code>
<code code="CCAT">
<editdetail attribution="Reuters BIP Coding Group" action="confirmed" date="1996-08-21"/>
</code>
<code code="M14">
<editdetail attribution="Reuters BIP Coding Group" action="confirmed" date="1996-08-21"/>
</code>
<code code="M141">
<editdetail attribution="Reuters BIP Coding Group" action="confirmed" date="1996-08-21"/>
</code>
<code code="MCAT">
<editdetail attribution="Reuters BIP Coding Group" action="confirmed" date="1996-08-21"/>
</code>
</codes>
</metadata>
</newsitem>
The RCV1 dataset is not released with a standard train, validation, test split. In this work, we separated 15% of samples as a test set for evaluation purposes. The remaining samples were further split leaving 70% and 15% for training and validation, respectively.
Regarding the categories distribution, we found that, from the original 126 categories, 23 categories were never assigned to any news; therefore, were disregarded. From the 103 classes left we found a high imbalance among the labels with a large number of underrepresented categories having less than 12 samples. The very low number of samples for these minority classes brings a great challenge to discriminate the very fine-grained categories. Aiming to alleviate this problem, we grouped into a same class all categories below the second hierarchical level. For example, given the root node CCAT (Corporate) we grouped C151 (ACCOUNTS/EARNINGS), C1511 (ANNUAL RESULTS) and C152 (COMMENT/FORECASTS) into the direct child node C15 (PERFORMANCE). Using this procedure the original 103 categories where reduced to 55. One of the benefits of this procedure was that the less represented classes end up having around thousand samples compared with only 12 samples in the original dataset.
fig:rcv1arch, shows the architecture for the end-to-end text categorization task. On the bottom of the architecture $S_e$ receives word embeddings and outputs a sentence vector $S$ . The $S$ vector pass through a fully connected (FC) layer with sigmoid activation function that outputs a vector $\hat{y} \in \mathbb {R}^{55}$ with each element $\hat{y}_j \in [0,1]$ .
The architecture described above is trained under the assumption that each category is independent but not mutually exclusive since a sample can have more than one category assigned (multilabel classification). The loss per sample is the average log loss across all labels:
$$\mathcal {L}(\hat{y}, y) = - \sum _{i=1}^{55}\left( y_i \log (\hat{y}_i) + (1-y_{i}) \log (1-\hat{y}_{i}) \right)$$ (Eq. 48)
where the index $i$ runs over the elements of the predicted and true vectors.
Given the high categories imbalance, during the training we monitor the $F_1$ metric of the validation set and choose the model with the highest value.
Stanford Natural Language Inference (SNLI) dataset BIBREF30 consist of 570,000 pairs of sentences. Each pair has a premise and a hypothesis, manually labeled with one of the three labels: entailment, contradiction, or neutral. The SNLI has many desired properties. The labels are equally balanced, as opposed to the RCV1 dataset. Additionally, language inference is a complex task that requires a deeper understanding of the sentence meaning making this dataset suitable for learning supervised sentence encoders that generalize well to other tasks BIBREF17 . tbl:snliexmaples, shows examples of SNLI dataset sentence pairs and its respective labels.
In order to learn sentence encoders that can be transfered to other tasks unambiguously, we consider a neural network architecture for the sentence encoder with shared parameters between the premise and hypothesis pairs as in BIBREF17 .
fig:snliarch, describes the neural network architecture. After each premise and hypothesis is encoded into $S_p$ and $S_h$ , respectively, we have a fusion layer. This layer has no trainable weights and just concatenate each sentence embedding. Following BIBREF17 , we add two more matching methods: the absolute difference $\vert S_p - S_h \vert $ and the element-wise $S_p \odot S_h$ . Finally, in order to learn the pair representation, $S_ph$ is feed into and FC layer with rectified linear unit (ReLU) activation function, which is expressed as $f(x) = \log (1 + e^x)$ . The last softmax layer outputs the probability of each class.
Finally, the NLI classifier weights are optimized in order to minimize the categorical log loss per sample
$$\mathcal {L}(\hat{y}, y) = - \sum _{j=1}^{3}y_i \log (\hat{y}_i)$$ (Eq. 52)
During the training, we monitor the validation set accuracy and choose the model with the highest metric value.
Sequence Models
We start this section by reviewing the Recurrent Neural Network (RNN) architecture and its application to encode a sequence of words.
RNN's are capable of handling variable-length sequences, this being a direct consequence of its recurrent cell, which shares the same parameters across all sequence elements. In this work, we adopt the Long Short-Term Memory (LSTM) cell BIBREF38 with forget gates $f_t$ BIBREF39 . The LSTM cell is endowed with a memory state that can learn representations that depend on the order of the words in a sentence. This makes LSTM more fit to find relations that could not be captured using standard bag-of-words representations.
Let $x_1, x_2, \cdots , x_T$ be a series of observations of length $T$ , where $x_t \in \mathbb {R}^{d_w}$ . In general terms, the LSTM cell receives a previous hidden state $h_{t-1}$ that is combined with the current observation $x_t$ and a memory state $C_t$ to output a new hidden state $h_t$ . This internal memory state $C_{t}$ is updated depending on its previous state and three modulating gates: input, forget, and output. Formally, for each step $t$ the updating process goes as follows (see fig:lstmcell for a high level schematic view): First, we calculate the input $i_t$ , forget $T$0 , and output $T$1 gates:
$$i_t &= \sigma _s\left(W_i x_t + U_i h_{t-1} + b_i\right) \\ f_t &= \sigma _s\left(W_f x_t + U_f h_{t-1} + b_f\right) \\ o_t &= \sigma _s\left(W_o x_t + U_o h_{t-1} + b_o\right)$$ (Eq. 54)
where $\sigma _s$ is the sigmoid activation. Second, a candidate memory state $\widetilde{C}_t$ is generated:
$$\widetilde{C}_t = \tanh \left(W_c x_t + U_c h_{t-1} + b_c\right)$$ (Eq. 55)
Now we are in a position to set the final memory state $C_t$ . Its value is modulated based on the input and forget gates of eq:inputforgetgates and is given by:
$$C_t = i_t \odot \widetilde{C}_t + f_t \odot C_{t-1}$$ (Eq. 56)
Finally, based on the memory state and output gate of eq:inputforgetgates, we have the output hidden state
$$h_t = o_t \odot \tanh \left(C_t\right)$$ (Eq. 57)
Regarding the trainable weights, let $n$ be the LSTM cell number of units. It follows that $W$ 's and $U$ 's matrices of the affine transformations have ${n \times d_w}$ and ${n \times n}$ dimensions, respectively. Its bias terms $b$ 's are vectors of size $n$ . Consequently, the total number of parameters is $4 (n d_w + n^2 + n)$ and does not depend on the sequence number of time steps $T$ .
We see that the LSTM networks are able to capture temporal dependencies in sequences of arbitrary length. One straightforward application is to model the Sentence encoder discussed in sec:transferlearning, which outputs a sentence vector representation using its words as input.
Given a sequence of words $\left\lbrace w_t\right\rbrace _{t=1}^{T}$ we aim to learn the words hidden state $\left\lbrace h_t\right\rbrace _{t=1}^{T}$ in a way that each word captures the influence of its past and future words. The Bidirectional LSTM (BiLSTM) proposed in BIBREF40 is an LSTM that “reads” a sentence, or any sequence in general, from the beginning to the end (forward) and the other way around (backward). The new state $h_t$ is the concatenation
$$h_t = [\overrightarrow{h_t}, \overleftarrow{h_t}]$$ (Eq. 59)
where
$$\overrightarrow{h_t} &= \text{LSTM}\left(w_1, \cdots , w_T\right) \\ \overleftarrow{h_t} &= \text{LSTM}\left(w_T, \cdots , w_1\right) \\$$ (Eq. 60)
Because sentences have different lengths, we need to convert the $T$ concatenated hidden states of the BiLSTM into a fixed-length sentence representation. One straightforward operation is to apply any form of pooling. Attention mechanism is an alternative approach where the sentence is represented as an weighted average of hidden states where the weights are learnt end-to-end.
In the next sections we describe the sentence encoders using pooling and attention layers.
The max-pooling layer aims to extract the most salient word features all over the sentence. Formally, it outputs a sentence vector representation $S_{MP} \in \mathbb {R}^{2n}$ such that
$$S_{MP} = \max _{t=1}^{T} h_t$$ (Eq. 62)
where $h_t$ is defined in eq:htconcat and the $\max $ operator is applied over the time steps dimension. fig:bilstmmaxpool illustrates the BiLSTM max-pooling (MP) sentence encoder.
The efficacy of the max-pooling layer was assessed in many NLP studies. BIBREF41 employed a max-pooling layer on top of word representations and argues that it performs better than mean pooling. Experimental results in BIBREF17 show that among three types of pooling (max, mean and last) the max-pooling provides the most universal sentence representations in terms of transferring performance to other tasks. Grounded on these studies, in this work, we choose the BiLSTM max-pooling as our pooling layer of choice.
Attention mechanisms were introduced in the deep learning literature to overcome some simplifications imposed by pooling operators. When we humans read a sentence, we are able to spot its most relevant parts in a given context and disregard information that is redundant or misleading. The attention model aims to mimic this behaviour.
Attention layers were proposed for different NLP tasks. For example, NLI, with cross-attention between premise and hypothesis, Question & Answering and Machine Translation (MT). Specifically in the Machine Translation task, each word in the target sentence learns to attend the relevant words of the source sentence in order to generate the sentence translation.
A sentence encoder with attention (or self-attentive) BIBREF42 , BIBREF43 , BIBREF44 assigns different weights to the own words of the sentence; therefore, converting the hidden states into a single sentence vector representation.
Considering the word hidden vectors set $\lbrace h_1, \cdots , h_T\rbrace $ where $h_t \in \mathbb {R}^n$ , the attention mechanism is defined by the equations:
$$\tilde{h}_t &= \sigma \left(W h_t + b \right) \\ \alpha _{t} &= \frac{\exp ({v^{\intercal } \cdot \tilde{h}_t} )}{\sum _{t} \exp ({v \cdot \tilde{h}_t})} \\ S_{A_w} &= \sum _{t} \alpha _{t} h_t$$ (Eq. 66)
where $W \in \mathbb {R}^{d_a \times n}$ , $b \in \mathbb {R}^{d_a \times 1}$ , and $v \in \mathbb {R}^{d_a \times 1}$ are trainable parameters.
We can see that the sentence representation $S_{A_w}$ is a weighted average of the hidden states. fig:bilstminneratt provides a schematic view of the BiLSTM attention, where we can account the attention described in eq:att as a two layer model with a dense layer ( $d_a$ units) followed by another dense that predicts $\alpha _t$ (single unit).
Methodology
In this section, we first introduce our problem in a deep multimodal learning framework. We then present our neural architecture, which is able to address the problems of news relevance and novelty. Finally, we review the methods applied to learn commonalities between stocks (global features).
Problem statement
Our problem is to predict the daily stock volatility. As discussed in subsub:rangevolestimators, the Gaman-Klass estimator $\widehat{\sigma _{GK,t}}$ in eq:volgk is a very efficient short-term volatility proxy, thus, it is adopted as our target variable.
Our goal is to learn a mapping between the next day volatility $\sigma _{t+1}$ and historical multimodal data available up to day $t$ . To this aim, we use a sliding window approach with window size $T$ . That is, for each stock $sc$ a sample on day $t$ is expressed as a sequence of historical prices $P^{sc}_t$ and corpus headlines $N^{sc}_t$ . The price sequence is a vector of Daily Prices (DP) and expressed as
$$P^{sc}_t = \left[DP^{sc}_{t-T}, DP^{sc}_{t-T+1}, \cdots , DP^{sc}_t \right]$$ (Eq. 69)
where $DP^{sc}_{t^{\prime }}$ is a vector of price features. In order to avoid task-specific feature engineering, the daily price features are expressed as the simple returns:
$$DP^{sc}_t = \left[ \frac{O^{sc}_{t}}{C^{sc}_{t-1}} - 1, \frac{H^{sc}_{t}}{C^{sc}_{t-1}} - 1, \frac{L^{sc}_{t}}{C^{sc}_{t-1}} - 1, \frac{C^{sc}_{t}}{C^{sc}_{t-1}} - 1 \right]$$ (Eq. 70)
The sequence of historical corpus headlines $N^{sc}_t$ is expressed as
$$N^{sc}_t = \left[n^{sc}_{t-T}, n^{sc}_{t-T+1}, \cdots , n^{sc}_{t} \right]$$ (Eq. 71)
where $n^{sc}_{t^{\prime }}$ is a set containing all headlines that influence the market on a given day $t^{\prime }$ .
Aiming to align prices and news modes, we consider the explicit alignment method discussed in subsec:stockheadlines. That is, $n^{sc}_{t^{\prime }}$ contains all stock headlines before the market opens ( $\texttt {before market}_{t}$ ), during the trading hours
( $\texttt {during market}_{t}$ ), and previous day after-markets
( $\texttt {after market}_{t-1}$ ).
As a text preprocessing step, we tokenize the headlines and convert each word to an integer that refers to its respective pre-trained word embedding. This process is described as follows: First, for all stocks of our corpus we tokenize each headline and extract the corpus vocabulary set $V$ . We then build the embedding matrix $E_w \in \mathbb {R}^{\vert V \vert \times d_w}$ , where each row is a word embedding vector $d_w$ dimensions. Words that do not have a corresponding embedding, i.e. out of vocabulary words, are skipped.
Finally, the input sample of the text mode is a tensor of integers with $T \times l_n \times l_s$ dimensions, where $l_n$ is the maximum number of news occurring in a given day and $l_s$ is the maximum length of a corpus sentence. Regarding the price mode, we have a $T \times 4$ tensor of floating numbers.
Global features and stock embedding
Given the price and news histories for each stock $sc$ we could directly learn one model per stock. However, this approach suffers from two main drawbacks. First, the market activity of one specific stock is expected to impact other stocks, which is a widely accepted pattern named “spillover effect”. Second, since our price data is sampled on a daily basis, we would train the stock model relying on a small number of samples. One possible solution to model the commonality among stocks would be feature enrichment. For example, when modeling a given stock $X$ we would enrich its news and price features by concatenating features from stock $Y$ and $Z$ . Although the feature enrichment is able to model the effect of other stocks, it still would consider only one sample per day.
In this work, we propose a method that learns an global model.
The global model is implemented using the following methods:
Multi-Stock batch samples: Since our models are trained using Stochastic Gradient Descent, we propose at each mini-batch iteration to sample from a batch set containing any stock of our stocks universe. As a consequence, the mapping between volatility and multimodal data is now able to learn common explanatory factors among stocks. Moreover, adopting this approach increases the total number of training samples, which is now the sum of the number of samples per stock.
Stock Embedding: Utilizing the Multi-Stock batch samples above, we tackle the problem of modeling commonality among stocks. However, it is reasonable to assume that stocks have part of its dynamic driven by idiosyncratic factors. Nevertheless, we could aggregate stocks per sector or rely on any measure of similarity among stocks. In order to incorporate information specific to each stock, we propose to equip our model with a “stock embedding” mode that is learnt jointly with price and news modes. That is to say, we leave the task of distinguishing the specific dynamic of each stock to be learnt by the neural network. Specifically, this stock embedding is modeled using a discrete encoding as input, i.e. $\mathcal {I}^{sc}_t$ is a vector with size equal to the number of stocks of the stocks universe and has element 1 for the i-th coordinate and 0 elsewhere, thus, indicating the stock of each sample.
Formally, we can express the one model per stock approach as the mapping
$$\begin{split} \sigma ^{sc}_{t+1} = f^{sc} ( DN^{sc}_{t-T}, DN^{sc}_{t-T+1}, \cdots , DN^{sc}_t ; \\ DP^{sc}_{t-T}, DP^{sc}_{t-T+1}, \cdots , DP^{sc}_t ) \end{split}$$ (Eq. 75)
where $DN^{sc}_{t^{\prime }}$ is a fixed-vector representing all news released on a given day for the stock $sc$ and $DP^{sc}_{t^{\prime }}$ is defined in eq:pricemodevec.
The global model attempts to learn a single mapping $f$ that at each mini-batch iteration randomly aggregates samples across all the universe of stocks, rather than one mapping $f^{sc}$ per stock. The global model is expressed as
$$\begin{split} \sigma ^{sc}_{t+1} = f ( DN^{sc}_{t-T}, DN^{sc}_{t-T+1}, \cdots , DN^{sc}_t ; \\ DP^{sc}_{t-T}, DP^{sc}_{t-T+1}, \cdots , DP^{sc}_t ; \\ \mathcal {I}^{sc}_t) \end{split}$$ (Eq. 77)
In the next section, we describe our hierarchical neural model and how the news, price and stock embedding are fused into a joint representation.
Our multimodal hierarchical network
In broad terms, our hierarchical neural architecture is described as follows. First, each headline released on a given day $t$ is encoded into a fixed-size vector $S_t$ using a sentence encoder. We then apply our daily New Relevance Attention (NRA) mechanism that attends each news based on its content and converts a variable size of news released on a given day into a single vector denoted by Daily News ( $DN$ ). We note that this representation take account of the overall effect of all news released on a given day. This process is illustrated in fig:DNencoder. We now are in a position to consider the temporal effect of the past $T$ days of market news and price features. fig:nntimeseriesarch illustrates the neural network architecture from the temporal sequence to the final volatility prediction. For each stock code $sc$ the temporal encoding for news is denoted by Market News $MN^{sc}_t$ and for the price by Market Price $MP^{sc}_t$ and are a function of the past $T$ Daily News representations ${\lbrace DN^{sc}_{t-T}, \cdots , DN^{sc}_t \rbrace }$ (Text mode) and Daily Prices features $S_t$0 (Price mode), where each Daily Price $S_t$1 feature is given by eq:pricemodevec and the $S_t$2 representation is calculated using Daily New Relevance Attention. After the temporal effects of $S_t$3 past days of market activity were already encoded into the Market News $S_t$4 and Market Price $S_t$5 , we concatenate feature-wise $S_t$6 , $S_t$7 and the Stock embedding $S_t$8 . The stock embedding $S_t$9 represents the stock code of the sample on a given day $t$ . Finally, we have a Fully Connected (FC) layer that learns the Joint Representation of all modes. This fixed-sized joint representation is fed into a FC layer with linear activation that predicts the next day volatility $\hat{\sigma }_{t+1}$ .
Below, we detail, for each mode separately, the layers of our hierarchical model.
– Text mode
Word Embedding Retrieval
Standard embedding layer with no trainable parameters. It receives a vector of word indices as input and returns a matrix of word embeddings.
News Encoder
This layer encodes all news on a given day and outputs a set news embeddings $\lbrace S^{1}_t, \cdots , S^{l_n}_t \rbrace $ . Each encoded sentence has dimension $d_S$ , which is a hyperparameter of our model. This layer constitutes a key component of our neural architectures and, as such, we evaluate our models considering sentence encoders trained end-to-end, using the BiLSTM attention (subsec:bilstminneratt) and BiLSTM max-pooling (subsec:bilstmmaxpool) architectures, and also transferred from the RCV1 and SNLI as fixed features.
Daily news relevance attention
Our proposed news relevance attention mechanism for all news released on a given day. The attention mechanism is introduced to tackle information overload. It was designed to “filter out” redundant or misleading news and focus on the relevant ones based solely on the news content. Formally, the layer outputs a Daily News (DN) embedding $DN^{sc}_t = \sum _{i=1}^{l_n} \beta _i S^{sc^{i}}_t$ , which is a linear combination of all encoded news on a given day $t$ . This news-level attention uses the same equations as in eq:att, but with trainable weights $\lbrace W_{R}, b_{R}, v_{R}\rbrace $ , i.e. the weights are segregated from the sentence encoder. fig:DNencoder, illustrates our relevance attention. Note that this layer was deliberately developed to be invariant to headlines permutation, as is the case with the linear combination formula above. The reason is that our price data is sampled daily and, as a consequence, we are not able to discriminate the market reaction for each intraday news.
News Temporal Context
Sequence layer with daily news embeddings $DN^{sc}_t$ as time steps. This layer aims to learn the temporal context of news, i.e. the relationship between the news at day $t$ and the $T$ past days. It receives as input a chronologically ordered sequence of $T$ past Daily News embeddings ${\lbrace DN^{sc}_{t-T}, \cdots , DN^{sc}_t \rbrace }$ and outputs the news mode encoding Market News $MN^{sc}_t \in d_{MN}$ . The sequence with $T$ time steps is encoded using a BiLSTM attention. The layer was designed to capture the temporal order that news are released and the current news novelty. i.e. news that were repeated in the past can be “forgotten” based on the modulating gates of the LSTM network.
– Price mode
Price Encoder
Sequence layer analogous to News Temporal Context, but for the price mode. The input is the ordered sequence Daily Prices ${\lbrace DP^{sc}_{t-T}, \cdots , DP^{sc}_t \rbrace }$ of size $T$ , where each element the price feature defined in eq:pricemodevec. Particularly, the architecture consists of two stacked LSTM's. The first one outputs for each price feature time step a hidden vector that takes the temporal context into account. Then these hidden vectors are again passed to a second independent LSTM. The layer outputs the price mode encoding Market Price $MP^{sc}_t \in d_{MP}$ . This encoding is the last hidden vector of the second LSTM Market.
– Stock embedding
Stock Encoder
Stock dense representation. The layer receives the discrete encoding $\mathcal {I}^{sc}_t$ indicating the sample stock code pass through a FC layer and outputs a stock embedding $E_{sc}$ .
– Joint Representation
Merging
Feature-wise News, Price, and Stock modes concatenation. No trainable parameters.
Joint Representation Encoder
FC layer of size $d_{JR}$ .
Multimodal learning with missing modes
During the training we feed into our neural model the price, news, and stock indicator data. The price and stock indicator modes data occur in all days. However, at the individual stock level we can have days that the company is not covered by the media. This feature imposes challenges to our multimodal training since neural networks are not able to handle missing modes without special intervention. A straightforward solution would be to consider only days with news released, disregarding the remaining samples. However, this approach has two main drawbacks. First, the “missing news” do not happen at random, or are attributed to measurement failure as is, for example, the case of multimodal tasks using mechanical sensors data. Conversely, as highlighted in BIBREF7 , BIBREF8 the same price behaviour results in distinct market reactions when accompanied or not by news. In other words, specifically to financial forecasting problems the absence or existence of news are highly informative.
Some methods were proposed in the multimodal literature to effectively treat informative missing modes or “informative missingness”, which is a characteristic refereed in the literature as learning with missing modalities BIBREF22 . In this work, we directly model the news missingness as a feature of our text model temporal sequence by using the method initially proposed in BIBREF45 , BIBREF46 for clinical data with missing measurements and applied in the context of financial forecasting in BIBREF47 . Specifically, we implement the Zeros & Imputation (ZI) method BIBREF46 in order to jointly learn the price mode and news relationship across all days of market activity.
The ZI implementation is described as follows: Before the daily news sequence is processed by the text temporal layer (described in itm:newstclayer) we input a 0 vector for all time steps with missing news and leave the news encoding unchanged otherwise. This step is called zero imputation. In addition, we concatenate feature-wise an indicator vector with value 1 for all vectors with zero imputation and 0 for the days with news.
As described in BIBREF47 , the ZI method endow a temporal sequence model with the ability to learn different representations depending on the news history and its relative time position. Moreover, it allows our model to predict the volatility for all days of our time series and, at the same time, take into account the current and past news informative missingness. Furthermore, the learnt positional news encoding works differently than a typical “masking”, where days without news are not passed through the LSTM cell. Masking the time steps would be losing information about the presence or absence of news concomitant with prices.
Experimental results and discussions
We aim to evaluate our hierarchical neural model in the light of three main aspects. First, we asses the importance of the different sentence encoders to our end-to-end models and how it compares to transferring the sentence encoder from our two auxiliary TL tasks. Second, we ablate our proposed news relevance attention (NRA) component to evaluate its importance. Finally, we consider a model that takes into consideration only the price mode (unimodal), i.e. ignoring any architecture related to the text mode.
Before we define the baselines to asses the three aspects described above, we review in the next section the scores of the trained TL tasks.
Auxiliary transfer learning tasks
This section reports the performance of the auxiliary TL tasks considered in this work. Our ultimate goal is to indicate that our scores are in line with previous works All the architectures presented in sec:transferlearning are trained for a maximum of 50 epochs using mini-batch SGD with Adam optimizer BIBREF48 . Moreover, at the end of each epoch, we evaluate the validation scores, which are accuracy (Stanfor SNLI dataset) and F1 (RCV1 dataset), and save the weights with the best values. Aiming to seeped up training, we implement early stopping with patience set to 8 epochs. That is, if the validation scores do not improve for more than 10 epochs we halt the training. Finally, we use Glove pre-trained word embeddings BIBREF37 as fixed features.
tbl:tlevaluation compares our test scores with state-of-the-art (SOTA) results reported in previous works. We can see that our scores for the SNLI task are very close to state-of-the-art.
Regarding the RCV1 dataset, our results consider only the headline content for training, while the refereed works consider both the news headline and message body. The reason for training using only the headlines is that both tasks are learnt with the sole purpose of transferring the sentence encoders to our main volatility prediction task, whose textual input is restricted to headlines.
Training setup
During the training of our hierarchical neural model described in sub:HAN we took special care to guard against overfitting. To this aim, we completely separate 2016 and 2017 as the test set and report our results on this “unseen” set. The remaining data is further split into training (2007 to 2013) and validation (2014 to 2015). The model convergence during training is monitored in the validation set. We monitor the validation score of our model at the end of each epoch and store the network weights if the validation scores improves between two consecutive epochs. Additionally, we use mini-batch SGD with Adam optimizer and early stopping with patience set to eight epochs. The hyperparameter tunning is performed using grid search.
All training is performed using the proposed global model approach described in sub:globalmodel, which learns a model that takes into account the features of all the 40 stocks of our corpus. Using this approach our training set has a total of 97,903 samples. Moreover, during the SGD mini-batch sampling the past $T$ days of price and news history tensors and each stock sample stock indicator are randomly selected from the set of all 40 stocks.
Stocks universe result
In order to evaluate the contributions of each component of our neural model described in sub:HAN and the effect of using textual data to predict the volatility, we report our results using the following baselines:
- News (unimodal price only): This baseline completely ablates (i.e. removes) any architecture related to the news mode, considering only the price encoding and the stock embedding components. Using this ablation we aim to evaluate the influence of news to the volatility prediction problem.
+ News (End-to-end Sentence Encoders) - NRA: This baseline ablates our proposed new relevance attention (NRA) component, and instead, makes use of the same Daily Averaging method in BIBREF26 , BIBREF27 , where all fixed-sized headline representations on a given day are averaged without taking into account the relevance of each news. We evaluate this baseline for both BiLSTM attention (Att) and BiLSTM max-pooling (MP) sentence encoders. Here, our goal is to asses the true contribution of our NRA component in the case SOTA sentence encoders are taken into account.
+ News (End-to-End W-L Att Sentence Encoder) + NRA: The Word-Level Attention (W-L Att) sentence encoder implements an attention mechanism directly on top of word embeddings, and, as such, does not consider the order of words in a sentence. This baseline complements the previous one, i.e. it evaluates the influence of the sentence encoder when our full specification is considered.
+ News (TL Sentence Encoders) + NRA: Makes use of sentence encoders of our two auxiliary TL tasks as fixed features. This baseline aims to address the following questions, namely: What dataset and models are more suitable to transfer to our specific volatility forecasting problem; How End-to-End models, which are trained on top of word embeddings, perform compared to sentence encoders transferred from other tasks.
tbl:comparativeallsectors summarizes the test scores for the ablations discussed above. Our best model is the + News (BiLSTM Att) + NRA, which is trained end-to-end and uses our full architecture. The second best model, i.e. + News (BiLSTM MP) + NRA, ranks slightly lower and only differs form the best model in terms of the sentence encoder. The former sentence encoder uses an attention layer (subsec:bilstminneratt) and the the last a max-pooling layer (subsec:bilstmmaxpool), where both layers are placed on top of the LSTM hidden states of each word.
Importantly, our experiments show that using news and price (multimodal) to predict the volatility improves the scores by 11% (MSE) and 9% (MAE) when compared with the – News (price only unimodal) model that considers only price features as explanatory variables.
When comparing the performance of End-to-End models and the TL auxiliary tasks the following can be observed: The end-to-end models trained with the two SOTA sentence encoders perform better than transferring sentence encoder from both auxiliary tasks. However, our experiments show that the same does not hold for models trained end-to-end relying on the simpler WL-Att sentence encoder, which ignores the order of words in a sentence. In other words, considering the appropriate TL task, it is preferable to transfer a SOTA sentence encoder trained on a larger dataset than learning a less robust sentence encoder in an end-to-end fashion. Moreover, initially, we thought that being the RCV1 a financial domain corpus it would demonstrate a superior performance when compared to the SNLI dataset. Still, the SNLI transfers better than RCV1. We hypothesize that the text categorization task (RCV1 dataset) is not able to capture complex sentence structures at the same level required to perform natural language inference. Particularly to the volatility forecasting problem, our TL results corroborates the same findings in BIBREF17 , where it was shown that SNLI dataset attains the best sentence encoding for a broad range of pure NLP tasks, including, among other, text categorization and sentiment analysis.
Significantly, experimental results in tbl:comparativeallsectors clearly demonstrate that our proposed news relevance attention (NRA) outperforms the News Averaging method proposed in previous studies BIBREF26 , BIBREF27 . Even when evaluating our NRA component in conjunction with the more elementary W-L Att sentence encoder it surpass the results of sophisticated sentence encoder using a News Averaging approach. In other words, our results strongly points to the advantage of discriminating noisy from impacting news and the effectiveness of learning to attend the most relevant news.
Having analyzed our best model, we now turn to its comparative performance with respect to the widely regarded GARCH(1,1) model described in sec:GARCH.
We asses our model performance relative to GARCH(1,1) using standard loss metrics (MSE and MAE) and the regression-based accuracy specified in eq:regressionloss and measured in terms of the coefficient of determination $R^2$ . In addition, we evaluate our model across two different volatility proxies: Garman-Klass ( $\widehat{\sigma _{GK}}$ ) (eq:volgk) and Parkinson ( $\widehat{\sigma _{PK}}$ ) (eq:volpk). We note that, as reviewed in sub:evalution, these two volatility proxies are statically efficient and proper estimators of the next day volatility.
tbl:garchallsectors reports the comparative performance among our best Price + News model (+ News BiLSTM (MP) + NRA), our Price only (unimodal) model and GARCH(1,1). The results clearly demonstrate the superiority of our model, being more accurate than GRACH for both volatility proxies. We note that evaluating the GARCH(1,1) model relying on standard MSE and MAE error metrics should be taken with a grain of salt. BIBREF35 provides the background theory and arguments supporting $R^2$ as the metric of choice to evaluate the predictive power of a volatility model. In any case, the outperformance or our model with respect to GARCH(1,1) permeates all three metrics, name $R^2$ , $MSE$ and $MAE$ .
Sector-level results
Company sectors are expected to have different risk levels, in the sense that each sector is driven by different types of news and economic cycles. Moreover, by performing a sector-level analysis we were initially interested in understanding if the outperformance of our model with respect to GARCH(1,1) was the result of a learning bias to a given sector or if, as turned out to be the case, the superior performance of our model spreads across a diversified portfolio of sectors.
In order to evaluate the performance per sector, we first separate the constituents stocks for each sector in tbl:stockuniverse. Then, we calculate the same metrics discussed in the previous section for each sector individually.
tbl:garcheachsector reports our experimental results segregated by sector. We observe that the GRACH model accuracy, measured using the $R^2$ score, has a high degree of variability among sectors. For example, the accuracy ranges from 0.15 to 0.44 for the HealthCare and Energy sector, respectively. This high degree of variability is in agreement with previous results reported in BIBREF16 , but in the context of long-term (quarterly) volatility predictions. Although the GARCH(1,1) accuracy is sector-dependent, without any exception, our model using price and news as input clearly outperforms GRACH sector-wise. This fact allow us to draw the following conclusions:
Our model outperformance is persistent across sectors, i.e. the characteristics of the results reported in tbl:garchallsectors permeates all sectors, rather than being composed of a mix of outperforming and underperforming sector contributions. This fact provides a strong evidence that our model is more accurate than GARCH(1,1).
The proposed Global model approach discussed in sub:globalmodel is able to generalize well, i.e. the patterns learnt are not biased to a given sector or stock.
One of the limitations of our work is to rely on proxies for the volatility estimation. Although these proxies are handy if only open, high, low and close daily price data is available, having high frequency price data we could estimate the daily volatility using the sum of squared intraday returns to measure the true daily latent volatility. For example, in evaluating the performance for the one-day-ahead GARCH(1,1) Yen/Dollar exchange rate BIBREF35 reports $R^2$ values of 0.237 and 0.392 using hourly and five minutes sampled intraday returns, respectively. However, we believe that utilizing intraday data would further improve our model performance.
Since our experimental results demonstrate the key aspect of the news relevance attention to model architecture we observe that intraday data would arguably ameliorate the learning process. Having intraday data would allow us to pair each individual news release with the instantaneous market price reaction. Using daily data we are losing part of this information by only measuring the aggregate effect of all news to the one-day-ahead prediction.
Conclusion
We study the joint effect of stock news and prices on the daily volatility forecasting problem. To the best of our knowledge, this work is one of the first studies aiming to predict short-term (daily) rather than long-term (quarterly or yearly) volatility taking news and price as explanatory variables and using a comprehensive dataset of news headlines at the individual stock level. Our hierarchical end-to-end model benefits from state-of-the-art approaches to encode text information and to deal with two main challenges in correlating news with market reaction: news relevance and novelty. That is, to address the problem of how to attend the most important news based purely on its content (news relevance attention) and to take into account the temporal information of past news (temporal context). Additionally, we propose a multi-stock mini-batch + stock embedding method suitable to model commonality among stocks.
The experimental results show that our multimodal approach outperforms the GARCH(1,1) volatility model, which is the most prevalent econometric model for daily volatility predictions. The outperformance being sector-wise and demonstrates the effectiveness of combining price and news for short-term volatility forecasting. The fact that we outperform GARCH(1,1) for all analyzed sectors confirms the robustness of our proposed architecture and evidences that our global model approach generalizes well.
We ablated (i.e. removed) different components of our neural architecture to assess its most relevant parts. To this aim, we replaced our proposed news relevance attention layer, which aims to attend the most important news on a given day, with a simpler architecture proposed in the literature, which averages the daily news. We found that our attention layer improves the results. Additionally, we ablated all the architecture related to the news mode and found that news enhances the forecasting accuracy.
Finally, we evaluated different sentence encoders, including those transfered from other NLP tasks, and concluded that they achieve better performance as compared to a plain Word-level attention sentence encoder trained end-to-end. However, they do not beat state-of-the-art sentence encoders trained end-to-end.
In order to contribute to the literature of Universal Sentence Encoders, we evaluated the performance of transferring sentence encoders from two different tasks to the volatility prediction problem. We showed that models trained on the Natural Language Inference (NLI) task are more suitable to forecasting problems than a financial domain dataset (Reuters RCV1). By analyzing different architectures, we showed that a BiLSTM with max-pooling for the SNLI dataset provides the best sentence encoder.
In the future, we plan to make use of intraday prices to better assess the predictive power of our proposed models. Additionally, we would further extend our analysis to other stock market sectors.
| [
"Energy with accuracy of 0.538",
"Energy"
] | 10,349 | qasper | en | null | c47ca982b6c4681c4741d4708801fa79a3e1cab17d0a2c4a | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Which stock market sector achieved the best performance?
Answer: | [
"Energy with accuracy of 0.538",
"Energy"
] | qasper | 128 | 16 | |
what NMT models did they compare with? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Ancient Chinese is the writing language in ancient China. It is a treasure of Chinese culture which brings together the wisdom and ideas of the Chinese nation and chronicles the ancient cultural heritage of China. Learning ancient Chinese not only helps people to understand and inherit the wisdom of the ancients, but also promotes people to absorb and develop Chinese culture.
However, it is difficult for modern people to read ancient Chinese. Firstly, compared with modern Chinese, ancient Chinese is more concise and shorter. The grammatical order of modern Chinese is also quite different from that of ancient Chinese. Secondly, most modern Chinese words are double syllables, while the most of the ancient Chinese words are monosyllabic. Thirdly, there is more than one polysemous phenomenon in ancient Chinese. In addition, manual translation has a high cost. Therefore, it is meaningful and useful to study the automatic translation from ancient Chinese to modern Chinese. Through ancient-modern Chinese translation, the wisdom, talent and accumulated experience of the predecessors can be passed on to more people.
Neural machine translation (NMT) BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 has achieved remarkable performance on many bilingual translation tasks. It is an end-to-end learning approach for machine translation, with the potential to show great advantages over the statistic machine translation (SMT) systems. However, NMT approach has not been widely applied to the ancient-modern Chinese translation task. One of the main reasons is the limited high-quality parallel data resource.
The most popular method of acquiring translation examples is bilingual text alignment BIBREF5 . This kind of method can be classified into two types: lexical-based and statistical-based. The lexical-based approaches BIBREF6 , BIBREF7 focus on lexical information, which utilize the bilingual dictionary BIBREF8 , BIBREF9 or lexical features. Meanwhile, the statistical-based approaches BIBREF10 , BIBREF11 rely on statistical information, such as sentence length ratio in two languages and align mode probability.
However, these methods are designed for other bilingual language pairs that are written in different language characters (e.g. English-French, Chinese-Japanese). The ancient-modern Chinese has some characteristics that are quite different from other language pairs. For example, ancient and modern Chinese are both written in Chinese characters, but ancient Chinese is highly concise and its syntactical structure is different from modern Chinese. The traditional methods do not take these characteristics into account. In this paper, we propose an effective ancient-modern Chinese text alignment method at the level of clause based on the characteristics of these two languages. The proposed method combines both lexical-based information and statistical-based information, which achieves 94.2 F1-score on Test set. Recently, a simple longest common subsequence based approach for ancient-modern Chinese sentence alignment is proposed in BIBREF12 . Our experiments showed that our proposed alignment approach performs much better than their method.
We apply the proposed method to create a large translation parallel corpus which contains INLINEFORM0 1.24M bilingual sentence pairs. To our best knowledge, this is the first large high-quality ancient-modern Chinese dataset. Furthermore, we test SMT models and various NMT models on the created dataset and provide a strong baseline for this task.
Overview
There are four steps to build the ancient-modern Chinese translation dataset: (i) The parallel corpus crawling and cleaning. (ii) The paragraph alignment. (iii) The clause alignment based on aligned paragraphs. (iv) Augmenting data by merging aligned adjacent clauses. The most critical step is the third step.
Clause Alignment
In the clause alignment step, we combine both statistical-based and lexical-based information to measure the score for each possible clause alignment between ancient and modern Chinese strings. The dynamic programming is employed to further find overall optimal alignment paragraph by paragraph. According to the characteristics of the ancient and modern Chinese languages, we consider the following factors to measure the alignment score INLINEFORM0 between a bilingual clause pair:
Lexical Matching. The lexical matching score is used to calculate the matching coverage of the ancient clause INLINEFORM0 . It contains two parts: exact matching and dictionary matching. An ancient Chinese character usually corresponds to one or more modern Chinese words. In the first part, we carry out Chinese Word segmentation to the modern Chinese clause INLINEFORM1 . Then we match the ancient characters and modern words in the order from left to right. In further matching, the words that have been matched will be deleted from the original clauses.
However, some ancient characters do not appear in its corresponding modern Chinese words. An ancient Chinese dictionary is employed to address this issue. We preprocess the ancient Chinese dictionary and remove the stop words. In this dictionary matching step, we retrieve the dictionary definition of each unmatched ancient character and use it to match the remaining modern Chinese words. To reduce the impact of universal word matching, we use Inverse Document Frequency (IDF) to weight the matching words. The lexical matching score is calculated as: DISPLAYFORM0
The above equation is used to calculate the matching coverage of the ancient clause INLINEFORM0 . The first term of equation ( EQREF8 ) represents exact matching score. INLINEFORM1 denotes the length of INLINEFORM2 , INLINEFORM3 denotes each ancient character in INLINEFORM4 , and the indicator function INLINEFORM5 indicates whether the character INLINEFORM6 can match the words in the clause INLINEFORM7 . The second term is dictionary matching score. Here INLINEFORM8 and INLINEFORM9 represent the remaining unmatched strings of INLINEFORM10 and INLINEFORM11 , respectively. INLINEFORM12 denotes the INLINEFORM13 -th character in the dictionary definition of the INLINEFORM14 and its IDF score is denoted as INLINEFORM15 . The INLINEFORM16 is a predefined parameter which is used to normalize the IDF score. We tuned the value of this parameter on the Dev set.
Statistical Information. Similar to BIBREF11 and BIBREF6 , the statistical information contains alignment mode and length information. There are many alignment modes between ancient and modern Chinese languages. If one ancient Chinese clause aligns two adjacent modern Chinese clauses, we call this alignment as 1-2 alignment mode. We show some examples of different alignment modes in Figure FIGREF9 . In this paper, we only consider 1-0, 0-1, 1-1, 1-2, 2-1 and 2-2 alignment modes which account for INLINEFORM0 of the Dev set. We estimate the probability Pr INLINEFORM1 n-m INLINEFORM2 of each alignment mode n-m on the Dev set. To utilize length information, we make an investigation on length correlation between these two languages. Based on the assumption of BIBREF11 that each character in one language gives rise to a random number of characters in the other language and those random variables INLINEFORM3 are independent and identically distributed with a normal distribution, we estimate the mean INLINEFORM4 and standard deviation INLINEFORM5 from the paragraph aligned parallel corpus. Given a clause pair INLINEFORM6 , the statistical information score can be calculated by: DISPLAYFORM0
where INLINEFORM0 denotes the normal distribution probability density function.
Edit Distance. Because ancient and modern Chinese are both written in Chinese characters, we also consider using the edit distance. It is a way of quantifying the dissimilarity between two strings by counting the minimum number of operations (insertion, deletion, and substitution) required to transform one string into the other. Here we define the edit distance score as: DISPLAYFORM0
Dynamic Programming. The overall alignment score for each possible clause alignment is as follows: DISPLAYFORM0
Here INLINEFORM0 and INLINEFORM1 are pre-defined interpolation factors. We use dynamic programming to find the overall optimal alignment paragraph by paragraph. Let INLINEFORM2 be total alignment scores of aligning the first to INLINEFORM3 -th ancient Chinese clauses with the first to to INLINEFORM4 -th modern Chinese clauses, and the recurrence then can be described as follows: DISPLAYFORM0
Where INLINEFORM0 denotes concatenate clause INLINEFORM1 to clause INLINEFORM2 . As we discussed above, here we only consider 1-0, 0-1, 1-1, 1-2, 2-1 and 2-2 alignment modes.
Ancient-Modern Chinese Dataset
Data Collection. To build the large ancient-modern Chinese dataset, we collected 1.7K bilingual ancient-modern Chinese articles from the internet. More specifically, a large part of the ancient Chinese data we used come from ancient Chinese history records in several dynasties (about 1000BC-200BC) and articles written by celebrities of that era. They used plain and accurate words to express what happened at that time, and thus ensure the generality of the translated materials.
Paragraph Alignment. To further ensure the quality of the new dataset, the work of paragraph alignment is manually completed. After data cleaning and manual paragraph alignment, we obtained 35K aligned bilingual paragraphs.
Clause Alignment. We applied our clause alignment algorithm on the 35K aligned bilingual paragraphs and obtained 517K aligned bilingual clauses. The reason we use clause alignment algorithm instead of sentence alignment is because we can construct more aligned sentences more flexibly and conveniently. To be specific, we can get multiple additional sentence level bilingual pairs by “data augmentation”.
Data Augmentation. We augmented the data in the following way: Given an aligned clause pair, we merged its adjacent clause pairs as a new sample pair. For example, suppose we have three adjacent clause level bilingual pairs: ( INLINEFORM0 , INLINEFORM1 ), ( INLINEFORM2 , INLINEFORM3 ), and ( INLINEFORM4 , INLINEFORM5 ). We can get some additional sentence level bilingual pairs, such as: ( INLINEFORM6 , INLINEFORM7 ) and ( INLINEFORM8 , INLINEFORM9 ). Here INLINEFORM10 , INLINEFORM11 , and INLINEFORM12 are adjacent clauses in the original paragraph, and INLINEFORM13 denotes concatenate clause INLINEFORM14 to clause INLINEFORM15 . The advantage of using this data augmentation method is that compared with only using ( INLINEFORM16 , INLINEFORM17 ) as the training data, we can also use ( INLINEFORM18 , INLINEFORM19 ) and ( INLINEFORM20 , INLINEFORM21 ) as the training data, which can provide richer supervision information for the model and make the model learn the align information between the source language and the target language better. After the data augmentation, we filtered the sentences which are longer than 50 or contain more than four clause pairs.
Dataset Creation. Finally, we split the dataset into three sets: training (Train), development (Dev) and testing (Test). Note that the unaugmented dataset contains 517K aligned bilingual clause pairs from 35K aligned bilingual paragraphs. To keep all the sentences in different sets come from different articles, we split the 35K aligned bilingual paragraphs into Train, Dev and Test sets following these ratios respectively: 80%, 10%, 10%. Before data augmentation, the unaugmented Train set contains INLINEFORM0 aligned bilingual clause pairs from 28K aligned bilingual paragraphs. Then we augmented the Train, Dev and Test sets respectively. Note that the augmented Train, Dev and Test sets also contain the unaugmented data. The statistical information of the three data sets is shown in Table TABREF17 . We show some examples of data in Figure FIGREF14 .
RNN-based NMT model
We first briefly introduce the RNN based Neural Machine Translation (RNN-based NMT) model. The RNN-based NMT with attention mechanism BIBREF0 has achieved remarkable performance on many translation tasks. It consists of encoder and decoder part.
We firstly introduce the encoder part. The input word sequence of source language are individually mapped into a INLINEFORM0 -dimensional vector space INLINEFORM1 . Then a bi-directional RNN BIBREF15 with GRU BIBREF16 or LSTM BIBREF17 cell converts these vectors into a sequences of hidden states INLINEFORM2 .
For the decoder part, another RNN is used to generate target sequence INLINEFORM0 . The attention mechanism BIBREF0 , BIBREF18 is employed to allow the decoder to refer back to the hidden state sequence and focus on a particular segment. The INLINEFORM1 -th hidden state INLINEFORM2 of decoder part is calculated as: DISPLAYFORM0
Here g INLINEFORM0 is a linear combination of attended context vector c INLINEFORM1 and INLINEFORM2 is the word embedding of (i-1)-th target word: DISPLAYFORM0
The attended context vector c INLINEFORM0 is computed as a weighted sum of the hidden states of the encoder: DISPLAYFORM0
The probability distribution vector of the next word INLINEFORM0 is generated according to the following: DISPLAYFORM0
We take this model as the basic RNN-based NMT model in the following experiments.
Transformer-NMT
Recently, the Transformer model BIBREF4 has made remarkable progress in machine translation. This model contains a multi-head self-attention encoder and a multi-head self-attention decoder.
As proposed by BIBREF4 , an attention function maps a query and a set of key-value pairs to an output, where the queries INLINEFORM0 , keys INLINEFORM1 , and values INLINEFORM2 are all vectors. The input consists of queries and keys of dimension INLINEFORM3 , and values of dimension INLINEFORM4 . The attention function is given by: DISPLAYFORM0
Multi-head attention mechanism projects queries, keys and values to INLINEFORM0 different representation subspaces and calculates corresponding attention. The attention function outputs are concatenated and projected again before giving the final output. Multi-head attention allows the model to attend to multiple features at different positions.
The encoder is composed of a stack of INLINEFORM0 identical layers. Each layer has two sub-layers: multi-head self-attention mechanism and position-wise fully connected feed-forward network. Similarly, the decoder is also composed of a stack of INLINEFORM1 identical layers. In addition to the two sub-layers in each encoder layer, the decoder contains a third sub-layer which performs multi-head attention over the output of the encoder stack (see more details in BIBREF4 ).
Experiments
Our experiments revolve around the following questions: Q1: As we consider three factors for clause alignment, do all these factors help? How does our method compare with previous methods? Q2: How does the NMT and SMT models perform on this new dataset we build?
Clause Alignment Results (Q1)
In order to evaluate our clause alignment algorithm, we manually aligned bilingual clauses from 37 bilingual ancient-modern Chinese articles, and finally got 4K aligned bilingual clauses as the Test set and 2K clauses as the Dev set.
Metrics. We used F1-score and precision score as the evaluation metrics. Suppose that we get INLINEFORM0 bilingual clause pairs after running the algorithm on the Test set, and there are INLINEFORM1 bilingual clause pairs of these INLINEFORM2 pairs are in the ground truth of the Test set, the precision score is defined as INLINEFORM3 (the algorithm gives INLINEFORM4 outputs, INLINEFORM5 of which are correct). And suppose that the ground truth of the Test set contains INLINEFORM6 bilingual clause pairs, the recall score is INLINEFORM7 (there are INLINEFORM8 ground truth samples, INLINEFORM9 of which are output by the algorithm), then the F1-score is INLINEFORM10 .
Baselines. Since the related work BIBREF10 , BIBREF11 can be seen as the ablation cases of our method (only statistical score INLINEFORM0 with dynamic programming), we compared the full proposed method with its variants on the Test set for ablation study. In addition, we also compared our method with the longest common subsequence (LCS) based approach proposed by BIBREF12 . To the best of our knowledge, BIBREF12 is the latest related work which are designed for Ancient-Modern Chinese alignment.
Hyper-parameters. For the proposed method, we estimated INLINEFORM0 and INLINEFORM1 on all aligned paragraphs. The probability Pr INLINEFORM2 n-m INLINEFORM3 of each alignment mode n-m was estimated on the Dev set. For the hyper-parameters INLINEFORM4 , INLINEFORM5 and INLINEFORM6 , the grid search was applied to tune them on the Dev set. In order to show the effect of hyper-parameters INLINEFORM7 , INLINEFORM8 , and INLINEFORM9 , we reported the results of various hyper-parameters on the Dev set in Table TABREF26 . Based on the results of grid search on the Dev set, we set INLINEFORM10 , INLINEFORM11 , and INLINEFORM12 in the following experiment. The Jieba Chinese text segmentation is employed for modern Chinese word segmentation.
Results. The results on the Test set are shown in Table TABREF28 , the abbreviation w/o means removing a particular part from the setting. From the results, we can see that the lexical matching score is the most important among these three factors, and statistical information score is more important than edit distance score. Moreover, the dictionary term in lexical matching score significantly improves the performance. From these results, we obtain the best setting that involves all these three factors. We used this setting for dataset creation. Furthermore, the proposed method performs much better than LCS BIBREF12 .
Translation Results (Q2)
In this experiment, we analyzed and compared the performance of the SMT and various NMT models on our built dataset. To verify the effectiveness of our data augmented method. We trained the NMT and SMT models on both unaugmented dataset (including 0.46M training pairs) and augmented dataset, and test all the models on the same Test set which is augmented. The models to be tested and their configurations are as follows:
SMT. The state-of-art Moses toolkit BIBREF19 was used to train SMT model. We used KenLM BIBREF20 to train a 5-gram language model, and the GIZA++ toolkit to align the data.
RNN-based NMT. The basic RNN-based NMT model is based on BIBREF0 which is introduced above. Both the encoder and decoder used 2-layer RNN with 1024 LSTM cells, and the encoder is a bi-directional RNN. The batch size, threshold of element-wise gradient clipping and initial learning rate of Adam optimizer BIBREF21 were set to 128, 5.0 and 0.001. When trained the model on augmented dataset, we used 4-layer RNN. Several techniques were investigated to train the model, including layer-normalization BIBREF22 , RNN-dropout BIBREF23 , and learning rate decay BIBREF1 . The hyper-parameters were chosen empirically and adjusted in the Dev set. Furthermore, we tested the basic NMT model with several techniques, such as target language reversal BIBREF24 (reversing the order of the words in all target sentences, but not source sentences), residual connection BIBREF25 and pre-trained word2vec BIBREF26 . For word embedding pre-training, we collected an external ancient corpus which contains INLINEFORM0 134M tokens.
Transformer-NMT. We also trained the Transformer model BIBREF4 which is a strong baseline of NMT on both augmented and unaugmented parallel corpus. The training configuration of the Transformer model is shown in Table TABREF32 . The hyper-parameters are set based on the settings in the paper BIBREF4 and the sizes of our training sets.
For the evaluation, we used the average of 1 to 4 gram BLEUs multiplied by a brevity penalty BIBREF27 which computed by multi-bleu.perl in Moses as metrics. The results are reported in Table TABREF34 . For RNN-based NMT, we can see that target language reversal, residual connection, and word2vec can further improve the performance of the basic RNN-based NMT model. However, we find that word2vec and reversal tricks seem no obvious improvement when trained the RNN-based NMT and Transformer models on augmented parallel corpus. For SMT, it performs better than NMT models when they were trained on the unaugmented dataset. Nevertheless, when trained on the augmented dataset, both the RNN-based NMT model and Transformer based NMT model outperform the SMT model. In addition, as with other translation tasks BIBREF4 , the Transformer also performs better than RNN-based NMT.
Because the Test set contains both augmented and unaugmented data, it is not surprising that the RNN-based NMT model and Transformer based NMT model trained on unaugmented data would perform poorly. In order to further verify the effect of data augmentation, we report the test results of the models on only unaugmented test data (including 48K test pairs) in Table TABREF35 . From the results, it can be seen that the data augmentation can still improve the models.
Analysis
The generated samples of various models are shown in Figure FIGREF36 . Besides BLEU scores, we analyze these examples from a human perspective and draw some conclusions. At the same time, we design different metrics and evaluate on the whole Test set to support our conclusions as follows:
On the one hand, we further compare the translation results from the perspective of people. We find that although the original meaning can be basically translated by SMT, its translation results are less smooth when compared with the other two NMT models (RNN-based NMT and Transformer). For example, the translations of SMT are usually lack of auxiliary words, conjunctions and function words, which is not consistent with human translation habits. To further confirm this conclusion, the average length of the translation results of the three models are measured (RNN-based NMT:17.12, SMT:15.50, Transformer:16.78, Reference:16.47). We can see that the average length of the SMT outputs is shortest, and the length gaps between the SMT outputs and the references are largest. Meanwhile, the average length of the sentences translated by Transformer is closest to the average length of references. These results indirectly verify our point of view, and show that the NMT models perform better than SMT in this task.
On the other hand, there still exists some problems to be solved. We observe that translating proper nouns and personal pronouns (such as names, place names and ancient-specific appellations) is very difficult for all of these models. For instance, the ancient Chinese appellation `Zhen' should be translated into `Wo' in modern Chinese. Unfortunately, we calculate the accurate rate of some special words (such as `Zhen',`Chen' and `Gua'), and find that this rate is very low (the accurate rate of translating `Zhen' are: RNN-based NMT:0.14, SMT:0.16, Transformer:0.05). We will focus on this issue in the future.
Conclusion and Future Work
We propose an effective ancient-modern Chinese clause alignment method which achieves 94.2 F1-score on Test set. Based on it, we build a large scale parallel corpus which contains INLINEFORM0 1.24M bilingual sentence pairs. To our best knowledge, this is the first large high-quality ancient-modern Chinese dataset. In addition, we test the performance of the SMT and various NMT models on our built dataset and provide a strong NMT baseline for this task which achieves 27.16 BLEU score (4-gram). We further analyze the performance of the SMT and various NMT models and summarize some specific problems that machine translation models will encounter when translating ancient Chinese.
For the future work, firstly, we are going to expand the dataset using the proposed method continually. Secondly, we will focus on solving the problem of proper noun translation and improve the translation system according to the features of ancient Chinese translation. Finally, we plan to introduce some techniques of statistical translation into neural machine translation to improve the performance.
This work is supported by National Natural Science Fund for Distinguished Young Scholar (Grant No. 61625204) and partially supported by the State Key Program of National Science Foundation of China (Grant Nos. 61836006 and 61432014).
| [
"RNN-based NMT model, Transformer-NMT"
] | 3,708 | qasper | en | null | 1614ea21f8d039debfc655015f8133d0ef220837804ddf62 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: what NMT models did they compare with?
Answer: | [
"RNN-based NMT model, Transformer-NMT"
] | qasper | 128 | 17 | |
What are the three regularization terms? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
We posses a wealth of prior knowledge about many natural language processing tasks. For example, in text categorization, we know that words such as NBA, player, and basketball are strong indicators of the sports category BIBREF0 , and words like terrible, boring, and messing indicate a negative polarity while words like perfect, exciting, and moving suggest a positive polarity in sentiment classification.
A key problem arisen here, is how to leverage such knowledge to guide the learning process, an interesting problem for both NLP and machine learning communities. Previous studies addressing the problem fall into several lines. First, to leverage prior knowledge to label data BIBREF1 , BIBREF2 . Second, to encode prior knowledge with a prior on parameters, which can be commonly seen in many Bayesian approaches BIBREF3 , BIBREF4 . Third, to formalise prior knowledge with additional variables and dependencies BIBREF5 . Last, to use prior knowledge to control the distributions over latent output variables BIBREF6 , BIBREF7 , BIBREF8 , which makes the output variables easily interpretable.
However, a crucial problem, which has rarely been addressed, is the bias in the prior knowledge that we supply to the learning model. Would the model be robust or sensitive to the prior knowledge? Or, which kind of knowledge is appropriate for the task? Let's see an example: we may be a baseball fan but unfamiliar with hockey so that we can provide a few number of feature words of baseball, but much less of hockey for a baseball-hockey classification task. Such prior knowledge may mislead the model with heavy bias to baseball. If the model cannot handle this situation appropriately, the performance may be undesirable.
In this paper, we investigate into the problem in the framework of Generalized Expectation Criteria BIBREF7 . The study aims to reveal the factors of reducing the sensibility of the prior knowledge and therefore to make the model more robust and practical. To this end, we introduce auxiliary regularization terms in which our prior knowledge is formalized as distribution over output variables. Recall the example just mentioned, though we do not have enough knowledge to provide features for class hockey, it is easy for us to provide some neutral words, namely words that are not strong indicators of any class, like player here. As one of the factors revealed in this paper, supplying neutral feature words can boost the performance remarkably, making the model more robust.
More attractively, we do not need manual annotation to label these neutral feature words in our proposed approach.
More specifically, we explore three regularization terms to address the problem: (1) a regularization term associated with neutral features; (2) the maximum entropy of class distribution regularization term; and (3) the KL divergence between reference and predicted class distribution. For the first manner, we simply use the most common features as neutral features and assume the neutral features are distributed uniformly over class labels. For the second and third one, we assume we have some knowledge about the class distribution which will be detailed soon later.
To summarize, the main contributions of this work are as follows:
The rest of the paper is structured as follows: In Section 2, we briefly describe the generalized expectation criteria and present the proposed regularization terms. In Section 3, we conduct extensive experiments to justify the proposed methods. We survey related work in Section 4, and summarize our work in Section 5.
Method
We address the robustness problem on top of GE-FL BIBREF0 , a GE method which leverages labeled features as prior knowledge. A labeled feature is a strong indicator of a specific class and is manually provided to the classifier. For example, words like amazing, exciting can be labeled features for class positive in sentiment classification.
Generalized Expectation Criteria
Generalized expectation (GE) criteria BIBREF7 provides us a natural way to directly constrain the model in the preferred direction. For example, when we know the proportion of each class of the dataset in a classification task, we can guide the model to predict out a pre-specified class distribution.
Formally, in a parameter estimation objective function, a GE term expresses preferences on the value of some constraint functions about the model's expectation. Given a constraint function $G({\rm x}, y)$ , a conditional model distribution $p_\theta (y|\rm x)$ , an empirical distribution $\tilde{p}({\rm x})$ over input samples and a score function $S$ , a GE term can be expressed as follows:
$$S(E_{\tilde{p}({\rm x})}[E_{p_\theta (y|{\rm x})}[G({\rm x}, y)]])$$ (Eq. 4)
Learning from Labeled Features
Druck et al. ge-fl proposed GE-FL to learn from labeled features using generalized expectation criteria. When given a set of labeled features $K$ , the reference distribution over classes of these features is denoted by $\hat{p}(y| x_k), k \in K$ . GE-FL introduces the divergence between this reference distribution and the model predicted distribution $p_\theta (y | x_k)$ , as a term of the objective function:
$$\mathcal {O} = \sum _{k \in K} KL(\hat{p}(y|x_k) || p_\theta (y | x_k)) + \sum _{y,i} \frac{\theta _{yi}^2}{2 \sigma ^2}$$ (Eq. 6)
where $\theta _{yi}$ is the model parameter which indicates the importance of word $i$ to class $y$ . The predicted distribution $p_\theta (y | x_k)$ can be expressed as follows: $ p_\theta (y | x_k) = \frac{1}{C_k} \sum _{\rm x} p_\theta (y|{\rm x})I(x_k) $
in which $I(x_k)$ is 1 if feature $k$ occurs in instance ${\rm x}$ and 0 otherwise, $C_k = \sum _{\rm x} I(x_k)$ is the number of instances with a non-zero value of feature $k$ , and $p_\theta (y|{\rm x})$ takes a softmax form as follows: $ p_\theta (y|{\rm x}) = \frac{1}{Z(\rm x)}\exp (\sum _i \theta _{yi}x_i). $
To solve the optimization problem, L-BFGS can be used for parameter estimation.
In the framework of GE, this term can be obtained by setting the constraint function $G({\rm x}, y) = \frac{1}{C_k} \vec{I} (y)I(x_k)$ , where $\vec{I}(y)$ is an indicator vector with 1 at the index corresponding to label $y$ and 0 elsewhere.
Regularization Terms
GE-FL reduces the heavy load of instance annotation and performs well when we provide prior knowledge with no bias. In our experiments, we observe that comparable numbers of labeled features for each class have to be supplied. But as mentioned before, it is often the case that we are not able to provide enough knowledge for some of the classes. For the baseball-hockey classification task, as shown before, GE-FL will predict most of the instances as baseball. In this section, we will show three terms to make the model more robust.
Neutral features are features that are not informative indicator of any classes, for instance, word player to the baseball-hockey classification task. Such features are usually frequent words across all categories. When we set the preference distribution of the neutral features to be uniform distributed, these neutral features will prevent the model from biasing to the class that has a dominate number of labeled features.
Formally, given a set of neutral features $K^{^{\prime }}$ , the uniform distribution is $\hat{p}_u(y|x_k) = \frac{1}{|C|}, k \in K^{^{\prime }}$ , where $|C|$ is the number of classes. The objective function with the new term becomes
$$\mathcal {O}_{NE} = \mathcal {O} + \sum _{k \in K^{^{\prime }}} KL(\hat{p}_u(y|x_k) || p_\theta (y | x_k)).$$ (Eq. 9)
Note that we do not need manual annotation to provide neutral features. One simple way is to take the most common features as neutral features. Experimental results show that this strategy works successfully.
Another way to prevent the model from drifting from the desired direction is to constrain the predicted class distribution on unlabeled data. When lacking knowledge about the class distribution of the data, one feasible way is to take maximum entropy principle, as below:
$$\mathcal {O}_{ME} = \mathcal {O} + \lambda \sum _{y} p(y) \log p(y)$$ (Eq. 11)
where $p(y)$ is the predicted class distribution, given by $ p(y) = \frac{1}{|X|} \sum _{\rm x} p_\theta (y | \rm x). $ To control the influence of this term on the overall objective function, we can tune $\lambda $ according to the difference in the number of labeled features of each class. In this paper, we simply set $\lambda $ to be proportional to the total number of labeled features, say $\lambda = \beta |K|$ .
This maximum entropy term can be derived by setting the constraint function to $G({\rm x}, y) = \vec{I}(y)$ . Therefore, $E_{p_\theta (y|{\rm x})}[G({\rm x}, y)]$ is just the model distribution $p_\theta (y|{\rm x})$ and its expectation with the empirical distribution $\tilde{p}(\rm x)$ is simply the average over input samples, namely $p(y)$ . When $S$ takes the maximum entropy form, we can derive the objective function as above.
Sometimes, we have already had much knowledge about the corpus, and can estimate the class distribution roughly without labeling instances. Therefore, we introduce the KL divergence between the predicted and reference class distributions into the objective function.
Given the preference class distribution $\hat{p}(y)$ , we modify the objective function as follows:
$$\mathcal {O}_{KL} &= \mathcal {O} + \lambda KL(\hat{p}(y) || p(y))$$ (Eq. 13)
Similarly, we set $\lambda = \beta |K|$ .
This divergence term can be derived by setting the constraint function to $G({\rm x}, y) = \vec{I}(y)$ and setting the score function to $S(\hat{p}, p) = \sum _i \hat{p}_i \log \frac{\hat{p}_i}{p_i}$ , where $p$ and $\hat{p}$ are distributions. Note that this regularization term involves the reference class distribution which will be discussed later.
Experiments
In this section, we first justify the approach when there exists unbalance in the number of labeled features or in class distribution. Then, to test the influence of $\lambda $ , we conduct some experiments with the method which incorporates the KL divergence of class distribution. Last, we evaluate our approaches in 9 commonly used text classification datasets. We set $\lambda = 5|K|$ by default in all experiments unless there is explicit declaration. The baseline we choose here is GE-FL BIBREF0 , a method based on generalization expectation criteria.
Data Preparation
We evaluate our methods on several commonly used datasets whose themes range from sentiment, web-page, science to medical and healthcare. We use bag-of-words feature and remove stopwords in the preprocess stage. Though we have labels of all documents, we do not use them during the learning process, instead, we use the label of features.
The movie dataset, in which the task is to classify the movie reviews as positive or negtive, is used for testing the proposed approaches with unbalanced labeled features, unbalanced datasets or different $\lambda $ parameters. All unbalanced datasets are constructed based on the movie dataset by randomly removing documents of the positive class. For each experiment, we conduct 10-fold cross validation.
As described in BIBREF0 , there are two ways to obtain labeled features. The first way is to use information gain. We first calculate the mutual information of all features according to the labels of the documents and select the top 20 as labeled features for each class as a feature pool. Note that using information gain requires the document label, but this is only to simulate how we human provide prior knowledge to the model. The second way is to use LDA BIBREF9 to select features. We use the same selection process as BIBREF0 , where they first train a LDA on the dataset, and then select the most probable features of each topic (sorted by $P(w_i|t_j)$ , the probability of word $w_i$ given topic $t_j$ ).
Similar to BIBREF10 , BIBREF0 , we estimate the reference distribution of the labeled features using a heuristic strategy. If there are $|C|$ classes in total, and $n$ classes are associated with a feature $k$ , the probability that feature $k$ is related with any one of the $n$ classes is $\frac{0.9}{n}$ and with any other class is $\frac{0.1}{|C| - n}$ .
Neutral features are the most frequent words after removing stop words, and their reference distributions are uniformly distributed. We use the top 10 frequent words as neutral features in all experiments.
With Unbalanced Labeled Features
In this section, we evaluate our approach when there is unbalanced knowledge on the categories to be classified. The labeled features are obtained through information gain. Two settings are chosen:
(a) We randomly select $t \in [1, 20]$ features from the feature pool for one class, and only one feature for the other. The original balanced movie dataset is used (positive:negative=1:1).
(b) Similar to (a), but the dataset is unbalanced, obtained by randomly removing 75% positive documents (positive:negative=1:4).
As shown in Figure 1 , Maximum entropy principle shows improvement only on the balanced case. An obvious reason is that maximum entropy only favors uniform distribution.
Incorporating Neutral features performs similarly to maximum entropy since we assume that neutral words are uniformly distributed. Its accuracy decreases slowly when the number of labeled features becomes larger ( $t>4$ ) (Figure 1 (a)), suggesting that the model gradually biases to the class with more labeled features, just like GE-FL.
Incorporating the KL divergence of class distribution performs much better than GE-FL on both balanced and unbalanced datasets. This shows that it is effective to control the unbalance in labeled features and in the dataset.
With Balanced Labeled Features
We also compare with the baseline when the labeled features are balanced. Similar to the experiment above, the labeled features are obtained by information gain. Two settings are experimented with:
(a) We randomly select $t \in [1, 20]$ features from the feature pool for each class, and conduct comparisons on the original balanced movie dataset (positive:negtive=1:1).
(b) Similar to (a), but the class distribution is unbalanced, by randomly removing 75% positive documents (positive:negative=1:4).
Results are shown in Figure 2 . When the dataset is balanced (Figure 2 (a)), there is little difference between GE-FL and our methods. The reason is that the proposed regularization terms provide no additional knowledge to the model and there is no bias in the labeled features. On the unbalanced dataset (Figure 2 (b)), incorporating KL divergence is much better than GE-FL since we provide additional knowledge(the true class distribution), but maximum entropy and neutral features are much worse because forcing the model to approach the uniform distribution misleads it.
With Unbalanced Class Distributions
Our methods are also evaluated on datasets with different unbalanced class distributions. We manually construct several movie datasets with class distributions of 1:2, 1:3, 1:4 by randomly removing 50%, 67%, 75% positive documents. The original balanced movie dataset is used as a control group. We test with both balanced and unbalanced labeled features. For the balanced case, we randomly select 10 features from the feature pool for each class, and for the unbalanced case, we select 10 features for one class, and 1 feature for the other. Results are shown in Figure 3 .
Figure 3 (a) shows that when the dataset and the labeled features are both balanced, there is little difference between our methods and GE-FL(also see Figure 2 (a)). But when the class distribution becomes more unbalanced, the difference becomes more remarkable. Performance of neutral features and maximum entropy decrease significantly but incorporating KL divergence increases remarkably. This suggests if we have more accurate knowledge about class distribution, KL divergence can guide the model to the right direction.
Figure 3 (b) shows that when the labeled features are unbalanced, our methods significantly outperforms GE-FL. Incorporating KL divergence is robust enough to control unbalance both in the dataset and in labeled features while the other three methods are not so competitive.
The Influence of λ\lambda
We present the influence of $\lambda $ on the method that incorporates KL divergence in this section. Since we simply set $\lambda = \beta |K|$ , we just tune $\beta $ here. Note that when $\beta = 0$ , the newly introduced regularization term is disappeared, and thus the model is actually GE-FL. Again, we test the method with different $\lambda $ in two settings:
(a) We randomly select $t \in [1, 20]$ features from the feature pool for one class, and only one feature for the other class. The original balanced movie dataset is used (positive:negative=1:1).
(b) Similar to (a), but the dataset is unbalanced, obtained by randomly removing 75% positive documents (positive:negative=1:4).
Results are shown in Figure 4 . As expected, $\lambda $ reflects how strong the regularization is. The model tends to be closer to our preferences with the increasing of $\lambda $ on both cases.
Using LDA Selected Features
We compare our methods with GE-FL on all the 9 datasets in this section. Instead of using features obtained by information gain, we use LDA to select labeled features. Unlike information gain, LDA does not employ any instance labels to find labeled features. In this setting, we can build classification models without any instance annotation, but just with labeled features.
Table 1 shows that our three methods significantly outperform GE-FL. Incorporating neutral features performs better than GE-FL on 7 of the 9 datasets, maximum entropy is better on 8 datasets, and KL divergence better on 7 datasets.
LDA selects out the most predictive features as labeled features without considering the balance among classes. GE-FL does not exert any control on such an issue, so the performance is severely suffered. Our methods introduce auxiliary regularization terms to control such a bias problem and thus promote the model significantly.
Related Work
There have been much work that incorporate prior knowledge into learning, and two related lines are surveyed here. One is to use prior knowledge to label unlabeled instances and then apply a standard learning algorithm. The other is to constrain the model directly with prior knowledge.
Liu et al.text manually labeled features which are highly predictive to unsupervised clustering assignments and use them to label unlabeled data. Chang et al.guiding proposed constraint driven learning. They first used constraints and the learned model to annotate unlabeled instances, and then updated the model with the newly labeled data. Daumé daume2008cross proposed a self training method in which several models are trained on the same dataset, and only unlabeled instances that satisfy the cross task knowledge constraints are used in the self training process.
MaCallum et al.gec proposed generalized expectation(GE) criteria which formalised the knowledge as constraint terms about the expectation of the model into the objective function.Graça et al.pr proposed posterior regularization(PR) framework which projects the model's posterior onto a set of distributions that satisfy the auxiliary constraints. Druck et al.ge-fl explored constraints of labeled features in the framework of GE by forcing the model's predicted feature distribution to approach the reference distribution. Andrzejewski et al.andrzejewski2011framework proposed a framework in which general domain knowledge can be easily incorporated into LDA. Altendorf et al.altendorf2012learning explored monotonicity constraints to improve the accuracy while learning from sparse data. Chen et al.chen2013leveraging tried to learn comprehensible topic models by leveraging multi-domain knowledge.
Mann and McCallum simple,generalized incorporated not only labeled features but also other knowledge like class distribution into the objective function of GE-FL. But they discussed only from the semi-supervised perspective and did not investigate into the robustness problem, unlike what we addressed in this paper.
There are also some active learning methods trying to use prior knowledge. Raghavan et al.feedback proposed to use feedback on instances and features interlacedly, and demonstrated that feedback on features boosts the model much. Druck et al.active proposed an active learning method which solicits labels on features rather than on instances and then used GE-FL to train the model.
Conclusion and Discussions
This paper investigates into the problem of how to leverage prior knowledge robustly in learning models. We propose three regularization terms on top of generalized expectation criteria. As demonstrated by the experimental results, the performance can be considerably improved when taking into account these factors. Comparative results show that our proposed methods is more effective and works more robustly against baselines. To the best of our knowledge, this is the first work to address the robustness problem of leveraging knowledge, and may inspire other research.
We then present more detailed discussions about the three regularization methods. Incorporating neutral features is the simplest way of regularization, which doesn't require any modification of GE-FL but just finding out some common features. But as Figure 1 (a) shows, only using neutral features are not strong enough to handle extremely unbalanced labeled features.
The maximum entropy regularization term shows the strong ability of controlling unbalance.
This method doesn't need any extra knowledge, and is thus suitable when we know nothing about the corpus. But this method assumes that the categories are uniformly distributed, which may not be the case in practice, and it will have a degraded performance if the assumption is violated (see Figure 1 (b), Figure 2 (b), Figure 3 (a)).
The KL divergence performs much better on unbalanced corpora than other methods. The reason is that KL divergence utilizes the reference class distribution and doesn't make any assumptions. The fact suggests that additional knowledge does benefit the model.
However, the KL divergence term requires providing the true class distribution. Sometimes, we may have the exact knowledge about the true distribution, but sometimes we may not. Fortunately, the model is insensitive to the true distribution and therefore a rough estimation of the true distribution is sufficient. In our experiments, when the true class distribution is 1:2, where the reference class distribution is set to 1:1.5/1:2/1:2.5, the accuracy is 0.755/0.756/0.760 respectively. This provides us the possibility to perform simple computing on the corpus to obtain the distribution in reality. Or, we can set the distribution roughly with domain expertise.
| [
"a regularization term associated with neutral features, the maximum entropy of class distribution regularization term, the KL divergence between reference and predicted class distribution",
"a regularization term associated with neutral features, the maximum entropy of class distribution, KL divergence between ... | 3,604 | qasper | en | null | 159a474e0a7155a2d0b984cbb5215cdc853d721605b90154 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What are the three regularization terms?
Answer: | [
"a regularization term associated with neutral features, the maximum entropy of class distribution regularization term, the KL divergence between reference and predicted class distribution",
"a regularization term associated with neutral features, the maximum entropy of class distribution, KL divergence between ... | qasper | 128 | 18 | |
What are the baselines? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
This work is licenced under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/ Deep neural networks have been widely used in text classification and have achieved promising results BIBREF0 , BIBREF1 , BIBREF2 . Most focus on content information and use models such as convolutional neural networks (CNN) BIBREF3 or recursive neural networks BIBREF4 . However, for user-generated posts on social media like Facebook or Twitter, there is more information that should not be ignored. On social media platforms, a user can act either as the author of a post or as a reader who expresses his or her comments about the post.
In this paper, we classify posts taking into account post authorship, likes, topics, and comments. In particular, users and their “likes” hold strong potential for text mining. For example, given a set of posts that are related to a specific topic, a user's likes and dislikes provide clues for stance labeling. From a user point of view, users with positive attitudes toward the issue leave positive comments on the posts with praise or even just the post's content; from a post point of view, positive posts attract users who hold positive stances. We also investigate the influence of topics: different topics are associated with different stance labeling tendencies and word usage. For example we discuss women's rights and unwanted babies on the topic of abortion, but we criticize medicine usage or crime when on the topic of marijuana BIBREF5 . Even for posts on a specific topic like nuclear power, a variety of arguments are raised: green energy, radiation, air pollution, and so on. As for comments, we treat them as additional text information. The arguments in the comments and the commenters (the users who leave the comments) provide hints on the post's content and further facilitate stance classification.
In this paper, we propose the user-topic-comment neural network (UTCNN), a deep learning model that utilizes user, topic, and comment information. We attempt to learn user and topic representations which encode user interactions and topic influences to further enhance text classification, and we also incorporate comment information. We evaluate this model on a post stance classification task on forum-style social media platforms. The contributions of this paper are as follows: 1. We propose UTCNN, a neural network for text in modern social media channels as well as legacy social media, forums, and message boards — anywhere that reveals users, their tastes, as well as their replies to posts. 2. When classifying social media post stances, we leverage users, including authors and likers. User embeddings can be generated even for users who have never posted anything. 3. We incorporate a topic model to automatically assign topics to each post in a single topic dataset. 4. We show that overall, the proposed method achieves the highest performance in all instances, and that all of the information extracted, whether users, topics, or comments, still has its contributions.
Extra-Linguistic Features for Stance Classification
In this paper we aim to use text as well as other features to see how they complement each other in a deep learning model. In the stance classification domain, previous work has showed that text features are limited, suggesting that adding extra-linguistic constraints could improve performance BIBREF6 , BIBREF7 , BIBREF8 . For example, Hasan and Ng as well as Thomas et al. require that posts written by the same author have the same stance BIBREF9 , BIBREF10 . The addition of this constraint yields accuracy improvements of 1–7% for some models and datasets. Hasan and Ng later added user-interaction constraints and ideology constraints BIBREF7 : the former models the relationship among posts in a sequence of replies and the latter models inter-topic relationships, e.g., users who oppose abortion could be conservative and thus are likely to oppose gay rights.
For work focusing on online forum text, since posts are linked through user replies, sequential labeling methods have been used to model relationships between posts. For example, Hasan and Ng use hidden Markov models (HMMs) to model dependent relationships to the preceding post BIBREF9 ; Burfoot et al. use iterative classification to repeatedly generate new estimates based on the current state of knowledge BIBREF11 ; Sridhar et al. use probabilistic soft logic (PSL) to model reply links via collaborative filtering BIBREF12 . In the Facebook dataset we study, we use comments instead of reply links. However, as the ultimate goal in this paper is predicting not comment stance but post stance, we treat comments as extra information for use in predicting post stance.
Deep Learning on Extra-Linguistic Features
In recent years neural network models have been applied to document sentiment classification BIBREF13 , BIBREF4 , BIBREF14 , BIBREF15 , BIBREF2 . Text features can be used in deep networks to capture text semantics or sentiment. For example, Dong et al. use an adaptive layer in a recursive neural network for target-dependent Twitter sentiment analysis, where targets are topics such as windows 7 or taylor swift BIBREF16 , BIBREF17 ; recursive neural tensor networks (RNTNs) utilize sentence parse trees to capture sentence-level sentiment for movie reviews BIBREF4 ; Le and Mikolov predict sentiment by using paragraph vectors to model each paragraph as a continuous representation BIBREF18 . They show that performance can thus be improved by more delicate text models.
Others have suggested using extra-linguistic features to improve the deep learning model. The user-word composition vector model (UWCVM) BIBREF19 is inspired by the possibility that the strength of sentiment words is user-specific; to capture this they add user embeddings in their model. In UPNN, a later extension, they further add a product-word composition as product embeddings, arguing that products can also show different tendencies of being rated or reviewed BIBREF20 . Their addition of user information yielded 2–10% improvements in accuracy as compared to the above-mentioned RNTN and paragraph vector methods. We also seek to inject user information into the neural network model. In comparison to the research of Tang et al. on sentiment classification for product reviews, the difference is two-fold. First, we take into account multiple users (one author and potentially many likers) for one post, whereas only one user (the reviewer) is involved in a review. Second, we add comment information to provide more features for post stance classification. None of these two factors have been considered previously in a deep learning model for text stance classification. Therefore, we propose UTCNN, which generates and utilizes user embeddings for all users — even for those who have not authored any posts — and incorporates comments to further improve performance.
Method
In this section, we first describe CNN-based document composition, which captures user- and topic-dependent document-level semantic representation from word representations. Then we show how to add comment information to construct the user-topic-comment neural network (UTCNN).
User- and Topic-dependent Document Composition
As shown in Figure FIGREF4 , we use a general CNN BIBREF3 and two semantic transformations for document composition . We are given a document with an engaged user INLINEFORM0 , a topic INLINEFORM1 , and its composite INLINEFORM2 words, each word INLINEFORM3 of which is associated with a word embedding INLINEFORM4 where INLINEFORM5 is the vector dimension. For each word embedding INLINEFORM6 , we apply two dot operations as shown in Equation EQREF6 : DISPLAYFORM0
where INLINEFORM0 models the user reading preference for certain semantics, and INLINEFORM1 models the topic semantics; INLINEFORM2 and INLINEFORM3 are the dimensions of transformed user and topic embeddings respectively. We use INLINEFORM4 to model semantically what each user prefers to read and/or write, and use INLINEFORM5 to model the semantics of each topic. The dot operation of INLINEFORM6 and INLINEFORM7 transforms the global representation INLINEFORM8 to a user-dependent representation. Likewise, the dot operation of INLINEFORM9 and INLINEFORM10 transforms INLINEFORM11 to a topic-dependent representation.
After the two dot operations on INLINEFORM0 , we have user-dependent and topic-dependent word vectors INLINEFORM1 and INLINEFORM2 , which are concatenated to form a user- and topic-dependent word vector INLINEFORM3 . Then the transformed word embeddings INLINEFORM4 are used as the CNN input. Here we apply three convolutional layers on the concatenated transformed word embeddings INLINEFORM5 : DISPLAYFORM0
where INLINEFORM0 is the index of words; INLINEFORM1 is a non-linear activation function (we use INLINEFORM2 ); INLINEFORM5 is the convolutional filter with input length INLINEFORM6 and output length INLINEFORM7 , where INLINEFORM8 is the window size of the convolutional operation; and INLINEFORM9 and INLINEFORM10 are the output and bias of the convolution layer INLINEFORM11 , respectively. In our experiments, the three window sizes INLINEFORM12 in the three convolution layers are one, two, and three, encoding unigram, bigram, and trigram semantics accordingly.
After the convolutional layer, we add a maximum pooling layer among convolutional outputs to obtain the unigram, bigram, and trigram n-gram representations. This is succeeded by an average pooling layer for an element-wise average of the three maximized convolution outputs.
UTCNN Model Description
Figure FIGREF10 illustrates the UTCNN model. As more than one user may interact with a given post, we first add a maximum pooling layer after the user matrix embedding layer and user vector embedding layer to form a moderator matrix embedding INLINEFORM0 and a moderator vector embedding INLINEFORM1 for moderator INLINEFORM2 respectively, where INLINEFORM3 is used for the semantic transformation in the document composition process, as mentioned in the previous section. The term moderator here is to denote the pseudo user who provides the overall semantic/sentiment of all the engaged users for one document. The embedding INLINEFORM4 models the moderator stance preference, that is, the pattern of the revealed user stance: whether a user is willing to show his preference, whether a user likes to show impartiality with neutral statements and reasonable arguments, or just wants to show strong support for one stance. Ideally, the latent user stance is modeled by INLINEFORM5 for each user. Likewise, for topic information, a maximum pooling layer is added after the topic matrix embedding layer and topic vector embedding layer to form a joint topic matrix embedding INLINEFORM6 and a joint topic vector embedding INLINEFORM7 for topic INLINEFORM8 respectively, where INLINEFORM9 models the semantic transformation of topic INLINEFORM10 as in users and INLINEFORM11 models the topic stance tendency. The latent topic stance is also modeled by INLINEFORM12 for each topic.
As for comments, we view them as short documents with authors only but without likers nor their own comments. Therefore we apply document composition on comments although here users are commenters (users who comment). It is noticed that the word embeddings INLINEFORM0 for the same word in the posts and comments are the same, but after being transformed to INLINEFORM1 in the document composition process shown in Figure FIGREF4 , they might become different because of their different engaged users. The output comment representation together with the commenter vector embedding INLINEFORM2 and topic vector embedding INLINEFORM3 are concatenated and a maximum pooling layer is added to select the most important feature for comments. Instead of requiring that the comment stance agree with the post, UTCNN simply extracts the most important features of the comment contents; they could be helpful, whether they show obvious agreement or disagreement. Therefore when combining comment information here, the maximum pooling layer is more appropriate than other pooling or merging layers. Indeed, we believe this is one reason for UTCNN's performance gains.
Finally, the pooled comment representation, together with user vector embedding INLINEFORM0 , topic vector embedding INLINEFORM1 , and document representation are fed to a fully connected network, and softmax is applied to yield the final stance label prediction for the post.
Experiment
We start with the experimental dataset and then describe the training process as well as the implementation of the baselines. We also implement several variations to reveal the effects of features: authors, likers, comment, and commenters. In the results section we compare our model with related work.
Dataset
We tested the proposed UTCNN on two different datasets: FBFans and CreateDebate. FBFans is a privately-owned, single-topic, Chinese, unbalanced, social media dataset, and CreateDebate is a public, multiple-topic, English, balanced, forum dataset. Results using these two datasets show the applicability and superiority for different topics, languages, data distributions, and platforms.
The FBFans dataset contains data from anti-nuclear-power Chinese Facebook fan groups from September 2013 to August 2014, including posts and their author and liker IDs. There are a total of 2,496 authors, 505,137 likers, 33,686 commenters, and 505,412 unique users. Two annotators were asked to take into account only the post content to label the stance of the posts in the whole dataset as supportive, neutral, or unsupportive (hereafter denoted as Sup, Neu, and Uns). Sup/Uns posts were those in support of or against anti-reconstruction; Neu posts were those evincing a neutral standpoint on the topic, or were irrelevant. Raw agreement between annotators is 0.91, indicating high agreement. Specifically, Cohen’s Kappa for Neu and not Neu labeling is 0.58 (moderate), and for Sup or Uns labeling is 0.84 (almost perfect). Posts with inconsistent labels were filtered out, and the development and testing sets were randomly selected from what was left. Posts in the development and testing sets involved at least one user who appeared in the training set. The number of posts for each stance is shown on the left-hand side of Table TABREF12 . About twenty percent of the posts were labeled with a stance, and the number of supportive (Sup) posts was much larger than that of the unsupportive (Uns) ones: this is thus highly skewed data, which complicates stance classification. On average, 161.1 users were involved in one post. The maximum was 23,297 and the minimum was one (the author). For comments, on average there were 3 comments per post. The maximum was 1,092 and the minimum was zero.
To test whether the assumption of this paper – posts attract users who hold the same stance to like them – is reliable, we examine the likes from authors of different stances. Posts in FBFans dataset are used for this analysis. We calculate the like statistics of each distinct author from these 32,595 posts. As the numbers of authors in the Sup, Neu and Uns stances are largely imbalanced, these numbers are normalized by the number of users of each stance. Table TABREF13 shows the results. Posts with stances (i.e., not neutral) attract users of the same stance. Neutral posts also attract both supportive and neutral users, like what we observe in supportive posts, but just the neutral posts can attract even more neutral likers. These results do suggest that users prefer posts of the same stance, or at least posts of no obvious stance which might cause annoyance when reading, and hence support the user modeling in our approach.
The CreateDebate dataset was collected from an English online debate forum discussing four topics: abortion (ABO), gay rights (GAY), Obama (OBA), and marijuana (MAR). The posts are annotated as for (F) and against (A). Replies to posts in this dataset are also labeled with stance and hence use the same data format as posts. The labeling results are shown in the right-hand side of Table TABREF12 . We observe that the dataset is more balanced than the FBFans dataset. In addition, there are 977 unique users in the dataset. To compare with Hasan and Ng's work, we conducted five-fold cross-validation and present the annotation results as the average number of all folds BIBREF9 , BIBREF5 .
The FBFans dataset has more integrated functions than the CreateDebate dataset; thus our model can utilize all linguistic and extra-linguistic features. For the CreateDebate dataset, on the other hand, the like and comment features are not available (as there is a stance label for each reply, replies are evaluated as posts as other previous work) but we still implemented our model using the content, author, and topic information.
Settings
In the UTCNN training process, cross-entropy was used as the loss function and AdaGrad as the optimizer. For FBFans dataset, we learned the 50-dimensional word embeddings on the whole dataset using GloVe BIBREF21 to capture the word semantics; for CreateDebate dataset we used the publicly available English 50-dimensional word embeddings, pre-trained also using GloVe. These word embeddings were fixed in the training process. The learning rate was set to 0.03. All user and topic embeddings were randomly initialized in the range of [-0.1 0.1]. Matrix embeddings for users and topics were sized at 250 ( INLINEFORM0 ); vector embeddings for users and topics were set to length 10.
We applied the LDA topic model BIBREF22 on the FBFans dataset to determine the latent topics with which to build topic embeddings, as there is only one general known topic: nuclear power plants. We learned 100 latent topics and assigned the top three topics for each post. For the CreateDebate dataset, which itself constitutes four topics, the topic labels for posts were used directly without additionally applying LDA.
For the FBFans data we report class-based f-scores as well as the macro-average f-score ( INLINEFORM0 ) shown in equation EQREF19 . DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are the average precision and recall of the three class. We adopted the macro-average f-score as the evaluation metric for the overall performance because (1) the experimental dataset is severely imbalanced, which is common for contentious issues; and (2) for stance classification, content in minor-class posts is usually more important for further applications. For the CreateDebate dataset, accuracy was adopted as the evaluation metric to compare the results with related work BIBREF7 , BIBREF9 , BIBREF12 .
Baselines
We pit our model against the following baselines: 1) SVM with unigram, bigram, and trigram features, which is a standard yet rather strong classifier for text features; 2) SVM with average word embedding, where a document is represented as a continuous representation by averaging the embeddings of the composite words; 3) SVM with average transformed word embeddings (the INLINEFORM0 in equation EQREF6 ), where a document is represented as a continuous representation by averaging the transformed embeddings of the composite words; 4) two mature deep learning models on text classification, CNN BIBREF3 and Recurrent Convolutional Neural Networks (RCNN) BIBREF0 , where the hyperparameters are based on their work; 5) the above SVM and deep learning models with comment information; 6) UTCNN without user information, representing a pure-text CNN model where we use the same user matrix and user embeddings INLINEFORM1 and INLINEFORM2 for each user; 7) UTCNN without the LDA model, representing how UTCNN works with a single-topic dataset; 8) UTCNN without comments, in which the model predicts the stance label given only user and topic information. All these models were trained on the training set, and parameters as well as the SVM kernel selections (linear or RBF) were fine-tuned on the development set. Also, we adopt oversampling on SVMs, CNN and RCNN because the FBFans dataset is highly imbalanced.
Results on FBFans Dataset
In Table TABREF22 we show the results of UTCNN and the baselines on the FBFans dataset. Here Majority yields good performance on Neu since FBFans is highly biased to the neutral class. The SVM models perform well on Sup and Neu but perform poorly for Uns, showing that content information in itself is insufficient to predict stance labels, especially for the minor class. With the transformed word embedding feature, SVM can achieve comparable performance as SVM with n-gram feature. However, the much fewer feature dimension of the transformed word embedding makes SVM with word embeddings a more efficient choice for modeling the large scale social media dataset. For the CNN and RCNN models, they perform slightly better than most of the SVM models but still, the content information is insufficient to achieve a good performance on the Uns posts. As to adding comment information to these models, since the commenters do not always hold the same stance as the author, simply adding comments and post contents together merely adds noise to the model.
Among all UTCNN variations, we find that user information is most important, followed by topic and comment information. UTCNN without user information shows results similar to SVMs — it does well for Sup and Neu but detects no Uns. Its best f-scores on both Sup and Neu among all methods show that with enough training data, content-based models can perform well; at the same time, the lack of user information results in too few clues for minor-class posts to either predict their stance directly or link them to other users and posts for improved performance. The 17.5% improvement when adding user information suggests that user information is especially useful when the dataset is highly imbalanced. All models that consider user information predict the minority class successfully. UCTNN without topic information works well but achieves lower performance than the full UTCNN model. The 4.9% performance gain brought by LDA shows that although it is satisfactory for single topic datasets, adding that latent topics still benefits performance: even when we are discussing the same topic, we use different arguments and supporting evidence. Lastly, we get 4.8% improvement when adding comment information and it achieves comparable performance to UTCNN without topic information, which shows that comments also benefit performance. For platforms where user IDs are pixelated or otherwise hidden, adding comments to a text model still improves performance. In its integration of user, content, and comment information, the full UTCNN produces the highest f-scores on all Sup, Neu, and Uns stances among models that predict the Uns class, and the highest macro-average f-score overall. This shows its ability to balance a biased dataset and supports our claim that UTCNN successfully bridges content and user, topic, and comment information for stance classification on social media text. Another merit of UTCNN is that it does not require a balanced training data. This is supported by its outperforming other models though no oversampling technique is applied to the UTCNN related experiments as shown in this paper. Thus we can conclude that the user information provides strong clues and it is still rich even in the minority class.
We also investigate the semantic difference when a user acts as an author/liker or a commenter. We evaluated a variation in which all embeddings from the same user were forced to be identical (this is the UTCNN shared user embedding setting in Table TABREF22 ). This setting yielded only a 2.5% improvement over the model without comments, which is not statistically significant. However, when separating authors/likers and commenters embeddings (i.e., the UTCNN full model), we achieved much greater improvements (4.8%). We attribute this result to the tendency of users to use different wording for different roles (for instance author vs commenter). This is observed when the user, acting as an author, attempts to support her argument against nuclear power by using improvements in solar power; when acting as a commenter, though, she interacts with post contents by criticizing past politicians who supported nuclear power or by arguing that the proposed evacuation plan in case of a nuclear accident is ridiculous. Based on this finding, in the final UTCNN setting we train two user matrix embeddings for one user: one for the author/liker role and the other for the commenter role.
Results on CreateDebate Dataset
Table TABREF24 shows the results of UTCNN, baselines as we implemented on the FBFans datset and related work on the CreateDebate dataset. We do not adopt oversampling on these models because the CreateDebate dataset is almost balanced. In previous work, integer linear programming (ILP) or linear-chain conditional random fields (CRFs) were proposed to integrate text features, author, ideology, and user-interaction constraints, where text features are unigram, bigram, and POS-dependencies; the author constraint tends to require that posts from the same author for the same topic hold the same stance; the ideology constraint aims to capture inferences between topics for the same author; the user-interaction constraint models relationships among posts via user interactions such as replies BIBREF7 , BIBREF9 .
The SVM with n-gram or average word embedding feature performs just similar to the majority. However, with the transformed word embedding, it achieves superior results. It shows that the learned user and topic embeddings really capture the user and topic semantics. This finding is not so obvious in the FBFans dataset and it might be due to the unfavorable data skewness for SVM. As for CNN and RCNN, they perform slightly better than most SVMs as we found in Table TABREF22 for FBFans.
Compared to the ILP BIBREF7 and CRF BIBREF9 methods, the UTCNN user embeddings encode author and user-interaction constraints, where the ideology constraint is modeled by the topic embeddings and text features are modeled by the CNN. The significant improvement achieved by UTCNN suggests the latent representations are more effective than overt model constraints.
The PSL model BIBREF12 jointly labels both author and post stance using probabilistic soft logic (PSL) BIBREF23 by considering text features and reply links between authors and posts as in Hasan and Ng's work. Table TABREF24 reports the result of their best AD setting, which represents the full joint stance/disagreement collective model on posts and is hence more relevant to UTCNN. In contrast to their model, the UTCNN user embeddings represent relationships between authors, but UTCNN models do not utilize link information between posts. Though the PSL model has the advantage of being able to jointly label the stances of authors and posts, its performance on posts is lower than the that for the ILP or CRF models. UTCNN significantly outperforms these models on posts and has the potential to predict user stances through the generated user embeddings.
For the CreateDebate dataset, we also evaluated performance when not using topic embeddings or user embeddings; as replies in this dataset are viewed as posts, the setting without comment embeddings is not available. Table TABREF24 shows the same findings as Table TABREF22 : the 21% improvement in accuracy demonstrates that user information is the most vital. This finding also supports the results in the related work: user constraints are useful and can yield 11.2% improvement in accuracy BIBREF7 . Further considering topic information yields 3.4% improvement, suggesting that knowing the subject of debates provides useful information. In sum, Table TABREF22 together with Table TABREF24 show that UTCNN achieves promising performance regardless of topic, language, data distribution, and platform.
Conclusion
We have proposed UTCNN, a neural network model that incorporates user, topic, content and comment information for stance classification on social media texts. UTCNN learns user embeddings for all users with minimum active degree, i.e., one post or one like. Topic information obtained from the topic model or the pre-defined labels further improves the UTCNN model. In addition, comment information provides additional clues for stance classification. We have shown that UTCNN achieves promising and balanced results. In the future we plan to explore the effectiveness of the UTCNN user embeddings for author stance classification.
Acknowledgements
Research of this paper was partially supported by Ministry of Science and Technology, Taiwan, under the contract MOST 104-2221-E-001-024-MY2.
| [
"SVM with unigram, bigram, and trigram features, SVM with average word embedding, SVM with average transformed word embeddings, CNN, ecurrent Convolutional Neural Networks, SVM and deep learning models with comment information",
"SVM with unigram, bigram, trigram features, with average word embedding, with averag... | 4,512 | qasper | en | null | d426a2d42f3dffc8771498ba64ed0e383b91939398e83dce | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What are the baselines?
Answer: | [
"SVM with unigram, bigram, and trigram features, SVM with average word embedding, SVM with average transformed word embeddings, CNN, ecurrent Convolutional Neural Networks, SVM and deep learning models with comment information",
"SVM with unigram, bigram, trigram features, with average word embedding, with averag... | qasper | 128 | 19 | |
By how much did they improve? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Automatic classification of sentiment has mainly focused on categorizing tweets in either two (binary sentiment analysis) or three (ternary sentiment analysis) categories BIBREF0 . In this work we study the problem of fine-grained sentiment classification where tweets are classified according to a five-point scale ranging from VeryNegative to VeryPositive. To illustrate this, Table TABREF3 presents examples of tweets associated with each of these categories. Five-point scales are widely adopted in review sites like Amazon and TripAdvisor, where a user's sentiment is ordered with respect to its intensity. From a sentiment analysis perspective, this defines a classification problem with five categories. In particular, Sebastiani et al. BIBREF1 defined such classification problems whose categories are explicitly ordered to be ordinal classification problems. To account for the ordering of the categories, learners are penalized according to how far from the true class their predictions are.
Although considering different scales, the various settings of sentiment classification are related. First, one may use the same feature extraction and engineering approaches to represent the text spans such as word membership in lexicons, morpho-syntactic statistics like punctuation or elongated word counts BIBREF2 , BIBREF3 . Second, one would expect that knowledge from one task can be transfered to the others and this would benefit the performance. Knowing that a tweet is “Positive” in the ternary setting narrows the classification decision between the VeryPositive and Positive categories in the fine-grained setting. From a research perspective this raises the question of whether and how one may benefit when tackling such related tasks and how one can transfer knowledge from one task to another during the training phase.
Our focus in this work is to exploit the relation between the sentiment classification settings and demonstrate the benefits stemming from combining them. To this end, we propose to formulate the different classification problems as a multitask learning problem and jointly learn them. Multitask learning BIBREF4 has shown great potential in various domains and its benefits have been empirically validated BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 using different types of data and learning approaches. An important benefit of multitask learning is that it provides an elegant way to access resources developed for similar tasks. By jointly learning correlated tasks, the amount of usable data increases. For instance, while for ternary classification one can label data using distant supervision with emoticons BIBREF9 , there is no straightforward way to do so for the fine-grained problem. However, the latter can benefit indirectly, if the ternary and fine-grained tasks are learned jointly.
The research question that the paper attempts to answer is the following: Can twitter sentiment classification problems, and fine-grained sentiment classification in particular, benefit from multitask learning? To answer the question, the paper brings the following two main contributions: (i) we show how jointly learning the ternary and fine-grained sentiment classification problems in a multitask setting improves the state-of-the-art performance, and (ii) we demonstrate that recurrent neural networks outperform models previously proposed without access to huge corpora while being flexible to incorporate different sources of data.
Multitask Learning for Twitter Sentiment Classification
In his work, Caruana BIBREF4 proposed a multitask approach in which a learner takes advantage of the multiplicity of interdependent tasks while jointly learning them. The intuition is that if the tasks are correlated, the learner can learn a model jointly for them while taking into account the shared information which is expected to improve its generalization ability. People express their opinions online on various subjects (events, products..), on several languages and in several styles (tweets, paragraph-sized reviews..), and it is exactly this variety that motivates the multitask approaches. Specifically for Twitter for instance, the different settings of classification like binary, ternary and fine-grained are correlated since their difference lies in the sentiment granularity of the classes which increases while moving from binary to fine-grained problems.
There are two main decisions to be made in our approach: the learning algorithm, which learns a decision function, and the data representation. With respect to the former, neural networks are particularly suitable as one can design architectures with different properties and arbitrary complexity. Also, as training neural network usually relies on back-propagation of errors, one can have shared parts of the network trained by estimating errors on the joint tasks and others specialized for particular tasks. Concerning the data representation, it strongly depends on the data type available. For the task of sentiment classification of tweets with neural networks, distributed embeddings of words have shown great potential. Embeddings are defined as low-dimensional, dense representations of words that can be obtained in an unsupervised fashion by training on large quantities of text BIBREF10 .
Concerning the neural network architecture, we focus on Recurrent Neural Networks (RNNs) that are capable of modeling short-range and long-range dependencies like those exhibited in sequence data of arbitrary length like text. While in the traditional information retrieval paradigm such dependencies are captured using INLINEFORM0 -grams and skip-grams, RNNs learn to capture them automatically BIBREF11 . To circumvent the problems with capturing long-range dependencies and preventing gradients from vanishing, the long short-term memory network (LSTM) was proposed BIBREF12 . In this work, we use an extended version of LSTM called bidirectional LSTM (biLSTM). While standard LSTMs access information only from the past (previous words), biLSTMs capture both past and future information effectively BIBREF13 , BIBREF11 . They consist of two LSTM networks, for propagating text forward and backwards with the goal being to capture the dependencies better. Indeed, previous work on multitask learning showed the effectiveness of biLSTMs in a variety of problems: BIBREF14 tackled sequence prediction, while BIBREF6 and BIBREF15 used biLSTMs for Named Entity Recognition and dependency parsing respectively.
Figure FIGREF2 presents the architecture we use for multitask learning. In the top-left of the figure a biLSTM network (enclosed by the dashed line) is fed with embeddings INLINEFORM0 that correspond to the INLINEFORM1 words of a tokenized tweet. Notice, as discussed above, the biLSTM consists of two LSTMs that are fed with the word sequence forward and backwards. On top of the biLSTM network one (or more) hidden layers INLINEFORM2 transform its output. The output of INLINEFORM3 is led to the softmax layers for the prediction step. There are INLINEFORM4 softmax layers and each is used for one of the INLINEFORM5 tasks of the multitask setting. In tasks such as sentiment classification, additional features like membership of words in sentiment lexicons or counts of elongated/capitalized words can be used to enrich the representation of tweets before the classification step BIBREF3 . The lower part of the network illustrates how such sources of information can be incorporated to the process. A vector “Additional Features” for each tweet is transformed from the hidden layer(s) INLINEFORM6 and then is combined by concatenation with the transformed biLSTM output in the INLINEFORM7 layer.
Experimental setup
Our goal is to demonstrate how multitask learning can be successfully applied on the task of sentiment classification of tweets. The particularities of tweets are to be short and informal text spans. The common use of abbreviations, creative language etc., makes the sentiment classification problem challenging. To validate our hypothesis, that learning the tasks jointly can benefit the performance, we propose an experimental setting where there are data from two different twitter sentiment classification problems: a fine-grained and a ternary. We consider the fine-grained task to be our primary task as it is more challenging and obtaining bigger datasets, e.g. by distant supervision, is not straightforward and, hence we report the performance achieved for this task.
Ternary and fine-grained sentiment classification were part of the SemEval-2016 “Sentiment Analysis in Twitter” task BIBREF16 . We use the high-quality datasets the challenge organizers released. The dataset for fine-grained classification is split in training, development, development_test and test parts. In the rest, we refer to these splits as train, development and test, where train is composed by the training and the development instances. Table TABREF7 presents an overview of the data. As discussed in BIBREF16 and illustrated in the Table, the fine-grained dataset is highly unbalanced and skewed towards the positive sentiment: only INLINEFORM0 of the training examples are labeled with one of the negative classes.
Feature representation We report results using two different feature sets. The first one, dubbed nbow, is a neural bag-of-words that uses text embeddings to generate low-dimensional, dense representations of the tweets. To construct the nbow representation, given the word embeddings dictionary where each word is associated with a vector, we apply the average compositional function that averages the embeddings of the words that compose a tweet. Simple compositional functions like average were shown to be robust and efficient in previous work BIBREF17 . Instead of training embeddings from scratch, we use the pre-trained on tweets GloVe embeddings of BIBREF10 . In terms of resources required, using only nbow is efficient as it does not require any domain knowledge. However, previous research on sentiment analysis showed that using extra resources, like sentiment lexicons, can benefit significantly the performance BIBREF3 , BIBREF2 . To validate this and examine at which extent neural networks and multitask learning benefit from such features we evaluate the models using an augmented version of nbow, dubbed nbow+. The feature space of the latter, is augmented using 1,368 extra features consisting mostly of counts of punctuation symbols ('!?#@'), emoticons, elongated words and word membership features in several sentiment lexicons. Due to space limitations, for a complete presentation of these features, we refer the interested reader to BIBREF2 , whose open implementation we used to extract them.
Evaluation measure To reproduce the setting of the SemEval challenges BIBREF16 , we optimize our systems using as primary measure the macro-averaged Mean Absolute Error ( INLINEFORM0 ) given by: INLINEFORM1
where INLINEFORM0 is the number of categories, INLINEFORM1 is the set of instances whose true class is INLINEFORM2 , INLINEFORM3 is the true label of the instance INLINEFORM4 and INLINEFORM5 the predicted label. The measure penalizes decisions far from the true ones and is macro-averaged to account for the fact that the data are unbalanced. Complementary to INLINEFORM6 , we report the performance achieved on the micro-averaged INLINEFORM7 measure, which is a commonly used measure for classification.
The models To evaluate the multitask learning approach, we compared it with several other models. Support Vector Machines (SVMs) are maximum margin classification algorithms that have been shown to achieve competitive performance in several text classification problems BIBREF16 . SVM INLINEFORM0 stands for an SVM with linear kernel and an one-vs-rest approach for the multi-class problem. Also, SVM INLINEFORM1 is an SVM with linear kernel that employs the crammer-singer strategy BIBREF18 for the multi-class problem. Logistic regression (LR) is another type of linear classification method, with probabilistic motivation. Again, we use two types of Logistic Regression depending on the multi-class strategy: LR INLINEFORM2 that uses an one-vs-rest approach and multinomial Logistic Regression also known as the MaxEnt classifier that uses a multinomial criterion.
Both SVMs and LRs as discussed above treat the problem as a multi-class one, without considering the ordering of the classes. For these four models, we tuned the hyper-parameter INLINEFORM0 that controls the importance of the L INLINEFORM1 regularization part in the optimization problem with grid-search over INLINEFORM2 using 10-fold cross-validation in the union of the training and development data and then retrained the models with the selected values. Also, to account for the unbalanced classification problem we used class weights to penalize more the errors made on the rare classes. These weights were inversely proportional to the frequency of each class. For the four models we used the implementations of Scikit-learn BIBREF19 .
For multitask learning we use the architecture shown in Figure FIGREF2 , which we implemented with Keras BIBREF20 . The embeddings are initialized with the 50-dimensional GloVe embeddings while the output of the biLSTM network is set to dimension 50. The activation function of the hidden layers is the hyperbolic tangent. The weights of the layers were initialized from a uniform distribution, scaled as described in BIBREF21 . We used the Root Mean Square Propagation optimization method. We used dropout for regularizing the network. We trained the network using batches of 128 examples as follows: before selecting the batch, we perform a Bernoulli trial with probability INLINEFORM0 to select the task to train for. With probability INLINEFORM1 we pick a batch for the fine-grained sentiment classification problem, while with probability INLINEFORM2 we pick a batch for the ternary problem. As shown in Figure FIGREF2 , the error is backpropagated until the embeddings, that we fine-tune during the learning process. Notice also that the weights of the network until the layer INLINEFORM3 are shared and therefore affected by both tasks.
To tune the neural network hyper-parameters we used 5-fold cross validation. We tuned the probability INLINEFORM0 of dropout after the hidden layers INLINEFORM1 and for the biLSTM for INLINEFORM2 , the size of the hidden layer INLINEFORM3 and the probability INLINEFORM4 of the Bernoulli trials from INLINEFORM5 . During training, we monitor the network's performance on the development set and apply early stopping if the performance on the validation set does not improve for 5 consecutive epochs.
Experimental results Table TABREF9 illustrates the performance of the models for the different data representations. The upper part of the Table summarizes the performance of the baselines. The entry “Balikas et al.” stands for the winning system of the 2016 edition of the challenge BIBREF2 , which to the best of our knowledge holds the state-of-the-art. Due to the stochasticity of training the biLSTM models, we repeat the experiment 10 times and report the average and the standard deviation of the performance achieved.
Several observations can be made from the table. First notice that, overall, the best performance is achieved by the neural network architecture that uses multitask learning. This entails that the system makes use of the available resources efficiently and improves the state-of-the-art performance. In conjunction with the fact that we found the optimal probability INLINEFORM0 , this highlights the benefits of multitask learning over single task learning. Furthermore, as described above, the neural network-based models have only access to the training data as the development are hold for early stopping. On the other hand, the baseline systems were retrained on the union of the train and development sets. Hence, even with fewer resources available for training on the fine-grained problem, the neural networks outperform the baselines. We also highlight the positive effect of the additional features that previous research proposed. Adding the features both in the baselines and in the biLSTM-based architectures improves the INLINEFORM1 scores by several points.
Lastly, we compare the performance of the baseline systems with the performance of the state-of-the-art system of BIBREF2 . While BIBREF2 uses n-grams (and character-grams) with INLINEFORM0 , the baseline systems (SVMs, LRs) used in this work use the nbow+ representation, that relies on unigrams. Although they perform on par, the competitive performance of nbow highlights the potential of distributed representations for short-text classification. Further, incorporating structure and distributed representations leads to the gains of the biLSTM network, in the multitask and single task setting.
Similar observations can be drawn from Figure FIGREF10 that presents the INLINEFORM0 scores. Again, the biLSTM network with multitask learning achieves the best performance. It is also to be noted that although the two evaluation measures are correlated in the sense that the ranking of the models is the same, small differences in the INLINEFORM1 have large effect on the scores of the INLINEFORM2 measure.
Conclusion
In this paper, we showed that by jointly learning the tasks of ternary and fine-grained classification with a multitask learning model, one can greatly improve the performance on the second. This opens several avenues for future research. Since sentiment is expressed in different textual types like tweets and paragraph-sized reviews, in different languages (English, German, ..) and in different granularity levels (binary, ternary,..) one can imagine multitask approaches that could benefit from combining such resources. Also, while we opted for biLSTM networks here, one could use convolutional neural networks or even try to combine different types of networks and tasks to investigate the performance effect of multitask learning. Lastly, while our approach mainly relied on the foundations of BIBREF4 , the internal mechanisms and the theoretical guarantees of multitask learning remain to be better understood.
Acknowledgements
This work is partially supported by the CIFRE N 28/2015.
| [
"They decrease MAE in 0.34"
] | 2,735 | qasper | en | null | 680ef767fe2a990615d0bc1cb1bf5f7d23c8766209138826 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: By how much did they improve?
Answer: | [
"They decrease MAE in 0.34"
] | qasper | 128 | 20 | |
How does their model improve interpretability compared to softmax transformers? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The Transformer architecture BIBREF0 for deep neural networks has quickly risen to prominence in NLP through its efficiency and performance, leading to improvements in the state of the art of Neural Machine Translation BIBREF1, BIBREF2, as well as inspiring other powerful general-purpose models like BERT BIBREF3 and GPT-2 BIBREF4. At the heart of the Transformer lie multi-head attention mechanisms: each word is represented by multiple different weighted averages of its relevant context. As suggested by recent works on interpreting attention head roles, separate attention heads may learn to look for various relationships between tokens BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9.
The attention distribution of each head is predicted typically using the softmax normalizing transform. As a result, all context words have non-zero attention weight. Recent work on single attention architectures suggest that using sparse normalizing transforms in attention mechanisms such as sparsemax – which can yield exactly zero probabilities for irrelevant words – may improve performance and interpretability BIBREF12, BIBREF13, BIBREF14. Qualitative analysis of attention heads BIBREF0 suggests that, depending on what phenomena they capture, heads tend to favor flatter or more peaked distributions.
Recent works have proposed sparse Transformers BIBREF10 and adaptive span Transformers BIBREF11. However, the “sparsity" of those models only limits the attention to a contiguous span of past tokens, while in this work we propose a highly adaptive Transformer model that is capable of attending to a sparse set of words that are not necessarily contiguous. Figure FIGREF1 shows the relationship of these methods with ours.
Our contributions are the following:
We introduce sparse attention into the Transformer architecture, showing that it eases interpretability and leads to slight accuracy gains.
We propose an adaptive version of sparse attention, where the shape of each attention head is learnable and can vary continuously and dynamically between the dense limit case of softmax and the sparse, piecewise-linear sparsemax case.
We make an extensive analysis of the added interpretability of these models, identifying both crisper examples of attention head behavior observed in previous work, as well as novel behaviors unraveled thanks to the sparsity and adaptivity of our proposed model.
Background ::: The Transformer
In NMT, the Transformer BIBREF0 is a sequence-to-sequence (seq2seq) model which maps an input sequence to an output sequence through hierarchical multi-head attention mechanisms, yielding a dynamic, context-dependent strategy for propagating information within and across sentences. It contrasts with previous seq2seq models, which usually rely either on costly gated recurrent operations BIBREF15, BIBREF16 or static convolutions BIBREF17.
Given $n$ query contexts and $m$ sequence items under consideration, attention mechanisms compute, for each query, a weighted representation of the items. The particular attention mechanism used in BIBREF0 is called scaled dot-product attention, and it is computed in the following way:
where $\mathbf {Q} \in \mathbb {R}^{n \times d}$ contains representations of the queries, $\mathbf {K}, \mathbf {V} \in \mathbb {R}^{m \times d}$ are the keys and values of the items attended over, and $d$ is the dimensionality of these representations. The $\mathbf {\pi }$ mapping normalizes row-wise using softmax, $\mathbf {\pi }(\mathbf {Z})_{ij} = \operatornamewithlimits{\mathsf {softmax}}(\mathbf {z}_i)_j$, where
In words, the keys are used to compute a relevance score between each item and query. Then, normalized attention weights are computed using softmax, and these are used to weight the values of each item at each query context.
However, for complex tasks, different parts of a sequence may be relevant in different ways, motivating multi-head attention in Transformers. This is simply the application of Equation DISPLAY_FORM7 in parallel $H$ times, each with a different, learned linear transformation that allows specialization:
In the Transformer, there are three separate multi-head attention mechanisms for distinct purposes:
Encoder self-attention: builds rich, layered representations of each input word, by attending on the entire input sentence.
Context attention: selects a representative weighted average of the encodings of the input words, at each time step of the decoder.
Decoder self-attention: attends over the partial output sentence fragment produced so far.
Together, these mechanisms enable the contextualized flow of information between the input sentence and the sequential decoder.
Background ::: Sparse Attention
The softmax mapping (Equation DISPLAY_FORM8) is elementwise proportional to $\exp $, therefore it can never assign a weight of exactly zero. Thus, unnecessary items are still taken into consideration to some extent. Since its output sums to one, this invariably means less weight is assigned to the relevant items, potentially harming performance and interpretability BIBREF18. This has motivated a line of research on learning networks with sparse mappings BIBREF19, BIBREF20, BIBREF21, BIBREF22. We focus on a recently-introduced flexible family of transformations, $\alpha $-entmax BIBREF23, BIBREF14, defined as:
where $\triangle ^d \lbrace \mathbf {p}\in \mathbb {R}^d:\sum _{i} p_i = 1\rbrace $ is the probability simplex, and, for $\alpha \ge 1$, $\mathsf {H}^{\textsc {T}}_\alpha $ is the Tsallis continuous family of entropies BIBREF24:
This family contains the well-known Shannon and Gini entropies, corresponding to the cases $\alpha =1$ and $\alpha =2$, respectively.
Equation DISPLAY_FORM14 involves a convex optimization subproblem. Using the definition of $\mathsf {H}^{\textsc {T}}_\alpha $, the optimality conditions may be used to derive the following form for the solution (Appendix SECREF83):
where $[\cdot ]_+$ is the positive part (ReLU) function, $\mathbf {1}$ denotes the vector of all ones, and $\tau $ – which acts like a threshold – is the Lagrange multiplier corresponding to the $\sum _i p_i=1$ constraint.
Background ::: Sparse Attention ::: Properties of @!START@$\alpha $@!END@-entmax.
The appeal of $\alpha $-entmax for attention rests on the following properties. For $\alpha =1$ (i.e., when $\mathsf {H}^{\textsc {T}}_\alpha $ becomes the Shannon entropy), it exactly recovers the softmax mapping (We provide a short derivation in Appendix SECREF89.). For all $\alpha >1$ it permits sparse solutions, in stark contrast to softmax. In particular, for $\alpha =2$, it recovers the sparsemax mapping BIBREF19, which is piecewise linear. In-between, as $\alpha $ increases, the mapping continuously gets sparser as its curvature changes.
To compute the value of $\alpha $-entmax, one must find the threshold $\tau $ such that the r.h.s. in Equation DISPLAY_FORM16 sums to one. BIBREF23 propose a general bisection algorithm. BIBREF14 introduce a faster, exact algorithm for $\alpha =1.5$, and enable using $\mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}$ with fixed $\alpha $ within a neural network by showing that the $\alpha $-entmax Jacobian w.r.t. $\mathbf {z}$ for $\mathbf {p}^\star = \mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}(\mathbf {z})$ is
Our work furthers the study of $\alpha $-entmax by providing a derivation of the Jacobian w.r.t. the hyper-parameter $\alpha $ (Section SECREF3), thereby allowing the shape and sparsity of the mapping to be learned automatically. This is particularly appealing in the context of multi-head attention mechanisms, where we shall show in Section SECREF35 that different heads tend to learn different sparsity behaviors.
Adaptively Sparse Transformers with @!START@$\alpha $@!END@-entmax
We now propose a novel Transformer architecture wherein we simply replace softmax with $\alpha $-entmax in the attention heads. Concretely, we replace the row normalization $\mathbf {\pi }$ in Equation DISPLAY_FORM7 by
This change leads to sparse attention weights, as long as $\alpha >1$; in particular, $\alpha =1.5$ is a sensible starting point BIBREF14.
Adaptively Sparse Transformers with @!START@$\alpha $@!END@-entmax ::: Different @!START@$\alpha $@!END@ per head.
Unlike LSTM-based seq2seq models, where $\alpha $ can be more easily tuned by grid search, in a Transformer, there are many attention heads in multiple layers. Crucial to the power of such models, the different heads capture different linguistic phenomena, some of them isolating important words, others spreading out attention across phrases BIBREF0. This motivates using different, adaptive $\alpha $ values for each attention head, such that some heads may learn to be sparser, and others may become closer to softmax. We propose doing so by treating the $\alpha $ values as neural network parameters, optimized via stochastic gradients along with the other weights.
Adaptively Sparse Transformers with @!START@$\alpha $@!END@-entmax ::: Derivatives w.r.t. @!START@$\alpha $@!END@.
In order to optimize $\alpha $ automatically via gradient methods, we must compute the Jacobian of the entmax output w.r.t. $\alpha $. Since entmax is defined through an optimization problem, this is non-trivial and cannot be simply handled through automatic differentiation; it falls within the domain of argmin differentiation, an active research topic in optimization BIBREF25, BIBREF26.
One of our key contributions is the derivation of a closed-form expression for this Jacobian. The next proposition provides such an expression, enabling entmax layers with adaptive $\alpha $. To the best of our knowledge, ours is the first neural network module that can automatically, continuously vary in shape away from softmax and toward sparse mappings like sparsemax.
Proposition 1 Let $\mathbf {p}^\star \mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}(\mathbf {z})$ be the solution of Equation DISPLAY_FORM14. Denote the distribution $\tilde{p}_i {(p_i^\star )^{2 - \alpha }}{ \sum _j(p_j^\star )^{2-\alpha }}$ and let $h_i -p^\star _i \log p^\star _i$. The $i$th component of the Jacobian $\mathbf {g} \frac{\partial \mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}(\mathbf {z})}{\partial \alpha }$ is
proof uses implicit function differentiation and is given in Appendix SECREF10.
Proposition UNKREF22 provides the remaining missing piece needed for training adaptively sparse Transformers. In the following section, we evaluate this strategy on neural machine translation, and analyze the behavior of the learned attention heads.
Experiments
We apply our adaptively sparse Transformers on four machine translation tasks. For comparison, a natural baseline is the standard Transformer architecture using the softmax transform in its multi-head attention mechanisms. We consider two other model variants in our experiments that make use of different normalizing transformations:
1.5-entmax: a Transformer with sparse entmax attention with fixed $\alpha =1.5$ for all heads. This is a novel model, since 1.5-entmax had only been proposed for RNN-based NMT models BIBREF14, but never in Transformers, where attention modules are not just one single component of the seq2seq model but rather an integral part of all of the model components.
$\alpha $-entmax: an adaptive Transformer with sparse entmax attention with a different, learned $\alpha _{i,j}^t$ for each head.
The adaptive model has an additional scalar parameter per attention head per layer for each of the three attention mechanisms (encoder self-attention, context attention, and decoder self-attention), i.e.,
and we set $\alpha _{i,j}^t = 1 + \operatornamewithlimits{\mathsf {sigmoid}}(a_{i,j}^t) \in ]1, 2[$. All or some of the $\alpha $ values can be tied if desired, but we keep them independent for analysis purposes.
Experiments ::: Datasets.
Our models were trained on 4 machine translation datasets of different training sizes:
[itemsep=.5ex,leftmargin=2ex]
IWSLT 2017 German $\rightarrow $ English BIBREF27: 200K sentence pairs.
KFTT Japanese $\rightarrow $ English BIBREF28: 300K sentence pairs.
WMT 2016 Romanian $\rightarrow $ English BIBREF29: 600K sentence pairs.
WMT 2014 English $\rightarrow $ German BIBREF30: 4.5M sentence pairs.
All of these datasets were preprocessed with byte-pair encoding BIBREF31, using joint segmentations of 32k merge operations.
Experiments ::: Training.
We follow the dimensions of the Transformer-Base model of BIBREF0: The number of layers is $L=6$ and number of heads is $H=8$ in the encoder self-attention, the context attention, and the decoder self-attention. We use a mini-batch size of 8192 tokens and warm up the learning rate linearly until 20k steps, after which it decays according to an inverse square root schedule. All models were trained until convergence of validation accuracy, and evaluation was done at each 10k steps for ro$\rightarrow $en and en$\rightarrow $de and at each 5k steps for de$\rightarrow $en and ja$\rightarrow $en. The end-to-end computational overhead of our methods, when compared to standard softmax, is relatively small; in training tokens per second, the models using $\alpha $-entmax and $1.5$-entmax are, respectively, $75\%$ and $90\%$ the speed of the softmax model.
Experiments ::: Results.
We report test set tokenized BLEU BIBREF32 results in Table TABREF27. We can see that replacing softmax by entmax does not hurt performance in any of the datasets; indeed, sparse attention Transformers tend to have slightly higher BLEU, but their sparsity leads to a better potential for analysis. In the next section, we make use of this potential by exploring the learned internal mechanics of the self-attention heads.
Analysis
We conduct an analysis for the higher-resource dataset WMT 2014 English $\rightarrow $ German of the attention in the sparse adaptive Transformer model ($\alpha $-entmax) at multiple levels: we analyze high-level statistics as well as individual head behavior. Moreover, we make a qualitative analysis of the interpretability capabilities of our models.
Analysis ::: High-Level Statistics ::: What kind of @!START@$\alpha $@!END@ values are learned?
Figure FIGREF37 shows the learning trajectories of the $\alpha $ parameters of a selected subset of heads. We generally observe a tendency for the randomly-initialized $\alpha $ parameters to decrease initially, suggesting that softmax-like behavior may be preferable while the model is still very uncertain. After around one thousand steps, some heads change direction and become sparser, perhaps as they become more confident and specialized. This shows that the initialization of $\alpha $ does not predetermine its sparsity level or the role the head will have throughout. In particular, head 8 in the encoder self-attention layer 2 first drops to around $\alpha =1.3$ before becoming one of the sparsest heads, with $\alpha \approx 2$.
The overall distribution of $\alpha $ values at convergence can be seen in Figure FIGREF38. We can observe that the encoder self-attention blocks learn to concentrate the $\alpha $ values in two modes: a very sparse one around $\alpha \rightarrow 2$, and a dense one between softmax and 1.5-entmax . However, the decoder self and context attention only learn to distribute these parameters in a single mode. We show next that this is reflected in the average density of attention weight vectors as well.
Analysis ::: High-Level Statistics ::: Attention weight density when translating.
For any $\alpha >1$, it would still be possible for the weight matrices in Equation DISPLAY_FORM9 to learn re-scalings so as to make attention sparser or denser. To visualize the impact of adaptive $\alpha $ values, we compare the empirical attention weight density (the average number of tokens receiving non-zero attention) within each module, against sparse Transformers with fixed $\alpha =1.5$.
Figure FIGREF40 shows that, with fixed $\alpha =1.5$, heads tend to be sparse and similarly-distributed in all three attention modules. With learned $\alpha $, there are two notable changes: (i) a prominent mode corresponding to fully dense probabilities, showing that our models learn to combine sparse and dense attention, and (ii) a distinction between the encoder self-attention – whose background distribution tends toward extreme sparsity – and the other two modules, who exhibit more uniform background distributions. This suggests that perhaps entirely sparse Transformers are suboptimal.
The fact that the decoder seems to prefer denser attention distributions might be attributed to it being auto-regressive, only having access to past tokens and not the full sentence. We speculate that it might lose too much information if it assigned weights of zero to too many tokens in the self-attention, since there are fewer tokens to attend to in the first place.
Teasing this down into separate layers, Figure FIGREF41 shows the average (sorted) density of each head for each layer. We observe that $\alpha $-entmax is able to learn different sparsity patterns at each layer, leading to more variance in individual head behavior, to clearly-identified dense and sparse heads, and overall to different tendencies compared to the fixed case of $\alpha =1.5$.
Analysis ::: High-Level Statistics ::: Head diversity.
To measure the overall disagreement between attention heads, as a measure of head diversity, we use the following generalization of the Jensen-Shannon divergence:
where $\mathbf {p}_j$ is the vector of attention weights assigned by head $j$ to each word in the sequence, and $\mathsf {H}^\textsc {S}$ is the Shannon entropy, base-adjusted based on the dimension of $\mathbf {p}$ such that $JS \le 1$. We average this measure over the entire validation set. The higher this metric is, the more the heads are taking different roles in the model.
Figure FIGREF44 shows that both sparse Transformer variants show more diversity than the traditional softmax one. Interestingly, diversity seems to peak in the middle layers of the encoder self-attention and context attention, while this is not the case for the decoder self-attention.
The statistics shown in this section can be found for the other language pairs in Appendix SECREF8.
Analysis ::: Identifying Head Specializations
Previous work pointed out some specific roles played by different heads in the softmax Transformer model BIBREF33, BIBREF5, BIBREF9. Identifying the specialization of a head can be done by observing the type of tokens or sequences that the head often assigns most of its attention weight; this is facilitated by sparsity.
Analysis ::: Identifying Head Specializations ::: Positional heads.
One particular type of head, as noted by BIBREF9, is the positional head. These heads tend to focus their attention on either the previous or next token in the sequence, thus obtaining representations of the neighborhood of the current time step. In Figure FIGREF47, we show attention plots for such heads, found for each of the studied models. The sparsity of our models allows these heads to be more confident in their representations, by assigning the whole probability distribution to a single token in the sequence. Concretely, we may measure a positional head's confidence as the average attention weight assigned to the previous token. The softmax model has three heads for position $-1$, with median confidence $93.5\%$. The $1.5$-entmax model also has three heads for this position, with median confidence $94.4\%$. The adaptive model has four heads, with median confidences $95.9\%$, the lowest-confidence head being dense with $\alpha =1.18$, while the highest-confidence head being sparse ($\alpha =1.91$).
For position $+1$, the models each dedicate one head, with confidence around $95\%$, slightly higher for entmax. The adaptive model sets $\alpha =1.96$ for this head.
Analysis ::: Identifying Head Specializations ::: BPE-merging head.
Due to the sparsity of our models, we are able to identify other head specializations, easily identifying which heads should be further analysed. In Figure FIGREF51 we show one such head where the $\alpha $ value is particularly high (in the encoder, layer 1, head 4 depicted in Figure FIGREF37). We found that this head most often looks at the current time step with high confidence, making it a positional head with offset 0. However, this head often spreads weight sparsely over 2-3 neighboring tokens, when the tokens are part of the same BPE cluster or hyphenated words. As this head is in the first layer, it provides a useful service to the higher layers by combining information evenly within some BPE clusters.
For each BPE cluster or cluster of hyphenated words, we computed a score between 0 and 1 that corresponds to the maximum attention mass assigned by any token to the rest of the tokens inside the cluster in order to quantify the BPE-merging capabilities of these heads. There are not any attention heads in the softmax model that are able to obtain a score over $80\%$, while for $1.5$-entmax and $\alpha $-entmax there are two heads in each ($83.3\%$ and $85.6\%$ for $1.5$-entmax and $88.5\%$ and $89.8\%$ for $\alpha $-entmax).
Analysis ::: Identifying Head Specializations ::: Interrogation head.
On the other hand, in Figure FIGREF52 we show a head for which our adaptively sparse model chose an $\alpha $ close to 1, making it closer to softmax (also shown in encoder, layer 1, head 3 depicted in Figure FIGREF37). We observe that this head assigns a high probability to question marks at the end of the sentence in time steps where the current token is interrogative, thus making it an interrogation-detecting head. We also observe this type of heads in the other models, which we also depict in Figure FIGREF52. The average attention weight placed on the question mark when the current token is an interrogative word is $98.5\%$ for softmax, $97.0\%$ for $1.5$-entmax, and $99.5\%$ for $\alpha $-entmax.
Furthermore, we can examine sentences where some tendentially sparse heads become less so, thus identifying sources of ambiguity where the head is less confident in its prediction. An example is shown in Figure FIGREF55 where sparsity in the same head differs for sentences of similar length.
Related Work ::: Sparse attention.
Prior work has developed sparse attention mechanisms, including applications to NMT BIBREF19, BIBREF12, BIBREF20, BIBREF22, BIBREF34. BIBREF14 introduced the entmax function this work builds upon. In their work, there is a single attention mechanism which is controlled by a fixed $\alpha $. In contrast, this is the first work to allow such attention mappings to dynamically adapt their curvature and sparsity, by automatically adjusting the continuous $\alpha $ parameter. We also provide the first results using sparse attention in a Transformer model.
Related Work ::: Fixed sparsity patterns.
Recent research improves the scalability of Transformer-like networks through static, fixed sparsity patterns BIBREF10, BIBREF35. Our adaptively-sparse Transformer can dynamically select a sparsity pattern that finds relevant words regardless of their position (e.g., Figure FIGREF52). Moreover, the two strategies could be combined. In a concurrent line of research, BIBREF11 propose an adaptive attention span for Transformer language models. While their work has each head learn a different contiguous span of context tokens to attend to, our work finds different sparsity patterns in the same span. Interestingly, some of their findings mirror ours – we found that attention heads in the last layers tend to be denser on average when compared to the ones in the first layers, while their work has found that lower layers tend to have a shorter attention span compared to higher layers.
Related Work ::: Transformer interpretability.
The original Transformer paper BIBREF0 shows attention visualizations, from which some speculation can be made of the roles the several attention heads have. BIBREF7 study the syntactic abilities of the Transformer self-attention, while BIBREF6 extract dependency relations from the attention weights. BIBREF8 find that the self-attentions in BERT BIBREF3 follow a sequence of processes that resembles a classical NLP pipeline. Regarding redundancy of heads, BIBREF9 develop a method that is able to prune heads of the multi-head attention module and make an empirical study of the role that each head has in self-attention (positional, syntactic and rare words). BIBREF36 also aim to reduce head redundancy by adding a regularization term to the loss that maximizes head disagreement and obtain improved results. While not considering Transformer attentions, BIBREF18 show that traditional attention mechanisms do not necessarily improve interpretability since softmax attention is vulnerable to an adversarial attack leading to wildly different model predictions for the same attention weights. Sparse attention may mitigate these issues; however, our work focuses mostly on a more mechanical aspect of interpretation by analyzing head behavior, rather than on explanations for predictions.
Conclusion and Future Work
We contribute a novel strategy for adaptively sparse attention, and, in particular, for adaptively sparse Transformers. We present the first empirical analysis of Transformers with sparse attention mappings (i.e., entmax), showing potential in both translation accuracy as well as in model interpretability.
In particular, we analyzed how the attention heads in the proposed adaptively sparse Transformer can specialize more and with higher confidence. Our adaptivity strategy relies only on gradient-based optimization, side-stepping costly per-head hyper-parameter searches. Further speed-ups are possible by leveraging more parallelism in the bisection algorithm for computing $\alpha $-entmax.
Finally, some of the automatically-learned behaviors of our adaptively sparse Transformers – for instance, the near-deterministic positional heads or the subword joining head – may provide new ideas for designing static variations of the Transformer.
Acknowledgments
This work was supported by the European Research Council (ERC StG DeepSPIN 758969), and by the Fundação para a Ciência e Tecnologia through contracts UID/EEA/50008/2019 and CMUPERI/TIC/0046/2014 (GoLocal). We are grateful to Ben Peters for the $\alpha $-entmax code and Erick Fonseca, Marcos Treviso, Pedro Martins, and Tsvetomila Mihaylova for insightful group discussion. We thank Mathieu Blondel for the idea to learn $\alpha $. We would also like to thank the anonymous reviewers for their helpful feedback.
Supplementary Material
Background ::: Regularized Fenchel-Young prediction functions
Definition 1 (BIBREF23)
Let $\Omega \colon \triangle ^d \rightarrow {\mathbb {R}}\cup \lbrace \infty \rbrace $ be a strictly convex regularization function. We define the prediction function $\mathbf {\pi }_{\Omega }$ as
Background ::: Characterizing the @!START@$\alpha $@!END@-entmax mapping
Lemma 1 (BIBREF14) For any $\mathbf {z}$, there exists a unique $\tau ^\star $ such that
Proof: From the definition of $\mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}$,
we may easily identify it with a regularized prediction function (Def. UNKREF81):
We first note that for all $\mathbf {p}\in \triangle ^d$,
From the constant invariance and scaling properties of $\mathbf {\pi }_{\Omega }$ BIBREF23,
Using BIBREF23, noting that $g^{\prime }(t) = t^{\alpha - 1}$ and $(g^{\prime })^{-1}(u) = u^{{1}{\alpha -1}}$, yields
Since $\mathsf {H}^{\textsc {T}}_\alpha $ is strictly convex on the simplex, $\mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}$ has a unique solution $\mathbf {p}^\star $. Equation DISPLAY_FORM88 implicitly defines a one-to-one mapping between $\mathbf {p}^\star $ and $\tau ^\star $ as long as $\mathbf {p}^\star \in \triangle $, therefore $\tau ^\star $ is also unique.
Background ::: Connections to softmax and sparsemax
The Euclidean projection onto the simplex, sometimes referred to, in the context of neural attention, as sparsemax BIBREF19, is defined as
The solution can be characterized through the unique threshold $\tau $ such that $\sum _i \operatornamewithlimits{\mathsf {sparsemax}}(\mathbf {z})_i = 1$ and BIBREF38
Thus, each coordinate of the sparsemax solution is a piecewise-linear function. Visibly, this expression is recovered when setting $\alpha =2$ in the $\alpha $-entmax expression (Equation DISPLAY_FORM85); for other values of $\alpha $, the exponent induces curvature.
On the other hand, the well-known softmax is usually defined through the expression
which can be shown to be the unique solution of the optimization problem
where $\mathsf {H}^\textsc {S}(\mathbf {p}) -\sum _i p_i \log p_i$ is the Shannon entropy. Indeed, setting the gradient to 0 yields the condition $\log p_i = z_j - \nu _i - \tau - 1$, where $\tau $ and $\nu > 0$ are Lagrange multipliers for the simplex constraints $\sum _i p_i = 1$ and $p_i \ge 0$, respectively. Since the l.h.s. is only finite for $p_i>0$, we must have $\nu _i=0$ for all $i$, by complementary slackness. Thus, the solution must have the form $p_i = {\exp (z_i)}{Z}$, yielding Equation DISPLAY_FORM92.
Jacobian of @!START@$\alpha $@!END@-entmax w.r.t. the shape parameter @!START@$\alpha $@!END@: Proof of Proposition @!START@UID22@!END@
Recall that the entmax transformation is defined as:
where $\alpha \ge 1$ and $\mathsf {H}^{\textsc {T}}_{\alpha }$ is the Tsallis entropy,
and $\mathsf {H}^\textsc {S}(\mathbf {p}):= -\sum _j p_j \log p_j$ is the Shannon entropy.
In this section, we derive the Jacobian of $\operatornamewithlimits{\mathsf {entmax }}$ with respect to the scalar parameter $\alpha $.
Jacobian of @!START@$\alpha $@!END@-entmax w.r.t. the shape parameter @!START@$\alpha $@!END@: Proof of Proposition @!START@UID22@!END@ ::: General case of @!START@$\alpha >1$@!END@
From the KKT conditions associated with the optimization problem in Eq. DISPLAY_FORM85, we have that the solution $\mathbf {p}^{\star }$ has the following form, coordinate-wise:
where $\tau ^{\star }$ is a scalar Lagrange multiplier that ensures that $\mathbf {p}^{\star }$ normalizes to 1, i.e., it is defined implicitly by the condition:
For general values of $\alpha $, Eq. DISPLAY_FORM98 lacks a closed form solution. This makes the computation of the Jacobian
non-trivial. Fortunately, we can use the technique of implicit differentiation to obtain this Jacobian.
The Jacobian exists almost everywhere, and the expressions we derive expressions yield a generalized Jacobian BIBREF37 at any non-differentiable points that may occur for certain ($\alpha $, $\mathbf {z}$) pairs. We begin by noting that $\frac{\partial p_i^{\star }}{\partial \alpha } = 0$ if $p_i^{\star } = 0$, because increasing $\alpha $ keeps sparse coordinates sparse. Therefore we need to worry only about coordinates that are in the support of $\mathbf {p}^\star $. We will assume hereafter that the $i$th coordinate of $\mathbf {p}^\star $ is non-zero. We have:
We can see that this Jacobian depends on $\frac{\partial \tau ^{\star }}{\partial \alpha }$, which we now compute using implicit differentiation.
Let $\mathcal {S} = \lbrace i: p^\star _i > 0 \rbrace $). By differentiating both sides of Eq. DISPLAY_FORM98, re-using some of the steps in Eq. DISPLAY_FORM101, and recalling Eq. DISPLAY_FORM97, we get
from which we obtain:
Finally, plugging Eq. DISPLAY_FORM103 into Eq. DISPLAY_FORM101, we get:
where we denote by
The distribution $\tilde{\mathbf {p}}(\alpha )$ can be interpreted as a “skewed” distribution obtained from $\mathbf {p}^{\star }$, which appears in the Jacobian of $\mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}(\mathbf {z})$ w.r.t. $\mathbf {z}$ as well BIBREF14.
Jacobian of @!START@$\alpha $@!END@-entmax w.r.t. the shape parameter @!START@$\alpha $@!END@: Proof of Proposition @!START@UID22@!END@ ::: Solving the indetermination for @!START@$\alpha =1$@!END@
We can write Eq. DISPLAY_FORM104 as
When $\alpha \rightarrow 1^+$, we have $\tilde{\mathbf {p}}(\alpha ) \rightarrow \mathbf {p}^{\star }$, which leads to a $\frac{0}{0}$ indetermination.
To solve this indetermination, we will need to apply L'Hôpital's rule twice. Let us first compute the derivative of $\tilde{p}_i(\alpha )$ with respect to $\alpha $. We have
therefore
Differentiating the numerator and denominator in Eq. DISPLAY_FORM107, we get:
with
and
When $\alpha \rightarrow 1^+$, $B$ becomes again a $\frac{0}{0}$ indetermination, which we can solve by applying again L'Hôpital's rule. Differentiating the numerator and denominator in Eq. DISPLAY_FORM112:
Finally, summing Eq. DISPLAY_FORM111 and Eq. DISPLAY_FORM113, we get
Jacobian of @!START@$\alpha $@!END@-entmax w.r.t. the shape parameter @!START@$\alpha $@!END@: Proof of Proposition @!START@UID22@!END@ ::: Summary
To sum up, we have the following expression for the Jacobian of $\mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}$ with respect to $\alpha $:
| [
"the attention heads in the proposed adaptively sparse Transformer can specialize more and with higher confidence",
"We introduce sparse attention into the Transformer architecture"
] | 4,902 | qasper | en | null | 8b6bf313950a892cbda035f2c7b3d8b01472ff34749f028d | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How does their model improve interpretability compared to softmax transformers?
Answer: | [
"the attention heads in the proposed adaptively sparse Transformer can specialize more and with higher confidence",
"We introduce sparse attention into the Transformer architecture"
] | qasper | 128 | 21 | |
what was the baseline? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Machine translation has made remarkable progress, and studies claiming it to reach a human parity are starting to appear BIBREF0. However, when evaluating translations of the whole documents rather than isolated sentences, human raters show a stronger preference for human over machine translation BIBREF1. These findings emphasize the need to shift towards context-aware machine translation both from modeling and evaluation perspective.
Most previous work on context-aware NMT assumed that either all the bilingual data is available at the document level BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10 or at least its fraction BIBREF11. But in practical scenarios, document-level parallel data is often scarce, which is one of the challenges when building a context-aware system.
We introduce an approach to context-aware machine translation using only monolingual document-level data. In our setting, a separate monolingual sequence-to-sequence model (DocRepair) is used to correct sentence-level translations of adjacent sentences. The key idea is to use monolingual data to imitate typical inconsistencies between context-agnostic translations of isolated sentences. The DocRepair model is trained to map inconsistent groups of sentences into consistent ones. The consistent groups come from the original training data; the inconsistent groups are obtained by sampling round-trip translations for each isolated sentence.
To validate the performance of our model, we use three kinds of evaluation: the BLEU score, contrastive evaluation of translation of several discourse phenomena BIBREF11, and human evaluation. We show strong improvements for all metrics.
We analyze which discourse phenomena are hard to capture using monolingual data only. Using contrastive test sets for targeted evaluation of several contextual phenomena, we compare the performance of the models trained on round-trip translations and genuine document-level parallel data. Among the four phenomena in the test sets we use (deixis, lexical cohesion, VP ellipsis and ellipsis which affects NP inflection) we find VP ellipsis to be the hardest phenomenon to be captured using round-trip translations.
Our key contributions are as follows:
we introduce the first approach to context-aware machine translation using only monolingual document-level data;
our approach shows substantial improvements in translation quality as measured by BLEU, targeted contrastive evaluation of several discourse phenomena and human evaluation;
we show which discourse phenomena are hard to capture using monolingual data only.
Our Approach: Document-level Repair
We propose a monolingual DocRepair model to correct inconsistencies between sentence-level translations of a context-agnostic MT system. It does not use any states of a trained MT model whose outputs it corrects and therefore can in principle be trained to correct translations from any black-box MT system.
The DocRepair model requires only monolingual document-level data in the target language. It is a monolingual sequence-to-sequence model that maps inconsistent groups of sentences into consistent ones. Consistent groups come from monolingual document-level data. To obtain inconsistent groups, each sentence in a group is replaced with its round-trip translation produced in isolation from context. More formally, forming a training minibatch for the DocRepair model involves the following steps (see also Figure FIGREF9):
sample several groups of sentences from the monolingual data;
for each sentence in a group, (i) translate it using a target-to-source MT model, (ii) sample a translation of this back-translated sentence in the source language using a source-to-target MT model;
using these round-trip translations of isolated sentences, form an inconsistent version of the initial groups;
use inconsistent groups as input for the DocRepair model, consistent ones as output.
At test time, the process of getting document-level translations is two-step (Figure FIGREF10):
produce translations of isolated sentences using a context-agnostic MT model;
apply the DocRepair model to a sequence of context-agnostic translations to correct inconsistencies between translations.
In the scope of the current work, the DocRepair model is the standard sequence-to-sequence Transformer. Sentences in a group are concatenated using a reserved token-separator between sentences. The Transformer is trained to correct these long inconsistent pseudo-sentences into consistent ones. The token-separator is then removed from corrected translations.
Evaluation of Contextual Phenomena
We use contrastive test sets for evaluation of discourse phenomena for English-Russian by BIBREF11. These test sets allow for testing different kinds of phenomena which, as we show, can be captured from monolingual data with varying success. In this section, we provide test sets statistics and briefly describe the tested phenomena. For more details, the reader is referred to BIBREF11.
Evaluation of Contextual Phenomena ::: Test sets
There are four test sets in the suite. Each test set contains contrastive examples. It is specifically designed to test the ability of a system to adapt to contextual information and handle the phenomenon under consideration. Each test instance consists of a true example (a sequence of sentences and their reference translation from the data) and several contrastive translations which differ from the true one only in one specific aspect. All contrastive translations are correct and plausible translations at the sentence level, and only context reveals the inconsistencies between them. The system is asked to score each candidate translation, and we compute the system accuracy as the proportion of times the true translation is preferred to the contrastive ones. Test set statistics are shown in Table TABREF15. The suites for deixis and lexical cohesion are split into development and test sets, with 500 examples from each used for validation purposes and the rest for testing. Convergence of both consistency scores on these development sets and BLEU score on a general development set are used as early stopping criteria in models training. For ellipsis, there is no dedicated development set, so we evaluate on all the ellipsis data and do not use it for development.
Evaluation of Contextual Phenomena ::: Phenomena overview
Deixis Deictic words or phrases, are referential expressions whose denotation depends on context. This includes personal deixis (“I”, “you”), place deixis (“here”, “there”), and discourse deixis, where parts of the discourse are referenced (“that's a good question”). The test set examples are all related to person deixis, specifically the T-V distinction between informal and formal you (Latin “tu” and “vos”) in the Russian translations, and test for consistency in this respect.
Ellipsis Ellipsis is the omission from a clause of one or more words that are nevertheless understood in the context of the remaining elements. In machine translation, elliptical constructions in the source language pose a problem in two situations. First, if the target language does not allow the same types of ellipsis, requiring the elided material to be predicted from context. Second, if the elided material affects the syntax of the sentence. For example, in Russian the grammatical function of a noun phrase, and thus its inflection, may depend on the elided verb, or, conversely, the verb inflection may depend on the elided subject.
There are two different test sets for ellipsis. One contains examples where a morphological form of a noun group in the last sentence can not be understood without context beyond the sentence level (“ellipsis (infl.)” in Table TABREF15). Another includes cases of verb phrase ellipsis in English, which does not exist in Russian, thus requires predicting the verb when translating into Russian (“ellipsis (VP)” in Table TABREF15).
Lexical cohesion The test set focuses on reiteration of named entities. Where several translations of a named entity are possible, a model has to prefer consistent translations over inconsistent ones.
Experimental Setup ::: Data preprocessing
We use the publicly available OpenSubtitles2018 corpus BIBREF12 for English and Russian. For a fair comparison with previous work, we train the baseline MT system on the data released by BIBREF11. Namely, our MT system is trained on 6m instances. These are sentence pairs with a relative time overlap of subtitle frames between source and target language subtitles of at least $0.9$.
We gathered 30m groups of 4 consecutive sentences as our monolingual data. We used only documents not containing groups of sentences from general development and test sets as well as from contrastive test sets. The main results we report are for the model trained on all 30m fragments.
We use the tokenization provided by the corpus and use multi-bleu.perl on lowercased data to compute BLEU score. We use beam search with a beam of 4.
Sentences were encoded using byte-pair encoding BIBREF13, with source and target vocabularies of about 32000 tokens. Translation pairs were batched together by approximate sequence length. Each training batch contained a set of translation pairs containing approximately 15000 source tokens. It has been shown that Transformer's performance depends heavily on batch size BIBREF14, and we chose a large batch size to ensure the best performance. In training context-aware models, for early stopping we use both convergence in BLEU score on the general development set and scores on the consistency development sets. After training, we average the 5 latest checkpoints.
Experimental Setup ::: Models
The baseline model, the model used for back-translation, and the DocRepair model are all Transformer base models BIBREF15. More precisely, the number of layers is $N=6$ with $h = 8$ parallel attention layers, or heads. The dimensionality of input and output is $d_{model} = 512$, and the inner-layer of a feed-forward networks has dimensionality $d_{ff}=2048$. We use regularization as described in BIBREF15.
As a second baseline, we use the two-pass CADec model BIBREF11. The first pass produces sentence-level translations. The second pass takes both the first-pass translation and representations of the context sentences as input and returns contextualized translations. CADec requires document-level parallel training data, while DocRepair only needs monolingual training data.
Experimental Setup ::: Generating round-trip translations
On the selected 6m instances we train sentence-level translation models in both directions. To create training data for DocRepair, we proceed as follows. The Russian monolingual data is first translated into English, using the Russian$\rightarrow $English model and beam search with beam size of 4. Then, we use the English$\rightarrow $Russian model to sample translations with temperature of $0{.}5$. For each sentence, we precompute 20 sampled translations and randomly choose one of them when forming a training minibatch for DocRepair. Also, in training, we replace each token in the input with a random one with the probability of $10\%$.
Experimental Setup ::: Optimizer
As in BIBREF15, we use the Adam optimizer BIBREF16, the parameters are $\beta _1 = 0{.}9$, $\beta _2 = 0{.}98$ and $\varepsilon = 10^{-9}$. We vary the learning rate over the course of training using the formula:
where $warmup\_steps = 16000$ and $scale=4$.
Results ::: General results
The BLEU scores are provided in Table TABREF24 (we evaluate translations of 4-sentence fragments). To see which part of the improvement is due to fixing agreement between sentences rather than simply sentence-level post-editing, we train the same repair model at the sentence level. Each sentence in a group is now corrected separately, then they are put back together in a group. One can see that most of the improvement comes from accounting for extra-sentential dependencies. DocRepair outperforms the baseline and CADec by 0.7 BLEU, and its sentence-level repair version by 0.5 BLEU.
Results ::: Consistency results
Scores on the phenomena test sets are provided in Table TABREF26. For deixis, lexical cohesion and ellipsis (infl.) we see substantial improvements over both the baseline and CADec. The largest improvement over CADec (22.5 percentage points) is for lexical cohesion. However, there is a drop of almost 5 percentage points for VP ellipsis. We hypothesize that this is because it is hard to learn to correct inconsistencies in translations caused by VP ellipsis relying on monolingual data alone. Figure FIGREF27(a) shows an example of inconsistency caused by VP ellipsis in English. There is no VP ellipsis in Russian, and when translating auxiliary “did” the model has to guess the main verb. Figure FIGREF27(b) shows steps of generating round-trip translations for the target side of the previous example. When translating from Russian, main verbs are unlikely to be translated as the auxiliary “do” in English, and hence the VP ellipsis is rarely present on the English side. This implies the model trained using the round-trip translations will not be exposed to many VP ellipsis examples in training. We discuss this further in Section SECREF34.
Table TABREF28 provides scores for deixis and lexical cohesion separately for different distances between sentences requiring consistency. It can be seen, that the performance of DocRepair degrades less than that of CADec when the distance between sentences requiring consistency gets larger.
Results ::: Human evaluation
We conduct a human evaluation on random 700 examples from our general test set. We picked only examples where a DocRepair translation is not a full copy of the baseline one.
The annotators were provided an original group of sentences in English and two translations: baseline context-agnostic one and the one corrected by the DocRepair model. Translations were presented in random order with no indication which model they came from. The task is to pick one of the three options: (1) the first translation is better, (2) the second translation is better, (3) the translations are of equal quality. The annotators were asked to avoid the third answer if they are able to give preference to one of the translations. No other guidelines were given.
The results are provided in Table TABREF30. In about $52\%$ of the cases annotators marked translations as having equal quality. Among the cases where one of the translations was marked better than the other, the DocRepair translation was marked better in $73\%$ of the cases. This shows a strong preference of the annotators for corrected translations over the baseline ones.
Varying Training Data
In this section, we discuss the influence of the training data chosen for document-level models. In all experiments, we used the DocRepair model.
Varying Training Data ::: The amount of training data
Table TABREF33 provides BLEU and consistency scores for the DocRepair model trained on different amount of data. We see that even when using a dataset of moderate size (e.g., 5m fragments) we can achieve performance comparable to the model trained on a large amount of data (30m fragments). Moreover, we notice that deixis scores are less sensitive to the amount of training data than lexical cohesion and ellipsis scores. The reason might be that, as we observed in our previous work BIBREF11, inconsistencies in translations due to the presence of deictic words and phrases are more frequent in this dataset than other types of inconsistencies. Also, as we show in Section SECREF7, this is the phenomenon the model learns faster in training.
Varying Training Data ::: One-way vs round-trip translations
In this section, we discuss the limitations of using only monolingual data to model inconsistencies between sentence-level translations. In Section SECREF25 we observed a drop in performance on VP ellipsis for DocRepair compared to CADec, which was trained on parallel data. We hypothesized that this is due to the differences between one-way and round-trip translations, and now we test this hypothesis. To do so, we fix the dataset and vary the way in which the input for DocRepair is generated: round-trip or one-way translations. The latter assumes that document-level data is parallel, and translations are sampled from the source side of the sentences in a group rather than from their back-translations. For parallel data, we take 1.5m parallel instances which were used for CADec training and add 1m instances from our monolingual data. For segments in the parallel part, we either sample translations from the source side or use round-trip translations. The results are provided in Table TABREF35.
The model trained on one-way translations is slightly better than the one trained on round-trip translations. As expected, VP ellipsis is the hardest phenomena to be captured using round-trip translations, and the DocRepair model trained on one-way translated data gains 6% accuracy on this test set. This shows that the DocRepair model benefits from having access to non-synthetic English data. This results in exposing DocRepair at training time to Russian translations which suffer from the same inconsistencies as the ones it will have to correct at test time.
Varying Training Data ::: Filtering: monolingual (no filtering) or parallel
Note that the scores of the DocRepair model trained on 2.5m instances randomly chosen from monolingual data (Table TABREF33) are different from the ones for the model trained on 2.5m instances combined from parallel and monolingual data (Table TABREF35). For convenience, we show these two in Table TABREF36.
The domain, the dataset these two data samples were gathered from, and the way we generated training data for DocRepair (round-trip translations) are all the same. The only difference lies in how the data was filtered. For parallel data, as in the previous work BIBREF6, we picked only sentence pairs with large relative time overlap of subtitle frames between source-language and target-language subtitles. This is necessary to ensure the quality of translation data: one needs groups of consecutive sentences in the target language where every sentence has a reliable translation.
Table TABREF36 shows that the quality of the model trained on data which came from the parallel part is worse than the one trained on monolingual data. This indicates that requiring each sentence in a group to have a reliable translation changes the distribution of the data, which might be not beneficial for translation quality and provides extra motivation for using monolingual data.
Learning Dynamics
Let us now look into how the process of DocRepair training progresses. Figure FIGREF38 shows how the BLEU scores with the reference translation and with the baseline context-agnostic translation (i.e. the input for the DocRepair model) are changing during training. First, the model quickly learns to copy baseline translations: the BLEU score with the baseline is very high. Then it gradually learns to change them, which leads to an improvement in BLEU with the reference translation and a drop in BLEU with the baseline. Importantly, the model is reluctant to make changes: the BLEU score between translations of the converged model and the baseline is 82.5. We count the number of changed sentences in every 4-sentence fragment in the test set and plot the histogram in Figure FIGREF38. In over than 20$\%$ of the cases the model has not changed base translations at all. In almost $40\%$, it modified only one sentence and left the remaining 3 sentences unchanged. The model changed more than half sentences in a group in only $14\%$ of the cases. Several examples of the DocRepair translations are shown in Figure FIGREF43.
Figure FIGREF42 shows how consistency scores are changing in training. For deixis, the model achieves the final quality quite quickly; for the rest, it needs a large number of training steps to converge.
Related Work
Our work is most closely related to two lines of research: automatic post-editing (APE) and document-level machine translation.
Related Work ::: Automatic post-editing
Our model can be regarded as an automatic post-editing system – a system designed to fix systematic MT errors that is decoupled from the main MT system. Automatic post-editing has a long history, including rule-based BIBREF17, statistical BIBREF18 and neural approaches BIBREF19, BIBREF20, BIBREF21.
In terms of architectures, modern approaches use neural sequence-to-sequence models, either multi-source architectures that consider both the original source and the baseline translation BIBREF19, BIBREF20, or monolingual repair systems, as in BIBREF21, which is concurrent work to ours. True post-editing datasets are typically small and expensive to create BIBREF22, hence synthetic training data has been created that uses original monolingual data as output for the sequence-to-sequence model, paired with an automatic back-translation BIBREF23 and/or round-trip translation as its input(s) BIBREF19, BIBREF21.
While previous work on automatic post-editing operated on the sentence level, the main novelty of this work is that our DocRepair model operates on groups of sentences and is thus able to fix consistency errors caused by the context-agnostic baseline MT system. We consider this strategy of sentence-level baseline translation and context-aware monolingual repair attractive when parallel document-level data is scarce.
For training, the DocRepair model only requires monolingual document-level data. While we create synthetic training data via round-trip translation similarly to earlier work BIBREF19, BIBREF21, note that we purposefully use sentence-level MT systems for this to create the types of consistency errors that we aim to fix with the context-aware DocRepair model. Not all types of consistency errors that we want to fix emerge from a round-trip translation, so access to parallel document-level data can be useful (Section SECREF34).
Related Work ::: Document-level NMT
Neural models of MT that go beyond the sentence-level are an active research area BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF10, BIBREF9, BIBREF11. Typically, the main MT system is modified to take additional context as its input. One limitation of these approaches is that they assume that parallel document-level training data is available.
Closest to our work are two-pass models for document-level NMT BIBREF24, BIBREF11, where a second, context-aware model takes the translation and hidden representations of the sentence-level first-pass model as its input. The second-pass model can in principle be trained on a subset of the parallel training data BIBREF11, somewhat relaxing the assumption that all training data is at the document level.
Our work is different from this previous work in two main respects. Firstly, we show that consistency can be improved with only monolingual document-level training data. Secondly, the DocRepair model is decoupled from the first-pass MT system, which improves its portability.
Conclusions
We introduce the first approach to context-aware machine translation using only monolingual document-level data. We propose a monolingual DocRepair model to correct inconsistencies between sentence-level translations. The model performs automatic post-editing on a sequence of sentence-level translations, refining translations of sentences in context of each other. Our approach results in substantial improvements in translation quality as measured by BLEU, targeted contrastive evaluation of several discourse phenomena and human evaluation. Moreover, we perform error analysis and detect which discourse phenomena are hard to capture using only monolingual document-level data. While in the current work we used text fragments of 4 sentences, in future work we would like to consider longer contexts.
Acknowledgments
We would like to thank the anonymous reviewers for their comments. The authors also thank David Talbot and Yandex Machine Translation team for helpful discussions and inspiration. Ivan Titov acknowledges support of the European Research Council (ERC StG BroadSem 678254) and the Dutch National Science Foundation (NWO VIDI 639.022.518). Rico Sennrich acknowledges support from the Swiss National Science Foundation (105212_169888), the European Union’s Horizon 2020 research and innovation programme (grant agreement no 825460), and the Royal Society (NAF\R1\180122).
| [
" MT system on the data released by BIBREF11",
"Transformer base, two-pass CADec model"
] | 3,716 | qasper | en | null | 04af9dc96013a1bc7faecbc589f7ea5c207e92c7d9a3495e | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: what was the baseline?
Answer: | [
" MT system on the data released by BIBREF11",
"Transformer base, two-pass CADec model"
] | qasper | 128 | 22 | |
What metrics are used for evaluation? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Pre-trained models BIBREF0, BIBREF1 have received much of attention recently thanks to their impressive results in many down stream NLP tasks. Additionally, multilingual pre-trained models enable many NLP applications for other languages via zero-short cross-lingual transfer. Zero-shot cross-lingual transfer has shown promising results for rapidly building applications for low resource languages. BIBREF2 show the potential of multilingual-BERT BIBREF0 in zero-shot transfer for a large number of languages from different language families on five NLP tasks, namely, natural language inference, document classification, named entity recognition, part-of-speech tagging, and dependency parsing.
Although multilingual models are an important ingredient for enhancing language technology in many languages, recent research on improving pre-trained models puts much emphasis on English BIBREF3, BIBREF4, BIBREF5. The current state of affairs makes it difficult to translate advancements in pre-training from English to non-English languages. To our best knowledge, there are only three available multilingual pre-trained models to date: (1) the multilingual-BERT (mBERT) that supports 104 languages, (2) cross-lingual language model BIBREF6 that supports 100 languages, and (3) Language Agnostic SEntence Representations BIBREF7 that supports 93 languages. Among the three models, LASER is based on neural machine translation approach and strictly requires parallel data to train.
Do multilingual models always need to be trained from scratch? Can we transfer linguistic knowledge learned by English pre-trained models to other languages? In this work, we develop a technique to rapidly transfer an existing pre-trained model from English to other languages in an energy efficient way BIBREF8. As the first step, we focus on building a bilingual language model (LM) of English and a target language. Starting from a pre-trained English LM, we learn the target language specific parameters (i.e., word embeddings), while keeping the encoder layers of the pre-trained English LM fixed. We then fine-tune both English and target model to obtain the bilingual LM. We apply our approach to autoencoding language models with masked language model objective and show the advantage of the proposed approach in zero-shot transfer. Our main contributions in this work are:
We propose a fast adaptation method for obtaining a bilingual BERT$_{\textsc {base}}$ of English and a target language within a day using one Tesla V100 16GB GPU.
We evaluate our bilingual LMs for six languages on two zero-shot cross-lingual transfer tasks, namely natural language inference BIBREF9 and universal dependency parsing. We show that our models offer competitive performance or even better that mBERT.
We illustrate that our bilingual LMs can serve as an excellent feature extractor in supervised dependency parsing task.
Bilingual Pre-trained LMs
We first provide some background of pre-trained language models. Let $_e$ be English word-embeddings and $\Psi ()$ be the Transformer BIBREF10 encoder with parameters $$. Let $_{w_i}$ denote the embedding of word $w_i$ (i.e., $_{w_i} = _e[w_1]$). We omit positional embeddings and bias for clarity. A pre-trained LM typically performs the following computations: (i) transform a sequence of input tokens to contextualized representations $[_{w_1},\dots ,_{w_n}] = \Psi (_{w_1}, \dots , _{w_n}; )$, and (ii) predict an output word $y_i$ at $i^{\text{th}}$ position $p(y_i | _{w_i}) \propto \exp (_{w_i}^\top _{y_i})$.
Autoencoding LM BIBREF0 corrupts some input tokens $w_i$ by replacing them with a special token [MASK]. It then predicts the original tokens $y_i = w_i$ from the corrupted tokens. Autoregressive LM BIBREF3 predicts the next token ($y_i = w_{i+1}$) given all the previous tokens. The recently proposed XLNet model BIBREF5 is an autoregressive LM that factorizes output with all possible permutations, which shows empirical performance improvement over GPT-2 due to the ability to capture bidirectional context. Here we assume that the encoder performs necessary masking with respect to each training objective.
Given an English pre-trained LM, we wish to learn a bilingual LM for English and a given target language $f$ under a limited computational resource budget. To quickly build a bilingual LM, we directly adapt the English pre-traind model to the target model. Our approach consists of three steps. First, we initialize target language word-embeddings $_f$ in the English embedding space such that embeddings of a target word and its English equivalents are close together (§SECREF8). Next, we create a target LM from the target embeddings and the English encoder $\Psi ()$. We then fine-tune target embeddings while keeping $\Psi ()$ fixed (§SECREF14). Finally, we construct a bilingual LM of $_e$, $_f$, and $\Psi ()$ and fine-tune all the parameters (§SECREF15). Figure FIGREF7 illustrates the last two steps in our approach.
Bilingual Pre-trained LMs ::: Initializing Target Embeddings
Our approach to learn the initial foreign word embeddings $_f \in ^{|V_f| \times d}$ is based on the idea of mapping the trained English word embeddings $_e \in ^{|V_e| \times d}$ to $_f$ such that if a foreign word and an English word are similar in meaning then their embeddings are similar. Borrowing the idea of universal lexical sharing from BIBREF11, we represent each foreign word embedding $_f[i] \in ^d$ as a linear combination of English word embeddings $_e[j] \in ^d$
where $_i\in ^{|V_e|}$ is a sparse vector and $\sum _j^{|V_e|} \alpha _{ij} = 1$.
In this step of initializing foreign embeddings, having a good estimation of $$ could speed of the convergence when tuning the foreign model and enable zero-shot transfer (§SECREF5). In the following, we discuss how to estimate $_i\;\forall i\in \lbrace 1,2, \dots , |V_f|\rbrace $ under two scenarios: (i) we have parallel data of English-foreign, and (ii) we only rely on English and foreign monolingual data.
Bilingual Pre-trained LMs ::: Initializing Target Embeddings ::: Learning from Parallel Corpus
Given an English-foreign parallel corpus, we can estimate word translation probability $p(e\,|\,f)$ for any (English-foreign) pair $(e, f)$ using popular word-alignment BIBREF12 toolkits such as fast-align BIBREF13. We then assign:
Since $_i$ is estimated from word alignment, it is a sparse vector.
Bilingual Pre-trained LMs ::: Initializing Target Embeddings ::: Learning from Monolingual Corpus
For low resource languages, parallel data may not be available. In this case, we rely only on monolingual data (e.g., Wikipedias). We estimate word translation probabilities from word embeddings of the two languages. Word vectors of these languages can be learned using fastText BIBREF14 and then are aligned into a shared space with English BIBREF15, BIBREF16. Unlike learning contextualized representations, learning word vectors is fast and computationally cheap. Given the aligned vectors $\bar{}_f$ of foreign and $\bar{}_e$ of English, we calculate the word translation matrix $\in ^{|V_f|\times |V_e|}$ as
Here, we use $\mathrm {sparsemax}$ BIBREF17 instead of softmax. Sparsemax is a sparse version of softmax and it puts zero probabilities on most of the word in the English vocabulary except few English words that are similar to a given foreign word. This property is desirable in our approach since it leads to a better initialization of the foreign embeddings.
Bilingual Pre-trained LMs ::: Fine-tuning Target Embeddings
After initializing foreign word-embeddings, we replace English word-embeddings in the English pre-trained LM with foreign word-embeddings to obtain the foreign LM. We then fine-tune only foreign word-embeddings on monolingual data. The training objective is the same as the training objective of the English pre-trained LM (i.e., masked LM for BERT). Since the trained encoder $\Psi ()$ is good at capturing association, the purpose of this step is to further optimize target embeddings such that the target LM can utilized the trained encoder for association task. For example, if the words Albert Camus presented in a French input sequence, the self-attention in the encoder more likely attends to words absurde and existentialisme once their embeddings are tuned.
Bilingual Pre-trained LMs ::: Fine-tuning Bilingual LM
We create a bilingual LM by plugging foreign language specific parameters to the pre-trained English LM (Figure FIGREF7). The new model has two separate embedding layers and output layers, one for English and one for foreign language. The encoder layer in between is shared. We then fine-tune this model using English and foreign monolingual data. Here, we keep tuning the model on English to ensure that it does not forget what it has learned in English and that we can use the resulting model for zero-shot transfer (§SECREF3). In this step, the encoder parameters are also updated so that in can learn syntactic aspects (i.e., word order, morphological agreement) of the target languages.
Zero-shot Experiments
We build our bilingual LMs, named RAMEN, starting from BERT$_{\textsc {base}}$, BERT$_{\textsc {large}}$, RoBERTa$_{\textsc {base}}$, and RoBERTa$_{\textsc {large}}$ pre-trained models. Using BERT$_{\textsc {base}}$ allows us to compare the results with mBERT model. Using BERT$_{\textsc {large}}$ and RoBERTa allows us to investigate whether the performance of the target LM correlates with the performance of the source LM. We evaluate our models on two cross-lingual zero-shot tasks: (1) Cross-lingual Natural Language Inference (XNLI) and (2) dependency parsing.
Zero-shot Experiments ::: Data
We evaluate our approach for six target languages: French (fr), Russian (ru), Arabic (ar), Chinese (zh), Hindi (hi), and Vietnamese (vi). These languages belong to four different language families. French, Russian, and Hindi are Indo-European languages, similar to English. Arabic, Chinese, and Vietnamese belong to Afro-Asiatic, Sino-Tibetan, and Austro-Asiatic family respectively. The choice of the six languages also reflects different training conditions depending on the amount of monolingual data. French and Russian, and Arabic can be regarded as high resource languages whereas Hindi has far less data and can be considered as low resource.
For experiments that use parallel data to initialize foreign specific parameters, we use the same datasets in the work of BIBREF6. Specifically, we use United Nations Parallel Corpus BIBREF18 for en-ru, en-ar, en-zh, and en-fr. We collect en-hi parallel data from IIT Bombay corpus BIBREF19 and en-vi data from OpenSubtitles 2018. For experiments that use only monolingual data to initialize foreign parameters, instead of training word-vectors from the scratch, we use the pre-trained word vectors from fastText BIBREF14 to estimate word translation probabilities (Eq. DISPLAY_FORM13). We align these vectors into a common space using orthogonal Procrustes BIBREF20, BIBREF15, BIBREF16. We only use identical words between the two languages as the supervised signal. We use WikiExtractor to extract extract raw sentences from Wikipedias as monolingual data for fine-tuning target embeddings and bilingual LMs (§SECREF15). We do not lowercase or remove accents in our data preprocessing pipeline.
We tokenize English using the provided tokenizer from pre-trained models. For target languages, we use fastBPE to learn 30,000 BPE codes and 50,000 codes when transferring from BERT and RoBERTa respectively. We truncate the BPE vocabulary of foreign languages to match the size of the English vocabulary in the source models. Precisely, the size of foreign vocabulary is set to 32,000 when transferring from BERT and 50,000 when transferring from RoBERTa.
We use XNLI dataset BIBREF9 for classification task and Universal Dependencies v2.4 BIBREF21 for parsing task. Since a language might have more than one treebank in Universal Dependencies, we use the following treebanks: en_ewt (English), fr_gsd (French), ru_syntagrus (Russian) ar_padt (Arabic), vi_vtb (Vietnamese), hi_hdtb (Hindi), and zh_gsd (Chinese).
Zero-shot Experiments ::: Data ::: Remark on BPE
BIBREF22 show that sharing subwords between languages improves alignments between embedding spaces. BIBREF2 observe a strong correlation between the percentage of overlapping subwords and mBERT's performances for cross-lingual zero-shot transfer. However, in our current approach, subwords between source and target are not shared. A subword that is in both English and foreign vocabulary has two different embeddings.
Zero-shot Experiments ::: Estimating translation probabilities
Since pre-trained models operate on subword level, we need to estimate subword translation probabilities. Therefore, we subsample 2M sentence pairs from each parallel corpus and tokenize the data into subwords before running fast-align BIBREF13.
Estimating subword translation probabilities from aligned word vectors requires an additional processing step since the provided vectors from fastText are not at subword level. We use the following approximation to obtain subword vectors: the vector $_s$ of subword $s$ is the weighted average of all the aligned word vectors $_{w_i}$ that have $s$ as an subword
where $p(w_j)$ is the unigram probability of word $w_j$ and $n_s = \sum _{w_j:\, s\in w_j} p(w_j)$. We take the top 50,000 words in each aligned word-vectors to compute subword vectors.
In both cases, not all the words in the foreign vocabulary can be initialized from the English word-embeddings. Those words are initialized randomly from a Gaussian $\mathcal {N}(0, {1}{d^2})$.
Zero-shot Experiments ::: Hyper-parameters
In all the experiments, we tune RAMEN$_{\textsc {base}}$ for 175,000 updates and RAMEN$_{\textsc {large}}$ for 275,000 updates where the first 25,000 updates are for language specific parameters. The sequence length is set to 256. The mini-batch size are 64 and 24 when tuning language specific parameters using RAMEN$_{\textsc {base}}$ and RAMEN$_{\textsc {large}}$ respectively. For tuning bilingual LMs, we use a mini-batch size of 64 for RAMEN$_{\textsc {base}}$ and 24 for RAMEN$_{\textsc {large}}$ where half of the batch are English sequences and the other half are foreign sequences. This strategy of balancing mini-batch has been used in multilingual neural machine translation BIBREF23, BIBREF24.
We optimize RAMEN$_{\textsc {base}}$ using Lookahead optimizer BIBREF25 wrapped around Adam with the learning rate of $10^{-4}$, the number of fast weight updates $k=5$, and interpolation parameter $\alpha =0.5$. We choose Lookahead optimizer because it has been shown to be robust to the initial parameters of the based optimizer (Adam). For Adam optimizer, we linearly increase the learning rate from $10^{-7}$ to $10^{-4}$ in the first 4000 updates and then follow an inverse square root decay. All RAMEN$_{\textsc {large}}$ models are optimized with Adam due to memory limit.
When fine-tuning RAMEN on XNLI and UD, we use a mini-batch size of 32, Adam's learning rate of $10^{-5}$. The number of epochs are set to 4 and 50 for XNLI and UD tasks respectively. All experiments are carried out on a single Tesla V100 16GB GPU. Each RAMEN$_{\textsc {base}}$ model is trained within a day and each RAMEN$_{\textsc {large}}$ is trained within two days.
Results
In this section, we present the results of out models for two zero-shot cross lingual transfer tasks: XNLI and dependency parsing.
Results ::: Cross-lingual Natural Language Inference
Table TABREF32 shows the XNLI test accuracy. For reference, we also include the scores from the previous work, notably the state-of-the-art system XLM BIBREF6. Before discussing the results, we spell out that the fairest comparison in this experiment is the comparison between mBERT and RAMEN$_{\textsc {base}}$+BERT trained with monolingual only.
We first discuss the transfer results from BERT. Initialized from fastText vectors, RAMEN$_{\textsc {base}}$ slightly outperforms mBERT by 1.9 points on average and widen the gap of 3.3 points on Arabic. RAMEN$_{\textsc {base}}$ gains extra 0.8 points on average when initialized from parallel data. With triple number of parameters, RAMEN$_{\textsc {large}}$ offers an additional boost in term of accuracy and initialization with parallel data consistently improves the performance. It has been shown that BERT$_{\textsc {large}}$ significantly outperforms BERT$_{\textsc {base}}$ on 11 English NLP tasks BIBREF0, the strength of BERT$_{\textsc {large}}$ also shows up when adapted to foreign languages.
Transferring from RoBERTa leads to better zero-shot accuracies. With the same initializing condition, RAMEN$_{\textsc {base}}$+RoBERTa outperforms RAMEN$_{\textsc {base}}$+BERT on average by 2.9 and 2.3 points when initializing from monolingual and parallel data respectively. This result show that with similar number of parameters, our approach benefits from a better English pre-trained model. When transferring from RoBERTa$_{\textsc {large}}$, we obtain state-of-the-art results for five languages.
Currently, RAMEN only uses parallel data to initialize foreign embeddings. RAMEN can also exploit parallel data through translation objective proposed in XLM. We believe that by utilizing parallel data during the fine-tuning of RAMEN would bring additional benefits for zero-shot tasks. We leave this exploration to future work. In summary, starting from BERT$_{\textsc {base}}$, our approach obtains competitive bilingual LMs with mBERT for zero-shot XNLI. Our approach shows the accuracy gains when adapting from a better pre-trained model.
Results ::: Universal Dependency Parsing
We build on top of RAMEN a graph-based dependency parser BIBREF27. For the purpose of evaluating the contextual representations learned by our model, we do not use part-of-speech tags. Contextualized representations are directly fed into Deep-Biaffine layers to predict arc and label scores. Table TABREF34 presents the Labeled Attachment Scores (LAS) for zero-shot dependency parsing.
We first look at the fairest comparison between mBERT and monolingually initialized RAMEN$_{\textsc {base}}$+BERT. The latter outperforms the former on five languages except Arabic. We observe the largest gain of +5.2 LAS for French. Chinese enjoys +3.1 LAS from our approach. With similar architecture (12 or 24 layers) and initialization (using monolingual or parallel data), RAMEN+RoBERTa performs better than RAMEN+BERT for most of the languages. Arabic and Hindi benefit the most from bigger models. For the other four languages, RAMEN$_{\textsc {large}}$ renders a modest improvement over RAMEN$_{\textsc {base}}$.
Analysis ::: Impact of initialization
Initializing foreign embeddings is the backbone of our approach. A good initialization leads to better zero-shot transfer results and enables fast adaptation. To verify the importance of a good initialization, we train a RAMEN$_{\textsc {base}}$+RoBERTa with foreign word-embeddings are initialized randomly from $\mathcal {N}(0, {1}{d^2})$. For a fair comparison, we use the same hyper-parameters in §SECREF27. Table TABREF36 shows the results of XNLI and UD parsing of random initialization. In comparison to the initialization using aligned fastText vectors, random initialization decreases the zero-shot performance of RAMEN$_{\textsc {base}}$ by 15.9% for XNLI and 27.8 points for UD parsing on average. We also see that zero-shot parsing of SOV languages (Arabic and Hindi) suffers random initialization.
Analysis ::: Are contextual representations from RAMEN also good for supervised parsing?
All the RAMEN models are built from English and tuned on English for zero-shot cross-lingual tasks. It is reasonable to expect RAMENs do well in those tasks as we have shown in our experiments. But are they also a good feature extractor for supervised tasks? We offer a partial answer to this question by evaluating our model for supervised dependency parsing on UD datasets.
We used train/dev/test splits provided in UD to train and evaluate our RAMEN-based parser. Table TABREF38 summarizes the results (LAS) of our supervised parser. For a fair comparison, we choose mBERT as the baseline and all the RAMEN models are initialized from aligned fastText vectors. With the same architecture of 12 Transformer layers, RAMEN$_{\textsc {base}}$+BERT performs competitive to mBERT and outshines mBERT by +1.2 points for Vietnamese. The best LAS results are obtained by RAMEN$_{\textsc {large}}$+RoBERTa with 24 Transformer layers. Overall, our results indicate the potential of using contextual representations from RAMEN for supervised tasks.
Analysis ::: How does linguistic knowledge transfer happen through each training stages?
We evaluate the performance of RAMEN+RoBERTa$_{\textsc {base}}$ (initialized from monolingual data) at each training steps: initialization of word embeddings (0K update), fine-tuning target embeddings (25K), and fine-tuning the model on both English and target language (at each 25K updates). The results are presented in Figure FIGREF40.
Without fine-tuning, the average accuracy of XLNI is 39.7% for a three-ways classification task, and the average LAS score is 3.6 for dependency parsing. We see the biggest leap in the performance after 50K updates. While semantic similarity task profits significantly at 25K updates of the target embeddings, syntactic task benefits with further fine-tuning the encoder. This is expected since the target languages might exhibit different syntactic structures than English and fine-tuning encoder helps to capture language specific structures. We observe a substantial gain of 19-30 LAS for all languages except French after 50K updates.
Language similarities have more impact on transferring syntax than semantics. Without tuning the English encoder, French enjoys 50.3 LAS for being closely related to English, whereas Arabic and Hindi, SOV languages, modestly reach 4.2 and 6.4 points using the SVO encoder. Although Chinese has SVO order, it is often seen as head-final while English is strong head-initial. Perhaps, this explains the poor performance for Chinese.
Limitations
While we have successfully adapted autoencoding pre-trained LMs from English to other languages, the question whether our approach can also be applied for autoregressive LM such as XLNet still remains. We leave the investigation to future work.
Conclusions
In this work, we have presented a simple and effective approach for rapidly building a bilingual LM under a limited computational budget. Using BERT as the starting point, we demonstrate that our approach produces better than mBERT on two cross-lingual zero-shot sentence classification and dependency parsing. We find that the performance of our bilingual LM, RAMEN, correlates with the performance of the original pre-trained English models. We also find that RAMEN is also a powerful feature extractor in supervised dependency parsing. Finally, we hope that our work sparks of interest in developing fast and effective methods for transferring pre-trained English models to other languages.
| [
"translation probabilities, Labeled Attachment Scores (LAS)",
"accuracy, Labeled Attachment Scores (LAS)"
] | 3,402 | qasper | en | null | 3b069a546d5417be546f57d02981cf5ac68b4d3d6c55624a | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What metrics are used for evaluation?
Answer: | [
"translation probabilities, Labeled Attachment Scores (LAS)",
"accuracy, Labeled Attachment Scores (LAS)"
] | qasper | 128 | 23 | |
What is the attention module pretrained on? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Speech-to-Text translation (ST) is essential for a wide range of scenarios: for example in emergency calls, where agents have to respond emergent requests in a foreign language BIBREF0; or in online courses, where audiences and speakers use different languages BIBREF1. To tackle this problem, existing approaches can be categorized into cascaded method BIBREF2, BIBREF3, where a machine translation (MT) model translates outputs of an automatic speech recognition (ASR) system into target language, and end-to-end method BIBREF4, BIBREF5, where a single model learns acoustic frames to target word sequence mappings in one step towards the final objective of interest. Although the cascaded model remains the dominant approach due to its better performance, the end-to-end method becomes more and more popular because it has lower latency by avoiding inferences with two models and rectifies the error propagation in theory.
Since it is hard to obtain a large-scale ST dataset, multi-task learning BIBREF5, BIBREF6 and pre-training techniques BIBREF7 have been applied to end-to-end ST model to leverage large-scale datasets of ASR and MT. A common practice is to pre-train two encoder-decoder models for ASR and MT respectively, and then initialize the ST model with the encoder of the ASR model and the decoder of the MT model. Subsequently, the ST model is optimized with the multi-task learning by weighing the losses of ASR, MT, and ST. This approach, however, causes a huge gap between pre-training and fine-tuning, which are summarized into three folds:
Subnet Waste: The ST system just reuses the ASR encoder and the MT decoder, while discards other pre-trained subnets, such as the MT encoder. Consequently, valuable semantic information captured by the MT encoder cannot be inherited by the final ST system.
Role Mismatch: The speech encoder plays different roles in pre-training and fine-tuning. The encoder is a pure acoustic model in pre-training, while it has to extract semantic and linguistic features additionally in fine-tuning, which significantly increases the learning difficulty.
Non-pre-trained Attention Module: Previous work BIBREF6 trains attention modules for ASR, MT and ST respectively, hence, the attention module of ST does not benefit from the pre-training.
To address these issues, we propose a Tandem Connectionist Encoding Network (TCEN), which is able to reuse all subnets in pre-training, keep the roles of subnets consistent, and pre-train the attention module. Concretely, the TCEN consists of three components, a speech encoder, a text encoder, and a target text decoder. Different from the previous work that pre-trains an encoder-decoder based ASR model, we only pre-train an ASR encoder by optimizing the Connectionist Temporal Classification (CTC) BIBREF8 objective function. In this way, the additional decoder of ASR is not required while keeping the ability to read acoustic features into the source language space by the speech encoder. Besides, the text encoder and decoder can be pre-trained on a large MT dataset. After that, we employ common used multi-task learning method to jointly learn ASR, MT and ST tasks.
Compared to prior works, the encoder of TCEN is a concatenation of an ASR encoder and an MT encoder and our model does not have an ASR decoder, so the subnet waste issue is solved. Furthermore, the two encoders work at tandem, disentangling acoustic feature extraction and linguistic feature extraction, ensuring the role consistency between pre-training and fine-tuning. Moreover, we reuse the pre-trained MT attention module in ST, so we can leverage the alignment information learned in pre-training.
Since the text encoder consumes word embeddings of plausible texts in MT task but uses speech encoder outputs in ST task, another question is how one guarantees the speech encoder outputs are consistent with the word embeddings. We further modify our model to achieve semantic consistency and length consistency. Specifically, (1) the projection matrix at the CTC classification layer for ASR is shared with the word embedding matrix, ensuring that they are mapped to the same latent space, and (2) the length of the speech encoder output is proportional to the length of the input frame, so it is much longer than a natural sentence. To bridge the length gap, source sentences in MT are lengthened by adding word repetitions and blank tokens to mimic the CTC output sequences.
We conduct comprehensive experiments on the IWSLT18 speech translation benchmark BIBREF1, demonstrating the effectiveness of each component. Our model is significantly better than previous methods by 3.6 and 2.2 BLEU scores for the subword-level decoding and character-level decoding strategies, respectively.
Our contributions are three-folds: 1) we shed light on why previous ST models cannot sufficiently utilize the knowledge learned from the pre-training process; 2) we propose a new ST model, which alleviates shortcomings in existing methods; and 3) we empirically evaluate the proposed model on a large-scale public dataset.
Background ::: Problem Formulation
End-to-end speech translation aims to translate a piece of audio into a target-language translation in one step. The raw speech signals are usually converted to sequences of acoustic features, e.g. Mel filterbank features. Here, we define the speech feature sequence as $\mathbf {x} = (x_1, \cdots , x_{T_x})$.The transcription and translation sequences are denoted as $\mathbf {y^{s}} = (y_1^{s}, \cdots , y_{T_s}^{s})$, and $\mathbf {y^{t}} = (y_1^{t}, \cdots , y_{T_t}^{t})$ repectively. Each symbol in $\mathbf {y^{s}}$ or $\mathbf {y^{t}}$ is an integer index of the symbol in a vocabulary $V_{src}$ or $V_{trg}$ respectively (e.g. $y^s_i=k, k\in [0, |V_{src}|-1]$). In this work, we suppose that an ASR dataset, an MT dataset, and a ST dataset are available, denoted as $\mathcal {A} = \lbrace (\mathbf {x_i}, \mathbf {y^{s}_i})\rbrace _{i=0}^I$, $\mathcal {M} =\lbrace (\mathbf {y^{s}_j}, \mathbf {y^{t}_j})\rbrace _{j=0}^J$ and $ \mathcal {S} =\lbrace (\mathbf {x_l}, \mathbf {y^{t}_l})\rbrace _{l=0}^L$ respectively. Given a new piece of audio $\mathbf {x}$, our goal is to learn an end to end model to generate a translation sentence $\mathbf {y^{t}}$ without generating an intermediate result $\mathbf {y^{s}}$.
Background ::: Multi-Task Learning and Pre-training for ST
To leverage large scale ASR and MT data, multi-task learning and pre-training techniques are widely employed to improve the ST system. As shown in Figure FIGREF4, there are three popular multi-task strategies for ST, including 1) one-to-many setting, in which a speech encoder is shared between ASR and ST tasks; 2) many-to-one setting in which a decoder is shared between MT and ST tasks; and 3) many-to-many setting where both the encoder and decoder are shared.
A many-to-many multi-task model contains two encoders as well as two decoders. It can be jointly trained on ASR, MT, and ST tasks. As the attention module is task-specific, three attentions are defined.
Usually, the size of $\mathcal {A}$ and $\mathcal {M}$ is much larger than $\mathcal {S}$. Therefore, the common training practice is to pre-train the model on ASR and MT tasks and then fine-tune it with a multi-task learning manner. However, as aforementioned, this method suffers from subnet waste, role mismatch and non-pre-trained attention issues, which severely limits the end-to-end ST performance.
Our method
In this section, we first introduce the architecture of TCEN, which consists of two encoders connected in tandem, and one decoder with an attention module. Then we give the pre-training and fine-tuning strategy for TCEN. Finally, we propose our solutions for semantic and length inconsistency problems, which are caused by multi-task learning.
Our method ::: TCEN Architecture
Figure FIGREF5 sketches the overall architecture of TCEN, including a speech encoder $enc_s$, a text encoder $enc_t$ and a decoder $dec$ with an attention module $att$. During training, the $enc_s$ acts like an acoustic model which reads the input $\mathbf {x}$ to word or subword representations $\mathbf {h^s}$, then $enc_t$ learns high-level linguistic knowledge into hidden representations $\mathbf {h^t}$. Finally, the $dec$ defines a distribution probability over target words. The advantage of our architecture is that two encoders disentangle acoustic feature extraction and linguistic feature extraction, making sure that valuable knowledge learned from ASR and MT tasks can be effectively leveraged for ST training. Besides, every module in pre-training can be utilized in fine-tuning, alleviating the subnet waste problem.
Follow BIBREF9 inaguma2018speech, we use CNN-BiLSTM architecture to build our model. Specifically, the input features $\mathbf {x}$ are organized as a sequence of feature vectors in length $T_x$. Then, $\mathbf {x}$ is passed into a stack of two convolutional layers followed by max-pooling:
where $\mathbf {v}^{(l-1)}$ is feature maps in last layer and $\mathbf {W}^{(l)}$ is the filter. The max-pooling layers downsample the sequence in length by a total factor of four. The down-sampled feature sequence is further fed into a stack of five bidirectional $d$-dimensional LSTM layers:
where $[;]$ denotes the vector concatenation. The final output representation from the speech encoder is denoted as $\mathbf {h^s}=(h^s_1, \cdots , h^s_{\frac{T_x}{4}})$, where $h_i^s \in \mathbb {R}^d$.
The text encoder $enc_t$ consists of two bidirectional LSTM layers. In ST task, $enc_t$ accepts speech encoder output $\mathbf {h}^s$ as input. While in MT, $enc_t$ consumes the word embedding representation $\mathbf {e^s}$ derived from $\mathbf {y^s}$, where each element $e^s_i$ is computed by choosing the $y_i^s$-th vector from the source embedding matrix $W_{E^s}$. The goal of $enc_t$ is to extract high-level linguistic features like syntactic features or semantic features from lower level subword representations $\mathbf {h}^s$ or $\mathbf {e}^s$. Since $\mathbf {h}^s$ and $\mathbf {e}^s$ belong to different latent space and have different lengths, there remain semantic and length inconsistency problems. We will provide our solutions in Section SECREF21. The output sequence of $enc_t$ is denoted as $\mathbf {h}^t$.
The decoder is defined as two unidirectional LSTM layers with an additive attention $att$. It predicts target sequence $\mathbf {y^{t}}$ by estimating conditional probability $P(\mathbf {y^{t}}|\mathbf {x})$:
Here, $z_k$ is the the hidden state of the deocder RNN at $k$ step and $c_k$ is a time-dependent context vector computed by the attention $att$.
Our method ::: Training Procedure
Following previous work, we split the training procedure to pre-training and fine-tuning stages. In pre-training stage, the speech encoder $enc_s$ is trained towards CTC objective using dataset $\mathcal {A}$, while the text encoder $enc_t$ and the decoder $dec$ are trained on MT dataset $\mathcal {M}$. In fine-tuning stage, we jointly train the model on ASR, MT, and ST tasks.
Our method ::: Training Procedure ::: Pre-training
To sufficiently utilize the large dataset $\mathcal {A}$ and $\mathcal {M}$, the model is pre-trained on CTC-based ASR task and MT task in the pre-training stage.
For ASR task, in order to get rid of the requirement for decoder and enable the $enc_s$ to generate subword representation, we leverage connectionist temporal classification (CTC) BIBREF8 loss to train the speech encoder.
Given an input $\mathbf {x}$, $enc_s$ emits a sequence of hidden vectors $\mathbf {h^s}$, then a softmax classification layer predicts a CTC path $\mathbf {\pi }$, where $\pi _t \in V_{src} \cup $ {`-'} is the observing label at particular RNN step $t$, and `-' is the blank token representing no observed labels:
where $W_{ctc} \in \mathbb {R}^{d \times (|V_{src}|+1)}$ is the weight matrix in the classification layer and $T$ is the total length of encoder RNN.
A legal CTC path $\mathbf {\pi }$ is a variation of the source transcription $\mathbf {y}^s$ by allowing occurrences of blank tokens and repetitions, as shown in Table TABREF14. For each transcription $\mathbf {y}$, there exist many legal CTC paths in length $T$. The CTC objective trains the model to maximize the probability of observing the golden sequence $\mathbf {y}^s$, which is calculated by summing the probabilities of all possible legal paths:
where $\Phi _T(y)$ is the set of all legal CTC paths for sequence $\mathbf {y}$ with length $T$. The loss can be easily computed using forward-backward algorithm. More details about CTC are provided in supplementary material.
For MT task, we use the cross-entropy loss as the training objective. During training, $\mathbf {y^s}$ is converted to embedding vectors $\mathbf {e^s}$ through embedding layer $W_{E^s}$, then $enc_t$ consumes $\mathbf {e^s}$ and pass the output $\mathbf {h^t}$ to decoder. The objective function is defined as:
Our method ::: Training Procedure ::: Fine-tune
In fine-tune stage, we jointly update the model on ASR, MT, and ST tasks. The training for ASR and MT follows the same process as it was in pre-training stage.
For ST task, the $enc_s$ reads the input $\mathbf {x}$ and generates $\mathbf {h^s}$, then $enc_t$ learns high-level linguistic knowledge into $\mathbf {h^t}$. Finally, the $dec$ predicts the target sentence. The ST loss function is defined as:
Following the update strategy proposed by BIBREF11 luong2015multi, we allocate a different training ratio $\alpha _i$ for each task. When switching between tasks, we select randomly a new task $i$ with probability $\frac{\alpha _i}{\sum _{j}\alpha _{j}}$.
Our method ::: Subnet-Consistency
Our model keeps role consistency between pre-training and fine-tuning by connecting two encoders for ST task. However, this leads to some new problems: 1) The text encoder consumes $\mathbf {e^s}$ during MT training, while it accepts $\mathbf {h^s}$ during ST training. However, $\mathbf {e^s}$ and $\mathbf {h^s}$ may not follow the same distribution, resulting in the semantic inconsistency. 2) Besides, the length of $\mathbf {h^s}$ is not the same order of magnitude with the length of $\mathbf {e^s}$, resulting in the length inconsistency.
In response to the above two challenges, we propose two countermeasures: 1) We share weights between CTC classification layer and source-end word embedding layer during training of ASR and MT, encouraging $\mathbf {e^s}$ and $\mathbf {h^s}$ in the same space. 2)We feed the text encoder source sentences in the format of CTC path, which are generated from a seq2seq model, making it more robust toward long inputs.
Our method ::: Subnet-Consistency ::: Semantic Consistency
As shown in Figure FIGREF5, during multi-task training, two different hidden features will be fed into the text encoder $enc_t$: the embedding representation $\mathbf {e}^s$ in MT task, and the $enc_s$ output $\mathbf {h^s}$ in ST task. Without any regularization, they may belong to different latent spaces. Due to the space gap, the $enc_t$ has to compromise between two tasks, limiting its performance on individual tasks.
To bridge the space gap, our idea is to pull $\mathbf {h^s}$ into the latent space where $\mathbf {e}^s$ belong. Specifically, we share the weight $W_{ctc}$ in CTC classification layer with the source embedding weights $W_{E^s}$, which means $W_{ctc} = W_{E^s}$. In this way, when predicting the CTC path $\mathbf {\pi }$, the probability of observing the particular label $w_i \in V_{src}\cup ${`-'} at time step $t$, $p(\pi _t=w_i|\mathbf {x})$, is computed by normalizing the product of hidden vector $h_t^s$ and the $i$-th vector in $W_{E^s}$:
The loss function closes the distance between $h^s_t$ and golden embedding vector, encouraging $\mathbf {h}^s$ have the same distribution with $\mathbf {e}^s$.
Our method ::: Subnet-Consistency ::: Length Consistency
Another existing problem is length inconsistency. The length of the sequence $\mathbf {h^s}$ is proportional to the length of the input frame $\mathbf {x}$, which is much longer than the length of $\mathbf {e^s}$. To solve this problem, we train an RNN-based seq2seq model to transform normal source sentences to noisy sentences in CTC path format, and replace standard MT with denoising MT for multi-tasking.
Specifically, we first train a CTC ASR model based on dataset $\mathcal {A} = \lbrace (\mathbf {x}_i, \mathbf {y}^s_i)\rbrace _{i=0}^{I}$, and generate a CTC-path $\mathbf {\pi }_i$ for each audio $\mathbf {x}_i$ by greedy decoding. Then we define an operation $S(\cdot )$, which converts a CTC path $\mathbf {\pi }$ to a sequence of the unique tokens $\mathbf {u}$ and a sequence of repetition times for each token $\mathbf {l}$, denoted as $S(\mathbf {\pi }) = (\mathbf {u}, \mathbf {l})$. Notably, the operation is reversible, meaning that $S^{-1} (\mathbf {u}, \mathbf {l})=\mathbf {\pi }$. We use the example $\mathbf {\pi _1}$ in Table TABREF14 and show the corresponding $\mathbf {u}$ and $\mathbf {l}$ in Table TABREF24.
Then we build a dataset $\mathcal {P} = \lbrace (\mathbf {y^s}_i, \mathbf {u}_i, \mathbf {l}_i)\rbrace _{i=0}^{I}$ by decoding all the audio pieces in $\mathcal {A}$ and transform the resulting path by the operation $S(\cdot )$. After that, we train a seq2seq model, as shown in Figure FIGREF25, which takes $ \mathbf {y^s}_i$ as input and decodes $\mathbf {u}_i, \mathbf {l}_i$ as outputs. With the seq2seq model, a noisy MT dataset $\mathcal {M}^{\prime }=\lbrace (\mathbf {\pi }_l, \mathbf {y^t}_l)\rbrace _{l=0}^{L}$ is obtained by converting every source sentence $\mathbf {y^s}_i \in \mathcal {M}$ to $\mathbf {\pi _i}$, where $\mathbf {\pi }_i = S^{-1}(\mathbf {u}_i, \mathbf {l}_i)$. We did not use the standard seq2seq model which takes $\mathbf {y^s}$ as input and generates $\mathbf {\pi }$ directly, since there are too many blank tokens `-' in $\mathbf {\pi }$ and the model tends to generate a long sequence with only blank tokens. During MT training, we randomly sample text pairs from $\mathcal {M}^{\prime }$ and $\mathcal {M}$ according to a hyper-parameter $k$. After tuning on the validation set, about $30\%$ pairs are sampled from $\mathcal {M}^{\prime }$. In this way, the $enc_t$ is more robust toward the longer inputs given by the $enc_s$.
Experiments
We conduct experiments on the IWSLT18 speech translation task BIBREF1. Since IWSLT participators use different data pre-processing methods, we reproduce several competitive baselines based on the ESPnet BIBREF12 for a fair comparison.
Experiments ::: Dataset ::: Speech translation data:
The organizer provides a speech translation corpus extracting from the TED talk (ST-TED), which consists of raw English wave files, English transcriptions, and aligned German translations. The corpus contains 272 hours of English speech with 171k segments. We split 2k segments from the corpus as dev set and tst2010, tst2013, tst2014, tst2015 are used as test sets.
Speech recognition data: Aside from ST-TED, TED-LIUM2 corpus BIBREF13 is provided as speech recognition data, which contains 207 hours of English speech and 93k transcript sentences.
Text translation data: We use transcription and translation pairs in the ST-TED corpus and WIT3 as in-domain MT data, which contains 130k and 200k sentence pairs respectively. WMT2018 is used as out-of-domain training data which consists of 41M sentence pairs.
Data preprocessing: For speech data, the utterances are segmented into multiple frames with a 25 ms window size and a 10 ms step size. Then we extract 80-channel log-Mel filter bank and 3-dimensional pitch features using Kaldi BIBREF14, resulting in 83-dimensional input features. We normalize them by the mean and the standard deviation on the whole training set. The utterances with more than 3000 frames are discarded. The transcripts in ST-TED are in true-case with punctuation while in TED-LIUM2, transcripts are in lower-case and unpunctuated. Thus, we lowercase all the sentences and remove the punctuation to keep consistent. To increase the amount of training data, we perform speed perturbation on the raw signals with speed factors 0.9 and 1.1. For the text translation data, sentences longer than 80 words or shorter than 10 words are removed. Besides, we discard pairs whose length ratio between source and target sentence is smaller than 0.5 or larger than 2.0. Word tokenization is performed using the Moses scripts and both English and German words are in lower-case.
We use two different sets of vocabulary for our experiments. For the subword experiments, both English and German vocabularies are generated using sentencepiece BIBREF15 with a fixed size of 5k tokens. BIBREF9 inaguma2018speech show that increasing the vocabulary size is not helpful for ST task. For the character experiments, both English and German sentences are represented in the character level.
For evaluation, we segment each audio with the LIUM SpkDiarization tool BIBREF16 and then perform MWER segmentation with RWTH toolkit BIBREF17. We use lowercase BLEU as evaluation metric.
Experiments ::: Baseline Models and Implementation
We compare our method with following baselines.
Vanilla ST baseline: The vanilla ST BIBREF9 has only a speech encoder and a decoder. It is trained from scratch on the ST-TED corpus.
Pre-training baselines: We conduct three pre-training baseline experiments: 1) encoder pre-training, in which the ST encoder is initialized from an ASR model; 2) decoder pre-training, in which the ST decoder is initialized from an MT model; and 3) encoder-decoder pre-training, where both the encoder and decoder are pre-trained. The ASR model has the same architecture with vanilla ST model, trained on the mixture of ST-TED and TED-LIUM2 corpus. The MT model has a text encoder and decoder with the same architecture of which in TCEN. It is first trained on WMT data (out-of-domain) and then fine-tuned on in-domain data.
Multi-task baselines: We also conduct three multi-task baseline experiments including one-to-many setting, many-to-one setting, and many-to-many setting. In the first two settings, we train the model with $\alpha _{st}=0.75$ while $\alpha _{asr}=0.25$ or $\alpha _{mt}=0.25$. For many-to-many setting, we use $\alpha _{st}=0.6, \alpha _{asr}=0.2$ and $\alpha _{mt}=0.2$.. For MT task, we use only in-domain data.
Many-to-many+pre-training: We train a many-to-many multi-task model where the encoders and decoders are derived from pre-trained ASR and MT models. Triangle+pre-train: BIBREF18 DBLP:conf/naacl/AnastasopoulosC18 proposed a triangle multi-task strategy for speech translation. Their model solves the subnet waste issue by concatenating an ST decoder to an ASR encoder-decoder model. Notably, their ST decoder can consume representations from the speech encoder as well as the ASR decoder. For a fair comparison, the speech encoder and the ASR decoder are initialized from the pre-trained ASR model. The Triangle model is fine-tuned under their multi-task manner.
All our baselines as well as TCEN are implemented based on ESPnet BIBREF12, the RNN size is set as $d=1024$ for all models. We use a dropout of 0.3 for embeddings and encoders, and train using Adadelta with initial learning rate of 1.0 for a maximum of 10 epochs.
For training of TCEN, we set $\alpha _{asr}=0.2$ and $\alpha _{mt}=0.8$ in the pre-training stage, since the MT dataset is much larger than ASR dataset. For fine-tune, we use $\alpha _{st}=0.6, \alpha _{asr}=0.2$ and $\alpha _{mt}=0.2$, same as the `many-to-many' baseline.
For testing, we select the model with the best accuracy on speech translation task on dev set. At inference time, we use a beam size of 10, and the beam scores include length normalization with a weight of 0.2.
Experiments ::: Experimental Results
Table TABREF29 shows the results on four test sets as well as the average performance. Our method significantly outperforms the strong `many-to-many+pretrain' baseline by 3.6 and 2.2 BLEU scores respectively, indicating the proposed method is very effective that substantially improves the translation quality. Besides, both pre-training and multi-task learning can improve translation quality, and the pre-training settings (2nd-4th rows) are more effective compared to multi-task settings (5th-8th rows). We observe a performance degradation in the `triangle+pretrain' baseline. Compared to our method, where the decoder receives higher-level syntactic and semantic linguistic knowledge extracted from text encoder, their ASR decoder can only provide lower word-level linguistic information. Besides, since their model lacks text encoder and the architecture of ST decoder is different from MT decoder, their model cannot utilize the large-scale MT data in all the training stages. Interestingly, we find that the char-level models outperform the subword-level models in all settings, especially in vanilla baseline. A similar phenomenon is observed by BIBREF6 berard2018end. A possible explanation is that learning the alignments between speech frames and subword units in another language is notoriously difficult. Our method can bring more gains in the subword setting since our model is good at learning the text-to-text alignment and the subword-level alignment is more helpful to the translation quality.
Experiments ::: Discussion ::: Ablation Study
To better understand the contribution of each component, we perform an ablation study on subword-level experiments. The results are shown in Table TABREF37. In `-MT noise' setting, we do not add noise to source sentences for MT. In `-weight sharing' setting, we use different parameters in CTC classification layer and source embedding layer. These two experiments prove that both weight sharing and using noisy MT input benefit to the final translation quality. Performance degrades more in `-weight sharing', indicating the semantic consistency contributes more to our model. In the `-pretrain' experiment, we remove the pre-training stage and directly update the model on three tasks, leading to a dramatic decrease on BLEU score, indicating the pre-training is an indispensable step for end-to-end ST.
Experiments ::: Discussion ::: Learning Curve
It is interesting to investigate why our method is superior to baselines. We find that TCEN achieves a higher final result owing to a better start-point in fine-tuning. Figure FIGREF39 provides learning curves of subword accuracy on validation set. The x-axis denotes the fine-tuning training steps. The vanilla model starts at a low accuracy, because its networks are not pre-trained on the ASR and MT data. The trends of our model and `many-to-many+pretrain' are similar, but our model outperforms it about five points in the whole fine-tuning process. It indicates that the gain comes from bridging the gap between pre-training and fine-tuning rather than a better fine-tuning process.
Experiments ::: Discussion ::: Compared with a Cascaded System
Table TABREF29 compares our model with end-to-end baselines. Here, we compare our model with cascaded systems. We build a cascaded system by combining the ASR model and MT model used in pre-training baseline. Word error rate (WER) of the ASR system and BLEU score of the MT system are reported in the supplementary material. In addition to a simple combination of the ASR and MT systems, we also re-segment the ASR outputs before feeding to the MT system, denoted as cascaded+re-seg. Specifically, we train a seq2seq model BIBREF19 on the MT dataset, where the source side is a no punctuation sentence and the target side is a natural sentence. After that, we use the seq2seq model to add sentence boundaries and punctuation on ASR outputs. Experimental results are shown in Table TABREF41. Our end-to-end model outperforms the simple cascaded model over 2 BLEU scores, and it achieves a comparable performance with the cascaded model combining with a sentence re-segment model.
Related Work
Early works conduct speech translation in a pipeline manner BIBREF2, BIBREF20, where the ASR output lattices are fed into an MT system to generate target sentences. HMM BIBREF21, DenseNet BIBREF22, TDNN BIBREF23 are commonly used ASR systems, while RNN with attention BIBREF19 and Transformer BIBREF10 are top choices for MT. To enhance the robustness of the NMT model towards ASR errors, BIBREF24 DBLP:conf/eacl/TsvetkovMD14 and BIBREF25 DBLP:conf/asru/ChenHHL17 propose to simulate the noise in training and inference.
To avoid error propagation and high latency issues, recent works propose translating the acoustic speech into text in target language without yielding the source transcription BIBREF4. Since ST data is scarce, pre-training BIBREF7, multi-task learning BIBREF4, BIBREF6, curriculum learning BIBREF26, attention-passing BIBREF27, and knowledge distillation BIBREF28, BIBREF29 strategies have been explored to utilize ASR data and MT data. Specifically, BIBREF5 DBLP:conf/interspeech/WeissCJWC17 show improvements of performance by training the ST model jointly with the ASR and the MT model. BIBREF6 berard2018end observe faster convergence and better results due to pre-training and multi-task learning on a larger dataset. BIBREF7 DBLP:conf/naacl/BansalKLLG19 show that pre-training a speech encoder on one language can improve ST quality on a different source language. All of them follow the traditional multi-task training strategies. BIBREF26 DBLP:journals/corr/abs-1802-06003 propose to use curriculum learning to improve ST performance on syntactically distant language pairs. To effectively leverage transcriptions in ST data, BIBREF18 DBLP:conf/naacl/AnastasopoulosC18 augment the multi-task model where the target decoder receives information from the source decoder and they show improvements on low-resource speech translation. Their model just consumes ASR and ST data, in contrast, our work sufficiently utilizes the large-scale MT data to capture the rich semantic knowledge. BIBREF30 DBLP:conf/icassp/JiaJMWCCALW19 use pre-trained MT and text-to-speech (TTS) synthesis models to convert weakly supervised data into ST pairs and demonstrate that an end-to-end MT model can be trained using only synthesised data.
Conclusion
This paper has investigated the end-to-end method for ST. It has discussed why there is a huge gap between pre-training and fine-tuning in previous methods. To alleviate these issues, we have proposed a method, which is capable of reusing every sub-net and keeping the role of sub-net consistent between pre-training and fine-tuning. Empirical studies have demonstrated that our model significantly outperforms baselines.
| [
"the model is pre-trained on CTC-based ASR task and MT task in the pre-training stage."
] | 4,656 | qasper | en | null | ebd4ae480fe1596841b2132e96f40eac8437c800db8ef59e | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What is the attention module pretrained on?
Answer: | [
"the model is pre-trained on CTC-based ASR task and MT task in the pre-training stage."
] | qasper | 128 | 24 | |
What kind of stylistic features are obtained? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Sarcasm is an intensive, indirect and complex construct that is often intended to express contempt or ridicule . Sarcasm, in speech, is multi-modal, involving tone, body-language and gestures along with linguistic artifacts used in speech. Sarcasm in text, on the other hand, is more restrictive when it comes to such non-linguistic modalities. This makes recognizing textual sarcasm more challenging for both humans and machines.
Sarcasm detection plays an indispensable role in applications like online review summarizers, dialog systems, recommendation systems and sentiment analyzers. This makes automatic detection of sarcasm an important problem. However, it has been quite difficult to solve such a problem with traditional NLP tools and techniques. This is apparent from the results reported by the survey from DBLP:journals/corr/JoshiBC16. The following discussion brings more insights into this.
Consider a scenario where an online reviewer gives a negative opinion about a movie through sarcasm: “This is the kind of movie you see because the theater has air conditioning”. It is difficult for an automatic sentiment analyzer to assign a rating to the movie and, in the absence of any other information, such a system may not be able to comprehend that prioritizing the air-conditioning facilities of the theater over the movie experience indicates a negative sentiment towards the movie. This gives an intuition to why, for sarcasm detection, it is necessary to go beyond textual analysis.
We aim to address this problem by exploiting the psycholinguistic side of sarcasm detection, using cognitive features extracted with the help of eye-tracking. A motivation to consider cognitive features comes from analyzing human eye-movement trajectories that supports the conjecture: Reading sarcastic texts induces distinctive eye movement patterns, compared to literal texts. The cognitive features, derived from human eye movement patterns observed during reading, include two primary feature types:
The cognitive features, along with textual features used in best available sarcasm detectors, are used to train binary classifiers against given sarcasm labels. Our experiments show significant improvement in classification accuracy over the state of the art, by performing such augmentation.
Related Work
Sarcasm, in general, has been the focus of research for quite some time. In one of the pioneering works jorgensen1984test explained how sarcasm arises when a figurative meaning is used opposite to the literal meaning of the utterance. In the word of clark1984pretense, sarcasm processing involves canceling the indirectly negated message and replacing it with the implicated one. giora1995irony, on the other hand, define sarcasm as a mode of indirect negation that requires processing of both negated and implicated messages. ivanko2003context define sarcasm as a six tuple entity consisting of a speaker, a listener, Context, Utterance, Literal Proposition and Intended Proposition and study the cognitive aspects of sarcasm processing.
Computational linguists have previously addressed this problem using rule based and statistical techniques, that make use of : (a) Unigrams and Pragmatic features BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 (b) Stylistic patterns BIBREF4 and patterns related to situational disparity BIBREF5 and (c) Hastag interpretations BIBREF6 , BIBREF7 .
Most of the previously done work on sarcasm detection uses distant supervision based techniques (ex: leveraging hashtags) and stylistic/pragmatic features (emoticons, laughter expressions such as “lol” etc). But, detecting sarcasm in linguistically well-formed structures, in absence of explicit cues or information (like emoticons), proves to be hard using such linguistic/stylistic features alone.
With the advent of sophisticated eye-trackers and electro/magneto-encephalographic (EEG/MEG) devices, it has been possible to delve deep into the cognitive underpinnings of sarcasm understanding. Filik2014, using a series of eye-tracking and EEG experiments try to show that for unfamiliar ironies, the literal interpretation would be computed first. They also show that a mismatch with context would lead to a re-interpretation of the statement, as being ironic. Camblin2007103 show that in multi-sentence passages, discourse congruence has robust effects on eye movements. This also implies that disrupted processing occurs for discourse incongruent words, even though they are perfectly congruous at the sentence level. In our previous work BIBREF8 , we augment cognitive features, derived from eye-movement patterns of readers, with textual features to detect whether a human reader has realized the presence of sarcasm in text or not.
The recent advancements in the literature discussed above, motivate us to explore gaze-based cognition for sarcasm detection. As far as we know, our work is the first of its kind.
Eye-tracking Database for Sarcasm Analysis
Sarcasm often emanates from incongruity BIBREF9 , which enforces the brain to reanalyze it BIBREF10 . This, in turn, affects the way eyes move through the text. Hence, distinctive eye-movement patterns may be observed in the case of successful processing of sarcasm in text in contrast to literal texts. This hypothesis forms the crux of our method for sarcasm detection and we validate this using our previously released freely available sarcasm dataset BIBREF8 enriched with gaze information.
Document Description
The database consists of 1,000 short texts, each having 10-40 words. Out of these, 350 are sarcastic and are collected as follows: (a) 103 sentences are from two popular sarcastic quote websites, (b) 76 sarcastic short movie reviews are manually extracted from the Amazon Movie Corpus BIBREF11 by two linguists. (c) 171 tweets are downloaded using the hashtag #sarcasm from Twitter. The 650 non-sarcastic texts are either downloaded from Twitter or extracted from the Amazon Movie Review corpus. The sentences do not contain words/phrases that are highly topic or culture specific. The tweets were normalized to make them linguistically well formed to avoid difficulty in interpreting social media lingo. Every sentence in our dataset carries positive or negative opinion about specific “aspects”. For example, the sentence “The movie is extremely well cast” has positive sentiment about the aspect “cast”.
The annotators were seven graduate students with science and engineering background, and possess good English proficiency. They were given a set of instructions beforehand and are advised to seek clarifications before they proceed. The instructions mention the nature of the task, annotation input method, and necessity of head movement minimization during the experiment.
Task Description
The task assigned to annotators was to read sentences one at a time and label them with with binary labels indicating the polarity (i.e., positive/negative). Note that, the participants were not instructed to annotate whether a sentence is sarcastic or not., to rule out the Priming Effect (i.e., if sarcasm is expected beforehand, processing incongruity becomes relatively easier BIBREF12 ). The setup ensures its “ecological validity” in two ways: (1) Readers are not given any clue that they have to treat sarcasm with special attention. This is done by setting the task to polarity annotation (instead of sarcasm detection). (2) Sarcastic sentences are mixed with non sarcastic text, which does not give prior knowledge about whether the forthcoming text will be sarcastic or not.
The eye-tracking experiment is conducted by following the standard norms in eye-movement research BIBREF13 . At a time, one sentence is displayed to the reader along with the “aspect” with respect to which the annotation has to be provided. While reading, an SR-Research Eyelink-1000 eye-tracker (monocular remote mode, sampling rate 500Hz) records several eye-movement parameters like fixations (a long stay of gaze) and saccade (quick jumping of gaze between two positions of rest) and pupil size.
The accuracy of polarity annotation varies between 72%-91% for sarcastic texts and 75%-91% for non-sarcastic text, showing the inherent difficulty of sentiment annotation, when sarcasm is present in the text under consideration. Annotation errors may be attributed to: (a) lack of patience/attention while reading, (b) issues related to text comprehension, and (c) confusion/indecisiveness caused due to lack of context.
For our analysis, we do not discard the incorrect annotations present in the database. Since our system eventually aims to involve online readers for sarcasm detection, it will be hard to segregate readers who misinterpret the text. We make a rational assumption that, for a particular text, most of the readers, from a fairly large population, will be able to identify sarcasm. Under this assumption, the eye-movement parameters, averaged across all readers in our setting, may not be significantly distorted by a few readers who would have failed to identify sarcasm. This assumption is applicable for both regular and multi-instance based classifiers explained in section SECREF6 .
Analysis of Eye-movement Data
We observe distinct behavior during sarcasm reading, by analyzing the “fixation duration on the text” (also referred to as “dwell time” in the literature) and “scanpaths” of the readers.
Variation in the Average Fixation Duration per Word
Since sarcasm in text can be expected to induce cognitive load, it is reasonable to believe that it would require more processing time BIBREF14 . Hence, fixation duration normalized over total word count should usually be higher for a sarcastic text than for a non-sarcastic one. We observe this for all participants in our dataset, with the average fixation duration per word for sarcastic texts being at least 1.5 times more than that of non-sarcastic texts. To test the statistical significance, we conduct a two-tailed t-test (assuming unequal variance) to compare the average fixation duration per word for sarcastic and non-sarcastic texts. The hypothesized mean difference is set to 0 and the error tolerance limit ( INLINEFORM0 ) is set to 0.05. The t-test analysis, presented in Table TABREF11 , shows that for all participants, a statistically significant difference exists between the average fixation duration per word for sarcasm (higher average fixation duration) and non-sarcasm (lower average fixation duration). This affirms that the presence of sarcasm affects the duration of fixation on words.
It is important to note that longer fixations may also be caused by other linguistic subtleties (such as difficult words, ambiguity and syntactically complex structures) causing delay in comprehension, or occulomotor control problems forcing readers to spend time adjusting eye-muscles. So, an elevated average fixation duration per word may not sufficiently indicate the presence of sarcasm. But we would also like to share that, for our dataset, when we considered readability (Flesch readability ease-score BIBREF15 ), number of words in a sentence and average character per word along with the sarcasm label as the predictors of average fixation duration following a linear mixed effect model BIBREF16 , sarcasm label turned out to be the most significant predictor with a maximum slope. This indicates that average fixation duration per word has a strong connection with the text being sarcastic, at least in our dataset.
We now analyze scanpaths to gain more insights into the sarcasm comprehension process.
Analysis of Scanpaths
Scanpaths are line-graphs that contain fixations as nodes and saccades as edges; the radii of the nodes represent the fixation duration. A scanpath corresponds to a participant's eye-movement pattern while reading a particular sentence. Figure FIGREF14 presents scanpaths of three participants for the sarcastic sentence S1 and the non-sarcastic sentence S2. The x-axis of the graph represents the sequence of words a reader reads, and the y-axis represents a temporal sequence in milliseconds.
Consider a sarcastic text containing incongruous phrases A and B. Our qualitative scanpath-analysis reveals that scanpaths with respect to sarcasm processing have two typical characteristics. Often, a long regression - a saccade that goes to a previously visited segment - is observed when a reader starts reading B after skimming through A. In a few cases, the fixation duration on A and B are significantly higher than the average fixation duration per word. In sentence S1, we see long and multiple regressions from the two incongruous phrases “misconception” and “cherish”, and a few instances where phrases “always cherish” and “original misconception” are fixated longer than usual. Such eye-movement behaviors are not seen for S2.
Though sarcasm induces distinctive scanpaths like the ones depicted in Figure FIGREF14 in the observed examples, presence of such patterns is not sufficient to guarantee sarcasm; such patterns may also possibly arise from literal texts. We believe that a combination of linguistic features, readability of text and features derived from scanpaths would help discriminative machine learning models learn sarcasm better.
Features for Sarcasm Detection
We describe the features used for sarcasm detection in Table . The features enlisted under lexical,implicit incongruity and explicit incongruity are borrowed from various literature (predominantly from joshi2015harnessing). These features are essential to separate sarcasm from other forms semantic incongruity in text (for example ambiguity arising from semantic ambiguity or from metaphors). Two additional textual features viz. readability and word count of the text are also taken under consideration. These features are used to reduce the effect of text hardness and text length on the eye-movement patterns.
Simple Gaze Based Features
Readers' eye-movement behavior, characterized by fixations, forward saccades, skips and regressions, can be directly quantified by simple statistical aggregation (i.e., either computing features for individual participants and then averaging or performing a multi-instance based learning as explained in section SECREF6 ). Since these eye-movement attributes relate to the cognitive process in reading BIBREF17 , we consider these as features in our model. Some of these features have been reported by sarcasmunderstandability for modeling sarcasm understandability of readers. However, as far as we know, these features are being introduced in NLP tasks like textual sarcasm detection for the first time. The values of these features are believed to increase with the increase in the degree of surprisal caused by incongruity in text (except skip count, which will decrease).
Complex Gaze Based Features
For these features, we rely on a graph structure, namely “saliency graphs", derived from eye-gaze information and word sequences in the text.
For each reader and each sentence, we construct a “saliency graph”, representing the reader's attention characteristics. A saliency graph for a sentence INLINEFORM0 for a reader INLINEFORM1 , represented as INLINEFORM2 , is a graph with vertices ( INLINEFORM3 ) and edges ( INLINEFORM4 ) where each vertex INLINEFORM5 corresponds to a word in INLINEFORM6 (may not be unique) and there exists an edge INLINEFORM7 between vertices INLINEFORM8 and INLINEFORM9 if R performs at least one saccade between the words corresponding to INLINEFORM10 and INLINEFORM11 .
Figure FIGREF15 shows an example of a saliency graph.A saliency graph may be weighted, but not necessarily connected, for a given text (as there may be words in the given text with no fixation on them). The “complex” gaze features derived from saliency graphs are also motivated by the theory of incongruity. For instance, Edge Density of a saliency graph increases with the number of distinct saccades, which could arise from the complexity caused by presence of sarcasm. Similarly, the highest weighted degree of a graph is expected to be higher, if the node corresponds to a phrase, incongruous to some other phrase in the text.
The Sarcasm Classifier
We interpret sarcasm detection as a binary classification problem. The training data constitutes 994 examples created using our eye-movement database for sarcasm detection. To check the effectiveness of our feature set, we observe the performance of multiple classification techniques on our dataset through a stratified 10-fold cross validation. We also compare the classification accuracy of our system and the best available systems proposed by riloff2013sarcasm and joshi2015harnessing on our dataset. Using Weka BIBREF18 and LibSVM BIBREF19 APIs, we implement the following classifiers:
Results
Table TABREF17 shows the classification results considering various feature combinations for different classifiers and other systems. These are:
Unigram (with principal components of unigram feature vectors),
Sarcasm (the feature-set reported by joshi2015harnessing subsuming unigram features and features from other reported systems)
Gaze (the simple and complex cognitive features we introduce, along with readability and word count features), and
Gaze+Sarcasm (the complete set of features).
For all regular classifiers, the gaze features are averaged across participants and augmented with linguistic and sarcasm related features. For the MILR classifier, the gaze features derived from each participant are augmented with linguistic features and thus, a multi instance “bag” of features is formed for each sentence in the training data. This multi-instance dataset is given to an MILR classifier, which follows the standard multi instance assumption to derive class-labels for each bag.
For all the classifiers, our feature combination outperforms the baselines (considering only unigram features) as well as BIBREF3 , with the MILR classifier getting an F-score improvement of 3.7% and Kappa difference of 0.08. We also achieve an improvement of 2% over the baseline, using SVM classifier, when we employ our feature set. We also observe that the gaze features alone, also capture the differences between sarcasm and non-sarcasm classes with a high-precision but a low recall.
To see if the improvement obtained is statistically significant over the state-of-the art system with textual sarcasm features alone, we perform McNemar test. The output of the SVM classifier using only linguistic features used for sarcasm detection by joshi2015harnessing and the output of the MILR classifier with the complete set of features are compared, setting threshold INLINEFORM0 . There was a significant difference in the classifier's accuracy with p(two-tailed) = 0.02 with an odds-ratio of 1.43, showing that the classification accuracy improvement is unlikely to be observed by chance in 95% confidence interval.
Considering Reading Time as a Cognitive Feature along with Sarcasm Features
One may argue that, considering simple measures of reading effort like “reading time” as cognitive feature instead of the expensive eye-tracking features for sarcasm detection may be a cost-effective solution. To examine this, we repeated our experiments with “reading time” considered as the only cognitive feature, augmented with the textual features. The F-scores of all the classifiers turn out to be close to that of the classifiers considering sarcasm feature alone and the difference in the improvement is not statistically significant ( INLINEFORM0 ). One the other hand, F-scores with gaze features are superior to the F-scores when reading time is considered as a cognitive feature.
How Effective are the Cognitive Features
We examine the effectiveness of cognitive features on the classification accuracy by varying the input training data size. To examine this, we create a stratified (keeping the class ratio constant) random train-test split of 80%:20%. We train our classifier with 100%, 90%, 80% and 70% of the training data with our whole feature set, and the feature combination from joshi2015harnessing. The goodness of our system is demonstrated by improvements in F-score and Kappa statistics, shown in Figure FIGREF22 .
We further analyze the importance of features by ranking the features based on (a) Chi squared test, and (b) Information Gain test, using Weka's attribute selection module. Figure FIGREF23 shows the top 20 ranked features produced by both the tests. For both the cases, we observe 16 out of top 20 features to be gaze features. Further, in each of the cases, Average Fixation Duration per Word and Largest Regression Position are seen to be the two most significant features.
Example Cases
Table TABREF21 shows a few example cases from the experiment with stratified 80%-20% train-test split.
Example sentence 1 is sarcastic, and requires extra-linguistic knowledge (about poor living conditions at Manchester). Hence, the sarcasm detector relying only on textual features is unable to detect the underlying incongruity. However, our system predicts the label successfully, possibly helped by the gaze features.
Similarly, for sentence 2, the false sense of presence of incongruity (due to phrases like “Helped me” and “Can't stop”) affects the system with only linguistic features. Our system, though, performs well in this case also.
Sentence 3 presents a false-negative case where it was hard for even humans to get the sarcasm. This is why our gaze features (and subsequently the complete set of features) account for erroneous prediction.
In sentence 4, gaze features alone false-indicate presence of incongruity, whereas the system predicts correctly when gaze and linguistic features are taken together.
From these examples, it can be inferred that, only gaze features would not have sufficed to rule out the possibility of detecting other forms of incongruity that do not result in sarcasm.
Error Analysis
Errors committed by our system arise from multiple factors, starting from limitations of the eye-tracker hardware to errors committed by linguistic tools and resources. Also, aggregating various eye-tracking parameters to extract the cognitive features may have caused information loss in the regular classification setting.
Conclusion
In the current work, we created a novel framework to detect sarcasm, that derives insights from human cognition, that manifests over eye movement patterns. We hypothesized that distinctive eye-movement patterns, associated with reading sarcastic text, enables improved detection of sarcasm. We augmented traditional linguistic features with cognitive features obtained from readers' eye-movement data in the form of simple gaze-based features and complex features derived from a graph structure. This extended feature-set improved the success rate of the sarcasm detector by 3.7%, over the best available system. Using cognitive features in an NLP Processing system like ours is the first proposal of its kind.
Our general approach may be useful in other NLP sub-areas like sentiment and emotion analysis, text summarization and question answering, where considering textual clues alone does not prove to be sufficient. We propose to augment this work in future by exploring deeper graph and gaze features. We also propose to develop models for the purpose of learning complex gaze feature representation, that accounts for the power of individual eye movement patterns along with the aggregated patterns of eye movements.
Acknowledgments
We thank the members of CFILT Lab, especially Jaya Jha and Meghna Singh, and the students of IIT Bombay for their help and support.
| [
"Unanswerable"
] | 3,543 | qasper | en | null | 036bfbdbfff8294f59afd5860663bbb4349924c9853b2151 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What kind of stylistic features are obtained?
Answer: | [
"Unanswerable"
] | qasper | 128 | 25 | |
What architecture does the encoder have? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
This paper describes our approach and results for Task 2 of the CoNLL–SIGMORPHON 2018 shared task on universal morphological reinflection BIBREF0 . The task is to generate an inflected word form given its lemma and the context in which it occurs.
Morphological (re)inflection from context is of particular relevance to the field of computational linguistics: it is compelling to estimate how well a machine-learned system can capture the morphosyntactic properties of a word given its context, and map those properties to the correct surface form for a given lemma.
There are two tracks of Task 2 of CoNLL–SIGMORPHON 2018: in Track 1 the context is given in terms of word forms, lemmas and morphosyntactic descriptions (MSD); in Track 2 only word forms are available. See Table TABREF1 for an example. Task 2 is additionally split in three settings based on data size: high, medium and low, with high-resource datasets consisting of up to 70K instances per language, and low-resource datasets consisting of only about 1K instances.
The baseline provided by the shared task organisers is a seq2seq model with attention (similar to the winning system for reinflection in CoNLL–SIGMORPHON 2016, BIBREF1 ), which receives information about context through an embedding of the two words immediately adjacent to the target form. We use this baseline implementation as a starting point and achieve the best overall accuracy of 49.87 on Task 2 by introducing three augmentations to the provided baseline system: (1) We use an LSTM to encode the entire available context; (2) We employ a multi-task learning approach with the auxiliary objective of MSD prediction; and (3) We train the auxiliary component in a multilingual fashion, over sets of two to three languages.
In analysing the performance of our system, we found that encoding the full context improves performance considerably for all languages: 11.15 percentage points on average, although it also highly increases the variance in results. Multi-task learning, paired with multilingual training and subsequent monolingual finetuning, scored highest for five out of seven languages, improving accuracy by another 9.86% on average.
System Description
Our system is a modification of the provided CoNLL–SIGMORPHON 2018 baseline system, so we begin this section with a reiteration of the baseline system architecture, followed by a description of the three augmentations we introduce.
Baseline
The CoNLL–SIGMORPHON 2018 baseline is described as follows:
The system is an encoder-decoder on character sequences. It takes a lemma as input and generates a word form. The process is conditioned on the context of the lemma [...] The baseline treats the lemma, word form and MSD of the previous and following word as context in track 1. In track 2, the baseline only considers the word forms of the previous and next word. [...] The baseline system concatenates embeddings for context word forms, lemmas and MSDs into a context vector. The baseline then computes character embeddings for each character in the input lemma. Each of these is concatenated with a copy of the context vector. The resulting sequence of vectors is encoded using an LSTM encoder. Subsequently, an LSTM decoder generates the characters in the output word form using encoder states and an attention mechanism.
To that we add a few details regarding model size and training schedule:
the number of LSTM layers is one;
embedding size, LSTM layer size and attention layer size is 100;
models are trained for 20 epochs;
on every epoch, training data is subsampled at a rate of 0.3;
LSTM dropout is applied at a rate 0.3;
context word forms are randomly dropped at a rate of 0.1;
the Adam optimiser is used, with a default learning rate of 0.001; and
trained models are evaluated on the development data (the data for the shared task comes already split in train and dev sets).
Our system
Here we compare and contrast our system to the baseline system. A diagram of our system is shown in Figure FIGREF4 .
The idea behind this modification is to provide the encoder with access to all morpho-syntactic cues present in the sentence. In contrast to the baseline, which only encodes the immediately adjacent context of a target word, we encode the entire context. All context word forms, lemmas, and MSD tags (in Track 1) are embedded in their respective high-dimensional spaces as before, and their embeddings are concatenated. However, we now reduce the entire past context to a fixed-size vector by encoding it with a forward LSTM, and we similarly represent the future context by encoding it with a backwards LSTM.
We introduce an auxiliary objective that is meant to increase the morpho-syntactic awareness of the encoder and to regularise the learning process—the task is to predict the MSD tag of the target form. MSD tag predictions are conditioned on the context encoding, as described in UID15 . Tags are generated with an LSTM one component at a time, e.g. the tag PRO;NOM;SG;1 is predicted as a sequence of four components, INLINEFORM0 PRO, NOM, SG, 1 INLINEFORM1 .
For every training instance, we backpropagate the sum of the main loss and the auxiliary loss without any weighting.
As MSD tags are only available in Track 1, this augmentation only applies to this track.
The parameters of the entire MSD (auxiliary-task) decoder are shared across languages.
Since a grouping of the languages based on language family would have left several languages in single-member groups (e.g. Russian is the sole representative of the Slavic family), we experiment with random groupings of two to three languages. Multilingual training is performed by randomly alternating between languages for every new minibatch. We do not pass any information to the auxiliary decoder as to the source language of the signal it is receiving, as we assume abstract morpho-syntactic features are shared across languages.
After 20 epochs of multilingual training, we perform 5 epochs of monolingual finetuning for each language. For this phase, we reduce the learning rate to a tenth of the original learning rate, i.e. 0.0001, to ensure that the models are indeed being finetuned rather than retrained.
We keep all hyperparameters the same as in the baseline. Training data is split 90:10 for training and validation. We train our models for 50 epochs, adding early stopping with a tolerance of five epochs of no improvement in the validation loss. We do not subsample from the training data.
We train models for 50 different random combinations of two to three languages in Track 1, and 50 monolingual models for each language in Track 2. Instead of picking the single model that performs best on the development set and thus risking to select a model that highly overfits that data, we use an ensemble of the five best models, and make the final prediction for a given target form with a majority vote over the five predictions.
Results and Discussion
Test results are listed in Table TABREF17 . Our system outperforms the baseline for all settings and languages in Track 1 and for almost all in Track 2—only in the high resource setting is our system not definitively superior to the baseline.
Interestingly, our results in the low resource setting are often higher for Track 2 than for Track 1, even though contextual information is less explicit in the Track 2 data and the multilingual multi-tasking approach does not apply to this track. We interpret this finding as an indicator that a simpler model with fewer parameters works better in a setting of limited training data. Nevertheless, we focus on the low resource setting in the analysis below due to time limitations. As our Track 1 results are still substantially higher than the baseline results, we consider this analysis valid and insightful.
Ablation Study
We analyse the incremental effect of the different features in our system, focusing on the low-resource setting in Track 1 and using development data.
Encoding the entire context with an LSTM highly increases the variance of the observed results. So we trained fifty models for each language and each architecture. Figure FIGREF23 visualises the means and standard deviations over the trained models. In addition, we visualise the average accuracy for the five best models for each language and architecture, as these are the models we use in the final ensemble prediction. Below we refer to these numbers only.
The results indicate that encoding the full context with an LSTM highly enhances the performance of the model, by 11.15% on average. This observation explains the high results we obtain also for Track 2.
Adding the auxiliary objective of MSD prediction has a variable effect: for four languages (de, en, es, and sv) the effect is positive, while for the rest it is negative. We consider this to be an issue of insufficient data for the training of the auxiliary component in the low resource setting we are working with.
We indeed see results improving drastically with the introduction of multilingual training, with multilingual results being 7.96% higher than monolingual ones on average.
We studied the five best models for each language as emerging from the multilingual training (listed in Table TABREF27 ) and found no strong linguistic patterns. The en–sv pairing seems to yield good models for these languages, which could be explained in terms of their common language family and similar morphology. The other natural pairings, however, fr–es, and de–sv, are not so frequent among the best models for these pairs of languages.
Finally, monolingual finetuning improves accuracy across the board, as one would expect, by 2.72% on average.
The final observation to be made based on this breakdown of results is that the multi-tasking approach paired with multilingual training and subsequent monolingual finetuning outperforms the other architectures for five out of seven languages: de, en, fr, ru and sv. For the other two languages in the dataset, es and fi, the difference between this approach and the approach that emerged as best for them is less than 1%. The overall improvement of the multilingual multi-tasking approach over the baseline is 18.30%.
Error analysis
Here we study the errors produced by our system on the English test set to better understand the remaining shortcomings of the approach. A small portion of the wrong predictions point to an incorrect interpretation of the morpho-syntactic conditioning of the context, e.g. the system predicted plan instead of plans in the context Our _ include raising private capital. The majority of wrong predictions, however, are nonsensical, like bomb for job, fify for fixing, and gnderrate for understand. This observation suggests that generally the system did not learn to copy the characters of lemma into inflected form, which is all it needs to do in a large number of cases. This issue could be alleviated with simple data augmentation techniques that encourage autoencoding BIBREF2 .
MSD prediction
Figure FIGREF32 summarises the average MSD-prediction accuracy for the multi-tasking experiments discussed above. Accuracy here is generally higher than on the main task, with the multilingual finetuned setup for Spanish and the monolingual setup for French scoring best: 66.59% and 65.35%, respectively. This observation illustrates the added difficulty of generating the correct surface form even when the morphosyntactic description has been identified correctly.
We observe some correlation between these numbers and accuracy on the main task: for de, en, ru and sv, the brown, pink and blue bars here pattern in the same way as the corresponding INLINEFORM0 's in Figure FIGREF23 . One notable exception to this pattern is fr where inflection gains a lot from multilingual training, while MSD prediction suffers greatly. Notice that the magnitude of change is not always the same, however, even when the general direction matches: for ru, for example, multilingual training benefits inflection much more than in benefits MSD prediction, even though the MSD decoder is the only component that is actually shared between languages. This observation illustrates the two-fold effect of multi-task training: an auxiliary task can either inform the main task through the parameters the two tasks share, or it can help the main task learning through its regularising effect.
Related Work
Our system is inspired by previous work on multi-task learning and multi-lingual learning, mainly building on two intuitions: (1) jointly learning related tasks tends to be beneficial BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 ; and (2) jointly learning related languages in an MTL-inspired framework tends to be beneficial BIBREF8 , BIBREF9 , BIBREF10 . In the context of computational morphology, multi-lingual approaches have previously been employed for morphological reinflection BIBREF2 and for paradigm completion BIBREF11 . In both of these cases, however, the available datasets covered more languages, 40 and 21, respectively, which allowed for linguistically-motivated language groupings and for parameter sharing directly on the level of characters. BIBREF10 explore parameter sharing between related languages for dependency parsing, and find that sharing is more beneficial in the case of closely related languages.
Conclusions
In this paper we described our system for the CoNLL–SIGMORPHON 2018 shared task on Universal Morphological Reinflection, Task 2, which achieved the best performance out of all systems submitted, an overall accuracy of 49.87. We showed in an ablation study that this is due to three core innovations, which extend a character-based encoder-decoder model: (1) a wide context window, encoding the entire available context; (2) multi-task learning with the auxiliary task of MSD prediction, which acts as a regulariser; (3) a multilingual approach, exploiting information across languages. In future work we aim to gain better understanding of the increase in variance of the results introduced by each of our modifications and the reasons for the varying effect of multi-task learning for different languages.
Acknowledgements
We gratefully acknowledge the support of the NVIDIA Corporation with the donation of the Titan Xp GPU used for this research.
| [
"LSTM",
"LSTM"
] | 2,289 | qasper | en | null | 8595332098adaabcfd8ae199f754a9b06cdde08cdd4cc64a | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What architecture does the encoder have?
Answer: | [
"LSTM",
"LSTM"
] | qasper | 128 | 26 | |
Is WordNet useful for taxonomic reasoning for this task? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Automatically answering questions, especially in the open-domain setting (i.e., where minimal or no contextual knowledge is explicitly provided), requires bringing to bear considerable amount of background knowledge and reasoning abilities. For example, knowing the answers to the two questions in Figure FIGREF1 requires identifying a specific ISA relation (i.e., that cooking is a type of learned behavior) as well as recalling the definition of a concept (i.e., that global warming is defined as a worldwide increase in temperature). In the multiple-choice setting, which is the variety of question-answering (QA) that we focus on in this paper, there is also pragmatic reasoning involved in selecting optimal answer choices (e.g., while greenhouse effect might in some other context be a reasonable answer to the second question in Figure FIGREF1, global warming is a preferable candidate).
Recent successes in QA, driven largely by the creation of new resources BIBREF2, BIBREF3, BIBREF4, BIBREF5 and advances in model pre-training BIBREF6, BIBREF7, raise a natural question: do state-of-the-art multiple-choice QA (MCQA) models that excel at standard tasks really have basic knowledge and reasoning skills?
Most existing MCQA datasets are constructed through either expensive crowd-sourcing BIBREF8 or hand engineering effort, in the former case making it possible to collect large amounts of data at the cost of losing systematic control over the semantics of the target questions. Hence, doing a controlled experiment to answer such a question for QA is difficult given a lack of targeted challenge datasets.
Having definitive empirical evidence of model competence on any given phenomenon requires constructing a wide range of systematic tests. For example, in measuring competence of definitions, not only do we want to see that the model can handle individual questions such as Figure FIGREF1.1 inside of benchmark tasks, but that it can answer a wider range of questions that exhaustively cover a broad set of concepts and question perturbations (i.e., systematic adjustments to how the questions are constructed). The same applies to ISA reasoning; not only is it important to recognize in the question in Figure FIGREF1.1 that cooking is a learned behavior, but also that cooking is a general type of behavior or, through a few more inferential steps, a type of human activity.
In this paper, we look at systematically constructing such tests by exploiting the vast amounts of structured information contained in various types of expert knowledge such as knowledge graphs and lexical taxonomies. Our general methodology works as illustrated in Figure FIGREF1: given any MCQA model trained on a set of benchmark tasks, we systematically generate a set of synthetic dataset probes (i.e., MCQA renderings of the target information) from information in expert knowledge sources. We then use these probes to ask two empirical questions: 1) how well do models trained on benchmark tasks perform on these probing tasks and; 2) can such models be re-trained to master new challenges with minimal performance loss on their original tasks?
While our methodology is amenable to any knowledge source and set of models/benchmark tasks, we focus on probing state-of-the-art transformer models BIBREF7, BIBREF9 in the domain of science MCQA. For sources of expert knowledge, we use WordNet, a comprehensive lexical ontology, and other publicly available dictionary resources. We devise probes that measure model competence in definition and taxonomic knowledge in different settings (including hypernymy, hyponymy, and synonymy detection, and word sense disambiguation). This choice is motivated by fact that the science domain is considered particularly challenging for QA BIBREF10, BIBREF11, BIBREF12, and existing science benchmarks are known to involve widespread use of such knowledge (see BIBREF1, BIBREF13 for analysis), which is also arguably fundamental to more complex forms of reasoning.
We show that accurately probing QA models via synthetic datasets is not straightforward, as unexpected artifacts can easily arise in such data. This motivates our carefully constructed baselines and close data inspection to ensure probe quality.
Our results confirm that transformer-based QA models have a remarkable ability to recognize certain types of knowledge captured in our probes—even without additional fine-tuning. Such models can even outperform strong task-specific models trained directly on our probing tasks (e.g., on definitions, our best model achieves 77% test accuracy without specialized training, as opposed to 51% for a task-specific LSTM-based model). We also show that the same models can be effectively re-fine-tuned on small samples (even 100 examples) of probe data, and that high performance on the probes tends to correlate with a smaller drop in the model's performance on the original QA task.
Our comprehensive assessment reveals several interesting nuances to the overall positive trend. For example, the performance of even the best QA models degrades substantially on our hyponym probes (by 8-15%) when going from 1-hop links to 2-hops. Further, the accuracy of even our best models on the WordNetQA probe drops by 14-44% under our cluster-based analysis, which assesses whether a model knows several facts about each individual concept, rather than just being good at answering isolated questions. State-of-the-art QA models thus have much room to improve even in some fundamental building blocks, namely definitions and taxonomic hierarchies, of more complex forms of reasoning.
Related Work
We follow recent work on constructing challenge datasets for probing neural models, which has primarily focused on the task of natural language inference (NLI) BIBREF14, BIBREF15, BIBREF16, BIBREF17, BIBREF18. Most of this work looks at constructing data through adversarial generation methods, which have also been found useful for creating stronger models BIBREF19. There has also been work on using synthetic data of the type we consider in this paper BIBREF20, BIBREF21, BIBREF22. We closely follow the methodology of BIBREF22, who use hand-constructed linguistic fragments to probe NLI models and study model re-training using a variant of the inoculation by fine-tuning strategy of BIBREF23. In contrast, we focus on probing open-domain MCQA models (see BIBREF24 for a related study in the reading comprehension setting) as well as constructing data from much larger sources of structured knowledge.
Our main study focuses on probing the BERT model and fine-tuning approach of BIBREF7, and other variants thereof, which are all based on the transformer architecture of BIBREF25. Related to our efforts, there have been recent studies into the types of relational knowledge contained in large-scale knowledge models BIBREF26, BIBREF27, BIBREF28, which, similar to our work, probe models using structured knowledge sources. This prior work, however, primarily focuses on unearthing the knowledge contained in the underlying language models as is without further training, using simple (single token) cloze-style probing tasks and templates (similar to what we propose in Section SECREF3). In contrast, we focus on understanding the knowledge contained in language models after they have been trained for a QA end-task using benchmark datasets in which such knowledge is expected to be widespread. Further, our evaluation is done before and after these models are fine-tuned on our probe QA tasks, using a more complex set of QA templates and target inferences.
The use of lexical resources and knowledge graphs such as WordNet to construct datasets has a long history, and has recently appeared in work on adversarial attacks BIBREF14, BIBREF29 and general task construction BIBREF30, BIBREF31. In the area of MCQA, there is related work on constructing questions from tuples BIBREF32, BIBREF3, both of which involve standard crowd annotation to elicit question-answer pairs (see also BIBREF33, BIBREF34). In contrast to this work, we focus on generating data in an entirely automatic fashion, which obviates the need for expensive annotation and gives us the flexibility to construct much larger datasets that control a rich set of semantic aspects of the target questions.
Dataset Probes and Construction
Our probing methodology starts by constructing challenge datasets (Figure FIGREF1, yellow box) from a target set of knowledge resources. Each of our probing datasets consists of multiple-choice questions that include a question $\textbf {q}$ and a set of answer choices or candidates $\lbrace a_{1},...a_{N}\rbrace $. This section describes in detail the 5 different datasets we build, which are drawn from two sources of expert knowledge, namely WordNet BIBREF35 and the GNU Collaborative International Dictionary of English (GCIDE). We describe each resource in turn, and explain how the resulting dataset probes, which we call WordNetQA and DictionaryQA, are constructed.
For convenience, we will describe each source of expert knowledge as a directed, edge-labeled graph $G$. The nodes of this graph are $\mathcal {V} = \mathcal {C} \cup \mathcal {W} \cup \mathcal {S} \cup \mathcal {D}$, where $\mathcal {C}$ is a set of atomic concepts, $\mathcal {W}$ a set of words, $\mathcal {S}$ a set of sentences, and $\mathcal {D}$ a set of definitions (see Table TABREF4 for details for WordNet and GCIDE). Each edge of $G$ is directed from an atomic concept in $\mathcal {C}$ to another node in $V$, and is labeled with a relation, such as hypernym or isa$^\uparrow $, from a set of relations $\mathcal {R}$ (see Table TABREF4).
When defining our probe question templates, it will be useful to view $G$ as a set of (relation, source, target) triples $\mathcal {T} \subseteq \mathcal {R} \times \mathcal {C} \times \mathcal {V}$. Due to their origin in an expert knowledge source, such triples preserve semantic consistency. For instance, when the relation in a triple is def, the corresponding edge maps a concept in $\mathcal {C}$ to a definition in $\mathcal {D}$.
To construct probe datasets, we rely on two heuristic functions, defined below for each individual probe: $\textsc {gen}_{\mathcal {Q}}(\tau )$, which generates gold question-answer pairs $(\textbf {q},\textbf {a})$ from a set of triples $\tau \subseteq \mathcal {T}$ and question templates $\mathcal {Q}$, and $\textsc {distr}(\tau ^{\prime })$, which generates distractor answers choices $\lbrace a^{\prime }_{1},...a^{\prime }_{N-1} \rbrace $ based on another set of triples $\tau ^{\prime }$ (where usually $\tau \subset \tau ^{\prime }$). For brevity, we will use $\textsc {gen}(\tau )$ to denote $\textsc {gen}_{\mathcal {Q}}(\tau )$, leaving question templates $\mathcal {Q}$ implicit.
Dataset Probes and Construction ::: WordNetQA
WordNet is an English lexical database consisting of around 117k concepts, which are organized into groups of synsets that each contain a gloss (i.e., a definition of the target concept), a set of representative English words (called lemmas), and, in around 33k synsets, example sentences. In addition, many synsets have ISA links to other synsets that express complex taxonomic relations. Figure FIGREF6 shows an example and Table TABREF4 summarizes how we formulate WordNet as a set of triples $\mathcal {T}$ of various types. These triples together represent a directed, edge-labeled graph $G$. Our main motivation for using WordNet, as opposed to a resource such as ConceptNet BIBREF36, is the availability of glosses ($\mathcal {D}$) and example sentences ($\mathcal {S}$), which allows us to construct natural language questions that contextualize the types of concepts we want to probe.
Dataset Probes and Construction ::: WordNetQA ::: Example Generation @!START@$\textsc {gen}(\tau )$@!END@.
We build 4 individual datasets based on semantic relations native to WordNet (see BIBREF37): hypernymy (i.e., generalization or ISA reasoning up a taxonomy, ISA$^\uparrow $), hyponymy (ISA$^{\downarrow }$), synonymy, and definitions. To generate a set of questions in each case, we employ a number of rule templates $\mathcal {Q}$ that operate over tuples. A subset of such templates is shown in Table TABREF8. The templates were designed to mimic naturalistic questions we observed in our science benchmarks.
For example, suppose we wish to create a question $\textbf {q}$ about the definition of a target concept $c \in \mathcal {C}$. We first select a question template from $\mathcal {Q}$ that first introduces the concept $c$ and its lemma $l \in \mathcal {W}$ in context using the example sentence $s \in \mathcal {S}$, and then asks to identify the corresponding WordNet gloss $d \in \mathcal {D}$, which serves as the gold answer $\textbf {a}$. The same is done for ISA reasoning; each question about a hypernym/hyponym relation between two concepts $c \rightarrow ^{\uparrow /\downarrow } c^{\prime } \in \mathcal {T}_{i}$ (e.g., $\texttt {dog} \rightarrow ^{\uparrow /\downarrow } \texttt {animal/terrier}$) first introduces a context for $c$ and then asks for an answer that identifies $c^{\prime }$ (which is also provided with a gloss so as to contain all available context).
In the latter case, the rules $(\texttt {isa}^{r},c,c^{\prime }) \in \mathcal {T}_i$ in Table TABREF8 cover only direct ISA links from $c$ in direction $r \in \lbrace \uparrow ,\downarrow \rbrace $. In practice, for each $c$ and direction $r$, we construct tests that cover the set HOPS$(c,r)$ of all direct as well as derived ISA relations of $c$:
This allows us to evaluate the extent to which models are able to handle complex forms of reasoning that require several inferential steps or hops.
Dataset Probes and Construction ::: WordNetQA ::: Distractor Generation: @!START@$\textsc {distr}(\tau ^{\prime })$@!END@.
An example of how distractors are generated is shown in Figure FIGREF6, which relies on similar principles as above. For each concept $c$, we choose 4 distractor answers that are close in the WordNet semantic space. For example, when constructing hypernymy tests for $c$ from the set hops$(c,\uparrow )$, we build distractors by drawing from $\textsc {hops}(c,\downarrow )$ (and vice versa), as well as from the $\ell $-deep sister family of $c$, defined as follows. The 1-deep sister family is simply $c$'s siblings or sisters, i.e., the other children $\tilde{c} \ne c$ of the parent node $c^{\prime }$ of $c$. For $\ell > 1$, the $\ell $-deep sister family also includes all descendants of each $\tilde{c}$ up to $\ell -1$ levels deep, denoted $\textsc {hops}_{\ell -1}(\tilde{c},\downarrow )$. Formally:
For definitions and synonyms we build distractors from all of these sets (with a similar restriction on the depth of sister distractors as noted above). In doing this, we can systematically investigate model performance on a wide range of distractor sets.
Dataset Probes and Construction ::: WordNetQA ::: Perturbations and Semantic Clusters
Based on how we generate data, for each concept $c$ (i.e., atomic WordNet synset) and probe type (i.e., definitions, hypernymy, etc.), we have a wide variety of questions related to $c$ that manipulate 1) the complexity of reasoning that is involved (e.g., the number of inferential hops) and; 2) the types of distractors (or distractor perturbations) that are employed. We call such sets semantic clusters. As we describe in the next section, semantic clusters allow us to devise new types of evaluation that reveal whether models have comprehensive and consistent knowledge of target concepts (e.g., evaluating whether a model can correctly answer several questions associated with a concept, as opposed to a few disjoint instances).
Details of the individual datasets are shown in Table TABREF12. From these sets, we follow BIBREF22 in allocating a maximum of 3k examples for training and reserve the rest for development and testing. Since we are interested in probing, having large held-out sets allows us to do detailed analysis and cluster-based evaluation.
Dataset Probes and Construction ::: DictionaryQA
The DictionaryQA dataset is created from the GCIDE dictionary, which is a comprehensive open-source English dictionary built largely from the Webster's Revised Unabridged Dictionary BIBREF38. Each entry consists of a word, its part-of-speech, its definition, and an optional example sentence (see Table TABREF14). Overall, 33k entries (out of a total of 155k) contain example sentences/usages. As with the WordNet probes, we focus on this subset so as to contextualize each word being probed. In contrast to WordNet, GCIDE does not have ISA relations or explicit synsets, so we take each unique entry to be a distinct sense. We then use the dictionary entries to create a probe that centers around word-sense disambiguation, as described below.
Dataset Probes and Construction ::: DictionaryQA ::: Example and Distractor Generation.
To generate gold questions and answers, we use the same generation templates for definitions exemplified in Figure TABREF8 for WordNetQA. To generate distractors, we simply take alternative definitions for the target words that represent a different word sense (e.g., the alternative definitions of gift shown in Table TABREF14), as well as randomly chosen definitions if needed to create a 5-way multiple choice question. As above, we reserve a maximum of 3k examples for training. Since we have only 9k examples in total in this dataset (see WordSense in Table TABREF12), we also reserve 3k each for development and testing.
We note that initial attempts to build this dataset through standard random splitting gave rise to certain systematic biases that were exploited by the choice-only baseline models described in the next section, and hence inflated overall model scores. After several efforts at filtering we found that, among other factors, using definitions from entries without example sentences as distractors (e.g., the first two entries in Table TABREF14) had a surprising correlation with such biases. This suggests that possible biases involving differences between dictionary entries with and without examples can taint the resulting automatically generated MCQA dataset (for more discussion on the pitfalls involved with automatic dataset construction, see Section SECREF5).
Probing Methodology and Modeling
Given the probes above, we now can start to answer the empirical questions posed at the beginning. Our main focus is on looking at transformer-based MCQA models trained in the science domain (using the benchmarks shown in Table TABREF21). In this section, we provide details of MCQA and the target models, as well as several baselines that we use to sanity check our new datasets. To evaluate model competence, we look at a combination of model performance after science pre-training and after additional model fine-tuning using the lossless inoculation strategy of BIBREF22 (Section SECREF22). In Section SECREF24, we also discuss a cluster-level accuracy metric for measuring performance over semantic clusters.
Probing Methodology and Modeling ::: Task Definition and Modeling
Given a dataset $D =\lbrace (\textbf {q}^{(d)}, \lbrace a_{1}^{(d)},..., a_{N}^{(d)}\rbrace ) \rbrace _{d}^{\mid D \mid }$ consisting of pairs of questions stems $\textbf {q}$ and answer choices $a_{i}$, the goal is to find the correct answer $a_{i^{*}}$ that correctly answers each $\textbf {q}$. Throughout this paper, we look at 5-way multiple-choice problems (i.e., where each $N=5$).
Probing Methodology and Modeling ::: Task Definition and Modeling ::: Question+Answer Encoder.
To model this, our investigation centers around the use of the transformer-based BIBREF25 BERT encoder and fine-tuning approach of BIBREF7 (see also BIBREF6). For each question and individual answer pair $q^{(j)}_{a_{i}}$, we assume the following rendering of this input:
which is run through the pre-trained BERT encoder to generate a representation for $ q^{(j)}_{a_{i}}$ using the hidden state representation for CLS (i.e., the classifier token) $\textbf {c}_{i}$:
The probability of a given answer $p^{(j)}_{i}$ is then computed as $p^{(j)}_{i} \propto e^{\textbf {v}\cdot \textbf {c}^{(j)}_{i}}$, which uses an additional set of classification parameters $\textbf {v} \in \mathbb {R}^{H}$ that are optimized (along with the full transformer network) by taking the final loss of the probability of each correct answer $p_{i^{*}}$ over all answer choices:
We specifically use BERT-large uncased with whole-word masking, as well as the RoBERTa-large model from BIBREF9, which is a more robustly trained version of the original BERT model. Our system uses the implementations provided in AllenNLP BIBREF39 and Huggingface BIBREF40.
Probing Methodology and Modeling ::: Task Definition and Modeling ::: Baselines and Sanity Checks.
When creating synthetic datasets, it is important to ensure that systematic biases, or annotation artifacts BIBREF41, are not introduced into the resulting probes and that the target datasets are sufficiently challenging (or good, in the sense of BIBREF42). To test for this, we use several of the MCQA baseline models first introduced in BIBREF0, which take inspiration from the LSTM-based models used in BIBREF43 for NLI and various partial-input baselines based on these models.
Following the notation from BIBREF0, for any given sequence $s$ of tokens in $\lbrace q^{(j)}, a_{1}^{(j)},...,a_{N}^{(j)}\rbrace $ in $D$, an encoding of $s$ is given as $h_{s}^{(j)} = \textbf {BiLSTM}(\textsc {embed}(s)) \in \mathbb {R}^{|s| \times 2h}$ (where $h$ is the dimension of the hidden state in each directional network, and embed$(\cdot )$ is an embedding function that assigns token-level embeddings to each token in $s$). A contextual representation for each $s$ is then built by applying an element-wise max operation over $h_{s}$ as follows:
With these contextual representations, different baseline models can be constructed. For example, a Choice-Only model, which is a variant of the well-known hypothesis-only baseline used in NLI BIBREF46, scores each choice $c_{i}$ in the following way:
for $\textbf {W}^{T} \in \mathbb {R}^{2h}$ independently of the question and assigns a probability to each answer $p_{i}^{(j)} \propto e^{\alpha _{i}^{(j)}}$.
A slight variant of this model, the Choice-to-choice model, tries to single out a given answer choice relative to other choices by scoring all choice pairs $\alpha _{i,i^{\prime }}^{(j)} = \textsc {Att}(r^{(j)}_{c_{i}},r^{(j)}_{c_{i^{\prime }}}) \in \mathbb {R}$ using a learned attention mechanism Att and finding the choice with the minimal similarity to other options (for full details, see their original paper). In using these partial-input baselines, which we train directly on each target probe, we can check whether systematic biases related to answer choices were introduced into the data creation process.
A Question-to-choice model, in contrast, uses the contextual representations for each question and individual choice and an attention model Att model to get a score $\alpha ^{(j)}_{q,i} = \textsc {Att}(r^{(j)}_{q},r^{(j)}_{c_{i}}) \in \mathbb {R}$ as above. Here we also experiment with using ESIM BIBREF47 to generate the contextual representations $r$, as well as a simpler VecSimilarity model that measures the average vector similarity between question and answer tokens: $\alpha ^{(j)}_{q,i} = \textsc {Sim}(\textsc {embed}(q^{(j)}),\textsc {embed}(c^{(j)}_{i}))$. In contrast to the models above, these sets of baselines are used to check for artifacts between questions and answers that are not captured in the partial-input baselines (see discussion in BIBREF49) and ensure that the overall MCQA tasks are sufficiently difficult for our transformer models.
Probing Methodology and Modeling ::: Inoculation and Pre-training
Using the various models introduced above, we train these models on benchmark tasks in the science domain and look at model performance on our probes with and without additional training on samples of probe data, building on the idea of inoculation from BIBREF23. Model inoculation is the idea of continuing to train models on new challenge tasks (in our cases, separately for each probe) using only a small amount of examples. Unlike in ordinary fine-tuning, the goal is not to learn an entirely re-purposed model, but to improve on (or vaccinate against) particular phenomena (e.g., our synthetic probes) that potentially deviate from a model's original training distribution (but that nonetheless might involve knowledge already contained in the model).
In the variant proposed in BIBREF22, for each pre-trained (science) model and architecture $M_{a}$ we continue training the model on $k$ new probe examples (with a maximum of $k=$ 3k) under a set of different hyper-parameter configurations $j \in \lbrace 1, ..., J\rbrace $ and identify, for each $k$, the model $M_{*}^{a,k}$ with the best aggregate performance $S$ on the original (orig) and new task:
As in BIBREF22, we found all models to be especially sensitive to different learning rates, and performed comprehensive hyper-parameters searches that also manipulate the number of iterations and random seeds used.
Using this methodology, we can see how much exposure to new data it takes for a given model to master a new task, and whether there are phenomena that stress particular models (e.g., lead to catastrophic forgetting of the original task). Given the restrictions on the number of fine-tuning examples, our assumption is that when models are able to maintain good performance on their original task during inoculation, the quickness with which they are able to learn the inoculated task provides evidence of prior competence, which is precisely what we aim to probe. To measure past performance, we define a model's inoculation cost as the difference in the performance of this model on its original task before and after inoculation.
We pre-train on an aggregated training set of the benchmark science exams detailed in Table TABREF21, and created an aggregate development set of around 4k science questions for evaluating overall science performance and inoculation costs. To handle the mismatch between number of answer choices in these sets, we made all sets 5-way by adding empty answers as needed. We also experimented with a slight variant of inoculation, called add-some inoculation, which involves balancing the inoculation training sets with naturalistic science questions. We reserve the MCQL dataset in Table TABREF21 for this purpose, and experiment with balancing each probe example with a science example (x1 matching) and adding twice as many science questions (x2 matching, up to 3k) for each new example.
Probing Methodology and Modeling ::: Evaluating Model Competence
The standard way to evaluate our MCQA models is by looking at the overall accuracy of the correct answer prediction, or what we call instance-level accuracy (as in Table TABREF25). Given the nature of our data and the existence of semantic clusters as detailed in Section SECREF11 (i.e., sets of questions and answers under different distractor choices and inference complexity), we also measure a model's cluster-level (or strict cluster) accuracy, which requires correctly answering all questions in a cluster. Example semantic clusters are shown in Table TABREF30; in the first case, there are 6 ISA$^\uparrow $ questions (including perturbations) about the concept trouser.n.01 (e.g., involving knowing that trousers are a type of consumer good and garment/clothing), which a model must answer in order to receive full credit.
Our cluster-based analysis is motivated by the idea that if a model truly knows the meaning of a given concept, such as the concept of trousers, then it should be able to answer arbitrary questions about this concept without sensitivity to varied distractors. While our strict cluster metric is simplistic, it takes inspiration from work on visual QA BIBREF53, and allows us to evaluate how consistent and robust models are across our different probes, and to get insight into whether errors are concentrated on a small set of concepts or widespread across clusters.
Results and Findings
In this section, we provide the results of the empirical questions first introduced in Figure FIGREF1, starting with the results of our baseline models.
Results and Findings ::: Are our Probes Sufficiently Challenging?
As shown in Table TABREF25, most of our partial-input baselines (i.e., Choice-Only and Choice-to-Choice models) failed to perform well on our dataset probes across a wide range of models, showing that such probes are generally immune from biases relating to how distractors were generated. As already discussed in Section SECREF13, however, initial versions of the DictionaryQA dataset had unforeseen biases partly related to whether distractors were sampled from entries without example sentences, which resulted in high Choice-Only-GloVe scores ranging around 56% accuracy before a filtering step was applied to remove these distractors.
We had similar issues with the hypernymy probe which, even after a filtering step that used our Choice-to-Choice-GloVe model, still leads to high results on the BERT and RoBERTa choice-only models. Given that several attempts were made to entirely de-duplicate the different splits (both in terms of gold answers and distractor types), the source of these biases is not at all obvious, which shows how easy it is for unintended biases in expert knowledge to appear in the resulting datasets and the importance of having rigorous baselines. We also note the large gap in some cases between the BERT and RoBERTa versus GloVe choice-only models, which highlights the need for having partial-input baselines that use the best available models.
Using a more conventional set of Task-Specific QA models (i.e., the LSTM-based Question-to-Choice models trained directly on the probes), we can see that results are not particularly strong on any of the datasets, suggesting that our probes are indeed sufficiently challenging and largely immune from overt artifacts. The poor performance of the VecSimilarity (which uses pre-trained Word2Vec embeddings without additional training) provides additional evidence that elementary lexical matching strategies are insufficient for solving any of the probing tasks.
Results and Findings ::: How well do pre-trained MCQA models do?
Science models that use non-transformer based encoders, such as the ESIM model with GloVe and ELMO, perform poorly across all probes, in many cases scoring near random chance, showing limits to how well they generalize from science to other tasks even with pre-trained GloVe and ELMO embeddings. In sharp contrast, the transformer models have mixed results, the most striking result being the RoBERTa models on the definitions and synonymy probes (achieving a test accuracy of 77% and 61%, respectively), which outperform several of the task-specific LSTM models trained directly on the probes. At first glance, this suggests that RoBERTa, which generally far outpaces even BERT across most probes, has high competence of definitions and synonyms even without explicit training on our new tasks.
Given the controlled nature of our probes, we can get a more detailed view of how well the science models are performing across different reasoning and distractor types, as shown in the first column of Figure FIGREF28 for ESIM and RoBERTa. The ESIM science model without training has uniformly poor performance across all categories, whereas the performance of RoBERTa is more varied. Across all datasets and number of hops (i.e., the rows in the heat maps), model performance for RoBERTa is consistently highest among examples with random distractors (i.e., the first column), and lowest in cases involving distractors that are closest in WordNet space (e.g., sister and ISA, or up/down, distractors of distance $k^{\prime }=1$). This is not surprising, given that, in the first case, random distractors are likely to be the easiest category (and the opposite for distractors close in space), but suggests that RoBERTa might only be getting the easiest cases correct.
Model performance also clearly degrades for hypernymy and hyponymy across all models as the number of hops $k$ increases (see red dashed boxes). For example, problems that involve hyponym reasoning with sister distractors of distance $k^{\prime }=1$ (i.e., the second column) degrades from 47% to 15% when the number of hops $k$ increases from 1 to 4. This general tendency persists even after additional fine-tuning, as we discuss next, and gives evidence that models are limited in their capacity for certain types of multi-hop inferences.
As discussed by BIBREF26, the choice of generation templates can have a significant effect on model performance. The results so far should therefore be regarded as a lower bound on model competence. It is possible that model performance is high for definitions, for example, because the associated templates best align with the science training distribution (which we know little about). For this reason, the subsequent inoculation step is important—it gives the model an opportunity to learn about our target templates and couple this learned knowledge with its general knowledge acquired during pre-training and science training (which is, again, what we aim to probe).
Results and Findings ::: Can Models Be Effectively Inoculated?
Model performance after additional fine-tuning, or inoculation, is shown in the last 3 rows of Table TABREF25, along with learning curves shown in Figure FIGREF29 for a selection of probes and models. In the former case, the performance represents the model (and inoculation amount) with the highest aggregate performance over the old task and new probe. Here we again see the transformer-based models outperform non-transformer models, and that better models correlate with lower inoculation costs. For example, when inoculating on synonymy, the cost for ESIM is around 7% reduced accuracy on its original task, as opposed to $< 1$% and around 1% for BERT and RoBERTa, respectively. This shows the high capacity for transformer models to absorb new tasks with minimal costs, as also observed in BIBREF22 for NLI.
As shown in Figure FIGREF29, transformer models tend to learn most tasks fairly quickly while keeping constant scores on their original tasks (i.e., the flat dashed lines observed in plots 1-4), which gives evidence of high competence. In both cases, add-some inoculation proves to be a cheap and easy way to 1) improve scores on the probing tasks (i.e., the solid black and blue lines in plot 1) and; 2) minimize loss on science (e.g., the blue and black dashed lines in plots 2-4). The opposite is the case for ESIM (plots 5-6); models are generally unable to simultaneously learn individual probes without degrading on their original task, and adding more science data during inoculation confuses models on both tasks.
As shown in Figure FIGREF28, RoBERTa is able to significantly improve performance across most categories even after inoculation with a mere 100 examples (the middle plot), which again provides strong evidence of prior competence. As an example, RoBERTa improves on 2-hop hyponymy inference with random distractors by 18% (from 59% to 77%). After 3k examples, the model has high performance on virtually all categories (the same score increases from 59% to 87%), however results still tends to degrade as a function of hop and distractor complexity, as discussed above.
Despite the high performance of our transformer models after inoculation, model performance on most probes (with the exception of Definitions) averages around 80% for our best models. This suggests that there is still considerable room for improvement, especially for synonymy and word sense, which is a topic that we discuss more in Section SECREF6.
Results and Findings ::: Are Models Consistent across Clusters?
Table TABREF32 shows cluster-level accuracies for the different WordNetQA probes. As with performance across the different inference/distractor categories, these results are mixed. For some probes, such as definitions, our best models appear to be rather robust; e.g., our RoBERTa model has a cluster accuracy of $75\%$, meaning that it can answer all questions perfectly for 75% of the target concepts and that errors are concentrated on a small minority (25%) of concepts. On synonymy and hypernymy, both BERT and RoBERTa appear robust on the majority of concepts, showing that errors are similarly concentrated. In contrast, our best model on hyponymy has an accuracy of 36%, meaning that its errors are spread across many concepts, thus suggesting less robustness.
Table TABREF30 shows a selection of semantic clusters involving ISA reasoning, as well as the model performance over different answers (shown symbolically) and perturbations. For example, in the the second case, the cluster is based around the concept/synset oppose.v.06 and involves 4 inferences and a total 24 questions (i.e., inferences with perturbations). Our weakest model, ESIM, answers only 5 out of 24 questions correctly, whereas RoBERTa gets 21/24. In the other cases, RoBERTa gets all clusters correct, whereas BERT and ESIM get none of them correct.
We emphasize that these results only provide a crude look into model consistency and robustness. Recalling again the details in Table TABREF12, probes differ in terms of average size of clusters. Hyponymy, in virtue of having many more questions per cluster, might simply be a much more difficult dataset. In addition, such a strict evaluation does not take into account potential errors inside of clusters, which is an important issue that we discuss in the next section. We leave addressing such issues and coming up with more insightful cluster-based metrics for future work.
Discussion and Conclusion
We presented several new challenge datasets and a novel methodology for automatically building such datasets from knowledge graphs and taxonomies. We used these to probe state-of-the-art open-domain QA models (centering around models based on variants of BERT). While our general methodology is amendable to any target knowledge resource or QA model/domain, we focus on probing definitions and ISA knowledge using open-source dictionaries and MCQA models trained in the science domain.
We find, consistent with recent probing studies BIBREF26, that transformer-based models have a remarkable ability to answer questions that involve complex forms of relational knowledge, both with and without explicit exposure to our new target tasks. In the latter case, a newer RoBERTa model trained only on benchmark science tasks is able to outperform several task-specific LSTM-based models trained directly on our probing data. When re-trained on small samples (e.g., 100 examples) of probing data using variations of the lossless inoculation strategy from BIBREF22, RoBERTa is able to master many aspects of our probes with virtually no performance loss on its original QA task.
These positive results suggest that transformer-based models, especially models additionally fine-tuned on small samples of synthetic data, can be used in place of task-specific models used for querying relational knowledge, as has already been done for targeted tasks such as word sense disambiguation BIBREF54. Since models seem to already contain considerable amounts of relational knowledge, our simple inoculation strategy, which tries to nudge models to bring out this knowledge explicitly, could serve as a cheaper alternative to recent attempts to build architectures that explicitly incorporate structured knowledge BIBREF55; we see many areas where our inoculation strategy could be improved for such purposes, including having more complex loss functions that manage old and new information, as well as using techniques that take into account network plasticity BIBREF56.
The main appeal of using automatically generate datasets is the ability to systematically manipulate and control the complexity of target questions, which allows for more controlled experimentation and new forms of evaluation. Despite the positive results described above, results that look directly at the effect of different types of distractors and the complexity of reasoning show that our best models, even after additional fine-tuning, struggle with certain categories of hard distractors and multi-hop inferences. For some probes, our cluster-based analysis also reveals that errors are widespread across concept clusters, suggesting that models are not always consistent and robust. These results, taken together with our findings about the vulnerability of synthetic datasets to systematic biases, suggest that there is much room for improvement and that the positive results should be taken with a grain of salt. Developing better ways to evaluate semantic clusters and model robustness would be a step in this direction.
We emphasize that using synthetic versus naturalistic QA data comes with important trade-offs. While we are able to generate large amounts of systematically controlled data at virtually no cost or need for manual annotation, it is much harder to validate the quality of such data at such a scale and such varying levels of complexity. Conversely, with benchmark QA datasets, it is much harder to perform the type of careful manipulations and cluster-based analyses we report here. While we assume that the expert knowledge we employ, in virtue of being hand-curated by human experts, is generally correct, we know that such resources are fallible and error-prone. Initial crowd-sourcing experiments that look at validating samples of our data show high agreement across probes and that human scores correlate with the model trends across the probe categories. More details of these studies are left for future work.
| [
"Unanswerable",
"Yes"
] | 6,391 | qasper | en | null | 3ca65b23b3cb316653709b469b4a9b42b790e0350c76ae1a | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Is WordNet useful for taxonomic reasoning for this task?
Answer: | [
"Unanswerable",
"Yes"
] | qasper | 128 | 27 | |
what were the baselines? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Conventional automatic speech recognition (ASR) systems typically consist of several independently learned components: an acoustic model to predict context-dependent sub-phoneme states (senones) from audio, a graph structure to map senones to phonemes, and a pronunciation model to map phonemes to words. Hybrid systems combine hidden Markov models to model state dependencies with neural networks to predict states BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 . Newer approaches such as end-to-end (E2E) systems reduce the overall complexity of the final system.
Our research builds on prior work that has explored using time-delay neural networks (TDNN), other forms of convolutional neural networks, and Connectionist Temporal Classification (CTC) loss BIBREF4 , BIBREF5 , BIBREF6 . We took inspiration from wav2letter BIBREF6 , which uses 1D-convolution layers. Liptchinsky et al. BIBREF7 improved wav2letter by increasing the model depth to 19 convolutional layers and adding Gated Linear Units (GLU) BIBREF8 , weight normalization BIBREF9 and dropout.
By building a deeper and larger capacity network, we aim to demonstrate that we can match or outperform non end-to-end models on the LibriSpeech and 2000hr Fisher+Switchboard tasks. Like wav2letter, our architecture, Jasper, uses a stack of 1D-convolution layers, but with ReLU and batch normalization BIBREF10 . We find that ReLU and batch normalization outperform other activation and normalization schemes that we tested for convolutional ASR. As a result, Jasper's architecture contains only 1D convolution, batch normalization, ReLU, and dropout layers – operators highly optimized for training and inference on GPUs.
It is possible to increase the capacity of the Jasper model by stacking these operations. Our largest version uses 54 convolutional layers (333M parameters), while our small model uses 34 (201M parameters). We use residual connections to enable this level of depth. We investigate a number of residual options and propose a new residual connection topology we call Dense Residual (DR).
Integrating our best acoustic model with a Transformer-XL BIBREF11 language model allows us to obtain new state-of-the-art (SOTA) results on LibriSpeech BIBREF12 test-clean of 2.95% WER and SOTA results among end-to-end models on LibriSpeech test-other. We show competitive results on Wall Street Journal (WSJ), and 2000hr Fisher+Switchboard (F+S). Using only greedy decoding without a language model we achieve 3.86% WER on LibriSpeech test-clean.
This paper makes the following contributions:
Jasper Architecture
Jasper is a family of end-to-end ASR models that replace acoustic and pronunciation models with a convolutional neural network. Jasper uses mel-filterbank features calculated from 20ms windows with a 10ms overlap, and outputs a probability distribution over characters per frame. Jasper has a block architecture: a Jasper INLINEFORM0 x INLINEFORM1 model has INLINEFORM2 blocks, each with INLINEFORM3 sub-blocks. Each sub-block applies the following operations: a 1D-convolution, batch norm, ReLU, and dropout. All sub-blocks in a block have the same number of output channels.
Each block input is connected directly into the last sub-block via a residual connection. The residual connection is first projected through a 1x1 convolution to account for different numbers of input and output channels, then through a batch norm layer. The output of this batch norm layer is added to the output of the batch norm layer in the last sub-block. The result of this sum is passed through the activation function and dropout to produce the output of the sub-block.
The sub-block architecture of Jasper was designed to facilitate fast GPU inference. Each sub-block can be fused into a single GPU kernel: dropout is not used at inference-time and is eliminated, batch norm can be fused with the preceding convolution, ReLU clamps the result, and residual summation can be treated as a modified bias term in this fused operation.
All Jasper models have four additional convolutional blocks: one pre-processing and three post-processing. See Figure FIGREF7 and Table TABREF8 for details.
We also build a variant of Jasper, Jasper Dense Residual (DR). Jasper DR follows DenseNet BIBREF15 and DenseRNet BIBREF16 , but instead of having dense connections within a block, the output of a convolution block is added to the inputs of all the following blocks. While DenseNet and DenseRNet concatenates the outputs of different layers, Jasper DR adds them in the same way that residuals are added in ResNet. As explained below, we find addition to be as effective as concatenation.
Normalization and Activation
In our study, we evaluate performance of models with:
3 types of normalization: batch norm BIBREF10 , weight norm BIBREF9 , and layer norm BIBREF17
3 types of rectified linear units: ReLU, clipped ReLU (cReLU), and leaky ReLU (lReLU)
2 types of gated units: gated linear units (GLU) BIBREF8 , and gated activation units (GAU) BIBREF18
All experiment results are shown in Table TABREF15 . We first experimented with a smaller Jasper5x3 model to pick the top 3 settings before training on larger Jasper models. We found that layer norm with GAU performed the best on the smaller model. Layer norm with ReLU and batch norm with ReLU came second and third in our tests. Using these 3, we conducted further experiments on a larger Jasper10x4. For larger models, we noticed that batch norm with ReLU outperformed other choices. Thus, leading us to decide on batch normalization and ReLU for our architecture.
During batching, all sequences are padded to match the longest sequence. These padded values caused issues when using layer norm. We applied a sequence mask to exclude padding values from the mean and variance calculation. Further, we computed mean and variance over both the time dimension and channels similar to the sequence-wise normalization proposed by Laurent et al. BIBREF19 . In addition to masking layer norm, we additionally applied masking prior to the convolution operation, and masking the mean and variance calculation in batch norm. These results are shown in Table TABREF16 . Interestingly, we found that while masking before convolution gives a lower WER, using masks for both convolutions and batch norm results in worse performance.
As a final note, we found that training with weight norm was very unstable leading to exploding activations.
Residual Connections
For models deeper than Jasper 5x3, we observe consistently that residual connections are necessary for training to converge. In addition to the simple residual and dense residual model described above, we investigated DenseNet BIBREF15 and DenseRNet BIBREF16 variants of Jasper. Both connect the outputs of each sub-block to the inputs of following sub-blocks within a block. DenseRNet, similar to Dense Residual, connects the output of each output of each block to the input of all following blocks. DenseNet and DenseRNet combine residual connections using concatenation whereas Residual and Dense Residual use addition. We found that Dense Residual and DenseRNet perform similarly with each performing better on specific subsets of LibriSpeech. We decided to use Dense Residual for subsequent experiments. The main reason is that due to concatenation, the growth factor for DenseNet and DenseRNet requires tuning for deeper models whereas Dense Residual simply just repeats a sub-blocks.
Language Model
A language model (LM) is a probability distribution over arbitrary symbol sequences INLINEFORM0 such that more likely sequences are assigned high probabilities. LMs are frequently used to condition beam search. During decoding, candidates are evaluated using both acoustic scores and LM scores. Traditional N-gram LMs have been augmented with neural LMs in recent work BIBREF20 , BIBREF21 , BIBREF22 .
We experiment with statistical N-gram language models BIBREF23 and neural Transformer-XL BIBREF11 models. Our best results use acoustic and word-level N-gram language models to generate a candidate list using beam search with a width of 2048. Next, an external Transformer-XL LM rescores the final list. All LMs were trained on datasets independently from acoustic models. We show results with the neural LM in our Results section. We observed a strong correlation between the quality of the neural LM (measured by perplexity) and WER as shown in Figure FIGREF20 .
NovoGrad
For training, we use either Stochastic Gradient Descent (SGD) with momentum or our own NovoGrad, an optimizer similar to Adam BIBREF14 , except that its second moments are computed per layer instead of per weight. Compared to Adam, it reduces memory consumption and we find it to be more numerically stable.
At each step INLINEFORM0 , NovoGrad computes the stochastic gradient INLINEFORM1 following the regular forward-backward pass. Then the second-order moment INLINEFORM2 is computed for each layer INLINEFORM3 similar to ND-Adam BIBREF27 : DISPLAYFORM0
The second-order moment INLINEFORM0 is used to re-scale gradients INLINEFORM1 before calculating the first-order moment INLINEFORM2 : DISPLAYFORM0
If L2-regularization is used, a weight decay INLINEFORM0 is added to the re-scaled gradient (as in AdamW BIBREF28 ): DISPLAYFORM0
Finally, new weights are computed using the learning rate INLINEFORM0 : DISPLAYFORM0
Using NovoGrad instead of SGD with momentum, we decreased the WER on dev-clean LibriSpeech from 4.00% to 3.64%, a relative improvement of 9% for Jasper DR 10x5. We will further analyze NovoGrad in forthcoming work.
Results
We evaluate Jasper across a number of datasets in various domains. In all experiments, we use dropout and weight decay as regularization. At training time, we use speed perturbation with fixed +/-10% BIBREF29 for LibriSpeech. For WSJ and Hub5'00, we use a random speed perturbation factor between [-10%, 10%] as each utterance is fed into the model. All models have been trained on NVIDIA DGX-1 in mixed precision BIBREF30 using OpenSeq2Seq BIBREF31 . Source code, training configurations, and pretrained models are available.
Read Speech
We evaluated the performance of Jasper on two read speech datasets: LibriSpeech and Wall Street Journal (WSJ). For LibriSpeech, we trained Jasper DR 10x5 using our NovoGrad optimizer for 400 epochs. We achieve SOTA performance on the test-clean subset and SOTA among end-to-end speech recognition models on test-other.
We trained a smaller Jasper 10x3 model with SGD with momentum optimizer for 400 epochs on a combined WSJ dataset (80 hours): LDC93S6A (WSJ0) and LDC94S13A (WSJ1). The results are provided in Table TABREF29 .
Conversational Speech
We also evaluate the Jasper model's performance on a conversational English corpus. The Hub5 Year 2000 (Hub5'00) evaluation (LDC2002S09, LDC2005S13) is widely used in academia. It is divided into two subsets: Switchboard (SWB) and Callhome (CHM). The training data for both the acoustic and language models consisted of the 2000hr Fisher+Switchboard training data (LDC2004S13, LDC2005S13, LDC97S62). Jasper DR 10x5 was trained using SGD with momentum for 50 epochs. We compare to other models trained using the same data and report Hub5'00 results in Table TABREF31 .
We obtain good results for SWB. However, there is work to be done to improve WER on harder tasks such as CHM.
Conclusions
We have presented a new family of neural architectures for end-to-end speech recognition. Inspired by wav2letter's convolutional approach, we build a deep and scalable model, which requires a well-designed residual topology, effective regularization, and a strong optimizer. As our architecture studies demonstrated, a combination of standard components leads to SOTA results on LibriSpeech and competitive results on other benchmarks. Our Jasper architecture is highly efficient for training and inference, and serves as a good baseline approach on top of which to explore more sophisticated regularization, data augmentation, loss functions, language models, and optimization strategies. We are interested to see if our approach can continue to scale to deeper models and larger datasets.
| [
"Unanswerable",
"LF-MMI Attention\nSeq2Seq \nRNN-T \nChar E2E LF-MMI \nPhone E2E LF-MMI \nCTC + Gram-CTC"
] | 1,856 | qasper | en | null | 39352d304bd452de69fcc492ad458b72614e329b4509eb48 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: what were the baselines?
Answer: | [
"Unanswerable",
"LF-MMI Attention\nSeq2Seq \nRNN-T \nChar E2E LF-MMI \nPhone E2E LF-MMI \nCTC + Gram-CTC"
] | qasper | 128 | 28 | |
How many users do they look at? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Over the past two decades, the emergence of social media has enabled the proliferation of traceable human behavior. The content posted by users can reflect who their friends are, what topics they are interested in, or which company they are working for. At the same time, users are listing a number of profile fields to define themselves to others. The utilization of such metadata has proven important in facilitating further developments of applications in advertising BIBREF0 , personalization BIBREF1 , and recommender systems BIBREF2 . However, profile information can be limited, depending on the platform, or it is often deliberately omitted BIBREF3 . To uncloak this information, a number of studies have utilized social media users' footprints to approximate their profiles.
This paper explores the potential of predicting a user's industry –the aggregate of enterprises in a particular field– by identifying industry indicative text in social media. The accurate prediction of users' industry can have a big impact on targeted advertising by minimizing wasted advertising BIBREF4 and improved personalized user experience. A number of studies in the social sciences have associated language use with social factors such as occupation, social class, education, and income BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 . An additional goal of this paper is to examine such findings, and in particular the link between language and occupational class, through a data-driven approach.
In addition, we explore how meaning changes depending on the occupational context. By leveraging word embeddings, we seek to quantify how, for example, cloud might mean a separate concept (e.g., condensed water vapor) in the text written by users that work in environmental jobs while it might be used differently by users in technology occupations (e.g., Internet-based computing).
Specifically, this paper makes four main contributions. First, we build a large, industry-annotated dataset that contains over 20,000 blog users. In addition to their posted text, we also link a number of user metadata including their gender, location, occupation, introduction and interests.
Second, we build content-based classifiers for the industry prediction task and study the effect of incorporating textual features from the users' profile metadata using various meta-classification techniques, significantly improving both the overall accuracy and the average per industry accuracy.
Next, after examining which words are indicative for each industry, we build vector-space representations of word meanings and calculate one deviation for each industry, illustrating how meaning is differentiated based on the users' industries. We qualitatively examine the resulting industry-informed semantic representations of words by listing the words per industry that are most similar to job related and general interest terms.
Finally, we rank the different industries based on the normalized relative frequencies of emotionally charged words (positive and negative) and, in addition, discover that, for both genders, these frequencies do not statistically significantly correlate with an industry's gender dominance ratio.
After discussing related work in Section SECREF2 , we present the dataset used in this study in Section SECREF3 . In Section SECREF4 we evaluate two feature selection methods and examine the industry inference problem using the text of the users' postings. We then augment our content-based classifier by building an ensemble that incorporates several metadata classifiers. We list the most industry indicative words and expose how each industrial semantic field varies with respect to a variety of terms in Section SECREF5 . We explore how the frequencies of emotionally charged words in each gender correlate with the industries and their respective gender dominance ratio and, finally, conclude in Section SECREF6 .
Related Work
Alongside the wide adoption of social media by the public, researchers have been leveraging the newly available data to create and refine models of users' behavior and profiling. There exists a myriad research that analyzes language in order to profile social media users. Some studies sought to characterize users' personality BIBREF9 , BIBREF10 , while others sequenced the expressed emotions BIBREF11 , studied mental disorders BIBREF12 , and the progression of health conditions BIBREF13 . At the same time, a number of researchers sought to predict the social media users' age and/or gender BIBREF14 , BIBREF15 , BIBREF16 , while others targeted and analyzed the ethnicity, nationality, and race of the users BIBREF17 , BIBREF18 , BIBREF19 . One of the profile fields that has drawn a great deal of attention is the location of a user. Among others, Hecht et al. Hecht11 predicted Twitter users' locations using machine learning on nationwide and state levels. Later, Han et al. Han14 identified location indicative words to predict the location of Twitter users down to the city level.
As a separate line of research, a number of studies have focused on discovering the political orientation of users BIBREF15 , BIBREF20 , BIBREF21 . Finally, Li et al. Li14a proposed a way to model major life events such as getting married, moving to a new place, or graduating. In a subsequent study, BIBREF22 described a weakly supervised information extraction method that was used in conjunction with social network information to identify the name of a user's spouse, the college they attended, and the company where they are employed.
The line of work that is most closely related to our research is the one concerned with understanding the relation between people's language and their industry. Previous research from the fields of psychology and economics have explored the potential for predicting one's occupation from their ability to use math and verbal symbols BIBREF23 and the relationship between job-types and demographics BIBREF24 . More recently, Huang et al. Huang15 used machine learning to classify Sina Weibo users to twelve different platform-defined occupational classes highlighting the effect of homophily in user interactions. This work examined only users that have been verified by the Sina Weibo platform, introducing a potential bias in the resulting dataset. Finally, Preotiuc-Pietro et al. Preoctiuc15 predicted the occupational class of Twitter users using the Standard Occupational Classification (SOC) system, which groups the different jobs based on skill requirements. In that work, the data collection process was limited to only users that specifically mentioned their occupation in their self-description in a way that could be directly mapped to a SOC occupational class. The mapping between a substring of their self-description and a SOC occupational class was done manually. Because of the manual annotation step, their method was not scalable; moreover, because they identified the occupation class inside a user self-description, only a very small fraction of the Twitter users could be included (in their case, 5,191 users).
Both of these recent studies are based on micro-blogging platforms, which inherently restrict the number of characters that a post can have, and consequently the way that users can express themselves.
Moreover, both studies used off-the-shelf occupational taxonomies (rather than self-declared occupation categories), resulting in classes that are either too generic (e.g., media, welfare and electronic are three of the twelve Sina Weibo categories), or too intermixed (e.g., an assistant accountant is in a different class from an accountant in SOC). To address these limitations, we investigate the industry prediction task in a large blog corpus consisting of over 20K American users, 40K web-blogs, and 560K blog posts.
Dataset
We compile our industry-annotated dataset by identifying blogger profiles located in the U.S. on the profile finder on http://www.blogger.com, and scraping only those users that had the industry profile element completed.
For each of these bloggers, we retrieve all their blogs, and for each of these blogs we download the 21 most recent blog postings. We then clean these blog posts of HTML tags and tokenize them, and drop those bloggers whose cumulative textual content in their posts is less than 600 characters. Following these guidelines, we identified all the U.S. bloggers with completed industry information.
Traditionally, standardized industry taxonomies organize economic activities into groups based on similar production processes, products or services, delivery systems or behavior in financial markets. Following such assumptions and regardless of their many similarities, a tomato farmer would be categorized into a distinct industry from a tobacco farmer. As demonstrated in Preotiuc-Pietro et al. Preoctiuc15 such groupings can cause unwarranted misclassifications.
The Blogger platform provides a total of 39 different industry options. Even though a completed industry value is an implicit text annotation, we acknowledge the same problem noted in previous studies: some categories are too broad, while others are very similar. To remedy this and following Guibert et al. Guibert71, who argued that the denominations used in a classification must reflect the purpose of the study, we group the different Blogger industries based on similar educational background and similar technical terminology. To do that, we exclude very general categories and merge conceptually similar ones. Examples of broad categories are the Education and the Student options: a teacher could be teaching in any concentration, while a student could be enrolled in any discipline. Examples of conceptually similar categories are the Investment Banking and the Banking options.
The final set of categories is shown in Table TABREF1 , along with the number of users in each category. The resulting dataset consists of 22,880 users, 41,094 blogs, and 561,003 posts. Table TABREF2 presents additional statistics of our dataset.
Text-based Industry Modeling
After collecting our dataset, we split it into three sets: a train set, a development set, and a test set. The sizes of these sets are 17,880, 2,500, and 2,500 users, respectively, with users randomly assigned to these sets. In all the experiments that follow, we evaluate our classifiers by training them on the train set, configure the parameters and measure performance on the development set, and finally report the prediction accuracy and results on the test set. Note that all the experiments are performed at user level, i.e., all the data for one user is compiled into one instance in our data sets.
To measure the performance of our classifiers, we use the prediction accuracy. However, as shown in Table TABREF1 , the available data is skewed across categories, which could lead to somewhat distorted accuracy numbers depending on how well a model learns to predict the most populous classes. Moreover, accuracy alone does not provide a great deal of insight into the individual performance per industry, which is one of the main objectives in this study. Therefore, in our results below, we report: (1) micro-accuracy ( INLINEFORM0 ), calculated as the percentage of correctly classified instances out of all the instances in the development (test) data; and (2) macro-accuracy ( INLINEFORM1 ), calculated as the average of the per-category accuracies, where the per-category accuracy is the percentage of correctly classified instances out of the instances belonging to one category in the development (test) data.
Leveraging Blog Content
In this section, we seek the effectiveness of using solely textual features obtained from the users' postings to predict their industry.
The industry prediction baseline Majority is set by discovering the most frequently featured class in our training set and picking that class in all predictions in the respective development or testing set.
After excluding all the words that are not used by at least three separate users in our training set, we build our AllWords model by counting the frequencies of all the remaining words and training a multinomial Naive Bayes classifier. As seen in Figure FIGREF3 , we can far exceed the Majority baseline performance by incorporating basic language signals into machine learning algorithms (173% INLINEFORM0 improvement).
We additionally explore the potential of improving our text classification task by applying a number of feature ranking methods and selecting varying proportions of top ranked features in an attempt to exclude noisy features. We start by ranking the different features, w, according to their Information Gain Ratio score (IGR) with respect to every industry, i, and training our classifier using different proportions of the top features. INLINEFORM0 INLINEFORM1
Even though we find that using the top 95% of all the features already exceeds the performance of the All Words model on the development data, we further experiment with ranking our features with a more aggressive formula that heavily promotes the features that are tightly associated with any industry category. Therefore, for every word in our training set, we define our newly introduced ranking method, the Aggressive Feature Ranking (AFR), as: INLINEFORM0
In Figure FIGREF3 we illustrate the performance of all four methods in our industry prediction task on the development data. Note that for each method, we provide both the accuracy ( INLINEFORM0 ) and the average per-class accuracy ( INLINEFORM1 ). The Majority and All Words methods apply to all the features; therefore, they are represented as a straight line in the figure. The IGR and AFR methods are applied to varying subsets of the features using a 5% step.
Our experiments demonstrate that the word choice that the users make in their posts correlates with their industry. The first observation in Figure FIGREF3 is that the INLINEFORM0 is proportional to INLINEFORM1 ; as INLINEFORM2 increases, so does INLINEFORM3 . Secondly, the best result on the development set is achieved by using the top 90% of the features using the AFR method. Lastly, the improvements of the IGR and AFR feature selections are not substantially better in comparison to All Words (at most 5% improvement between All Words and AFR), which suggest that only a few noisy features exist and most of the words play some role in shaping the “language" of an industry.
As a final evaluation, we apply on the test data the classifier found to work best on the development data (AFR feature selection, top 90% features), for an INLINEFORM0 of 0.534 and INLINEFORM1 of 0.477.
Leveraging User Metadata
Together with the industry information and the most recent postings of each blogger, we also download a number of accompanying profile elements. Using these additional elements, we explore the potential of incorporating users' metadata in our classifiers.
Table TABREF7 shows the different user metadata we consider together with their coverage percentage (not all users provide a value for all of the profile elements). With the exception of the gender field, the remaining metadata elements shown in Table TABREF7 are completed by the users as a freely editable text field. This introduces a considerable amount of noise in the set of possible metadata values. Examples of noise in the occupation field include values such as “Retired”, “I work.”, or “momma” which are not necessarily informative for our industry prediction task.
To examine whether the metadata fields can help in the prediction of a user's industry, we build classifiers using the different metadata elements. For each metadata element that has a textual value, we use all the words in the training set for that field as features. The only two exceptions are the state field, which is encoded as one feature that can take one out of 50 different values representing the 50 U.S. states; and the gender field, which is encoded as a feature with a distinct value for each user gender option: undefined, male, or female.
As shown in Table TABREF9 , we build four different classifiers using the multinomial NB algorithm: Occu (which uses the words found in the occupation profile element), Intro (introduction), Inter (interests), and Gloc (combined gender, city, state).
In general, all the metadata classifiers perform better than our majority baseline ( INLINEFORM0 of 18.88%). For the Gloc classifier, this result is in alignment with previous studies BIBREF24 . However, the only metadata classifier that outperforms the content classifier is the Occu classifier, which despite missing and noisy occupation values exceeds the content classifier's performance by an absolute 3.2%.
To investigate the promise of combining the five different classifiers we have built so far, we calculate their inter-prediction agreement using Fleiss's Kappa BIBREF25 , as well as the lower prediction bounds using the double fault measure BIBREF26 . The Kappa values, presented in the lower left side of Table TABREF10 , express the classification agreement for categorical items, in this case the users' industry. Lower values, especially values below 30%, mean smaller agreement. Since all five classifiers have better-than-baseline accuracy, this low agreement suggests that their predictions could potentially be combined to achieve a better accumulated result.
Moreover, the double fault measure values, which are presented in the top-right hand side of Table TABREF10 , express the proportion of test cases for which both of the two respective classifiers make false predictions, essentially providing the lowest error bound for the pairwise ensemble classifier performance. The lower those numbers are, the greater the accuracy potential of any meta-classification scheme that combines those classifiers. Once again, the low double fault measure values suggest potential gain from a combination of the base classifiers into an ensemble of models.
After establishing the promise of creating an ensemble of classifiers, we implement two meta-classification approaches. First, we combine our classifiers using features concatenation (or early fusion). Starting with our content-based classifier (Text), we successively add the features derived from each metadata element. The results, both micro- and macro-accuracy, are presented in Table TABREF12 . Even though all these four feature concatenation ensembles outperform the content-based classifier in the development set, they fail to outperform the Occu classifier.
Second, we explore the potential of using stacked generalization (or late fusion) BIBREF27 . The base classifiers, referred to as L0 classifiers, are trained on different folds of the training set and used to predict the class of the remaining instances. Those predictions are then used together with the true label of the training instances to train a second classifier, referred to as the L1 classifier: this L1 is used to produce the final prediction on both the development data and the test data. Traditionally, stacking uses different machine learning algorithms on the same training data. However in our case, we use the same algorithm (multinomial NB) on heterogeneous data (i.e., different types of data such as content, occupation, introduction, interests, gender, city and state) in order to exploit all available sources of information.
The ensemble learning results on the development set are shown in Table TABREF12 . We notice a constant improvement for both metrics when adding more classifiers to our ensemble except for the Gloc classifier, which slightly reduces the performance. The best result is achieved using an ensemble of the Text, Occu, Intro, and Inter L0 classifiers; the respective performance on the test set is an INLINEFORM0 of 0.643 and an INLINEFORM1 of 0.564. Finally, we present in Figure FIGREF11 the prediction accuracy for the final classifier for each of the different industries in our test dataset. Evidently, some industries are easier to predict than others. For example, while the Real Estate and Religion industries achieve accuracy figures above 80%, other industries, such as the Banking industry, are predicted correctly in less than 17% of the time. Anecdotal evidence drawn from the examination of the confusion matrix does not encourage any strong association of the Banking class with any other. The misclassifications are roughly uniform across all other classes, suggesting that the users in the Banking industry use language in a non-distinguishing way.
Qualitative Analysis
In this section, we provide a qualitative analysis of the language of the different industries.
Top-Ranked Words
To conduct a qualitative exploration of which words indicate the industry of a user, Table TABREF14 shows the three top-ranking content words for the different industries using the AFR method.
Not surprisingly, the top ranked words align well with what we would intuitively expect for each industry. Even though most of these words are potentially used by many users regardless of their industry in our dataset, they are still distinguished by the AFR method because of the different frequencies of these words in the text of each industry.
Industry-specific Word Similarities
Next, we examine how the meaning of a word is shaped by the context in which it is uttered. In particular, we qualitatively investigate how the speakers' industry affects meaning by learning vector-space representations of words that take into account such contextual information. To achieve this, we apply the contextualized word embeddings proposed by Bamman et al. Bamman14, which are based on an extension of the “skip-gram" language model BIBREF28 .
In addition to learning a global representation for each word, these contextualized embeddings compute one deviation from the common word embedding representation for each contextual variable, in this case, an industry option. These deviations capture the terms' meaning variations (shifts in the INLINEFORM0 -dimensional space of the representations, where INLINEFORM1 in our experiments) in the text of the different industries, however all the embeddings are in the same vector space to allow for comparisons to one another.
Using the word representations learned for each industry, we present in Table TABREF16 the terms in the Technology and the Tourism industries that have the highest cosine similarity with a job-related word, customers. Similarly, Table TABREF17 shows the words in the Environment and the Tourism industries that are closest in meaning to a general interest word, food. More examples are given in the Appendix SECREF8 .
The terms that rank highest in each industry are noticeably different. For example, as seen in Table TABREF17 , while food in the Environment industry is similar to nutritionally and locally, in the Tourism industry the same word relates more to terms such as delicious and pastries. These results not only emphasize the existing differences in how people in different industries perceive certain terms, but they also demonstrate that those differences can effectively be captured in the resulting word embeddings.
Emotional Orientation per Industry and Gender
As a final analysis, we explore how words that are emotionally charged relate to different industries. To quantify the emotional orientation of a text, we use the Positive Emotion and Negative Emotion categories in the Linguistic Inquiry and Word Count (LIWC) dictionary BIBREF29 . The LIWC dictionary contains lists of words that have been shown to correlate with the psychological states of people that use them; for example, the Positive Emotion category contains words such as “happy,” “pretty,” and “good.”
For the text of all the users in each industry we measure the frequencies of Positive Emotion and Negative Emotion words normalized by the text's length. Table TABREF20 presents the industries' ranking for both categories of words based on their relative frequencies in the text of each industry.
We further perform a breakdown per-gender, where we once again calculate the proportion of emotionally charged words in each industry, but separately for each gender. We find that the industry rankings of the relative frequencies INLINEFORM0 of emotionally charged words for the two genders are statistically significantly correlated, which suggests that regardless of their gender, users use positive (or negative) words with a relative frequency that correlates with their industry. (In other words, even if e.g., Fashion has a larger number of women users, both men and women working in Fashion will tend to use more positive words than the corresponding gender in another industry with a larger number of men users such as Automotive.)
Finally, motivated by previous findings of correlations between job satisfaction and gender dominance in the workplace BIBREF30 , we explore the relationship between the usage of Positive Emotion and Negative Emotion words and the gender dominance in an industry. Although we find that there are substantial gender imbalances in each industry (Appendix SECREF9 ), we did not find any statistically significant correlation between the gender dominance ratio in the different industries and the usage of positive (or negative) emotional words in either gender in our dataset.
Conclusion
In this paper, we examined the task of predicting a social media user's industry. We introduced an annotated dataset of over 20,000 blog users and applied a content-based classifier in conjunction with two feature selection methods for an overall accuracy of up to 0.534, which represents a large improvement over the majority class baseline of 0.188.
We also demonstrated how the user metadata can be incorporated in our classifiers. Although concatenation of features drawn both from blog content and profile elements did not yield any clear improvements over the best individual classifiers, we found that stacking improves the prediction accuracy to an overall accuracy of 0.643, as measured on our test dataset. A more in-depth analysis showed that not all industries are equally easy to predict: while industries such as Real Estate and Religion are clearly distinguishable with accuracy figures over 0.80, others such as Banking are much harder to predict.
Finally, we presented a qualitative analysis to provide some insights into the language of different industries, which highlighted differences in the top-ranked words in each industry, word semantic similarities, and the relative frequency of emotionally charged words.
Acknowledgments
This material is based in part upon work supported by the National Science Foundation (#1344257) and by the John Templeton Foundation (#48503). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation or the John Templeton Foundation.
Additional Examples of Word Similarities
| [
"22,880 users",
"20,000"
] | 4,160 | qasper | en | null | 18576b9ee9994a46dc0c7d916a009ff6d0964991541010ee | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How many users do they look at?
Answer: | [
"22,880 users",
"20,000"
] | qasper | 128 | 29 | |
What metrics are used for evaluation? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
In the kitchen, we increasingly rely on instructions from cooking websites: recipes. A cook with a predilection for Asian cuisine may wish to prepare chicken curry, but may not know all necessary ingredients apart from a few basics. These users with limited knowledge cannot rely on existing recipe generation approaches that focus on creating coherent recipes given all ingredients and a recipe name BIBREF0. Such models do not address issues of personal preference (e.g. culinary tastes, garnish choices) and incomplete recipe details. We propose to approach both problems via personalized generation of plausible, user-specific recipes using user preferences extracted from previously consumed recipes.
Our work combines two important tasks from natural language processing and recommender systems: data-to-text generation BIBREF1 and personalized recommendation BIBREF2. Our model takes as user input the name of a specific dish, a few key ingredients, and a calorie level. We pass these loose input specifications to an encoder-decoder framework and attend on user profiles—learned latent representations of recipes previously consumed by a user—to generate a recipe personalized to the user's tastes. We fuse these `user-aware' representations with decoder output in an attention fusion layer to jointly determine text generation. Quantitative (perplexity, user-ranking) and qualitative analysis on user-aware model outputs confirm that personalization indeed assists in generating plausible recipes from incomplete ingredients.
While personalized text generation has seen success in conveying user writing styles in the product review BIBREF3, BIBREF4 and dialogue BIBREF5 spaces, we are the first to consider it for the problem of recipe generation, where output quality is heavily dependent on the content of the instructions—such as ingredients and cooking techniques.
To summarize, our main contributions are as follows:
We explore a new task of generating plausible and personalized recipes from incomplete input specifications by leveraging historical user preferences;
We release a new dataset of 180K+ recipes and 700K+ user reviews for this task;
We introduce new evaluation strategies for generation quality in instructional texts, centering on quantitative measures of coherence. We also show qualitatively and quantitatively that personalized models generate high-quality and specific recipes that align with historical user preferences.
Related Work
Large-scale transformer-based language models have shown surprising expressivity and fluency in creative and conditional long-text generation BIBREF6, BIBREF7. Recent works have proposed hierarchical methods that condition on narrative frameworks to generate internally consistent long texts BIBREF8, BIBREF9, BIBREF10. Here, we generate procedurally structured recipes instead of free-form narratives.
Recipe generation belongs to the field of data-to-text natural language generation BIBREF1, which sees other applications in automated journalism BIBREF11, question-answering BIBREF12, and abstractive summarization BIBREF13, among others. BIBREF14, BIBREF15 model recipes as a structured collection of ingredient entities acted upon by cooking actions. BIBREF0 imposes a `checklist' attention constraint emphasizing hitherto unused ingredients during generation. BIBREF16 attend over explicit ingredient references in the prior recipe step. Similar hierarchical approaches that infer a full ingredient list to constrain generation will not help personalize recipes, and would be infeasible in our setting due to the potentially unconstrained number of ingredients (from a space of 10K+) in a recipe. We instead learn historical preferences to guide full recipe generation.
A recent line of work has explored user- and item-dependent aspect-aware review generation BIBREF3, BIBREF4. This work is related to ours in that it combines contextual language generation with personalization. Here, we attend over historical user preferences from previously consumed recipes to generate recipe content, rather than writing styles.
Approach
Our model's input specification consists of: the recipe name as a sequence of tokens, a partial list of ingredients, and a caloric level (high, medium, low). It outputs the recipe instructions as a token sequence: $\mathcal {W}_r=\lbrace w_{r,0}, \dots , w_{r,T}\rbrace $ for a recipe $r$ of length $T$. To personalize output, we use historical recipe interactions of a user $u \in \mathcal {U}$.
Encoder: Our encoder has three embedding layers: vocabulary embedding $\mathcal {V}$, ingredient embedding $\mathcal {I}$, and caloric-level embedding $\mathcal {C}$. Each token in the (length $L_n$) recipe name is embedded via $\mathcal {V}$; the embedded token sequence is passed to a two-layered bidirectional GRU (BiGRU) BIBREF17, which outputs hidden states for names $\lbrace \mathbf {n}_{\text{enc},j} \in \mathbb {R}^{2d_h}\rbrace $, with hidden size $d_h$. Similarly each of the $L_i$ input ingredients is embedded via $\mathcal {I}$, and the embedded ingredient sequence is passed to another two-layered BiGRU to output ingredient hidden states as $\lbrace \mathbf {i}_{\text{enc},j} \in \mathbb {R}^{2d_h}\rbrace $. The caloric level is embedded via $\mathcal {C}$ and passed through a projection layer with weights $W_c$ to generate calorie hidden representation $\mathbf {c}_{\text{enc}} \in \mathbb {R}^{2d_h}$.
Ingredient Attention: We apply attention BIBREF18 over the encoded ingredients to use encoder outputs at each decoding time step. We define an attention-score function $\alpha $ with key $K$ and query $Q$:
with trainable weights $W_{\alpha }$, bias $\mathbf {b}_{\alpha }$, and normalization term $Z$. At decoding time $t$, we calculate the ingredient context $\mathbf {a}_{t}^{i} \in \mathbb {R}^{d_h}$ as:
Decoder: The decoder is a two-layer GRU with hidden state $h_t$ conditioned on previous hidden state $h_{t-1}$ and input token $w_{r, t}$ from the original recipe text. We project the concatenated encoder outputs as the initial decoder hidden state:
To bias generation toward user preferences, we attend over a user's previously reviewed recipes to jointly determine the final output token distribution. We consider two different schemes to model preferences from user histories: (1) recipe interactions, and (2) techniques seen therein (defined in data). BIBREF19, BIBREF20, BIBREF21 explore similar schemes for personalized recommendation.
Prior Recipe Attention: We obtain the set of prior recipes for a user $u$: $R^+_u$, where each recipe can be represented by an embedding from a recipe embedding layer $\mathcal {R}$ or an average of the name tokens embedded by $\mathcal {V}$. We attend over the $k$-most recent prior recipes, $R^{k+}_u$, to account for temporal drift of user preferences BIBREF22. These embeddings are used in the `Prior Recipe' and `Prior Name' models, respectively.
Given a recipe representation $\mathbf {r} \in \mathbb {R}^{d_r}$ (where $d_r$ is recipe- or vocabulary-embedding size depending on the recipe representation) the prior recipe attention context $\mathbf {a}_{t}^{r_u}$ is calculated as
Prior Technique Attention: We calculate prior technique preference (used in the `Prior Tech` model) by normalizing co-occurrence between users and techniques seen in $R^+_u$, to obtain a preference vector $\rho _{u}$. Each technique $x$ is embedded via a technique embedding layer $\mathcal {X}$ to $\mathbf {x}\in \mathbb {R}^{d_x}$. Prior technique attention is calculated as
where, inspired by copy mechanisms BIBREF23, BIBREF24, we add $\rho _{u,x}$ for technique $x$ to emphasize the attention by the user's prior technique preference.
Attention Fusion Layer: We fuse all contexts calculated at time $t$, concatenating them with decoder GRU output and previous token embedding:
We then calculate the token probability:
and maximize the log-likelihood of the generated sequence conditioned on input specifications and user preferences. fig:ex shows a case where the Prior Name model attends strongly on previously consumed savory recipes to suggest the usage of an additional ingredient (`cilantro').
Recipe Dataset: Food.com
We collect a novel dataset of 230K+ recipe texts and 1M+ user interactions (reviews) over 18 years (2000-2018) from Food.com. Here, we restrict to recipes with at least 3 steps, and at least 4 and no more than 20 ingredients. We discard users with fewer than 4 reviews, giving 180K+ recipes and 700K+ reviews, with splits as in tab:recipeixnstats.
Our model must learn to generate from a diverse recipe space: in our training data, the average recipe length is 117 tokens with a maximum of 256. There are 13K unique ingredients across all recipes. Rare words dominate the vocabulary: 95% of words appear $<$100 times, accounting for only 1.65% of all word usage. As such, we perform Byte-Pair Encoding (BPE) tokenization BIBREF25, BIBREF26, giving a training vocabulary of 15K tokens across 19M total mentions. User profiles are similarly diverse: 50% of users have consumed $\le $6 recipes, while 10% of users have consumed $>$45 recipes.
We order reviews by timestamp, keeping the most recent review for each user as the test set, the second most recent for validation, and the remainder for training (sequential leave-one-out evaluation BIBREF27). We evaluate only on recipes not in the training set.
We manually construct a list of 58 cooking techniques from 384 cooking actions collected by BIBREF15; the most common techniques (bake, combine, pour, boil) account for 36.5% of technique mentions. We approximate technique adherence via string match between the recipe text and technique list.
Experiments and Results
For training and evaluation, we provide our model with the first 3-5 ingredients listed in each recipe. We decode recipe text via top-$k$ sampling BIBREF7, finding $k=3$ to produce satisfactory results. We use a hidden size $d_h=256$ for both the encoder and decoder. Embedding dimensions for vocabulary, ingredient, recipe, techniques, and caloric level are 300, 10, 50, 50, and 5 (respectively). For prior recipe attention, we set $k=20$, the 80th %-ile for the number of user interactions. We use the Adam optimizer BIBREF28 with a learning rate of $10^{-3}$, annealed with a decay rate of 0.9 BIBREF29. We also use teacher-forcing BIBREF30 in all training epochs.
In this work, we investigate how leveraging historical user preferences can improve generation quality over strong baselines in our setting. We compare our personalized models against two baselines. The first is a name-based Nearest-Neighbor model (NN). We initially adapted the Neural Checklist Model of BIBREF0 as a baseline; however, we ultimately use a simple Encoder-Decoder baseline with ingredient attention (Enc-Dec), which provides comparable performance and lower complexity. All personalized models outperform baseline in BPE perplexity (tab:metricsontest) with Prior Name performing the best. While our models exhibit comparable performance to baseline in BLEU-1/4 and ROUGE-L, we generate more diverse (Distinct-1/2: percentage of distinct unigrams and bigrams) and acceptable recipes. BLEU and ROUGE are not the most appropriate metrics for generation quality. A `correct' recipe can be written in many ways with the same main entities (ingredients). As BLEU-1/4 capture structural information via n-gram matching, they are not correlated with subjective recipe quality. This mirrors observations from BIBREF31, BIBREF8.
We observe that personalized models make more diverse recipes than baseline. They thus perform better in BLEU-1 with more key entities (ingredient mentions) present, but worse in BLEU-4, as these recipes are written in a personalized way and deviate from gold on the phrasal level. Similarly, the `Prior Name' model generates more unigram-diverse recipes than other personalized models and obtains a correspondingly lower BLEU-1 score.
Qualitative Analysis: We present sample outputs for a cocktail recipe in tab:samplerecipes, and additional recipes in the appendix. Generation quality progressively improves from generic baseline output to a blended cocktail produced by our best performing model. Models attending over prior recipes explicitly reference ingredients. The Prior Name model further suggests the addition of lemon and mint, which are reasonably associated with previously consumed recipes like coconut mousse and pork skewers.
Personalization: To measure personalization, we evaluate how closely the generated text corresponds to a particular user profile. We compute the likelihood of generated recipes using identical input specifications but conditioned on ten different user profiles—one `gold' user who consumed the original recipe, and nine randomly generated user profiles. Following BIBREF8, we expect the highest likelihood for the recipe conditioned on the gold user. We measure user matching accuracy (UMA)—the proportion where the gold user is ranked highest—and Mean Reciprocal Rank (MRR) BIBREF32 of the gold user. All personalized models beat baselines in both metrics, showing our models personalize generated recipes to the given user profiles. The Prior Name model achieves the best UMA and MRR by a large margin, revealing that prior recipe names are strong signals for personalization. Moreover, the addition of attention mechanisms to capture these signals improves language modeling performance over a strong non-personalized baseline.
Recipe Level Coherence: A plausible recipe should possess a coherent step order, and we evaluate this via a metric for recipe-level coherence. We use the neural scoring model from BIBREF33 to measure recipe-level coherence for each generated recipe. Each recipe step is encoded by BERT BIBREF34. Our scoring model is a GRU network that learns the overall recipe step ordering structure by minimizing the cosine similarity of recipe step hidden representations presented in the correct and reverse orders. Once pretrained, our scorer calculates the similarity of a generated recipe to the forward and backwards ordering of its corresponding gold label, giving a score equal to the difference between the former and latter. A higher score indicates better step ordering (with a maximum score of 2). tab:coherencemetrics shows that our personalized models achieve average recipe-level coherence scores of 1.78-1.82, surpassing the baseline at 1.77.
Recipe Step Entailment: Local coherence is also crucial to a user following a recipe: it is crucial that subsequent steps are logically consistent with prior ones. We model local coherence as an entailment task: predicting the likelihood that a recipe step follows the preceding. We sample several consecutive (positive) and non-consecutive (negative) pairs of steps from each recipe. We train a BERT BIBREF34 model to predict the entailment score of a pair of steps separated by a [SEP] token, using the final representation of the [CLS] token. The step entailment score is computed as the average of scores for each set of consecutive steps in each recipe, averaged over every generated recipe for a model, as shown in tab:coherencemetrics.
Human Evaluation: We presented 310 pairs of recipes for pairwise comparison BIBREF8 (details in appendix) between baseline and each personalized model, with results shown in tab:metricsontest. On average, human evaluators preferred personalized model outputs to baseline 63% of the time, confirming that personalized attention improves the semantic plausibility of generated recipes. We also performed a small-scale human coherence survey over 90 recipes, in which 60% of users found recipes generated by personalized models to be more coherent and preferable to those generated by baseline models.
Conclusion
In this paper, we propose a novel task: to generate personalized recipes from incomplete input specifications and user histories. On a large novel dataset of 180K recipes and 700K reviews, we show that our personalized generative models can generate plausible, personalized, and coherent recipes preferred by human evaluators for consumption. We also introduce a set of automatic coherence measures for instructional texts as well as personalization metrics to support our claims. Our future work includes generating structured representations of recipes to handle ingredient properties, as well as accounting for references to collections of ingredients (e.g. “dry mix").
Acknowledgements. This work is partly supported by NSF #1750063. We thank all reviewers for their constructive suggestions, as well as Rei M., Sujoy P., Alicia L., Eric H., Tim S., Kathy C., Allen C., and Micah I. for their feedback.
Appendix ::: Food.com: Dataset Details
Our raw data consists of 270K recipes and 1.4M user-recipe interactions (reviews) scraped from Food.com, covering a period of 18 years (January 2000 to December 2018). See tab:int-stats for dataset summary statistics, and tab:samplegk for sample information about one user-recipe interaction and the recipe involved.
Appendix ::: Generated Examples
See tab:samplechx for a sample recipe for chicken chili and tab:samplewaffle for a sample recipe for sweet waffles.
Human Evaluation
We prepared a set of 15 pairwise comparisons per evaluation session, and collected 930 pairwise evaluations (310 per personalized model) over 62 sessions. For each pair, users were given a partial recipe specification (name and 3-5 key ingredients), as well as two generated recipes labeled `A' and `B'. One recipe is generated from our baseline encoder-decoder model and one recipe is generated by one of our three personalized models (Prior Tech, Prior Name, Prior Recipe). The order of recipe presentation (A/B) is randomly selected for each question. A screenshot of the user evaluation interface is given in fig:exeval. We ask the user to indicate which recipe they find more coherent, and which recipe best accomplishes the goal indicated by the recipe name. A screenshot of this survey interface is given in fig:exeval2.
| [
"Byte-Pair Encoding perplexity (BPE PPL),\nBLEU-1,\nBLEU-4,\nROUGE-L,\npercentage of distinct unigram (D-1),\npercentage of distinct bigrams(D-2),\nuser matching accuracy(UMA),\nMean Reciprocal Rank(MRR)\nPairwise preference over baseline(PP)",
"BLEU-1/4 and ROUGE-L, likelihood of generated recipes using identic... | 2,673 | qasper | en | null | 86c75d2a9157cb601a49f8424fa1e6e90fcd724132490bbd | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What metrics are used for evaluation?
Answer: | [
"Byte-Pair Encoding perplexity (BPE PPL),\nBLEU-1,\nBLEU-4,\nROUGE-L,\npercentage of distinct unigram (D-1),\npercentage of distinct bigrams(D-2),\nuser matching accuracy(UMA),\nMean Reciprocal Rank(MRR)\nPairwise preference over baseline(PP)",
"BLEU-1/4 and ROUGE-L, likelihood of generated recipes using identic... | qasper | 128 | 30 | |
What labels do they create on their dataset? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Problem Statement
Spoken conversations still remain the most natural and effortless means of human communication. Thus a lot of valuable information is conveyed and exchanged in such an unstructured form. In telehealth settings, nurses might call discharged patients who have returned home to continue to monitor their health status. Human language technology that can efficiently and effectively extract key information from such conversations is clinically useful, as it can help streamline workflow processes and digitally document patient medical information to increase staff productivity. In this work, we design and prototype a dialogue comprehension system in the question-answering manner, which is able to comprehend spoken conversations between nurses and patients to extract clinical information.
Motivation of Approach
Machine comprehension of written passages has made tremendous progress recently. Large quantities of supervised training data for reading comprehension (e.g. SQuAD BIBREF0 ), the wide adoption and intense experimentation of neural modeling BIBREF1 , BIBREF2 , and the advancements in vector representations of word embeddings BIBREF3 , BIBREF4 all contribute significantly to the achievements obtained so far. The first factor, the availability of large scale datasets, empowers the latter two factors. To date, there is still very limited well-annotated large-scale data suitable for modeling human-human spoken dialogues. Therefore, it is not straightforward to directly port over the recent endeavors in reading comprehension to dialogue comprehension tasks.
In healthcare, conversation data is even scarcer due to privacy issues. Crowd-sourcing is an efficient way to annotate large quantities of data, but less suitable for healthcare scenarios, where domain knowledge is required to guarantee data quality. To demonstrate the feasibility of a dialogue comprehension system used for extracting key clinical information from symptom monitoring conversations, we developed a framework to construct a simulated human-human dialogue dataset to bootstrap such a prototype. Similar efforts have been conducted for human-machine dialogues for restaurant or movie reservations BIBREF5 . To the best of our knowledge, no one to date has done so for human-human conversations in healthcare.
Human-human Spoken Conversations
Human-human spoken conversations are a dynamic and interactive flow of information exchange. While developing technology to comprehend such spoken conversations presents similar technical challenges as machine comprehension of written passages BIBREF6 , the challenges are further complicated by the interactive nature of human-human spoken conversations:
(1) Zero anaphora is more common: Co-reference resolution of spoken utterances from multiple speakers is needed. For example, in Figure FIGREF5 (a) headaches, the pain, it, head bulging all refer to the patient's headache symptom, but they were uttered by different speakers and across multiple utterances and turns. In addition, anaphors are more likely to be omitted (see Figure FIGREF5 (a) A4) as this does not affect the human listener’s understanding, but it might be challenging for computational models.
(2) Thinking aloud more commonly occurs: Since it is more effortless to speak than to type, one is more likely to reveal her running thoughts when talking. In addition, one cannot retract what has been uttered, while in text communications, one is more likely to confirm the accuracy of the information in a written response and revise if necessary before sending it out. Thinking aloud can lead to self-contradiction, requiring more context to fully understand the dialogue; e.g., in A6 in Figure FIGREF5 (a), the patient at first says he has none of the symptoms asked, but later revises his response saying that he does get dizzy after running.
(3) Topic drift is more common and harder to detect in spoken conversations: An example is shown in Figure FIGREF5 (a) in A3, where No is actually referring to cough in the previous question, and then the topic is shifted to headache. In spoken conversations, utterances are often incomplete sentences so traditional linguistic features used in written passages such as punctuation marks indicating syntactic boundaries or conjunction words suggesting discourse relations might no longer exist.
Dialogue Comprehension Task
Figure FIGREF5 (b) illustrates the proposed dialogue comprehension task using a question answering (QA) model. The input are a multi-turn symptom checking dialogue INLINEFORM0 and a query INLINEFORM1 specifying a symptom with one of its attributes; the output is the extracted answer INLINEFORM2 from the given dialogue. A training or test sample is defined as INLINEFORM3 . Five attributes, specifying certain details of clinical significance, are defined to characterize the answer types of INLINEFORM4 : (1) the time the patient has been experiencing the symptom, (2) activities that trigger the symptom (to occur or worsen), (3) the extent of seriousness, (4) the frequency occurrence of the symptom, and (5) the location of symptom. For each symptom/attribute, it can take on different linguistic expressions, defined as entities. Note that if the queried symptom or attribute is not mentioned in the dialogue, the groundtruth output is “No Answer”, as in BIBREF6 .
Reading Comprehension
Large-scale reading comprehension tasks like SQuAD BIBREF0 and MARCO BIBREF7 provide question-answer pairs from a vast range of written passages, covering different kinds of factual answers involving entities such as location and numerical values. Furthermore, HotpotQA BIBREF8 requires multi-step inference and provides numerous answer types. CoQA BIBREF9 and QuAC BIBREF10 are designed to mimic multi-turn information-seeking discussions of the given material. In these tasks, contextual reasoning like coreference resolution is necessary to grasp rich linguistic patterns, encouraging semantic modeling beyond naive lexical matching. Neural networks contribute to impressive progress in semantic modeling: distributional semantic word embeddings BIBREF3 , contextual sequence encoding BIBREF11 , BIBREF12 and the attention mechanism BIBREF13 , BIBREF14 are widely adopted in state-of-the-art comprehension models BIBREF1 , BIBREF2 , BIBREF4 .
While language understanding tasks in dialogue such as domain identification BIBREF15 , slot filling BIBREF16 and user intent detection BIBREF17 have attracted much research interest, work in dialogue comprehension is still limited, if any. It is labor-intensive and time-consuming to obtain a critical mass of annotated conversation data for computational modeling. Some propose to collect text data from human-machine or machine-machine dialogues BIBREF18 , BIBREF5 . In such cases, as human speakers are aware of current limitations of dialogue systems or due to pre-defined assumptions of user simulators, there are fewer cases of zero anaphora, thinking aloud, and topic drift, which occur more often in human-human spoken interactions.
NLP for Healthcare
There is emerging interest in research and development activities at the intersection of machine learning and healthcare , of which much of the NLP related work are centered around social media or online forums (e.g., BIBREF19 , BIBREF20 ), partially due to the world wide web as a readily available source of information. Other work in this area uses public data sources such as MIMIC in electronic health records: text classification approaches have been applied to analyze unstructured clinical notes for ICD code assignment BIBREF21 and automatic intensive emergency prediction BIBREF22 . Sequence-to-sequence textual generation has been used for readable notes based on medical and demographic recordings BIBREF23 . For mental health, there has been more focus on analyzing dialogues. For example, sequential modeling of audio and text have helped detect depression from human-machine interviews BIBREF24 . However, few studies have examined human-human spoken conversations in healthcare settings.
Data Preparation
We used recordings of nurse-initiated telephone conversations for congestive heart failure patients undergoing telemonitoring, post-discharge from the hospital. The clinical data was acquired by the Health Management Unit at Changi General Hospital. This research study was approved by the SingHealth Centralised Institutional Review Board (Protocol 1556561515). The patients were recruited during 2014-2016 as part of their routine care delivery, and enrolled into the telemonitoring health management program with consent for use of anonymized versions of their data for research.
The dataset comprises a total of 353 conversations from 40 speakers (11 nurses, 16 patients, and 13 caregivers) with consent to the use of anonymized data for research. The speakers are 38 to 88 years old, equally distributed across gender, and comprise a range of ethnic groups (55% Chinese, 17% Malay, 14% Indian, 3% Eurasian, and 11% unspecified). The conversations cover 11 topics (e.g., medication compliance, symptom checking, education, greeting) and 9 symptoms (e.g., chest pain, cough) and amount to 41 hours.
Data preprocessing and anonymization were performed by a data preparation team, separate from the data analysis team to maintain data confidentiality. The data preparation team followed standard speech recognition transcription guidelines, where words are transcribed verbatim to include false starts, disfluencies, mispronunciations, and private self-talk. Confidential information were marked and clipped off from the audio and transcribed with predefined tags in the annotation. Conversation topics and clinical symptoms were also annotated and clinically validated by certified telehealth nurses.
Linguistic Characterization on Seed Data
To analyze the linguistic structure of the inquiry-response pairs in the entire 41-hour dataset, we randomly sampled a seed dataset consisting of 1,200 turns and manually categorized them to different types, which are summarized in Table TABREF14 along with the corresponding occurrence frequency statistics. Note that each given utterance could be categorized to more than one type. We elaborate on each utterance type below.
Open-ended Inquiry: Inquiries about general well-being or a particular symptom; e.g., “How are you feeling?” and “Do you cough?”
Detailed Inquiry: Inquiries with specific details that prompt yes/no answers or clarifications; e.g., “Do you cough at night?”
Multi-Intent Inquiry: Inquiring more than one symptom in a question; e.g., “Any cough, chest pain, or headache?”
Reconfirmation Inquiry: The nurse reconfirms particular details; e.g., “Really? At night?” and “Serious or mild?”. This case is usually related to explicit or implicit coreferencing.
Inquiry with Transitional Clauses: During spoken conversations, one might repeat what the other party said, but it is unrelated to the main clause of the question. This is usually due to private self-talk while thinking aloud, and such utterances form a transitional clause before the speaker starts a new topic; e.g., “Chest pain... no chest pain, I see... any cough?”.
Yes/No Response: Yes/No responses seem straightforward, but sometimes lead to misunderstanding if one does not interpret the context appropriately. One case is tag questions: A:“You don't cough at night, do you?” B:`Yes, yes” A:“cough at night?” B:“No, no cough”. Usually when the answer is unclear, clarifying inquiries will be asked for reconfirmation purposes.
Detailed Response: Responses that contain specific information of one symptom, like “I felt tightness in my chest”.
Response with Revision: Revision is infrequent but can affect comprehension significantly. One cause is thinking aloud so a later response overrules the previous one; e.g., “No dizziness, oh wait... last week I felt a bit dizzy when biking”.
Response with Topic Drift: When a symptom/topic like headache is inquired, the response might be: “Only some chest pain at night”, not referring to the original symptom (headache) at all.
Response with Transitional Clauses: Repeating some of the previous content, but often unrelated to critical clinical information and usually followed by topic drift. For example, “Swelling... swelling... I don't cough at night”.
Simulating Symptom Monitoring Dataset for Training
We divide the construction of data simulation into two stages. In Section SECREF16 , we build templates and expression pools using linguistic analysis followed by manual verification. In Section SECREF20 , we present our proposed framework for generating simulated training data. The templates and framework are verified for logical correctness and clinical soundness.
Template Construction
Each utterance in the seed data is categorized according to Table TABREF14 and then abstracted into templates by replacing entity phrases like cough and often with respective placeholders “#symptom#” and “#frequency#”. The templates are refined through verifying logical correctness and injecting expression diversity by linguistically trained researchers. As these replacements do not alter the syntactic structure, we interchange such placeholders with various verbal expressions to enlarge the simulated training set in Section SECREF20 . Clinical validation was also conducted by certified telehealth nurses.
For the 9 symptoms (e.g. chest pain, cough) and 5 attributes (e.g., extent, frequency), we collect various expressions from the seed data, and expand them through synonym replacement. Some attributes are unique to a particular symptom; e.g., “left leg” in #location# is only suitable to describe the symptom swelling, but not the symptom headache. Therefore, we only reuse general expressions like “slight” in #extent# across different symptoms to diversify linguistic expressions.
Two linguistically trained researchers constructed expression pools for each symptom and each attribute to account for different types of paraphrasing and descriptions. These expression pools are used in Section SECREF20 (c).
Simulated Data Generation Framework
Figure FIGREF15 shows the five steps we use to generate multi-turn symptom monitoring dialogue samples.
(a) Topic Selection: While nurses might prefer to inquire the symptoms in different orders depending on the patient's history, our preliminary analysis shows that modeling results do not differ noticeably if topics are of equal prior probabilities. Thus we adopt this assumption for simplicity.
(b) Template Selection: For each selected topic, one inquiry template and one response template are randomly chosen to compose a turn. To minimize adverse effects of underfitting, we redistributed the frequency distribution in Table TABREF14 : For utterance types that are below 15%, we boosted them to 15%, and the overall relative distribution ranking is balanced and consistent with Table TABREF14 .
(c) Enriching Linguistic Expressions: The placeholders in the selected templates are substituted with diverse expressions from the expression pools in Section UID19 to characterize the symptoms and their corresponding attributes.
(d) Multi-Turn Dialogue State Tracking: A greedy algorithm is applied to complete conversations. A “completed symptoms” list and a “to-do symptoms” list are used for symptom topic tracking. We also track the “completed attributes" and “to-do attributes". For each symptom, all related attributes are iterated. A dialogue ends only when all possible entities are exhausted, generating a multi-turn dialogue sample, which encourages the model to learn from the entire discussion flow rather than a single turn to comprehend contextual dependency. The average length of a simulated dialogue is 184 words, which happens to be twice as long as an average dialogue from the real-world evaluation set. Moreover, to model the roles of the respondents, we set the ratio between patients and caregivers to 2:1; this statistic is inspired by the real scenarios in the seed dataset. For both the caregivers and patients, we assume equal probability of both genders. The corresponding pronouns in the conversations are thus determined by the role and gender of these settings.
(e) Multi-Turn Sample Annotation: For each multi-turn dialogue, a query is specified by a symptom and an attribute. The groundtruth output of the QA system is automatically labeled based on the template generation rules, but also manually verified to ensure annotation quality. Moreover, we adopt the unanswerable design in BIBREF6 : when the patient does not mention a particular symptom, the answer is defined as “No Answer”. This process is repeated until all logical permutations of symptoms and attributes are exhausted.
Experiments
Model Design
We implemented an established model in reading comprehension, a bi-directional attention pointer network BIBREF1 , and equipped it with an answerable classifier, as depicted in Figure FIGREF21 . First, tokens in the given dialogue INLINEFORM0 and query INLINEFORM1 are converted into embedding vectors. Then the dialogue embeddings are fed to a bi-directional LSTM encoding layer, generating a sequence of contextual hidden states. Next, the hidden states and query embeddings are processed by a bi-directional attention layer, fusing attention information in both context-to-query and query-to-context directions. The following two bi-directional LSTM modeling layers read the contextual sequence with attention. Finally, two respective linear layers with softmax functions are used to estimate token INLINEFORM2 's INLINEFORM3 and INLINEFORM4 probability of the answer span INLINEFORM5 .
In addition, we add a special tag “[SEQ]” at the head of INLINEFORM0 to account for the case of “No answer” BIBREF4 and adopt an answerable classifier as in BIBREF25 . More specifically, when the queried symptom or attribute is not mentioned in the dialogue, the answer span should point to the tag “[SEQ]” and answerable probability should be predicted as 0.
Implementation Details
The model was trained via gradient backpropagation with the cross-entropy loss function of answer span prediction and answerable classification, optimized by Adam algorithm BIBREF26 with initial learning rate of INLINEFORM0 . Pre-trained GloVe BIBREF3 embeddings (size INLINEFORM1 ) were used. We re-shuffled training samples at each epoch (batch size INLINEFORM2 ). Out-of-vocabulary words ( INLINEFORM3 ) were replaced with a fixed random vector. L2 regularization and dropout (rate INLINEFORM4 ) were used to alleviate overfitting BIBREF27 .
Evaluation Setup
To evaluate the effectiveness of our linguistically-inspired simulation approach, the model is trained on the simulated data (see Section SECREF20 ). We designed 3 evaluation sets: (1) Base Set (1,264 samples) held out from the simulated data. (2) Augmented Set (1,280 samples) built by adding two out-of-distribution symptoms, with corresponding dialogue contents and queries, to the Base Set (“bleeding” and “cold”, which never appeared in training data). (3) Real-World Set (944 samples) manually delineated from the the symptom checking portions (approximately 4 hours) of real-world dialogues, and annotated as evaluation samples.
Results
Evaluation results are in Table TABREF25 with exact match (EM) and F1 score in BIBREF0 metrics. To distinguish the correct answer span from the plausible ones which contain the same words, we measure the scores on the position indices of tokens. Our results show that both EM and F1 score increase with training sample size growing and the optimal size in our setting is 100k. The best-trained model performs well on both the Base Set and the Augmented Set, indicating that out-of-distribution symptoms do not affect the comprehension of existing symptoms and outputs reasonable answers for both in- and out-of-distribution symptoms. On the Real-World Set, we obtained 78.23 EM score and 80.18 F1 score respectively.
Error analysis suggests the performance drop from the simulated test sets is due to the following: 1) sparsity issues resulting from the expression pools excluding various valid but sporadic expressions. 2) nurses and patients occasionally chit-chat in the Real-World Set, which is not simulated in the training set. At times, these chit-chats make the conversations overly lengthy, causing the information density to be lower. These issues could potentially distract and confuse the comprehension model. 3) an interesting type of infrequent error source, caused by patients elaborating on possible causal relations of two symptoms. For example, a patient might say “My giddiness may be due to all this cough”. We are currently investigating how to close this performance gap efficiently.
Ablation Analysis
To assess the effectiveness of bi-directional attention, we bypassed the bi-attention layer by directly feeding the contextual hidden states and query embeddings to the modeling layer. To evaluate the pre-trained GloVe embeddings, we randomly initialized and trained the embeddings from scratch. These two procedures lead to 10% and 18% performance degradation on the Augmented Set and Real-World Set, respectively (see Table TABREF27 ).
Conclusion
We formulated a dialogue comprehension task motivated by the need in telehealth settings to extract key clinical information from spoken conversations between nurses and patients. We analyzed linguistic characteristics of real-world human-human symptom checking dialogues, constructed a simulated dataset based on linguistically inspired and clinically validated templates, and prototyped a QA system. The model works effectively on a simulated test set using symptoms excluded during training and on real-world conversations between nurses and patients. We are currently improving the model's dialogue comprehension capability in complex reasoning and context understanding and also applying the QA model to summarization and virtual nurse applications.
Acknowledgements
Research efforts were supported by funding for Digital Health and Deep Learning I2R (DL2 SSF Project No: A1718g0045) and the Science and Engineering Research Council (SERC Project No: A1818g0044), A*STAR, Singapore. In addition, this work was conducted using resources and infrastructure provided by the Human Language Technology unit at I2R. The telehealth data acquisition was funded by the Economic Development Board (EDB), Singapore Living Lab Fund and Philips Electronics – Hospital to Home Pilot Project (EDB grant reference number: S14-1035-RF-LLF H and W).
We acknowledge valuable support and assistance from Yulong Lin, Yuanxin Xiang, and Ai Ti Aw at the Institute for Infocomm Research (I2R); Weiliang Huang at the Changi General Hospital (CGH) Department of Cardiology, Hong Choon Oh at CGH Health Services Research, and Edris Atikah Bte Ahmad, Chiu Yan Ang, and Mei Foon Yap of the CGH Integrated Care Department.
We also thank Eduard Hovy and Bonnie Webber for insightful discussions and the anonymous reviewers for their precious feedback to help improve and extend this piece of work.
| [
"(1) the time the patient has been experiencing the symptom, (2) activities that trigger the symptom (to occur or worsen), (3) the extent of seriousness, (4) the frequency occurrence of the symptom, and (5) the location of symptom, No Answer",
"the time the patient has been experiencing the symptom, activities th... | 3,424 | qasper | en | null | b72f9154e71c03d0403e06e50063325961ea2ad27c245763 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What labels do they create on their dataset?
Answer: | [
"(1) the time the patient has been experiencing the symptom, (2) activities that trigger the symptom (to occur or worsen), (3) the extent of seriousness, (4) the frequency occurrence of the symptom, and (5) the location of symptom, No Answer",
"the time the patient has been experiencing the symptom, activities th... | qasper | 128 | 31 | |
How much data is needed to train the task-specific encoder? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Assembling training corpora of annotated natural language examples in specialized domains such as biomedicine poses considerable challenges. Experts with the requisite domain knowledge to perform high-quality annotation tend to be expensive, while lay annotators may not have the necessary knowledge to provide high-quality annotations. A practical approach for collecting a sufficiently large corpus would be to use crowdsourcing platforms like Amazon Mechanical Turk (MTurk). However, crowd workers in general are likely to provide noisy annotations BIBREF0 , BIBREF1 , BIBREF2 , an issue exacerbated by the technical nature of specialized content. Some of this noise may reflect worker quality and can be modeled BIBREF0 , BIBREF1 , BIBREF3 , BIBREF4 , but for some instances lay people may simply lack the domain knowledge to provide useful annotation.
In this paper we report experiments on the EBM-NLP corpus comprising crowdsourced annotations of medical literature BIBREF5 . We operationalize the concept of annotation difficulty and show how it can be exploited during training to improve information extraction models. We then obtain expert annotations for the abstracts predicted to be most difficult, as well as for a similar number of randomly selected abstracts. The annotation of highly specialized data and the use of lay and expert annotators allow us to examine the following key questions related to lay and expert annotations in specialized domains:
Can we predict item difficulty? We define a training instance as difficult if a lay annotator or an automated model disagree on its labeling. We show that difficulty can be predicted, and that it is distinct from inter-annotator agreement. Further, such predictions can be used during training to improve information extraction models.
Are there systematic differences between expert and lay annotations? We observe decidedly lower agreement between lay workers as compared to domain experts. Lay annotations have high precision but low recall with respect to expert annotations in the new data that we collected. More generally, we expect lay annotations to be lower quality, which may translate to lower precision, recall, or both, compared to expert annotations. Can one rely solely on lay annotations? Reasonable models can be trained using lay annotations alone, but similar performance can be achieved using markedly less expert data. This suggests that the optimal ratio of expert to crowd annotations for specialized tasks will depend on the cost and availability of domain experts. Expert annotations are preferable whenever its collection is practical. But in real-world settings, a combination of expert and lay annotations is better than using lay data alone.
Does it matter what data is annotated by experts? We demonstrate that a system trained on combined data achieves better predictive performance when experts annotate difficult examples rather than instances selected at i.i.d. random.
Our contributions in this work are summarized as follows. We define a task difficulty prediction task and show how this is related to, but distinct from, inter-worker agreement. We introduce a new model for difficulty prediction combining learned representations induced via a pre-trained `universal' sentence encoder BIBREF6 , and a sentence encoder learned from scratch for this task. We show that predicting annotation difficulty can be used to improve the task routing and model performance for a biomedical information extraction task. Our results open up a new direction for ensuring corpus quality. We believe that item difficulty prediction will likely be useful in other, non-specialized tasks as well, and that the most effective data collection in specialized domains requires research addressing the fundamental questions we examine here.
Related Work
Crowdsourcing annotation is now a well-studied problem BIBREF7 , BIBREF0 , BIBREF1 , BIBREF2 . Due to the noise inherent in such annotations, there have also been considerable efforts to develop aggregation models that minimize noise BIBREF0 , BIBREF1 , BIBREF3 , BIBREF4 .
There are also several surveys of crowdsourcing in biomedicine specifically BIBREF8 , BIBREF9 , BIBREF10 . Some work in this space has contrasted model performance achieved using expert vs. crowd annotated training data BIBREF11 , BIBREF12 , BIBREF13 . Dumitrache et al. Dumitrache:2018:CGT:3232718.3152889 concluded that performance is similar under these supervision types, finding no clear advantage from using expert annotators. This differs from our findings, perhaps owing to differences in design. The experts we used already hold advanced medical degrees, for instance, while those in prior work were medical students. Furthermore, the task considered here would appear to be of greater difficulty: even a system trained on $\sim $ 5k instances performs reasonably, but far from perfect. By contrast, in some of the prior work where experts and crowd annotations were deemed equivalent, a classifier trained on 300 examples can achieve very high accuracy BIBREF12 .
More relevant to this paper, prior work has investigated methods for `task routing' in active learning scenarios in which supervision is provided by heterogeneous labelers with varying levels of expertise BIBREF14 , BIBREF15 , BIBREF16 , BIBREF17 , BIBREF14 . The related question of whether effort is better spent collecting additional annotations for already labeled (but potentially noisily so) examples or novel instances has also been addressed BIBREF18 . What distinguishes the work here is our focus on providing an operational definition of instance difficulty, showing that this can be predicted, and then using this to inform task routing.
Application Domain
Our specific application concerns annotating abstracts of articles that describe the conduct and results of randomized controlled trials (RCTs). Experimentation in this domain has become easy with the recent release of the EBM-NLP BIBREF5 corpus, which includes a reasonably large training dataset annotated via crowdsourcing, and a modest test set labeled by individuals with advanced medical training. More specifically, the training set comprises 4,741 medical article abstracts with crowdsourced annotations indicating snippets (sequences) that describe the Participants (p), Interventions (i), and Outcome (o) elements of the respective RCT, and the test set is composed of 191 abstracts with p, i, o sequence annotations from three medical experts.
Table 1 shows an example of difficult and easy examples according to our definition of difficulty. The underlined text demarcates the (consensus) reference label provided by domain experts. In the difficult examples, crowd workers marked text distinct from these reference annotations; whereas in the easy cases they reproduced them with reasonable fidelity. The difficult sentences usually exhibit complicated structure and feature jargon.
An abstract may contain some `easy' and some `difficult' sentences. We thus perform our analysis at the sentence level. We split abstracts into sentences using spaCy. We excluded sentences that comprise fewer than two tokens, as these are likely an artifact of errors in sentence splitting. In total, this resulted in 57,505 and 2,428 sentences in the train and test set abstracts, respectively.
Quantifying Task Difficulty
The test set includes annotations from both crowd workers and domain experts. We treat the latter as ground truth and then define the difficulty of sentences in terms of the observed agreement between expert and lay annotators. Formally, for annotation task $t$ and instance $i$ :
$$\text{Difficulty}_{ti} = \frac{\sum _{j=1}^n{f(\text{label}_{ij}, y_i})}{n}$$ (Eq. 3)
where $f$ is a scoring function that measures the quality of the label from worker $j$ for sentence $i$ , as compared to a ground truth annotation, $y_i$ . The difficulty score of sentence $i$ is taken as an average over the scores for all $n$ layworkers. We use Spearmans' correlation coefficient as a scoring function. Specifically, for each sentence we create two vectors comprising counts of how many times each token was annotated by crowd and expert workers, respectively, and calculate the correlation between these. Sentences with no labels are treated as maximally easy; those with only either crowd worker or expert label(s) are assumed maximally difficult.
The training set contains only crowdsourced annotations. To label the training data, we use a 10-fold validation like setting. We iteratively retrain the LSTM-CRF-Pattern sequence tagger of Patel et al. patel2018syntactic on 9 folds of the training data and use that trained model to predict labels for the 10th. In this way we obtain predictions on the full training set. We then use predicted spans as proxy `ground truth' annotations to calculate the difficulty score of sentences as described above; we normalize these to the [ $0, 1$ ] interval. We validate this approximation by comparing the proxy scores against reference scores over the test set, the Pearson's correlation coefficients are 0.57 for Population, 0.71 for Intervention and 0.68 for Outcome.
There exist many sentences that contain neither manual nor predicted annotations. We treat these as maximally easy sentences (with difficulty scores of 0). Such sentences comprise 51%, 42% and 36% for Population, Interventions and Outcomes data respectively, indicating that it is easier to identify sentences that have no Population spans, but harder to identify sentences that have no Interventions or Outcomes spans. This is intuitive as descriptions of the latter two tend to be more technical and dense with medical jargon.
We show the distribution of the automatically labeled scores for sentences that do contain spans in Figure 1 . The mean of the Population (p) sentence scores is significantly lower than that for other types of sentences (i and o), again indicating that they are easier on average to annotate. This aligns with a previous finding that annotating Interventions and Outcomes is more difficult than annotating Participants BIBREF5 .
Many sentences contain spans tagged by the LSTM-CRF-Pattern model, but missed by all crowd workers, resulting in a maximally difficult score (1). Inspection of such sentences revealed that some are truly difficult examples, but others are tagging model errors. In either case, such sentences have confused workers and/or the model, and so we retain them all as `difficult' sentences.
Content describing the p, i and o, respectively, is quite different. As such, one sentence usually contains (at most) only one of these three content types. We thus treat difficulty prediction for the respective label types as separate tasks.
Difficulty is not Worker Agreement
Our definition of difficulty is derived from agreement between expert and crowd annotations for the test data, and agreement between a predictive model and crowd annotations in the training data. It is reasonable to ask if these measures are related to inter-annotator agreement, a metric often used in language technology research to identify ambiguous or difficult items. Here we explicitly verify that our definition of difficulty only weakly correlates with inter-annotator agreement.
We calculate inter-worker agreement between crowd and expert annotators using Spearman's correlation coefficient. As shown in Table 2 , average agreement between domain experts are considerably higher than agreements between crowd workers for all three label types. This is a clear indication that the crowd annotations are noisier.
Furthermore, we compare the correlation between inter-annotator agreement and difficulty scores in the training data. Given that the majority of sentences do not contain a PICO span, we only include in these calculations those that contain a reference label. Pearson's r are 0.34, 0.30 and 0.31 for p, i and o, respectively, confirming that inter-worker agreement and our proposed difficulty score are quite distinct.
Predicting Annotation Difficulty
We treat difficulty prediction as a regression problem, and propose and evaluate neural model variants for the task. We first train RNN BIBREF19 and CNN BIBREF20 models.
We also use the universal sentence encoder (USE) BIBREF6 to induce sentence representations, and train a model using these as features. Following BIBREF6 , we then experiment with an ensemble model that combines the `universal' and task-specific representations to predict annotation difficulty. We expect these universal embeddings to capture general, high-level semantics, and the task specific representations to capture more granular information. Figure 2 depicts the model architecture. Sentences are fed into both the universal sentence encoder and, separately, a task specific neural encoder, yielding two representations. We concatenate these and pass the combined vector to the regression layer.
Experimental Setup and Results
We trained models for each label type separately. Word embeddings were initialized to 300d GloVe vectors BIBREF21 trained on common crawl data; these are fine-tuned during training. We used the Adam optimizer BIBREF22 with learning rate and decay set to 0.001 and 0.99, respectively. We used batch sizes of 16.
We used the large version of the universal sentence encoder with a transformer BIBREF23 . We did not update the pretrained sentence encoder parameters during training. All hyperparamaters for all models (including hidden layers, hidden sizes, and dropout) were tuned using Vizier BIBREF24 via 10-fold cross validation on the training set maximizing for F1.
As a baseline, we also trained a linear Support-Vector Regression BIBREF25 model on $n$ -gram features ( $n$ ranges from 1 to 3).
Table 3 reports Pearson correlation coefficients between the predictions with each of the neural models and the ground truth difficulty scores. Rows 1-4 correspond to individual models, and row 5 reports the ensemble performance. Columns correspond to label type. Results from all models outperform the baseline SVR model: Pearson's correlation coefficients range from 0.550 to 0.622. The regression correlations are the lowest.
The RNN model realizes the strongest performance among the stand-alone (non-ensemble) models, outperforming variants that exploit CNN and USE representations. Combining the RNN and USE further improves results. We hypothesize that this is due to complementary sentence information encoded in universal representations.
For all models, correlations for Intervention and Outcomes are higher than for Population, which is expected given the difficulty distributions in Figure 1 . In these, the sentences are more uniformly distributed, with a fair number of difficult and easier sentences. By contrast, in Population there are a greater number of easy sentences and considerably fewer difficult sentences, which makes the difficulty ranking task particularly challenging.
Better IE with Difficulty Prediction
We next present experiments in which we attempt to use the predicted difficulty during training to improve models for information extraction of descriptions of Population, Interventions and Outcomes from medical article abstracts. We investigate two uses: (1) simply removing the most difficult sentences from the training set, and, (2) re-weighting the most difficult sentences.
We again use LSTM-CRF-Pattern as the base model and experimenting on the EBM-NLP corpus BIBREF5 . This is trained on either (1) the training set with difficult sentences removed, or (2) the full training set but with instances re-weighted in proportion to their predicted difficulty score. Following BIBREF5 , we use the Adam optimizer with learning rate of 0.001, decay 0.9, batch size 20 and dropout 0.5. We use pretrained 200d GloVe vectors BIBREF21 to initialize word embeddings, and use 100d hidden char representations. Each word is thus represented with 300 dimensions in total. The hidden size is 100 for the LSTM in the character representation component, and 200 for the LSTM in the information extraction component. We train for 15 epochs, saving parameters that achieve the best F1 score on a nested development set.
Removing Difficult Examples
We first evaluate changes in performance induced by training the sequence labeling model using less data by removing difficult sentences prior to training. The hypothesis here is that these difficult instances are likely to introduce more noise than signal. We used a cross-fold approach to predict sentence difficulties, training on 9/10ths of the data and scoring the remaining 1/10th at a time. We then sorted sentences by predicted difficulty scores, and experimented with removing increasing numbers of these (in order of difficulty) prior to training the LSTM-CRF-Pattern model.
Figure 3 shows the results achieved by the LSTM-CRF-Pattern model after discarding increasing amounts of the training data: the $x$ and $y$ axes correspond to the the percentage of data removed and F1 scores, respectively. We contrast removing sentences predicted to be difficult with removing them (a) randomly (i.i.d.), and, (b) in inverse order of predicted inter-annotator agreement. The agreement prediction model is trained exactly the same like difficult prediction model, with simply changing the difficult score to annotation agreement. F1 scores actually improve (marginally) when we remove the most difficult sentences, up until we drop 4% of the data for Population and Interventions, and 6% for Outcomes. Removing training points at i.i.d. random degrades performance, as expected. Removing sentences in order of disagreement seems to have similar effect as removing them by difficulty score when removing small amount of the data, but the F1 scores drop much faster when removing more data. These findings indicate that sentences predicted to be difficult are indeed noisy, to the extent that they do not seem to provide the model useful signal.
Re-weighting by Difficulty
We showed above that removing a small number of the most difficult sentences does not harm, and in fact modestly improves, medical IE model performance. However, using the available data we are unable to test if this will be useful in practice, as we would need additional data to determine how many difficult sentences should be dropped.
We instead explore an alternative, practical means of exploiting difficulty predictions: we re-weight sentences during training inversely to their predicted difficulty. Formally, we weight sentence $i$ with difficulty scores above $\tau $ according to: $1-a\cdot (d_i-\tau )/(1-\tau )$ , where $d_i$ is the difficulty score for sentence $i$ , and $a$ is a parameter codifying the minimum weight value. We set $\tau $ to 0.8 so as to only re-weight sentences with difficulty in the top 20th percentile, and we set $a$ to 0.5. The re-weighting is equivalent to down-sampling the difficult sentences. LSTM-CRF-Pattern is our base model.
Table 4 reports the precision, recall and F1 achieved both with and without sentence re-weighting. Re-weighting improves all metrics modestly but consistently. All F1 differences are statistically significant under a sign test ( $p<0.01$ ). The model with best precision is different for Patient, Intervention and Outcome labels. However re-weighting by difficulty does consistently yield the best recall for all three extraction types, with the most notable improvement for i and o, where recall improved by 10 percentage points. This performance increase translated to improvements in F1 across all types, as compared to the base model and to re-weighting by agreement.
Involving Expert Annotators
The preceding experiments demonstrate that re-weighting difficult sentences annotated by the crowd generally improves the extraction models. Presumably the performance is influenced by the annotation quality.
We now examine the possibility that the higher quality and more consistent annotations of domain experts on the difficult instances will benefit the extraction model. This simulates an annotation strategy in which we route difficult instances to domain experts and easier ones to crowd annotators. We also contrast the value of difficult data to that of an i.i.d. random sample of the same size, both annotated by experts.
Expert annotations of Random and Difficult Instances
We re-annotate by experts a subset of most difficult instances and the same number of random instances. As collecting annotations from experts is slow and expensive, we only re-annotate the difficult instances for the interventions extraction task. We re-annotate the abstracts which cover the sentences with predicted difficulty scores in the top 5 percentile. We rank the abstracts from the training set by the count of difficult sentences, and re-annotate the abstracts that contain the most difficult sentences. Constrained by time and budget, we select only 2000 abstracts for re-annotation; 1000 of these are top-ranked, and 1000 are randomly sampled. This re-annotation cost $3,000. We have released the new annotation data at: https://github.com/bepnye/EBM-NLP.
Following BIBREF5 , we recruited five medical experts via Up-work with advanced medical training and strong technical reading/writing skills. The expert annotator were asked to read the entire abstract and highlight, using the BRAT toolkit BIBREF26 , all spans describing medical Interventions. Each abstract is only annotated by one expert. We examined 30 re-annotated abstracts to ensure the annotation quality before hiring the annotator.
Table 5 presents the results of LSTM-CRF-Pattern model trained on the reannotated difficult subset and the random subset. The first two rows show the results for models trained with expert annotations. The model trained on random data has a slightly better F1 than that trained on the same amount of difficult data. The model trained on random data has higher precision but lower recall.
Rows 3 and 4 list the results for models trained on the same data but with crowd annotation. Models trained with expert-annotated data are clearly superior to those trained with crowd labels with respect to F1, indicating that the experts produced higher quality annotations. For crowdsourced annotations, training the model with data sampled at i.i.d. random achieves 2% higher F1 than when difficult instances are used. When expert annotations are used, this difference is less than 1%. This trend in performance may be explained by differences in annotation quality: the randomly sampled set was more consistently annotated by both experts and crowd because the difficult set is harder. However, in both cases expert annotations are better, with a bigger difference between the expert and crowd models on the difficult set.
The last row is the model trained on all 5k abstracts with crowd annotations. Its F1 score is lower than either expert model trained on only 20% of data, suggesting that expert annotations should be collected whenever possible. Again the crowd model on complete data has higher precision than expert models but its recall is much lower.
Routing To Experts or Crowd
So far a system was trained on one type of data, either labeled by crowd or experts. We now examine the performance of a system trained on data that was routed to either experts or crowd annotators depending on their predicted difficult. Given the results presented so far mixing annotators may be beneficial given their respective trade-offs of precision and recall. We use the annotations from experts for an abstract if it exists otherwise use crowd annotations. The results are presented in Table 6 .
Rows 1 and 2 repeat the performance of the models trained on difficult subset and random subset with expert annotations only respectively. The third row is the model trained by combining difficult and random subsets with expert annotations. There are around 250 abstracts in the overlap of these two sets, so there are total 1.75k abstracts used for training the D+R model. Rows 4 to 6 are the models trained on all 5k abstracts with mixed annotations, where Other means the rest of the abstracts with crowd annotation only.
The results show adding more training data with crowd annotation still improves at least 1 point F1 score in all three extraction tasks. The improvement when the difficult subset with expert annotations is mixed with the remaining crowd annotation is 3.5 F1 score, much larger than when a random set of expert annotations are added. The model trained with re-annotating the difficult subset (D+Other) also outperforms the model with re-annotating the random subset (R+Other) by 2 points in F1. The model trained with re-annotating both of difficult and random subsets (D+R+Other), however, achieves only marginally higher F1 than the model trained with the re-annotated difficult subset (D+Other). In sum, the results clearly indicate that mixing expert and crowd annotations leads to better models than using solely crowd data, and better than using expert data alone. More importantly, there is greater gain in performance when instances are routed according to difficulty, as compared to randomly selecting the data for expert annotators. These findings align with our motivating hypothesis that annotation quality for difficult instances is important for final model performance. They also indicate that mixing annotations from expert and crowd could be an effective way to achieve acceptable model performance given a limited budget.
How Many Expert Annotations?
We established that crowd annotation are still useful in supplementing expert annotations for medical IE. Obtaining expert annotations for the one thousand most difficult instances greatly improved the model performance. However the choice of how many difficult instances to annotate was an uninformed choice. Here we check if less expert data would have yielded similar gains. Future work will need to address how best to choose this parameter for a routing system.
We simulate a routing scenario in which we send consecutive batches of the most difficult examples to the experts for annotation. We track changes in performance as we increase the number of most-difficult-articles sent to domain experts. As shown in Figure 4 , adding expert annotations for difficult articles consistently increases F1 scores. The performance gain is mostly from increased recall; the precision changes only a bit with higher quality annotation. This observation implies that crowd workers often fail to mark target tokens, but do not tend to produce large numbers of false positives. We suspect such failures to identify relevant spans/tokens are due to insufficient domain knowledge possessed by crowd workers.
The F1 score achieved after re-annotating the 600 most-difficult articles reaches 68.1%, which is close to the performance when re-annotating 1000 random articles. This demonstrates the effectiveness of recognizing difficult instances. The trend when we use up all expert data is still upward, so adding even more expert data is likely to further improve performance. Unfortunately we exhausted our budget and were not able to obtain additional expert annotations. It is likely that as the size of the expert annotations increases, the value of crowd annotations will diminish. This investigation is left for future work.
Conclusions
We have introduced the task of predicting annotation difficulty for biomedical information extraction (IE). We trained neural models using different learned representations to score texts in terms of their difficulty. Results from all models were strong with Pearson’s correlation coefficients higher than 0.45 in almost all evaluations, indicating the feasibility of this task. An ensemble model combining universal and task specific feature sentence vectors yielded the best results.
Experiments on biomedical IE tasks show that removing up to $\sim $ 10% of the sentences predicted to be most difficult did not decrease model performance, and that re-weighting sentences inversely to their difficulty score during training improves predictive performance. Simulations in which difficult examples are routed to experts and other instances to crowd annotators yields the best results, outperforming the strategy of randomly selecting data for expert annotation, and substantially improving upon the approach of relying exclusively on crowd annotations. In future work, routing strategies based on instance difficulty could be further investigated for budget-quality trade-off.
Acknowledgements
This work has been partially supported by NSF1748771 grant. Wallace was support in part by NIH/NLM R01LM012086.
| [
"57,505 sentences",
"57,505 sentences"
] | 4,371 | qasper | en | null | 2ecd2808617740f222a4e2f3b2df546da046d876b1580952 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How much data is needed to train the task-specific encoder?
Answer: | [
"57,505 sentences",
"57,505 sentences"
] | qasper | 128 | 32 | |
What tasks are used for evaluation? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The Transformer architecture BIBREF0 for deep neural networks has quickly risen to prominence in NLP through its efficiency and performance, leading to improvements in the state of the art of Neural Machine Translation BIBREF1, BIBREF2, as well as inspiring other powerful general-purpose models like BERT BIBREF3 and GPT-2 BIBREF4. At the heart of the Transformer lie multi-head attention mechanisms: each word is represented by multiple different weighted averages of its relevant context. As suggested by recent works on interpreting attention head roles, separate attention heads may learn to look for various relationships between tokens BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9.
The attention distribution of each head is predicted typically using the softmax normalizing transform. As a result, all context words have non-zero attention weight. Recent work on single attention architectures suggest that using sparse normalizing transforms in attention mechanisms such as sparsemax – which can yield exactly zero probabilities for irrelevant words – may improve performance and interpretability BIBREF12, BIBREF13, BIBREF14. Qualitative analysis of attention heads BIBREF0 suggests that, depending on what phenomena they capture, heads tend to favor flatter or more peaked distributions.
Recent works have proposed sparse Transformers BIBREF10 and adaptive span Transformers BIBREF11. However, the “sparsity" of those models only limits the attention to a contiguous span of past tokens, while in this work we propose a highly adaptive Transformer model that is capable of attending to a sparse set of words that are not necessarily contiguous. Figure FIGREF1 shows the relationship of these methods with ours.
Our contributions are the following:
We introduce sparse attention into the Transformer architecture, showing that it eases interpretability and leads to slight accuracy gains.
We propose an adaptive version of sparse attention, where the shape of each attention head is learnable and can vary continuously and dynamically between the dense limit case of softmax and the sparse, piecewise-linear sparsemax case.
We make an extensive analysis of the added interpretability of these models, identifying both crisper examples of attention head behavior observed in previous work, as well as novel behaviors unraveled thanks to the sparsity and adaptivity of our proposed model.
Background ::: The Transformer
In NMT, the Transformer BIBREF0 is a sequence-to-sequence (seq2seq) model which maps an input sequence to an output sequence through hierarchical multi-head attention mechanisms, yielding a dynamic, context-dependent strategy for propagating information within and across sentences. It contrasts with previous seq2seq models, which usually rely either on costly gated recurrent operations BIBREF15, BIBREF16 or static convolutions BIBREF17.
Given $n$ query contexts and $m$ sequence items under consideration, attention mechanisms compute, for each query, a weighted representation of the items. The particular attention mechanism used in BIBREF0 is called scaled dot-product attention, and it is computed in the following way:
where $\mathbf {Q} \in \mathbb {R}^{n \times d}$ contains representations of the queries, $\mathbf {K}, \mathbf {V} \in \mathbb {R}^{m \times d}$ are the keys and values of the items attended over, and $d$ is the dimensionality of these representations. The $\mathbf {\pi }$ mapping normalizes row-wise using softmax, $\mathbf {\pi }(\mathbf {Z})_{ij} = \operatornamewithlimits{\mathsf {softmax}}(\mathbf {z}_i)_j$, where
In words, the keys are used to compute a relevance score between each item and query. Then, normalized attention weights are computed using softmax, and these are used to weight the values of each item at each query context.
However, for complex tasks, different parts of a sequence may be relevant in different ways, motivating multi-head attention in Transformers. This is simply the application of Equation DISPLAY_FORM7 in parallel $H$ times, each with a different, learned linear transformation that allows specialization:
In the Transformer, there are three separate multi-head attention mechanisms for distinct purposes:
Encoder self-attention: builds rich, layered representations of each input word, by attending on the entire input sentence.
Context attention: selects a representative weighted average of the encodings of the input words, at each time step of the decoder.
Decoder self-attention: attends over the partial output sentence fragment produced so far.
Together, these mechanisms enable the contextualized flow of information between the input sentence and the sequential decoder.
Background ::: Sparse Attention
The softmax mapping (Equation DISPLAY_FORM8) is elementwise proportional to $\exp $, therefore it can never assign a weight of exactly zero. Thus, unnecessary items are still taken into consideration to some extent. Since its output sums to one, this invariably means less weight is assigned to the relevant items, potentially harming performance and interpretability BIBREF18. This has motivated a line of research on learning networks with sparse mappings BIBREF19, BIBREF20, BIBREF21, BIBREF22. We focus on a recently-introduced flexible family of transformations, $\alpha $-entmax BIBREF23, BIBREF14, defined as:
where $\triangle ^d \lbrace \mathbf {p}\in \mathbb {R}^d:\sum _{i} p_i = 1\rbrace $ is the probability simplex, and, for $\alpha \ge 1$, $\mathsf {H}^{\textsc {T}}_\alpha $ is the Tsallis continuous family of entropies BIBREF24:
This family contains the well-known Shannon and Gini entropies, corresponding to the cases $\alpha =1$ and $\alpha =2$, respectively.
Equation DISPLAY_FORM14 involves a convex optimization subproblem. Using the definition of $\mathsf {H}^{\textsc {T}}_\alpha $, the optimality conditions may be used to derive the following form for the solution (Appendix SECREF83):
where $[\cdot ]_+$ is the positive part (ReLU) function, $\mathbf {1}$ denotes the vector of all ones, and $\tau $ – which acts like a threshold – is the Lagrange multiplier corresponding to the $\sum _i p_i=1$ constraint.
Background ::: Sparse Attention ::: Properties of @!START@$\alpha $@!END@-entmax.
The appeal of $\alpha $-entmax for attention rests on the following properties. For $\alpha =1$ (i.e., when $\mathsf {H}^{\textsc {T}}_\alpha $ becomes the Shannon entropy), it exactly recovers the softmax mapping (We provide a short derivation in Appendix SECREF89.). For all $\alpha >1$ it permits sparse solutions, in stark contrast to softmax. In particular, for $\alpha =2$, it recovers the sparsemax mapping BIBREF19, which is piecewise linear. In-between, as $\alpha $ increases, the mapping continuously gets sparser as its curvature changes.
To compute the value of $\alpha $-entmax, one must find the threshold $\tau $ such that the r.h.s. in Equation DISPLAY_FORM16 sums to one. BIBREF23 propose a general bisection algorithm. BIBREF14 introduce a faster, exact algorithm for $\alpha =1.5$, and enable using $\mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}$ with fixed $\alpha $ within a neural network by showing that the $\alpha $-entmax Jacobian w.r.t. $\mathbf {z}$ for $\mathbf {p}^\star = \mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}(\mathbf {z})$ is
Our work furthers the study of $\alpha $-entmax by providing a derivation of the Jacobian w.r.t. the hyper-parameter $\alpha $ (Section SECREF3), thereby allowing the shape and sparsity of the mapping to be learned automatically. This is particularly appealing in the context of multi-head attention mechanisms, where we shall show in Section SECREF35 that different heads tend to learn different sparsity behaviors.
Adaptively Sparse Transformers with @!START@$\alpha $@!END@-entmax
We now propose a novel Transformer architecture wherein we simply replace softmax with $\alpha $-entmax in the attention heads. Concretely, we replace the row normalization $\mathbf {\pi }$ in Equation DISPLAY_FORM7 by
This change leads to sparse attention weights, as long as $\alpha >1$; in particular, $\alpha =1.5$ is a sensible starting point BIBREF14.
Adaptively Sparse Transformers with @!START@$\alpha $@!END@-entmax ::: Different @!START@$\alpha $@!END@ per head.
Unlike LSTM-based seq2seq models, where $\alpha $ can be more easily tuned by grid search, in a Transformer, there are many attention heads in multiple layers. Crucial to the power of such models, the different heads capture different linguistic phenomena, some of them isolating important words, others spreading out attention across phrases BIBREF0. This motivates using different, adaptive $\alpha $ values for each attention head, such that some heads may learn to be sparser, and others may become closer to softmax. We propose doing so by treating the $\alpha $ values as neural network parameters, optimized via stochastic gradients along with the other weights.
Adaptively Sparse Transformers with @!START@$\alpha $@!END@-entmax ::: Derivatives w.r.t. @!START@$\alpha $@!END@.
In order to optimize $\alpha $ automatically via gradient methods, we must compute the Jacobian of the entmax output w.r.t. $\alpha $. Since entmax is defined through an optimization problem, this is non-trivial and cannot be simply handled through automatic differentiation; it falls within the domain of argmin differentiation, an active research topic in optimization BIBREF25, BIBREF26.
One of our key contributions is the derivation of a closed-form expression for this Jacobian. The next proposition provides such an expression, enabling entmax layers with adaptive $\alpha $. To the best of our knowledge, ours is the first neural network module that can automatically, continuously vary in shape away from softmax and toward sparse mappings like sparsemax.
Proposition 1 Let $\mathbf {p}^\star \mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}(\mathbf {z})$ be the solution of Equation DISPLAY_FORM14. Denote the distribution $\tilde{p}_i {(p_i^\star )^{2 - \alpha }}{ \sum _j(p_j^\star )^{2-\alpha }}$ and let $h_i -p^\star _i \log p^\star _i$. The $i$th component of the Jacobian $\mathbf {g} \frac{\partial \mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}(\mathbf {z})}{\partial \alpha }$ is
proof uses implicit function differentiation and is given in Appendix SECREF10.
Proposition UNKREF22 provides the remaining missing piece needed for training adaptively sparse Transformers. In the following section, we evaluate this strategy on neural machine translation, and analyze the behavior of the learned attention heads.
Experiments
We apply our adaptively sparse Transformers on four machine translation tasks. For comparison, a natural baseline is the standard Transformer architecture using the softmax transform in its multi-head attention mechanisms. We consider two other model variants in our experiments that make use of different normalizing transformations:
1.5-entmax: a Transformer with sparse entmax attention with fixed $\alpha =1.5$ for all heads. This is a novel model, since 1.5-entmax had only been proposed for RNN-based NMT models BIBREF14, but never in Transformers, where attention modules are not just one single component of the seq2seq model but rather an integral part of all of the model components.
$\alpha $-entmax: an adaptive Transformer with sparse entmax attention with a different, learned $\alpha _{i,j}^t$ for each head.
The adaptive model has an additional scalar parameter per attention head per layer for each of the three attention mechanisms (encoder self-attention, context attention, and decoder self-attention), i.e.,
and we set $\alpha _{i,j}^t = 1 + \operatornamewithlimits{\mathsf {sigmoid}}(a_{i,j}^t) \in ]1, 2[$. All or some of the $\alpha $ values can be tied if desired, but we keep them independent for analysis purposes.
Experiments ::: Datasets.
Our models were trained on 4 machine translation datasets of different training sizes:
[itemsep=.5ex,leftmargin=2ex]
IWSLT 2017 German $\rightarrow $ English BIBREF27: 200K sentence pairs.
KFTT Japanese $\rightarrow $ English BIBREF28: 300K sentence pairs.
WMT 2016 Romanian $\rightarrow $ English BIBREF29: 600K sentence pairs.
WMT 2014 English $\rightarrow $ German BIBREF30: 4.5M sentence pairs.
All of these datasets were preprocessed with byte-pair encoding BIBREF31, using joint segmentations of 32k merge operations.
Experiments ::: Training.
We follow the dimensions of the Transformer-Base model of BIBREF0: The number of layers is $L=6$ and number of heads is $H=8$ in the encoder self-attention, the context attention, and the decoder self-attention. We use a mini-batch size of 8192 tokens and warm up the learning rate linearly until 20k steps, after which it decays according to an inverse square root schedule. All models were trained until convergence of validation accuracy, and evaluation was done at each 10k steps for ro$\rightarrow $en and en$\rightarrow $de and at each 5k steps for de$\rightarrow $en and ja$\rightarrow $en. The end-to-end computational overhead of our methods, when compared to standard softmax, is relatively small; in training tokens per second, the models using $\alpha $-entmax and $1.5$-entmax are, respectively, $75\%$ and $90\%$ the speed of the softmax model.
Experiments ::: Results.
We report test set tokenized BLEU BIBREF32 results in Table TABREF27. We can see that replacing softmax by entmax does not hurt performance in any of the datasets; indeed, sparse attention Transformers tend to have slightly higher BLEU, but their sparsity leads to a better potential for analysis. In the next section, we make use of this potential by exploring the learned internal mechanics of the self-attention heads.
Analysis
We conduct an analysis for the higher-resource dataset WMT 2014 English $\rightarrow $ German of the attention in the sparse adaptive Transformer model ($\alpha $-entmax) at multiple levels: we analyze high-level statistics as well as individual head behavior. Moreover, we make a qualitative analysis of the interpretability capabilities of our models.
Analysis ::: High-Level Statistics ::: What kind of @!START@$\alpha $@!END@ values are learned?
Figure FIGREF37 shows the learning trajectories of the $\alpha $ parameters of a selected subset of heads. We generally observe a tendency for the randomly-initialized $\alpha $ parameters to decrease initially, suggesting that softmax-like behavior may be preferable while the model is still very uncertain. After around one thousand steps, some heads change direction and become sparser, perhaps as they become more confident and specialized. This shows that the initialization of $\alpha $ does not predetermine its sparsity level or the role the head will have throughout. In particular, head 8 in the encoder self-attention layer 2 first drops to around $\alpha =1.3$ before becoming one of the sparsest heads, with $\alpha \approx 2$.
The overall distribution of $\alpha $ values at convergence can be seen in Figure FIGREF38. We can observe that the encoder self-attention blocks learn to concentrate the $\alpha $ values in two modes: a very sparse one around $\alpha \rightarrow 2$, and a dense one between softmax and 1.5-entmax . However, the decoder self and context attention only learn to distribute these parameters in a single mode. We show next that this is reflected in the average density of attention weight vectors as well.
Analysis ::: High-Level Statistics ::: Attention weight density when translating.
For any $\alpha >1$, it would still be possible for the weight matrices in Equation DISPLAY_FORM9 to learn re-scalings so as to make attention sparser or denser. To visualize the impact of adaptive $\alpha $ values, we compare the empirical attention weight density (the average number of tokens receiving non-zero attention) within each module, against sparse Transformers with fixed $\alpha =1.5$.
Figure FIGREF40 shows that, with fixed $\alpha =1.5$, heads tend to be sparse and similarly-distributed in all three attention modules. With learned $\alpha $, there are two notable changes: (i) a prominent mode corresponding to fully dense probabilities, showing that our models learn to combine sparse and dense attention, and (ii) a distinction between the encoder self-attention – whose background distribution tends toward extreme sparsity – and the other two modules, who exhibit more uniform background distributions. This suggests that perhaps entirely sparse Transformers are suboptimal.
The fact that the decoder seems to prefer denser attention distributions might be attributed to it being auto-regressive, only having access to past tokens and not the full sentence. We speculate that it might lose too much information if it assigned weights of zero to too many tokens in the self-attention, since there are fewer tokens to attend to in the first place.
Teasing this down into separate layers, Figure FIGREF41 shows the average (sorted) density of each head for each layer. We observe that $\alpha $-entmax is able to learn different sparsity patterns at each layer, leading to more variance in individual head behavior, to clearly-identified dense and sparse heads, and overall to different tendencies compared to the fixed case of $\alpha =1.5$.
Analysis ::: High-Level Statistics ::: Head diversity.
To measure the overall disagreement between attention heads, as a measure of head diversity, we use the following generalization of the Jensen-Shannon divergence:
where $\mathbf {p}_j$ is the vector of attention weights assigned by head $j$ to each word in the sequence, and $\mathsf {H}^\textsc {S}$ is the Shannon entropy, base-adjusted based on the dimension of $\mathbf {p}$ such that $JS \le 1$. We average this measure over the entire validation set. The higher this metric is, the more the heads are taking different roles in the model.
Figure FIGREF44 shows that both sparse Transformer variants show more diversity than the traditional softmax one. Interestingly, diversity seems to peak in the middle layers of the encoder self-attention and context attention, while this is not the case for the decoder self-attention.
The statistics shown in this section can be found for the other language pairs in Appendix SECREF8.
Analysis ::: Identifying Head Specializations
Previous work pointed out some specific roles played by different heads in the softmax Transformer model BIBREF33, BIBREF5, BIBREF9. Identifying the specialization of a head can be done by observing the type of tokens or sequences that the head often assigns most of its attention weight; this is facilitated by sparsity.
Analysis ::: Identifying Head Specializations ::: Positional heads.
One particular type of head, as noted by BIBREF9, is the positional head. These heads tend to focus their attention on either the previous or next token in the sequence, thus obtaining representations of the neighborhood of the current time step. In Figure FIGREF47, we show attention plots for such heads, found for each of the studied models. The sparsity of our models allows these heads to be more confident in their representations, by assigning the whole probability distribution to a single token in the sequence. Concretely, we may measure a positional head's confidence as the average attention weight assigned to the previous token. The softmax model has three heads for position $-1$, with median confidence $93.5\%$. The $1.5$-entmax model also has three heads for this position, with median confidence $94.4\%$. The adaptive model has four heads, with median confidences $95.9\%$, the lowest-confidence head being dense with $\alpha =1.18$, while the highest-confidence head being sparse ($\alpha =1.91$).
For position $+1$, the models each dedicate one head, with confidence around $95\%$, slightly higher for entmax. The adaptive model sets $\alpha =1.96$ for this head.
Analysis ::: Identifying Head Specializations ::: BPE-merging head.
Due to the sparsity of our models, we are able to identify other head specializations, easily identifying which heads should be further analysed. In Figure FIGREF51 we show one such head where the $\alpha $ value is particularly high (in the encoder, layer 1, head 4 depicted in Figure FIGREF37). We found that this head most often looks at the current time step with high confidence, making it a positional head with offset 0. However, this head often spreads weight sparsely over 2-3 neighboring tokens, when the tokens are part of the same BPE cluster or hyphenated words. As this head is in the first layer, it provides a useful service to the higher layers by combining information evenly within some BPE clusters.
For each BPE cluster or cluster of hyphenated words, we computed a score between 0 and 1 that corresponds to the maximum attention mass assigned by any token to the rest of the tokens inside the cluster in order to quantify the BPE-merging capabilities of these heads. There are not any attention heads in the softmax model that are able to obtain a score over $80\%$, while for $1.5$-entmax and $\alpha $-entmax there are two heads in each ($83.3\%$ and $85.6\%$ for $1.5$-entmax and $88.5\%$ and $89.8\%$ for $\alpha $-entmax).
Analysis ::: Identifying Head Specializations ::: Interrogation head.
On the other hand, in Figure FIGREF52 we show a head for which our adaptively sparse model chose an $\alpha $ close to 1, making it closer to softmax (also shown in encoder, layer 1, head 3 depicted in Figure FIGREF37). We observe that this head assigns a high probability to question marks at the end of the sentence in time steps where the current token is interrogative, thus making it an interrogation-detecting head. We also observe this type of heads in the other models, which we also depict in Figure FIGREF52. The average attention weight placed on the question mark when the current token is an interrogative word is $98.5\%$ for softmax, $97.0\%$ for $1.5$-entmax, and $99.5\%$ for $\alpha $-entmax.
Furthermore, we can examine sentences where some tendentially sparse heads become less so, thus identifying sources of ambiguity where the head is less confident in its prediction. An example is shown in Figure FIGREF55 where sparsity in the same head differs for sentences of similar length.
Related Work ::: Sparse attention.
Prior work has developed sparse attention mechanisms, including applications to NMT BIBREF19, BIBREF12, BIBREF20, BIBREF22, BIBREF34. BIBREF14 introduced the entmax function this work builds upon. In their work, there is a single attention mechanism which is controlled by a fixed $\alpha $. In contrast, this is the first work to allow such attention mappings to dynamically adapt their curvature and sparsity, by automatically adjusting the continuous $\alpha $ parameter. We also provide the first results using sparse attention in a Transformer model.
Related Work ::: Fixed sparsity patterns.
Recent research improves the scalability of Transformer-like networks through static, fixed sparsity patterns BIBREF10, BIBREF35. Our adaptively-sparse Transformer can dynamically select a sparsity pattern that finds relevant words regardless of their position (e.g., Figure FIGREF52). Moreover, the two strategies could be combined. In a concurrent line of research, BIBREF11 propose an adaptive attention span for Transformer language models. While their work has each head learn a different contiguous span of context tokens to attend to, our work finds different sparsity patterns in the same span. Interestingly, some of their findings mirror ours – we found that attention heads in the last layers tend to be denser on average when compared to the ones in the first layers, while their work has found that lower layers tend to have a shorter attention span compared to higher layers.
Related Work ::: Transformer interpretability.
The original Transformer paper BIBREF0 shows attention visualizations, from which some speculation can be made of the roles the several attention heads have. BIBREF7 study the syntactic abilities of the Transformer self-attention, while BIBREF6 extract dependency relations from the attention weights. BIBREF8 find that the self-attentions in BERT BIBREF3 follow a sequence of processes that resembles a classical NLP pipeline. Regarding redundancy of heads, BIBREF9 develop a method that is able to prune heads of the multi-head attention module and make an empirical study of the role that each head has in self-attention (positional, syntactic and rare words). BIBREF36 also aim to reduce head redundancy by adding a regularization term to the loss that maximizes head disagreement and obtain improved results. While not considering Transformer attentions, BIBREF18 show that traditional attention mechanisms do not necessarily improve interpretability since softmax attention is vulnerable to an adversarial attack leading to wildly different model predictions for the same attention weights. Sparse attention may mitigate these issues; however, our work focuses mostly on a more mechanical aspect of interpretation by analyzing head behavior, rather than on explanations for predictions.
Conclusion and Future Work
We contribute a novel strategy for adaptively sparse attention, and, in particular, for adaptively sparse Transformers. We present the first empirical analysis of Transformers with sparse attention mappings (i.e., entmax), showing potential in both translation accuracy as well as in model interpretability.
In particular, we analyzed how the attention heads in the proposed adaptively sparse Transformer can specialize more and with higher confidence. Our adaptivity strategy relies only on gradient-based optimization, side-stepping costly per-head hyper-parameter searches. Further speed-ups are possible by leveraging more parallelism in the bisection algorithm for computing $\alpha $-entmax.
Finally, some of the automatically-learned behaviors of our adaptively sparse Transformers – for instance, the near-deterministic positional heads or the subword joining head – may provide new ideas for designing static variations of the Transformer.
Acknowledgments
This work was supported by the European Research Council (ERC StG DeepSPIN 758969), and by the Fundação para a Ciência e Tecnologia through contracts UID/EEA/50008/2019 and CMUPERI/TIC/0046/2014 (GoLocal). We are grateful to Ben Peters for the $\alpha $-entmax code and Erick Fonseca, Marcos Treviso, Pedro Martins, and Tsvetomila Mihaylova for insightful group discussion. We thank Mathieu Blondel for the idea to learn $\alpha $. We would also like to thank the anonymous reviewers for their helpful feedback.
Supplementary Material
Background ::: Regularized Fenchel-Young prediction functions
Definition 1 (BIBREF23)
Let $\Omega \colon \triangle ^d \rightarrow {\mathbb {R}}\cup \lbrace \infty \rbrace $ be a strictly convex regularization function. We define the prediction function $\mathbf {\pi }_{\Omega }$ as
Background ::: Characterizing the @!START@$\alpha $@!END@-entmax mapping
Lemma 1 (BIBREF14) For any $\mathbf {z}$, there exists a unique $\tau ^\star $ such that
Proof: From the definition of $\mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}$,
we may easily identify it with a regularized prediction function (Def. UNKREF81):
We first note that for all $\mathbf {p}\in \triangle ^d$,
From the constant invariance and scaling properties of $\mathbf {\pi }_{\Omega }$ BIBREF23,
Using BIBREF23, noting that $g^{\prime }(t) = t^{\alpha - 1}$ and $(g^{\prime })^{-1}(u) = u^{{1}{\alpha -1}}$, yields
Since $\mathsf {H}^{\textsc {T}}_\alpha $ is strictly convex on the simplex, $\mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}$ has a unique solution $\mathbf {p}^\star $. Equation DISPLAY_FORM88 implicitly defines a one-to-one mapping between $\mathbf {p}^\star $ and $\tau ^\star $ as long as $\mathbf {p}^\star \in \triangle $, therefore $\tau ^\star $ is also unique.
Background ::: Connections to softmax and sparsemax
The Euclidean projection onto the simplex, sometimes referred to, in the context of neural attention, as sparsemax BIBREF19, is defined as
The solution can be characterized through the unique threshold $\tau $ such that $\sum _i \operatornamewithlimits{\mathsf {sparsemax}}(\mathbf {z})_i = 1$ and BIBREF38
Thus, each coordinate of the sparsemax solution is a piecewise-linear function. Visibly, this expression is recovered when setting $\alpha =2$ in the $\alpha $-entmax expression (Equation DISPLAY_FORM85); for other values of $\alpha $, the exponent induces curvature.
On the other hand, the well-known softmax is usually defined through the expression
which can be shown to be the unique solution of the optimization problem
where $\mathsf {H}^\textsc {S}(\mathbf {p}) -\sum _i p_i \log p_i$ is the Shannon entropy. Indeed, setting the gradient to 0 yields the condition $\log p_i = z_j - \nu _i - \tau - 1$, where $\tau $ and $\nu > 0$ are Lagrange multipliers for the simplex constraints $\sum _i p_i = 1$ and $p_i \ge 0$, respectively. Since the l.h.s. is only finite for $p_i>0$, we must have $\nu _i=0$ for all $i$, by complementary slackness. Thus, the solution must have the form $p_i = {\exp (z_i)}{Z}$, yielding Equation DISPLAY_FORM92.
Jacobian of @!START@$\alpha $@!END@-entmax w.r.t. the shape parameter @!START@$\alpha $@!END@: Proof of Proposition @!START@UID22@!END@
Recall that the entmax transformation is defined as:
where $\alpha \ge 1$ and $\mathsf {H}^{\textsc {T}}_{\alpha }$ is the Tsallis entropy,
and $\mathsf {H}^\textsc {S}(\mathbf {p}):= -\sum _j p_j \log p_j$ is the Shannon entropy.
In this section, we derive the Jacobian of $\operatornamewithlimits{\mathsf {entmax }}$ with respect to the scalar parameter $\alpha $.
Jacobian of @!START@$\alpha $@!END@-entmax w.r.t. the shape parameter @!START@$\alpha $@!END@: Proof of Proposition @!START@UID22@!END@ ::: General case of @!START@$\alpha >1$@!END@
From the KKT conditions associated with the optimization problem in Eq. DISPLAY_FORM85, we have that the solution $\mathbf {p}^{\star }$ has the following form, coordinate-wise:
where $\tau ^{\star }$ is a scalar Lagrange multiplier that ensures that $\mathbf {p}^{\star }$ normalizes to 1, i.e., it is defined implicitly by the condition:
For general values of $\alpha $, Eq. DISPLAY_FORM98 lacks a closed form solution. This makes the computation of the Jacobian
non-trivial. Fortunately, we can use the technique of implicit differentiation to obtain this Jacobian.
The Jacobian exists almost everywhere, and the expressions we derive expressions yield a generalized Jacobian BIBREF37 at any non-differentiable points that may occur for certain ($\alpha $, $\mathbf {z}$) pairs. We begin by noting that $\frac{\partial p_i^{\star }}{\partial \alpha } = 0$ if $p_i^{\star } = 0$, because increasing $\alpha $ keeps sparse coordinates sparse. Therefore we need to worry only about coordinates that are in the support of $\mathbf {p}^\star $. We will assume hereafter that the $i$th coordinate of $\mathbf {p}^\star $ is non-zero. We have:
We can see that this Jacobian depends on $\frac{\partial \tau ^{\star }}{\partial \alpha }$, which we now compute using implicit differentiation.
Let $\mathcal {S} = \lbrace i: p^\star _i > 0 \rbrace $). By differentiating both sides of Eq. DISPLAY_FORM98, re-using some of the steps in Eq. DISPLAY_FORM101, and recalling Eq. DISPLAY_FORM97, we get
from which we obtain:
Finally, plugging Eq. DISPLAY_FORM103 into Eq. DISPLAY_FORM101, we get:
where we denote by
The distribution $\tilde{\mathbf {p}}(\alpha )$ can be interpreted as a “skewed” distribution obtained from $\mathbf {p}^{\star }$, which appears in the Jacobian of $\mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}(\mathbf {z})$ w.r.t. $\mathbf {z}$ as well BIBREF14.
Jacobian of @!START@$\alpha $@!END@-entmax w.r.t. the shape parameter @!START@$\alpha $@!END@: Proof of Proposition @!START@UID22@!END@ ::: Solving the indetermination for @!START@$\alpha =1$@!END@
We can write Eq. DISPLAY_FORM104 as
When $\alpha \rightarrow 1^+$, we have $\tilde{\mathbf {p}}(\alpha ) \rightarrow \mathbf {p}^{\star }$, which leads to a $\frac{0}{0}$ indetermination.
To solve this indetermination, we will need to apply L'Hôpital's rule twice. Let us first compute the derivative of $\tilde{p}_i(\alpha )$ with respect to $\alpha $. We have
therefore
Differentiating the numerator and denominator in Eq. DISPLAY_FORM107, we get:
with
and
When $\alpha \rightarrow 1^+$, $B$ becomes again a $\frac{0}{0}$ indetermination, which we can solve by applying again L'Hôpital's rule. Differentiating the numerator and denominator in Eq. DISPLAY_FORM112:
Finally, summing Eq. DISPLAY_FORM111 and Eq. DISPLAY_FORM113, we get
Jacobian of @!START@$\alpha $@!END@-entmax w.r.t. the shape parameter @!START@$\alpha $@!END@: Proof of Proposition @!START@UID22@!END@ ::: Summary
To sum up, we have the following expression for the Jacobian of $\mathop {\mathsf {\alpha }\textnormal {-}\mathsf {entmax }}$ with respect to $\alpha $:
| [
"four machine translation tasks: German -> English, Japanese -> English, Romanian -> English, English -> German",
" four machine translation tasks, IWSLT 2017 German $\\rightarrow $ English BIBREF27, KFTT Japanese $\\rightarrow $ English BIBREF28, WMT 2016 Romanian $\\rightarrow $ English BIBREF29, WMT 2014 Engli... | 4,898 | qasper | en | null | 11be2f14f540e957e9797cc962203b8186ca10561228f81f | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What tasks are used for evaluation?
Answer: | [
"four machine translation tasks: German -> English, Japanese -> English, Romanian -> English, English -> German",
" four machine translation tasks, IWSLT 2017 German $\\rightarrow $ English BIBREF27, KFTT Japanese $\\rightarrow $ English BIBREF28, WMT 2016 Romanian $\\rightarrow $ English BIBREF29, WMT 2014 Engli... | qasper | 128 | 33 | |
What is the improvement in performance for Estonian in the NER task? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Word embeddings are representations of words in numerical form, as vectors of typically several hundred dimensions. The vectors are used as an input to machine learning models; for complex language processing tasks these are typically deep neural networks. The embedding vectors are obtained from specialized learning tasks, based on neural networks, e.g., word2vec BIBREF0, GloVe BIBREF1, FastText BIBREF2, ELMo BIBREF3, and BERT BIBREF4. For training, the embeddings algorithms use large monolingual corpora that encode important information about word meaning as distances between vectors. In order to enable downstream machine learning on text understanding tasks, the embeddings shall preserve semantic relations between words, and this is true even across languages.
Probably the best known word embeddings are produced by the word2vec method BIBREF5. The problem with word2vec embeddings is their failure to express polysemous words. During training of an embedding, all senses of a given word (e.g., paper as a material, as a newspaper, as a scientific work, and as an exam) contribute relevant information in proportion to their frequency in the training corpus. This causes the final vector to be placed somewhere in the weighted middle of all words' meanings. Consequently, rare meanings of words are poorly expressed with word2vec and the resulting vectors do not offer good semantic representations. For example, none of the 50 closest vectors of the word paper is related to science.
The idea of contextual embeddings is to generate a different vector for each context a word appears in and the context is typically defined sentence-wise. To a large extent, this solves the problems with word polysemy, i.e. the context of a sentence is typically enough to disambiguate different meanings of a word for humans and so it is for the learning algorithms. In this work, we describe high-quality models for contextual embeddings, called ELMo BIBREF3, precomputed for seven morphologically rich, less-resourced languages: Slovenian, Croatian, Finnish, Estonian, Latvian, Lithuanian, and Swedish. ELMo is one of the most successful approaches to contextual word embeddings. At time of its creation, ELMo has been shown to outperform previous word embeddings BIBREF3 like word2vec and GloVe on many NLP tasks, e.g., question answering, named entity extraction, sentiment analysis, textual entailment, semantic role labeling, and coreference resolution.
This report is split into further five sections. In section SECREF2, we describe the contextual embeddings ELMo. In Section SECREF3, we describe the datasets used and in Section SECREF4 we describe preprocessing and training of the embeddings. We describe the methodology for evaluation of created vectors and results in Section SECREF5. We present conclusion in Section SECREF6 where we also outline plans for further work.
ELMo
Typical word embeddings models or representations, such as word2vec BIBREF0, GloVe BIBREF1, or FastText BIBREF2, are fast to train and have been pre-trained for a number of different languages. They do not capture the context, though, so each word is always given the same vector, regardless of its context or meaning. This is especially problematic for polysemous words. ELMo (Embeddings from Language Models) embedding BIBREF3 is one of the state-of-the-art pretrained transfer learning models, that remedies the problem and introduces a contextual component.
ELMo model`s architecture consists of three neural network layers. The output of the model after each layer gives one set of embeddings, altogether three sets. The first layer is a CNN layer, which operates on a character level. It is context independent, so each word always gets the same embedding, regardless of its context. It is followed by two biLM layers. A biLM layer consists of two concatenated LSTMs. In the first LSTM, we try to predict the following word, based on the given past words, where each word is represented by the embeddings from the CNN layer. In the second LSTM, we try to predict the preceding word, based on the given following words. It is equivalent to the first LSTM, just reading the text in reverse.
In NLP tasks, any set of these embeddings may be used; however, a weighted average is usually used. The weights of the average are learned during the training of the model for the specific task. Additionally, an entire ELMo model can be fine-tuned on a specific end task.
Although ELMo is trained on character level and is able to handle out-of-vocabulary words, a vocabulary file containing most common tokens is used for efficiency during training and embedding generation. The original ELMo model was trained on a one billion word large English corpus, with a given vocabulary file of about 800,000 words. Later, ELMo models for other languages were trained as well, but limited to larger languages with many resources, like German and Japanese.
ELMo ::: ELMoForManyLangs
Recently, ELMoForManyLangs BIBREF6 project released pre-trained ELMo models for a number of different languages BIBREF7. These models, however, were trained on a significantly smaller datasets. They used 20-million-words data randomly sampled from the raw text released by the CoNLL 2017 Shared Task - Automatically Annotated Raw Texts and Word Embeddings BIBREF8, which is a combination of Wikipedia dump and common crawl. The quality of these models is questionable. For example, we compared the Latvian model by ELMoForManyLangs with a model we trained on a complete (wikidump + common crawl) Latvian corpus, which has about 280 million tokens. The difference of each model on the word analogy task is shown in Figure FIGREF16 in Section SECREF5. As the results of the ELMoForManyLangs embeddings are significantly worse than using the full corpus, we can conclude that these embeddings are not of sufficient quality. For that reason, we computed ELMo embeddings for seven languages on much larger corpora. As this effort requires access to large amount of textual data and considerable computational resources, we made the precomputed models publicly available by depositing them to Clarin repository.
Training Data
We trained ELMo models for seven languages: Slovenian, Croatian, Finnish, Estonian, Latvian, Lithuanian and Swedish. To obtain high-quality embeddings, we used large monolingual corpora from various sources for each language. Some corpora are available online under permissive licences, others are available only for research purposes or have limited availability. The corpora used in training datasets are a mix of news articles and general web crawl, which we preprocessed and deduplicated. Below we shortly describe the used corpora in alphabetical order of the involved languages. Their names and sizes are summarized in Table TABREF3.
Croatian dataset include hrWaC 2.1 corpus BIBREF9, Riznica BIBREF10, and articles of Croatian branch of Styria media house, made available to us through partnership in a joint project. hrWaC was built by crawling the .hr internet domain in 2011 and 2014. Riznica is composed of Croatian fiction and non-fiction prose, poetry, drama, textbooks, manuals, etc. The Styria dataset consists of 570,219 news articles published on the Croatian 24sata news portal and niche portals related to 24sata.
Estonian dataset contains texts from two sources, CoNLL 2017 Shared Task - Automatically Annotated Raw Texts and Word Embeddings BIBREF8, and news articles made available to us by Ekspress Meedia due to partnership in the project. Ekspress Meedia dataset is composed of Estonian news articles between years 2009 and 2019. The CoNLL 2017 corpus is composed of Estonian Wikipedia and webcrawl.
Finnish dataset contains articles by Finnish news agency STT, Finnish part of the CoNLL 2017 dataset, and Ylilauta downloadable version BIBREF11. STT news articles were published between years 1992 and 2018. Ylilauta is a Finnish online discussion board; the corpus contains parts of the discussions from 2012 to 2014.
Latvian dataset consists only of the Latvian portion of the ConLL 2017 corpus.
Lithuanian dataset is composed of Lithuanian Wikipedia articles from 2018, DGT-UD corpus, and LtTenTen. DGT-UD is a parallel corpus of 23 official languages of the EU, composed of JRC DGT translation memory of European law, automatically annotated with UD-Pipe 1.2. LtTenTen is Lithuanian web corpus made up of texts collected from the internet in April 2014 BIBREF12.
Slovene dataset is formed from the Gigafida 2.0 corpus BIBREF13. It is a general language corpus composed of various sources, mostly newspapers, internet pages, and magazines, but also fiction and non-fiction prose, textbooks, etc.
Swedish dataset is composed of STT Swedish articles and Swedish part of CoNLL 2017. The Finnish news agency STT publishes some of its articles in Swedish language. They were made available to us through partnership in a joint project. The corpus contains those articles from 1992 to 2017.
Preprocessing and Training
Prior to training the ELMo models, we sentence and word tokenized all the datasets. The text was formatted in such a way that each sentence was in its own line with tokens separated by white spaces. CoNLL 2017, DGT-UD and LtTenTen14 corpora were already pre-tokenized. We tokenized the others using the NLTK library and its tokenizers for each of the languages. There is no tokenizer for Croatian in NLTK library, so we used Slovene tokenizer instead. After tokenization, we deduplicated the datasets for each language separately, using the Onion (ONe Instance ONly) tool for text deduplication. We applied the tool on paragraph level for corpora that did not have sentences shuffled and on sentence level for the rest. We considered 9-grams with duplicate content threshold of 0.9.
For each language we prepared a vocabulary file, containing roughly one million most common tokens, i.e. tokens that appear at least $n$ times in the corpus, where $n$ is between 15 and 25, depending on the dataset size. We included the punctuation marks among the tokens. We trained each ELMo model using default values used to train the original English ELMo (large) model.
Evaluation
We evaluated the produced ELMo models for all languages using two evaluation tasks: a word analogy task and named entity recognition (NER) task. Below, we first shortly describe each task, followed by the evaluation results.
Evaluation ::: Word Analogy Task
The word analogy task was popularized by mikolov2013distributed. The goal is to find a term $y$ for a given term $x$ so that the relationship between $x$ and $y$ best resembles the given relationship $a : b$. There are two main groups of categories: 5 semantic and 10 syntactic. To illustrate a semantic relationship, consider for example that the word pair $a : b$ is given as “Finland : Helsinki”. The task is to find the term $y$ corresponding to the relationship “Sweden : $y$”, with the expected answer being $y=$ Stockholm. In syntactic categories, the two words in a pair have a common stem (in some cases even same lemma), with all the pairs in a given category having the same morphological relationship. For example, given the word pair “long : longer”, we see that we have an adjective in its base form and the same adjective in a comparative form. That task is then to find the term $y$ corresponding to the relationship “dark : $y$”, with the expected answer being $y=$ darker, that is a comparative form of the adjective dark.
In the vector space, the analogy task is transformed into vector arithmetic and search for nearest neighbours, i.e. we compute the distance between vectors: d(vec(Finland), vec(Helsinki)) and search for word $y$ which would give the closest result in distance d(vec(Sweden), vec($y$)). In the analogy dataset the analogies are already pre-specified, so we are measuring how close are the given pairs. In the evaluation below, we use analogy datasets for all tested languages based on the English dataset by BIBREF14 . Due to English-centered bias of this dataset, we used a modified dataset which was first written in Slovene language and then translated into other languages BIBREF15.
As each instance of analogy contains only four words, without any context, the contextual models (such as ELMo) do not have enough context to generate sensible embeddings. We therefore used some additional text to form simple sentences using the four analogy words, while taking care that their noun case stays the same. For example, for the words "Rome", "Italy", "Paris" and "France" (forming the analogy Rome is to Italy as Paris is to $x$, where the correct answer is $x=$France), we formed the sentence "If the word Rome corresponds to the word Italy, then the word Paris corresponds to the word France". We generated embeddings for those four words in the constructed sentence, substituted the last word with each word in our vocabulary and generated the embeddings again. As typical for non-contextual analogy task, we measure the cosine distance ($d$) between the last word ($w_4$) and the combination of the first three words ($w_2-w_1+w_3$). We use the CSLS metric BIBREF16 to find the closest candidate word ($w_4$). If we find the correct word among the five closest words, we consider that entry as successfully identified. The proportion of correctly identified words forms a statistic called accuracy@5, which we report as the result.
We first compare existing Latvian ELMo embeddings from ELMoForManyLangs project with our Latvian embeddings, followed by the detailed analysis of our ELMo embeddings. We trained Latvian ELMo using only CoNLL 2017 corpora. Since this is the only language, where we trained the embedding model on exactly the same corpora as ELMoForManyLangs models, we chose it for comparison between our ELMo model with ELMoForManyLangs. In other languages, additional or other corpora were used, so a direct comparison would also reflect the quality of the corpora used for training. In Latvian, however, only the size of the training dataset is different. ELMoForManyLangs uses only 20 million tokens and we use the whole corpus of 270 million tokens.
The Latvian ELMo model from ELMoForManyLangs project performs significantly worse than EMBEDDIA ELMo Latvian model on all categories of word analogy task (Figure FIGREF16). We also include the comparison with our Estonian ELMo embeddings in the same figure. This comparison shows that while differences between our Latvian and Estonian embeddings can be significant for certain categories, the accuracy score of ELMoForManyLangs is always worse than either of our models. The comparison of Estonian and Latvian models leads us to believe that a few hundred million tokens is a sufficiently large corpus to train ELMo models (at least for word analogy task), but 20-million token corpora used in ELMoForManyLangs are too small.
The results for all languages and all ELMo layers, averaged over semantic and syntactic categories, are shown in Table TABREF17. The embeddings after the first LSTM layer perform best in semantic categories. In syntactic categories, the non-contextual CNN layer performs the best. Syntactic categories are less context dependent and much more morphology and syntax based, so it is not surprising that the non-contextual layer performs well. The second LSTM layer embeddings perform the worst in syntactic categories, though still outperforming CNN layer embeddings in semantic categories. Latvian ELMo performs worse compared to other languages we trained, especially in semantic categories, presumably due to smaller training data size. Surprisingly, the original English ELMo performs very poorly in syntactic categories and only outperforms Latvian in semantic categories. The low score can be partially explained by English model scoring $0.00$ in one syntactic category “opposite adjective”, which we have not been able to explain.
Evaluation ::: Named Entity Recognition
For evaluation of ELMo models on a relevant downstream task, we used named entity recognition (NER) task. NER is an information extraction task that seeks to locate and classify named entity mentions in unstructured text into pre-defined categories such as the person names, organizations, locations, medical codes, time expressions, quantities, monetary values, percentages, etc. To allow comparison of results between languages, we used an adapted version of this task, which uses a reduced set of labels, available in NER datasets for all processed languages. The labels in the used NER datasets are simplified to a common label set of three labels (person - PER, location - LOC, organization - ORG). Each word in the NER dataset is labeled with one of the three mentioned labels or a label 'O' (other, i.e. not a named entity) if it does not fit any of the other three labels. The number of words having each label is shown in Table TABREF19.
To measure the performance of ELMo embeddings on the NER task we proceeded as follows. We embedded the text in the datasets sentence by sentence, producing three vectors (one from each ELMo layer) for each token in a sentence. We calculated the average of the three vectors and used it as the input of our recognition model. The input layer was followed by a single LSTM layer with 128 LSTM cells and a dropout layer, randomly dropping 10% of the neurons on both the output and the recurrent branch. The final layer of our model was a time distributed softmax layer with 4 neurons.
We used ADAM optimiser BIBREF17 with the learning rate 0.01 and $10^{-5}$ learning rate decay. We used categorical cross-entropy as a loss function and trained the model for 3 epochs. We present the results using the Macro $F_1$ score, that is the average of $F_1$-scores for each of the three NE classes (the class Other is excluded).
Since the differences between the tested languages depend more on the properties of the NER datasets than on the quality of embeddings, we can not directly compare ELMo models. For this reason, we take the non-contextual fastText embeddings as a baseline and predict named entities using them. The architecture of the model using fastText embeddings is the same as the one using ELMo embeddings, except that the input uses 300 dimensional fastText embedding vectors, and the model was trained for 5 epochs (instead of 3 as for ELMo). In both cases (ELMo and fastText) we trained and evaluated the model five times, because there is some random component involved in initialization of the neural network model. By training and evaluating multiple times, we minimise this random component.
The results are presented in Table TABREF21. We included the evaluation of the original ELMo English model in the same table. NER models have little difficulty distinguishing between types of named entities, but recognizing whether a word is a named entity or not is more difficult. For languages with the smallest NER datasets, Croatian and Lithuanian, ELMo embeddings show the largest improvement over fastText embeddings. However, we can observe significant improvements with ELMo also on English and Finnish, which are among the largest datasets (English being by far the largest). Only on Slovenian dataset did ELMo perform slightly worse than fastText, on all other EMBEDDIA languages, the ELMo embeddings improve the results.
Conclusion
We prepared precomputed ELMo contextual embeddings for seven languages: Croatian, Estonian, Finnish, Latvian, Lithuanian, Slovenian, and Swedish. We present the necessary background on embeddings and contextual embeddings, the details of training the embedding models, and their evaluation. We show that the size of used training sets importantly affects the quality of produced embeddings, and therefore the existing publicly available ELMo embeddings for the processed languages are inadequate. We trained new ELMo embeddings on larger training sets and analysed their properties on the analogy task and on the NER task. The results show that the newly produced contextual embeddings produce substantially better results compared to the non-contextual fastText baseline. In future work, we plan to use the produced contextual embeddings on the problems of news media industry. The pretrained ELMo models will be deposited to the CLARIN repository by the time of the final version of this paper.
Acknowledgments
The work was partially supported by the Slovenian Research Agency (ARRS) core research programme P6-0411. This paper is supported by European Union's Horizon 2020 research and innovation programme under grant agreement No 825153, project EMBEDDIA (Cross-Lingual Embeddings for Less-Represented Languages in European News Media). The results of this publication reflects only the authors' view and the EU Commission is not responsible for any use that may be made of the information it contains.
| [
"5 percent points.",
"0.05 F1"
] | 3,290 | qasper | en | null | bfc5d4d72997fdcc0107cd5cab9b9718777c43b86be45eb2 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What is the improvement in performance for Estonian in the NER task?
Answer: | [
"5 percent points.",
"0.05 F1"
] | qasper | 128 | 34 | |
What background do they have? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
In June 2015, the operators of the online discussion site Reddit banned several communities under new anti-harassment rules. BIBREF0 used this opportunity to combine rich online data with computational methods to study a current question: Does eliminating these “echo chambers” diminish the amount of hate speech overall? Exciting opportunities like these, at the intersection of “thick” cultural and societal questions on the one hand, and the computational analysis of rich textual data on larger-than-human scales on the other, are becoming increasingly common.
Indeed, computational analysis is opening new possibilities for exploring challenging questions at the heart of some of the most pressing contemporary cultural and social issues. While a human reader is better equipped to make logical inferences, resolve ambiguities, and apply cultural knowledge than a computer, human time and attention are limited. Moreover, many patterns are not obvious in any specific context, but only stand out in the aggregate. For example, in a landmark study, BIBREF1 analyzed the authorship of The Federalist Papers using a statistical text analysis by focusing on style, based on the distribution of function words, rather than content. As another example, BIBREF2 studied what defines English haiku and showed how computational analysis and close reading can complement each other. Computational approaches are valuable precisely because they help us identify patterns that would not otherwise be discernible.
Yet these approaches are not a panacea. Examining thick social and cultural questions using computational text analysis carries significant challenges. For one, texts are culturally and socially situated. They reflect the ideas, values and beliefs of both their authors and their target audiences, and such subtleties of meaning and interpretation are difficult to incorporate in computational approaches. For another, many of the social and cultural concepts we seek to examine are highly contested — hate speech is just one such example. Choices regarding how to operationalize and analyze these concepts can raise serious concerns about conceptual validity and may lead to shallow or obvious conclusions, rather than findings that reflect the depth of the questions we seek to address.
These are just a small sample of the many opportunities and challenges faced in computational analyses of textual data. New possibilities and frustrating obstacles emerge at every stage of research, from identification of the research question to interpretation of the results. In this article, we take the reader through a typical research process that involves measuring social or cultural concepts using computational methods, discussing both the opportunities and complications that often arise. In the Reddit case, for example, hate speech is measured, however imperfectly, by the presence of particular words semi-automatically extracted from a machine learning algorithm. Operationalizations are never perfect translations, and are often refined over the course of an investigation, but they are crucial.
We begin our exploration with the identification of research questions, proceed through data selection, conceptualization, and operationalization, and end with analysis and the interpretation of results. The research process sounds more or less linear this way, but each of these phases overlaps, and in some instances turns back upon itself. The analysis phase, for example, often feeds back into the original research questions, which may continue to evolve for much of the project. At each stage, our discussion is critically informed by insights from the humanities and social sciences, fields that have focused on, and worked to tackle, the challenges of textual analysis—albeit at smaller scales—since their inception.
In describing our experiences with computational text analysis, we hope to achieve three primary goals. First, we aim to shed light on thorny issues not always at the forefront of discussions about computational text analysis methods. Second, we hope to provide a set of best practices for working with thick social and cultural concepts. Our guidance is based on our own experiences and is therefore inherently imperfect. Still, given our diversity of disciplinary backgrounds and research practices, we hope to capture a range of ideas and identify commonalities that will resonate for many. And this leads to our final goal: to help promote interdisciplinary collaborations. Interdisciplinary insights and partnerships are essential for realizing the full potential of any computational text analysis that involves social and cultural concepts, and the more we are able to bridge these divides, the more fruitful we believe our work will be.
Research questions
We typically start by identifying the questions we wish to explore. Can text analysis provide a new perspective on a “big question” that has been attracting interest for years? Or can we raise new questions that have only recently emerged, for example about social media? For social scientists working in computational analysis, the questions are often grounded in theory, asking: How can we explain what we observe? These questions are also influenced by the availability and accessibility of data sources. For example, the choice to work with data from a particular social media platform may be partly determined by the fact that it is freely available, and this will in turn shape the kinds of questions that can be asked. A key output of this phase are the concepts to measure, for example: influence; copying and reproduction; the creation of patterns of language use; hate speech. Computational analysis of text motivated by these questions is insight driven: we aim to describe a phenomenon or explain how it came about. For example, what can we learn about how and why hate speech is used or how this changes over time? Is hate speech one thing, or does it comprise multiple forms of expression? Is there a clear boundary between hate speech and other types of speech, and what features make it more or less ambiguous? In these cases, it is critical to communicate high-level patterns in terms that are recognizable.
This contrasts with much of the work in computational text analysis, which tends to focus on automating tasks that humans perform inefficiently. These tasks range from core linguistically motivated tasks that constitute the backbone of natural language processing, such as part-of-speech tagging and parsing, to filtering spam and detecting sentiment. Many tasks are motivated by applications, for example to automatically block online trolls. Success, then, is often measured by performance, and communicating why a certain prediction was made—for example, why a document was labeled as positive sentiment, or why a word was classified as a noun—is less important than the accuracy of the prediction itself. The approaches we use and what we mean by `success' are thus guided by our research questions.
Domain experts and fellow researchers can provide feedback on questions and help with dynamically revising them. For example, they may say “we already think we know that”, “that's too naïve”, “that doesn't reflect social reality” (negative); “two major camps in the field would give different answers to that question” (neutral); “we tried to look at that back in the 1960s, but we didn't have the technology” (positive); and “that sounds like something that people who made that archive would love”, “that's a really fundamental question” (very positive).
Sometimes we also hope to connect to multiple disciplines. For example, while focusing on the humanistic concerns of an archive, we could also ask social questions such as “is this archive more about collaborative processes, culture-building or norm creation?” or “how well does this archive reflect the society in which it is embedded?" BIBREF3 used quantitative methods to tell a story about Darwin's intellectual development—an essential biographical question for a key figure in the history of science. At the same time, their methods connected Darwin's development to the changing landscape of Victorian scientific culture, allowing them to contrast Darwin's “foraging” in the scientific literature of his time to the ways in which that literature was itself produced. Finally, their methods provided a case study, and validation of technical approaches, for cognitive scientists who are interested in how people explore and exploit sources of knowledge.
Questions about potential “dual use” may also arise. Returning to our introductory example, BIBREF0 started with a deceptively simple question: if an internet platform eliminates forums for hate speech, does this impact hate speech in other forums? The research was motivated by the belief that a rising tide of online hate speech was (and is) making the internet increasingly unfriendly for disempowered groups, including minorities, women, and LBGTQ individuals. Yet the possibility of dual use troubled the researchers from the onset. Could the methodology be adopted to target the speech of groups like Black Lives Matter? Could it be adopted by repressive governments to minimize online dissent? While these concerns remained, they concluded that hypothetical dual use scenarios did not outweigh the tangible contribution this research could offer towards making the online environment more equal and just.
Data
The next step involves deciding on the data sources, collecting and compiling the dataset, and inspecting its metadata.
Data acquisition
Many scholars in the humanities and the social sciences work with sources that are not available in digital form, and indeed may never be digitized. Others work with both analogue and digitized materials, and the increasing digitization of archives has opened opportunities to study these archives in new ways. We can go to the canonical archive or open up something that nobody has studied before. For example, we might focus on major historical moments (French Revolution, post-Milosevic Serbia) or critical epochs (Britain entering the Victorian era, the transition from Latin to proto-Romance). Or, we could look for records of how people conducted science, wrote and consumed literature, and worked out their philosophies.
A growing number of researchers work with born-digital sources or data. Born-digital data, e.g., from social media, generally do not involve direct elicitation from participants and therefore enable unobtrusive measurements BIBREF5 , BIBREF6 . In contrast, methods like surveys sometimes elicit altered responses from participants, who might adapt their responses to what they think is expected. Moreover, born-digital data is often massive, enabling large-scale studies of language and behavior in a variety of social contexts.
Still, many scholars in the social sciences and humanities work with multiple data sources. The variety of sources typically used means that more than one data collection method is often required. For example, a project examining coverage of a UK General Election, could draw data from traditional media, web archives, Twitter and Facebook, campaign manifestos, etc. and might combine textual analysis of these materials with surveys, laboratory experiments, or field observations offline. In contrast, many computational studies based on born-digital data have focused on one specific source, such as Twitter.
The use of born-digital data raises ethical concerns. Although early studies often treated privacy as a binary construct, many now acknowledge its complexity BIBREF7 . Conversations on private matters can be posted online, visible for all, but social norms regarding what should be considered public information may differ from the data's explicit visibility settings. Often no informed consent has been obtained, raising concerns and challenges regarding publishing content and potentially harmful secondary uses BIBREF8 , BIBREF4 .
Recently, concerns about potential harms stemming from secondary uses have led a number of digital service providers to restrict access to born-digital data. Facebook and Twitter, for example, have reduced or eliminated public access to their application programming interfaces (APIs) and expressed hesitation about allowing academic researchers to use data from their platforms to examine certain sensitive or controversial topics. Despite the seeming abundance of born-digital data, we therefore cannot take its availability for granted.
Working with data that someone else has acquired presents additional problems related to provenance and contextualisation. It may not always be possible to determine the criteria applied during the creation process. For example, why were certain newspapers digitized but not others, and what does this say about the collection? Similar questions arise with the use of born-digital data. For instance, when using the Internet Archive’s Wayback Machine to gather data from archived web pages, we need to consider what pages were captured, which are likely missing, and why.
We must often repurpose born-digital data (e.g., Twitter was not designed to measure public opinion), but data biases may lead to spurious results and limit justification for generalization. In particular, data collected via black box APIs designed for commercial, not research, purposes are likely to introduce biases into the inferences we draw, and the closed nature of these APIs means we rarely know what biases are introduced, let alone how severely they might impact our research BIBREF10 . These, however, are not new problems. Historians, for example, have always understood that their sources were produced within particular contexts and for particular purposes, which are not always apparent to us.
Non-representative data can still be useful for making comparisons within a sample. In the introductory example on hate speech BIBREF0 , the Reddit forums do not present a comprehensive or balanced picture of hate speech: the writing is almost exclusively in English, the targets of hate speech are mainly restricted (e.g., to black people, or women), and the population of writers is shaped by Reddit's demographics, which skew towards young white men. These biases limit the generalizability of the findings, which cannot be extrapolated to other languages, other types of hate speech, and other demographic groups. However, because the findings are based on measurements on the same sort of hate speech and the same population of writers, as long as the collected data are representative of this specific population, these biases do not pose an intractable validity problem if claims are properly restricted.
The size of many newly available datasets is one of their most appealing characteristics. Bigger datasets often make statistics more robust. The size needed for a computational text analysis depends on the research goal: When it involves studying rare events, bigger datasets are needed. However, larger is not always better. Some very large archives are “secretly” collections of multiple and distinct processes that no in-field scholar would consider related. For example, Google Books is frequently used to study cultural patterns, but the over-representation of scientific articles in Google books can be problematic BIBREF11 . Even very large born-digital datasets usually cover limited timespans compared to, e.g., the Gutenberg archive of British novels.
This stage of the research also raises important questions about fairness. Are marginalized groups, for example, represented in the tweets we have collected? If not, what types of biases might result from analyses relying on those tweets?
Local experts and “informants” can help navigate the data. They can help understand the role an archive plays in the time and place. They might tell us: Is this the central archive, or a peripheral one? What makes it unusual? Or they might tell us how certain underrepresented communities use a social media platform and advise us on strategies for ensuring our data collection includes their perspectives.
However, when it is practically infeasible to navigate the data in this way—for instance, when we cannot determine what is missing from Twitter's Streaming API or what webpages are left out of the Internet Archive—we should be open about the limitations of our analyses, acknowledging the flaws in our data and drawing cautious and reasonable conclusions from them. In all cases, we should report the choices we have made when creating or re-using any dataset.
Compiling data
After identifying the data source(s), the next step is compiling the data. This step is fundamental: if the sources cannot support a convincing result, no result will be convincing. In many cases, this involves defining a “core" set of documents and a “comparison" set. We often have a specific set of documents in mind: an author's work, a particular journal, a time period. But if we want to say that this “core" set has some distinctive property, we need a “comparison" set. Expanding the collection beyond the documents that we would immediately think of has the beneficial effect of increasing our sample size. Having more sources increases the chance that we will notice something consistent across many individually varying contexts.
Comparing sets of documents can sometimes support causal inference, presented as a contrast between a treatment group and a control. In BIBREF0 , the treatment consisted of the text written in the two forums that were eventually closed by Reddit. However, identifying a control group required a considerable amount of time and effort. Reddit is a diverse platform, with a wide variety of interactional and linguistic styles; it would be pointless to compare hate speech forums against forums dedicated to, say, pictures of wrecked bicycles. Chandrasekharan et al. used a matching design, populating the control group with forums that were as similar as possible to the treatment group, but were not banned from Reddit. The goal is to estimate the counterfactual scenario: in this case, what would have happened had the site not taken action against these specific forums? An ideal control would make it possible to distinguish the effect of the treatment — closing the forums — from other idiosyncratic properties of texts that were treated.
We also look for categories of documents that might not be useful. We might remove documents that are meta-discourse, like introductions and notes, or documents that are in a language that is not the primary language of the collection, or duplicates when we are working with archived web pages. However, we need to carefully consider the potential consequences of information we remove. Does its removal alter the data, or the interpretation of the data, we are analyzing? Are we losing anything that might be valuable at a later stage?
Labels and metadata
Sometimes all we have is documents, but often we want to look at documents in the context of some additional information, or metadata. This additional information could tell us about the creation of documents (date, author, forum), or about the reception of documents (flagged as hate speech, helpful review). Information about text segments can be extremely valuable, but it is also prone to errors, inconsistencies, bias, and missing information. Examining metadata is a good way to check a collection's balance and representativeness. Are sources disproportionately of one form? Is the collection missing a specific time window? This type of curation can be extremely time consuming as it may require expert labeling, but it often leads to the most compelling results. Sometimes metadata are also used as target labels to develop machine learning models. But using them as a “ground truth” requires caution. Labels sometimes mean something different than we expect. For example, a down vote for a social media post could indicate that the content is offensive, or that the voter simply disagreed with the expressed view.
Conceptualization
A core step in many analyses is translating social and cultural concepts (such as hate speech, rumor, or conversion) into measurable quantities. Before we can develop measurements for these concepts (the operationalization step, or the “implementation” step as denoted by BIBREF12 ), we need to define them. In the conceptualization phase we often start with questions such as: who are the domain experts, and how have they approached the topic? We are looking for a definition of the concept that is flexible enough to apply on our dataset, yet formal enough for computational research. For example, our introductory study on hate speech BIBREF0 used a statement on hate speech produced by the European Union Court of Human Rights. The goal was not to implement this definition directly in software but to use it as a reference point to anchor subsequent analyses.
If we want to move beyond the use of ad hoc definitions, it can be useful to distinguish between what political scientists Adcock and Collier call the “background concept” and the “systematized concept” BIBREF13 . The background concept comprises the full and diverse set of meanings that might be associated with a particular term. This involves delving into theoretical, conceptual, and empirical studies to assess how a concept has been defined by other scholars and, most importantly, to determine which definition is most appropriate for the particular research question and the theoretical framework in which it is situated. That definition, in turn, represents the systematized concept: the formulation that is adopted for the study.
It is important to consider that for social and cultural concepts there is no absolute ground truth. There are often multiple valid definitions for a concept (the “background” concept in the terms of Adcock and Collier), and definitions might be contested over time. This may be uncomfortable for computer scientists, whose primary measure of success is often based on comparing a model's output against “ground truth” or a “gold standard”, e.g., by comparing a sentiment classifier's output against manual annotations. However, the notion of ground truth is uncommon in the humanities and the social sciences and it is often taken too far in machine learning. BIBREF14 notes that in literary criticism and the digital humanities more broadly “interpretation, ambiguity, and argumentation are prized far above ground truth and definitive conclusions". BIBREF15 draw attention to the different attitudes of literary scholars and computational linguists towards ambiguity, stating that “In Computational Linguistics [..] ambiguity is almost uniformly treated as a problem to be solved; the focus is on disambiguation, with the assumption that one true, correct interpretation exists." The latter is probably true for tasks such as spam filtering, but in the social sciences and the humanities many relevant concepts are fundamentally unobservable, such as latent traits of political actors BIBREF16 or cultural fit in organizations BIBREF17 , leading to validation challenges. Moreover, when the ground truth comes from people, it may be influenced by ideological priors, priming, simple differences of opinion or perspective, and many other factors BIBREF18 . We return to this issue in our discussions on validation and analysis.
Operationalization
In this phase we develop measures (or, “operationalizations”, or “indicators”) for the concepts of interest, a process called “operationalization”. Regardless of whether we are working with computers, the output produced coincides with Adcock and Collier's “scores”—the concrete translation and output of the systematized concept into numbers or labels BIBREF13 . Choices made during this phase are always tied to the question “Are we measuring what we intend to measure?” Does our operationalization match our conceptual definition? To ensure validity we must recognize gaps between what is important and what is easy to measure. We first discuss modeling considerations. Next, we describe several frequently used computational approaches and their limitations and strengths.
Modeling considerations
The variables (both predictors and outcomes) are rarely simply binary or categorical. For example, a study on language use and age could focus on chronological age (instead of, e.g., social age BIBREF19 ). However, even then, age can be modeled in different ways. Discretization can make the modeling easier and various NLP studies have modeled age as a categorical variable BIBREF20 . But any discretization raises questions: How many categories? Where to place the boundaries? Fine distinctions might not always be meaningful for the analysis we are interested in, but categories that are too broad can threaten validity. Other interesting variables include time, space, and even the social network position of the author. It is often preferable to keep the variable in its most precise form. For example, BIBREF21 perform exploration in the context of hypothesis testing by using latitude and longitude coordinates — the original metadata attached to geotagged social media such as tweets — rather than aggregating into administrative units such as counties or cities. This is necessary when such administrative units are unlikely to be related to the target concept, as is the case in their analysis of dialect differences. Focusing on precise geographical coordinates also makes it possible to recognize fine-grained effects, such as language variation across the geography of a city.
Using a particular classification scheme means deciding which variations are visible, and which ones are hidden BIBREF22 . We are looking for a categorization scheme for which it is feasible to collect a large enough labeled document collection (e.g., to train supervised models), but which is also fine-grained enough for our purposes. Classification schemes rarely exhibit the ideal properties, i.e., that they are consistent, their categories are mutually exclusive, and that the system is complete BIBREF22 . Borderline cases are challenging, especially with social and cultural concepts, where the boundaries are often not clear-cut. The choice of scheme can also have ethical implications BIBREF22 . For example, gender is usually represented as a binary variable in NLP and computational models tend to learn gender-stereotypical patterns. The operationalization of gender in NLP has been challenged only recently BIBREF23 , BIBREF24 , BIBREF25 .
Supervised and unsupervised learning are the most common approaches to learning from data. With supervised learning, a model learns from labeled data (e.g., social media messages labeled by sentiment) to infer (or predict) these labels from unlabeled texts. In contrast, unsupervised learning uses unlabeled data. Supervised approaches are especially suitable when we have a clear definition of the concept of interest and when labels are available (either annotated or native to the data). Unsupervised approaches, such as topic models, are especially useful for exploration. In this setting, conceptualization and operationalization may occur simultaneously, with theory emerging from the data BIBREF26 . Unsupervised approaches are also used when there is a clear way of measuring a concept, often based on strong assumptions. For example, BIBREF3 measure “surprise” in an analysis of Darwin's reading decisions based on the divergence between two probability distributions.
From an analysis perspective, the unit of text that we are labeling (or annotating, or coding), either automatic or manual, can sometimes be different than one's final unit of analysis. For example, if in a study on media frames in news stories, the theoretical framework and research question point toward frames at the story level (e.g., what is the overall causal analysis of the news article?), the story must be the unit of analysis. Yet it is often difficult to validly and reliably code a single frame at the story level. Multiple perspectives are likely to sit side-by-side in a story. Thus, an article on income inequality might point to multiple causes, such as globalization, education, and tax policies. Coding at the sentence level would detect each of these causal explanations individually, but this information would need to be somehow aggregated to determine the overall story-level frame. Sometimes scholars solve this problem by only examining headlines and lead paragraphs, arguing that based on journalistic convention, the most important information can be found at the beginning of a story. However, this leads to a return to a shorter, less nuanced analysis.
From a computational perspective, the unit of text can also make a huge difference, especially when we are using bag-of-words models, where word order within a unit does not matter. Small segments, like tweets, sometimes do not have enough information to make their semantic context clear. In contrast, larger segments, like novels, have too much variation, making it difficult to train focused models. Finding a good segmentation sometimes means combining short documents and subdividing long documents. The word “document" can therefore be misleading. But it is so ingrained in the common NLP lexicon that we use it anyway in this article.
For insight-driven text analysis, it is often critical that high-level patterns can be communicated. Furthermore, interpretable models make it easier to find spurious features, to do error analysis, and to support interpretation of results. Some approaches are effective for prediction, but harder to interpret. The value we place on interpretability can therefore influence the approach we choose. There is an increasing interest in developing interpretable or transparent models in the NLP and machine learning communities.
Annotation
Many studies involve human coders. Sometimes the goal is to fully code the data, but in a computational analysis we often use the labels (or annotations) to train machine learning models to automatically recognize them, and to identify language patterns that are associated with these labels. For example, for a project analyzing rumors online BIBREF27 , conversation threads were annotated along different dimensions, including rumor versus non-rumor and stance towards a rumor.
The collection of annotation choices make up an annotation scheme (or “codebook”). Existing schemes and annotations can be useful as starting points. Usually settling on an annotation scheme requires several iterations, in which the guidelines are updated and annotation examples are added. For example, a political scientist could use a mixed deductive-inductive strategy for developing a codebook. She starts by laying out a set of theory-driven deductive coding rules, which means that the broad principles of the coding rules are laid out without examining examples first. These are then tested (and possibly adjusted) based on a sample of the data. In line with Adcock and Collier's notion of “content validity” BIBREF13 , the goal is to assess whether the codebook adequately captures the systematized concept. By looking at the data themselves, she gains a better sense of whether some things have been left out of the coding rules and whether anything is superfluous, misleading, or confusing. Adjustments are made and the process is repeated, often with another researcher involved.
The final annotations can be collected using a crowdsourcing platform, a smaller number of highly-trained annotators, or a group of experts. Which type of annotator to use should be informed by the complexity and specificity of the concept. For more complex concepts, highly-trained or expert annotators tend to produce more reliable results. However, complex concepts can sometimes be broken down into micro-tasks that can be performed independently in parallel by crowdsourced annotators. Concepts from highly specialized domains may require expert annotators. In all cases, however, some training will be required, and the training phase should involve continual checks of inter-annotator agreement (i.e. intercoder reliability) or checks against a gold standard (e.g. quizzes in crowdsourcing platforms).
We also need to decide how inter-annotator agreement will be measured and what an acceptable level of agreement would be. Krippendorff's alpha is frequently used in the social sciences, but the right measure depends on the type of data and task. For manual coding, we can continually check inter-annotator agreement and begin introducing checks of intra-annotator agreement, too. For most communication scholars using only manual content analysis, an acceptable rate of agreement is achieved when Krippendorf's alpha reaches 0.80 or above. When human-coded data are used to validate machine learning algorithms, the reliability of the human-coded data is even more important. Disagreement between annotators can signal weaknesses of the annotation scheme, or highlight the inherent ambiguity in what we are trying to measure. Disagreement itself can be meaningful and can be integrated in subsequent analyses BIBREF28 , BIBREF29 .
Data pre-processing
Preparing the data can be a complex and time-consuming process, often involving working with partially or wholly unstructured data. The pre-processing steps have a big impact on the operationalizations, subsequent analyses and reproducibility efforts BIBREF30 , and they are usually tightly linked to what we intend to measure. Unfortunately, these steps tend to be underreported, but documenting the pre-processing choices made is essential and is analogous to recording the decisions taken during the production of a scholarly edition or protocols in biomedical research. Data may also vary enormously in quality, depending on how it has been generated. Many historians, for example, work with text produced from an analogue original using Optical Character Recognition (OCR). Often, there will be limited information available regarding the accuracy of the OCR, and the degree of accuracy may even vary within a single corpus (e.g. where digitized text has been produced over a period of years, and the software has gradually improved). The first step, then, is to try to correct for common OCR errors. These will vary depending on the type of text, the date at which the `original' was produced, and the nature of the font and typesetting.
One step that almost everyone takes is to tokenize the original character sequence into the words and word-like units. Tokenization is a more subtle and more powerful process than people expect. It is often done using regular expressions or scripts that have been circulating within the NLP community. Tokenization heuristics, however, can be badly confused by emoticons, creative orthography (e.g., U$A, sh!t), and missing whitespace. Multi-word terms are also challenging. Treating them as a single unit can dramatically alter the patterns in text. Many words that are individually ambiguous have clear, unmistakable meanings as terms, like “black hole" or “European Union". However, deciding what constitutes a multi-word term is a difficult problem. In writing systems like Chinese, tokenization is a research problem in its own right.
Beyond tokenization, common steps include lowercasing, removing punctuation, stemming (removing suffixes), lemmatization (converting inflections to a base lemma), and normalization, which has never been clearly defined, but often includes grouping abbreviations like “U.S.A." and “USA", ordinals like “1st" and “first", and variant spellings like “noooooo". The main goal of these steps is to improve the ratio of tokens (individual occurrences) to types (the distinct things in a corpus). Each step requires making additional assumptions about which distinctions are relevant: is “apple” different from “Apple”? Is “burnt” different from “burned”? Is “cool" different from “coooool"? Sometimes these steps can actively hide useful patterns, like social meaning BIBREF32 . Some of us therefore try do as little modification as possible.
From a multilingual perspective, English and Chinese have an unusually simple inflectional system, and so it is statistically reasonable to treat each inflection as a unique word type. Romance languages have considerably more inflections than English; many indigenous North American languages have still more. For these languages, unseen data is far more likely to include previously-unseen inflections, and therefore, dealing with inflections is more important. On the other hand, the resources for handling inflections vary greatly by language, with European languages dominating the attention of the computational linguistics community thus far.
We sometimes also remove words that are not relevant to our goals, for example by calculating vocabulary frequencies. We construct a “stoplist” of words that we are not interested in. If we are looking for semantic themes we might remove function words like determiners and prepositions. If we are looking for author-specific styles, we might remove all words except function words. Some words are generally meaningful but too frequent to be useful within a specific collection. We sometimes also remove very infrequent words. Their occurrences are too low for robust patterns and removing them helps reducing the vocabulary size.
The choice of processing steps can be guided by theory or knowledge about the domain as well as experimental investigation. When we have labels, predictive accuracy of a model is a way to assess the effect of the processing steps. In unsupervised settings, it is more challenging to understand the effects of different steps. Inferences drawn from unsupervised settings can be sensitive to pre-processing choices BIBREF33 . Stemming has been found to provide little measurable benefits for topic modeling and can sometimes even be harmful BIBREF34 . All in all, this again highlights the need to document these steps.
Finally, we can also mark up the data, e.g., by identifying entities (people, places, organizations, etc.) or parts of speech. Although many NLP tools are available for such tasks, they are often challenged by linguistic variation, such as orthographic variation in historical texts BIBREF35 and social media BIBREF32 . Moreover, the performance of NLP tools often drops when applying them outside the training domain, such as applying tools developed on newswire texts to texts written by younger authors BIBREF36 . Problems (e.g., disambiguation in named entity recognition) are sometimes resolved using considerable manual intervention. This combination of the automated and the manual, however, becomes more difficult as the scale of the data increases, and the `certainty' brought by the latter may have to be abandoned.
Dictionary-based approaches
Dictionaries are frequently used to code texts in content analyses BIBREF37 . Dictionaries consist of one or more categories (i.e. word lists). Sometimes the output is simply the number of category occurrences (e.g., positive sentiment), thus weighting words within a category equally. In some other cases, words are assigned continuous scores. The high transparency of dictionaries makes them sometimes more suitable than supervised machine learning models. However, dictionaries should only be used if the scores assigned to words match how the words are used in the data (see BIBREF38 for a detailed discussion on limitations). There are many off-the-shelf dictionaries available (e.g., LIWC BIBREF39 ). These are often well-validated, but applying them on a new domain may not be appropriate without additional validation. Corpus- or domain-specific dictionaries can overcome limitations of general-purpose dictionaries.
The dictionaries are often manually compiled, but increasingly they are constructed semi-automatically (e.g., BIBREF40 ). When we semi-automatically create a word list, we use automation to identify an initial word list, and human insight to filter it. By automatically generating the initial words lists, words can be identified that human annotators might have difficulty intuiting. By manually filtering the lists, we use our theoretical understanding of the target concept to remove spurious features.
In the introduction study, SAGE BIBREF41 was used to obtain a list of words that distinguished the text in the treatment group (subreddits that were closed by Reddit) from text in the control group (similar subreddits that were not closed). The researchers then returned to the hate speech definition provided by the European Court of Human Rights, and manually filtered the top SAGE words based on this definition. Not all identified words fitted the definition. The others included: the names of the subreddits themselves, names of related subreddits, community-specific jargon that was not directly related to hate speech, and terms such as IQ and welfare, which were frequently used in discourses of hate speech, but had significant other uses. The word lists provided the measurement instrument for their main result, which is that the use of hate speech throughout Reddit declined after the two treatment subreddits were closed.
Supervised models
Supervised learning is frequently used to scale up analyses. For example, BIBREF42 wanted to analyze the motivations of Movember campaign participants. By developing a classifier based on a small set of annotations, they were able to expand the analysis to over 90k participants.
The choice of supervised learning model is often guided by the task definition and the label types. For example, to identify stance towards rumors based on sequential annotations, an algorithm for learning from sequential BIBREF43 or time series data BIBREF44 could be used. The features (sometimes called variables or predictors) are used by the model to make the predictions. They may vary from content-based features such as single words, sequences of words, or information about their syntactic structure, to meta-information such as user or network information. Deciding on the features requires experimentation and expert insight and is often called feature engineering. For insight-driven analysis, we are often interested in why a prediction has been made and features that can be interpreted by humans may be preferred. Recent neural network approaches often use simple features as input (such as word embeddings or character sequences), which requires less feature engineering but make interpretation more difficult.
Supervised models are powerful, but they can latch on to spurious features of the dataset. This is particularly true for datasets that are not well-balanced, and for annotations that are noisy. In our introductory example on hate speech in Reddit BIBREF0 , the annotations are automatically derived from the forum in which each post appears, and indeed, many of the posts in the forums (subreddits) that were banned by Reddit would be perceived by many as hate speech. But even in banned subreddits, not all of the content is hate speech (e.g., some of the top features were self-referential like the name of the subreddit) but a classifier would learn a high weight for these features.
Even when expert annotations are available on the level of individual posts, spurious features may remain. BIBREF45 produced expert annotations of hate speech on Twitter. They found that one of the strongest features for sexism is the name of an Australian TV show, because people like to post sexist comments about the contestants. If we are trying to make claims about what inhibits or encourages hate speech, we would not want those claims to be tied to the TV show's popularity. Such problems are inevitable when datasets are not well-balanced over time, across genres, topics, etc. Especially with social media data, we lack a clear and objective definition of `balance' at this time.
The risk of supervised models latching on to spurious features reinforces the need for interpretability. Although the development of supervised models is usually performance driven, placing more emphasis on interpretability could increase the adoption of these models in insight-driven analyses. One way would be to only use models that are already somewhat interpretable, for example models that use a small number of human-interpretable features. Rather than imposing such restrictions, there is also work on generating post-hoc explanations for individual predictions (e.g., BIBREF46 ), even when the underlying model itself is very complex.
Topic modeling
Topic models (e.g., LDA BIBREF47 ) are usually unsupervised and therefore less biased towards human-defined categories. They are especially suited for insight-driven analysis, because they are constrained in ways that make their output interpretable. Although there is no guarantee that a “topic” will correspond to a recognizable theme or event or discourse, they often do so in ways that other methods do not. Their easy applicability without supervision and ready interpretability make topic models good for exploration. Topic models are less successful for many performance-driven applications. Raw word features are almost always better than topics for search and document classification. LSTMs and other neural network models are better as language models. Continuous word embeddings have more expressive power to represent fine-grained semantic similarities between words.
A topic model provides a different perspective on a collection. It creates a set of probability distributions over the vocabulary of the collection, which, when combined together in different proportions, best match the content of the collection. We can sort the words in each of these distributions in descending order by probability, take some arbitrary number of most-probable words, and get a sense of what (if anything) the topic is “about”. Each of the text segments also has its own distribution over the topics, and we can sort these segments by their probability within a given topic to get a sense of how that topic is used.
One of the most common questions about topic models is how many topics to use, usually with the implicit assumption that there is a “right” number that is inherent in the collection. We prefer to think of this parameter as more like the scale of a map or the magnification of a microscope. The “right” number is determined by the needs of the user, not by the collection. If the analyst is looking for a broad overview, a relatively small number of topics may be best. If the analyst is looking for fine-grained phenomena, a larger number is better.
After fitting the model, it may be necessary to circle back to an earlier phase. Topic models find consistent patterns. When authors repeatedly use a particular theme or discourse, that repetition creates a consistent pattern. But other factors can also create similar patterns, which look as good to the algorithm. We might notice a topic that has highest probability on French stopwords, indicating that we need to do a better job of filtering by language. We might notice a topic of word fragments, such as “ing”, “tion”, “inter”, indicating that we are not handling end-of-line hyphenation correctly. We may need to add to our stoplist or change how we curate multi-word terms.
Validation
The output of our measurement procedures (in the social sciences often called the “scores”) must now be assessed in terms of their reliability and validity with regard to the (systemized) concept. Reliability aims to capture repeatability, i.e. the extent to which a given tool provides consistent results.
Validity assesses the extent to which a given measurement tool measures what it is supposed to measure. In NLP and machine learning, most models are primarily evaluated by comparing the machine-generated labels against an annotated sample. This approach presumes that the human output is the “gold standard" against which performance should be tested. In contrast, when the reliability is measured based on the output of different annotators, no coder is taken as the standard and the likelihood of coders reaching agreement by chance (rather than because they are “correct") is factored into the resulting statistic. Comparing against a “gold standard” suggests that the threshold for human inter- and intra-coder reliability should be particularly high.
Accuracy, as well as other measures such as precision, recall and F-score, are sometimes presented as a measure of validity, but if we do not have a genuinely objective determination of what something is supposed measure—as is often the case in text analysis—then accuracy is perhaps a better indication of reliability than of validity. In that case, validity needs to be assessed based on other techniques like those we discuss later in this section. It is also worth asking what level of accuracy is sufficient for our analysis and to what extent there may be an upper bound, especially when the labels are native to the data or when the notion of a “gold standard” is not appropriate.
For some in the humanities, validation takes the form of close reading, not designed to confirm whether the model output is correct, but to present what BIBREF48 refers to as a form of “further discovery in two directions”. Model outputs tell us something about the texts, while a close reading of the texts alongside those outputs tells us something about the models that can be used for more effective model building. Applying this circular, iterative process to 450 18th-century novels written in three languages, Piper was able to uncover a new form of “conversional novel” that was not previously captured in “literary history's received critical categories” BIBREF48 .
Along similar lines, we can subject both the machine-generated output and the human annotations to another round of content validation. That is, take a stratified random sample, selecting observations from the full range of scores, and ask: Do these make sense in light of the systematized concept? If not, what seems to be missing? Or is something extraneous being captured? This is primarily a qualitative process that requires returning to theory and interrogating the systematized concept, indicators, and scores together. This type of validation is rarely done in NLP, but it is especially important when it is difficult to assess what drives a given machine learning model. If there is a mismatch between the scores and systematized concept at this stage, the codebook may need to be adjusted, human coders retrained, more training data prepared, algorithms adjusted, or in some instances, even a new analytical method adopted.
Other types of validation are also possible, such as comparing with other approaches that aim to capture the same concept, or comparing the output with external measures (e.g., public opinion polls, the occurrence of future events). We can also go beyond only evaluating the labels (or point estimates). BIBREF16 used human judgments to not only assess the positional estimates from a scaling method of latent political traits but also to assess uncertainty intervals. Using different types of validation can increase our confidence in the approach, especially when there is no clear notion of ground truth.
Besides focusing on rather abstract evaluation measures, we could also assess the models in task-based settings using human experts. Furthermore, for insight-driven analyses, it can be more useful to focus on improving explanatory power than making small improvements in predictive performance.
Analysis
In this phase, we use our models to explore or answer our research questions. For example, given a topic model we can look at the connection between topics and metadata elements. Tags such as “hate speech" or metadata information imply a certain way of organizing the collection. Computational models provide another organization, which may differ in ways that provide more insight into how these categories manifest themselves, or fail to do so.
Moreover, when using a supervised approach, the “errors”, i.e. disagreement between the system output and human-provided labels, can point towards interesting cases for closer analysis and help us reflect on our conceptualizations. In the words of BIBREF2 , they can be “opportunities for interpretation”. Other types of “failures” can be insightful as well. Sometimes there is a “dog that didn't bark” BIBREF49 –i.e., something that everyone thinks we should have found, but we did not. Or, sometimes the failures are telling us about the existence of something in the data that nobody noticed, or thought important, until then (e.g., the large number of travel journals in Darwin's reading lists).
Computational text analysis is not a replacement for but rather an addition to the approaches one can take to analyze social and cultural phenomena using textual data. By moving back and forth between large-scale computational analyses and small-scale qualitative analyses, we can combine their strengths so that we can identify large-scale and long-term trends, but also tell individual stories. For example, the Reddit study on hate speech BIBREF0 raised various follow-up questions: Can we distinguish hate speech from people talking about hate speech? Did people find new ways to express hate speech? If so, did the total amount of online hate speech decrease after all? As possible next steps, a qualitative discourse analyst might examine a smaller corpus to investigate whether commenters were indeed expressing hate speech in new ways; a specialist in interview methodologies might reach out to commenters to better understand the role of online hate speech in their lives. Computational text analysis represents a step towards better understanding social and cultural phenomena, and it is in many cases better suited towards opening questions rather than closing them.
Conclusion
Insight-driven computational analysis of text is becoming increasingly common. It not only helps us see more broadly, it helps us see subtle patterns more clearly and allows us to explore radical new questions about culture and society. In this article we have consolidated our experiences, as scholars from very different disciplines, in analyzing text as social and cultural data and described how the research process often unfolds. Each of the steps in the process is time-consuming and labor-intensive. Each presents challenges. And especially when working across disciplines, the research often involves a fair amount of discussion—even negotiation—about what means of operationalization and approaches to analysis are appropriate and feasible. And yet, with a bit of perseverance and mutual understanding, conceptually sound and meaningful work results so that we can truly make use of the exciting opportunities rich textual data offers.
Acknowledgements
This work was supported by The Alan Turing Institute under the EPSRC grant EP/N510129/1. Dong Nguyen is supported with an Alan Turing Institute Fellowship (TU/A/000006). Maria Liakata is a Turing fellow at 40%. We would also like to thank the participants of the “Bridging disciplines in analysing text as social and cultural data” workshop held at the Turing Institute (2017) for insightful discussions. The workshop was funded by a Turing Institute seed funding award to Nguyen and Liakata.
| [
"Unanswerable"
] | 8,506 | qasper | en | null | 65f7bdb541fd6f01fe866dcf694891f92533186085fcba20 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What background do they have?
Answer: | [
"Unanswerable"
] | qasper | 128 | 35 | |
LDA is an unsupervised method; is this paper introducing an unsupervised approach to spam detection? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Microblogging such as Twitter and Weibo is a popular social networking service, which allows users to post messages up to 140 characters. There are millions of active users on the platform who stay connected with friends. Unfortunately, spammers also use it as a tool to post malicious links, send unsolicited messages to legitimate users, etc. A certain amount of spammers could sway the public opinion and cause distrust of the social platform. Despite the use of rigid anti-spam rules, human-like spammers whose homepages having photos, detailed profiles etc. have emerged. Unlike previous "simple" spammers, whose tweets contain only malicious links, those "smart" spammers are more difficult to distinguish from legitimate users via content-based features alone BIBREF0 .
There is a considerable amount of previous work on spammer detection on social platforms. Researcher from Twitter Inc. BIBREF1 collect bot accounts and perform analysis on the user behavior and user profile features. Lee et al. lee2011seven use the so-called social honeypot by alluring social spammers' retweet to build a benchmark dataset, which has been extensively explored in our paper. Some researchers focus on the clustering of urls in tweets and network graph of social spammers BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , showing the power of social relationship features.As for content information modeling, BIBREF6 apply improved sparse learning methods. However, few studies have adopted topic-based features. Some researchers BIBREF7 discuss topic characteristics of spamming posts, indicating that spammers are highly likely to dwell on some certain topics such as promotion. But this may not be applicable to the current scenario of smart spammers.
In this paper, we propose an efficient feature extraction method. In this method, two new topic-based features are extracted and used to discriminate human-like spammers from legitimate users. We consider the historical tweets of each user as a document and use the Latent Dirichlet Allocation (LDA) model to compute the topic distribution for each user. Based on the calculated topic probability, two topic-based features, the Local Outlier Standard Score (LOSS) which captures the user's interests on different topics and the Global Outlier Standard Score (GOSS) which reveals the user's interests on specific topic in comparison with other users', are extracted. The two features contain both local and global information, and the combination of them can distinguish human-like spammers effectively.
To the best of our knowledge, it is the first time that features based on topic distributions are used in spammer classification. Experimental results on one public dataset and one self-collected dataset further validate that the two sets of extracted topic-based features get excellent performance on human-like spammer classification problem compared with other state-of-the-art methods. In addition, we build a Weibo dataset, which contains both legitimate users and spammers.
To summarize, our major contributions are two-fold:
In the following sections, we first propose the topic-based features extraction method in Section 2, and then introduce the two datasets in Section 3. Experimental results are discussed in Section 4, and we conclude the paper in Section 5. Future work is presented in Section 6.
Methodology
In this section, we first provide some observations we obtained after carefully exploring the social network, then the LDA model is introduced. Based on the LDA model, the ways to obtain the topic probability vector for each user and the two topic-based features are provided.
Observation
After exploring the homepages of a substantial number of spammers, we have two observations. 1) social spammers can be divided into two categories. One is content polluters, and their tweets are all about certain kinds of advertisement and campaign. The other is fake accounts, and their tweets resemble legitimate users' but it seems they are simply random copies of others to avoid being detected by anti-spam rules. 2) For legitimate users, content polluters and fake accounts, they show different patterns on topics which interest them.
Legitimate users mainly focus on limited topics which interest him. They seldom post contents unrelated to their concern.
Content polluters concentrate on certain topics.
Fake accounts focus on a wide range of topics due to random copying and retweeting of other users' tweets.
Spammers and legitimate users show different interests on some topics e.g. commercial, weather, etc.
To better illustrate our observation, Figure. 1 shows the topic distribution of spammers and legitimate users in two employed datasets(the Honeypot dataset and Weibo dataset). We can see that on both topics (topic-3 and topic-11) there exists obvious difference between the red bars and green bars, representing spammers and legitimate users. On the Honeypot dataset, spammers have a narrower shape of distribution (the outliers on the red bar tail are not counted) than that of legitimate users. This is because there are more content polluters than fake accounts. In other word, spammers in this dataset tend to concentrate on limited topics. While on the Weibo dataset, fake accounts who are interested in different topics take large proportion of spammers. Their distribution is more flat (i.e. red bars) than that of the legitimate users. Therefore we can detect spammers by means of the difference of their topic distribution patterns.
LDA model
Blei et al.blei2003latent first presented Latent Dirichlet Allocation(LDA) as an example of topic model.
Each document $i$ is deemed as a bag of words $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $ and $M$ is the number of words. Each word is attributable to one of the document's topics $Z=\left\lbrace z_{i1},z_{i2},...,z_{iK}\right\rbrace $ and $K$ is the number of topics. $\psi _{k}$ is a multinomial distribution over words for topic $k$ . $\theta _i$ is another multinomial distribution over topics for document $i$ . The smoothed generative model is illustrated in Figure. 2 . $\alpha $ and $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $0 are hyper parameter that affect scarcity of the document-topic and topic-word distributions. In this paper, $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $1 , $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $2 and $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $3 are empirically set to 0.3, 0.01 and 15. The entire content of each Twitter user is regarded as one document. We adopt Gibbs Sampling BIBREF8 to speed up the inference of LDA. Based on LDA, we can get the topic probabilities for all users in the employed dataset as: $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $4 , where $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $5 is the number of users. Each element $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $6 is a topic probability vector for the $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $7 document. $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $8 is the raw topic probability vector and our features are developed on top of it.
Topic-based Features
Using the LDA model, each person in the dataset is with a topic probability vector $X_i$ . Assume $x_{ik}\in X_{i}$ denotes the likelihood that the $\emph {i}^{th}$ tweet account favors $\emph {k}^{th}$ topic in the dataset. Our topic based features can be calculated as below.
Global Outlier Standard Score measures the degree that a user's tweet content is related to a certain topic compared to the other users. Specifically, the "GOSS" score of user $i$ on topic $k$ can be calculated as Eq.( 12 ):
$$\centering \begin{array}{ll} \mu \left(x_{k}\right)=\frac{\sum _{i=1}^{n} x_{ik}}{n},\\ GOSS\left(x_{ik}\right)=\frac{x_{ik}-\mu \left(x_k\right)}{\sqrt{\underset{i}{\sum }\left(x_{ik}-\mu \left(x_{k}\right)\right)^{2}}}. \end{array}$$ (Eq. 12)
The value of $GOSS\left(x_{ik}\right)$ indicates the interesting degree of this person to the $\emph {k}^{th}$ topic. Specifically, if $GOSS\left(x_{ik}\right)$ > $GOSS\left(x_{jk}\right)$ , it means that the $\emph {i}^{th}$ person has more interest in topic $k$ than the $\emph {j}^{th}$ person. If the value $GOSS\left(x_{ik}\right)$ is extremely high or low, the $\emph {i}^{th}$ person showing extreme interest or no interest on topic $k$ which will probably be a distinctive pattern in the fowllowing classfication. Therefore, the topics interested or disliked by the $\emph {k}^{th}$0 person can be manifested by $\emph {k}^{th}$1 , from which the pattern of the interested topics with regarding to this person is found. Denote $\emph {k}^{th}$2 our first topic-based feature, and it hopefully can get good performance on spammer detection.
Local Outlier Standard Score measures the degree of interest someone shows to a certain topic by considering his own homepage content only. For instance, the "LOSS" score of account $i$ on topic $k$ can be calculated as Eq.( 13 ):
$$\centering \begin{array}{ll} \mu \left(x_{i}\right)=\frac{\sum _{k=1}^{K} x_{ik}}{K},\\ LOSS\left(x_{ik}\right)=\frac{x_{ik}-\mu \left(x_i\right)}{\sqrt{\underset{k}{\sum }\left(x_{ik}-\mu \left(x_{i}\right)\right)^{2}}}. \end{array}$$ (Eq. 13)
$\mu (x_i)$ represents the averaged interesting degree for all topics with regarding to $\emph {i}^{th}$ user and his tweet content. Similarly to $GOSS$ , the topics interested or disliked by the $\emph {i}^{th}$ person via considering his single post information can be manifested by $f_{LOSS}^{i}=[LOSS(x_{i1})\cdots LOSS(x_{iK})]$ , and $LOSS$ becomes our second topic-based features for the $\emph {i}^{th}$ person.
Dataset
We use one public dataset Social Honeypot dataset and one self-collected dataset Weibo dataset to validate the effectiveness of our proposed features.
Social Honeypot Dataset: Lee et al. lee2010devils created and deployed 60 seed social accounts on Twitter to attract spammers by reporting back what accounts interact with them. They collected 19,276 legitimate users and 22,223 spammers in their datasets along with their tweet content in 7 months. This is our first test dataset.
Our Weibo Dataset: Sina Weibo is one of the most famous social platforms in China. It has implemented many features from Twitter. The 2197 legitimate user accounts in this dataset are provided by the Tianchi Competition held by Sina Weibo. The spammers are all purchased commercially from multiple vendors on the Internet. We checked them manually and collected 802 suitable "smart" spammers accounts.
Preprocessing: Before directly performing the experiments on the employed datasets, we first delete some accounts with few posts in the two employed since the number of tweets is highly indicative of spammers. For the English Honeypot dataset, we remove stopwords, punctuations, non-ASCII words and apply stemming. For the Chinese Weibo dataset, we perform segmentation with "Jieba", a Chinese text segmentation tool. After preprocessing steps, the Weibo dataset contains 2197 legitimate users and 802 spammers, and the honeypot dataset contains 2218 legitimate users and 2947 spammers. It is worth mentioning that the Honeypot dataset has been slashed because most of the Twitter accounts only have limited number of posts, which are not enough to show their interest inclination.
Evaluation Metrics
The evaluating indicators in our model are show in 2 . We calculate precision, recall and F1-score (i.e. F1 score) as in Eq. ( 19 ). Precision is the ratio of selected accounts that are spammers. Recall is the ratio of spammers that are detected so. F1-score is the harmonic mean of precision and recall.
$$precision =\frac{TP}{TP+FP}, recall =\frac{TP}{TP+FN}\nonumber \\ F1-score = \frac{2\times precision \times recall}{precision + recall}$$ (Eq. 19)
Performance Comparisons with Baseline
Three baseline classification methods: Support Vector Machines (SVM), Adaboost, and Random Forests are adopted to evaluate our extracted features. We test each classification algorithm with scikit-learn BIBREF9 and run a 10-fold cross validation. On each dataset, the employed classifiers are trained with individual feature first, and then with the combination of the two features. From 1 , we can see that GOSS+LOSS achieves the best performance on F1-score among all others. Besides, the classification by combination of LOSS and GOSS can increase accuracy by more than 3% compared with raw topic distribution probability.
Comparison with Other Features
To compare our extracted features with previously used features for spammer detection, we use three most discriminative feature sets according to Lee et al. lee2011seven( 4 ). Two classifiers (Adaboost and SVM) are selected to conduct feature performance comparisons. Using Adaboost, our LOSS+GOSS features outperform all other features except for UFN which is 2% higher than ours with regard to precision on the Honeypot dataset. It is caused by the incorrectly classified spammers who are mostly news source after our manual check. They keep posting all kinds of news pieces covering diverse topics, which is similar to the behavior of fake accounts. However, UFN based on friendship networks is more useful for public accounts who possess large number of followers. The best recall value of our LOSS+GOSS features using SVM is up to 6% higher than the results by other feature groups. Regarding F1-score, our features outperform all other features. To further show the advantages of our proposed features, we compare our combined LOSS+GOSS with the combination of all the features from Lee et al. lee2011seven (UFN+UC+UH). It's obvious that LOSS+GOSS have a great advantage over UFN+UC+UH in terms of recall and F1-score. Moreover, by combining our LOSS+GOSS features and UFN+UC+UH features together, we obtained another 7.1% and 2.3% performance gain with regard to precision and F1-score by Adaboost. Though there is a slight decline in terms of recall. By SVM, we get comparative results on recall and F1-score but about 3.5% improvement on precision.
Conclusion
In this paper, we propose a novel feature extraction method to effectively detect "smart" spammers who post seemingly legitimate tweets and are thus difficult to identify by existing spammer classification methods. Using the LDA model, we obtain the topic probability for each Twitter user. By utilizing the topic probability result, we extract our two topic-based features: GOSS and LOSS which represent the account with global and local information. Experimental results on a public dataset and a self-built Chinese microblog dataset validate the effectiveness of the proposed features.
Future Work
In future work, the combination method of local and global information can be further improved to maximize their individual strengths. We will also apply decision theory to enhancing the performance of our proposed features. Moreover, we are also building larger datasets on both Twitter and Weibo to validate our method. Moreover, larger datasets on both Twitter and Weibo will be built to further validate our method.
| [
"No",
"No"
] | 2,239 | qasper | en | null | 091b4028a3b5e9d8248c58f17a62fd16c878da69693cfbfb | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: LDA is an unsupervised method; is this paper introducing an unsupervised approach to spam detection?
Answer: | [
"No",
"No"
] | qasper | 128 | 36 | |
Which languages are similar to each other? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Accurate language identification (LID) is the first step in many natural language processing and machine comprehension pipelines. If the language of a piece of text is known then the appropriate downstream models like parts of speech taggers and language models can be applied as required.
LID is further also an important step in harvesting scarce language resources. Harvested data can be used to bootstrap more accurate LID models and in doing so continually improve the quality of the harvested data. Availability of data is still one of the big roadblocks for applying data driven approaches like supervised machine learning in developing countries.
Having 11 official languages of South Africa has lead to initiatives (discussed in the next section) that have had positive effect on the availability of language resources for research. However, many of the South African languages are still under resourced from the point of view of building data driven models for machine comprehension and process automation.
Table TABREF2 shows the percentages of first language speakers for each of the official languages of South Africa. These are four conjunctively written Nguni languages (zul, xho, nbl, ssw), Afrikaans (afr) and English (eng), three disjunctively written Sotho languages (nso, sot, tsn), as well as tshiVenda (ven) and Xitsonga (tso). The Nguni languages are similar to each other and harder to distinguish. The same is true of the Sotho languages.
This paper presents a hierarchical naive Bayesian and lexicon based classifier for LID of short pieces of text of 15-20 characters long. The algorithm is evaluated against recent approaches using existing test sets from previous works on South African languages as well as the Discriminating between Similar Languages (DSL) 2015 and 2017 shared tasks.
Section SECREF2 reviews existing works on the topic and summarises the remaining research problems. Section SECREF3 of the paper discusses the proposed algorithm and Section SECREF4 presents comparative results.
Related Works
The focus of this section is on recently published datasets and LID research applicable to the South African context. An in depth survey of algorithms, features, datasets, shared tasks and evaluation methods may be found in BIBREF0.
The datasets for the DSL 2015 & DSL 2017 shared tasks BIBREF1 are often used in LID benchmarks and also available on Kaggle . The DSL datasets, like other LID datasets, consists of text sentences labelled by language. The 2017 dataset, for example, contains 14 languages over 6 language groups with 18000 training samples and 1000 testing samples per language.
The recently published JW300 parallel corpus BIBREF2 covers over 300 languages with around 100 thousand parallel sentences per language pair on average. In South Africa, a multilingual corpus of academic texts produced by university students with different mother tongues is being developed BIBREF3. The WiLI-2018 benchmark dataset BIBREF4 for monolingual written natural language identification includes around 1000 paragraphs of 235 languages. A possibly useful link can also be made BIBREF5 between Native Language Identification (NLI) (determining the native language of the author of a text) and Language Variety Identification (LVI) (classification of different varieties of a single language) which opens up more datasets. The Leipzig Corpora Collection BIBREF6, the Universal Declaration of Human Rights and Tatoeba are also often used sources of data.
The NCHLT text corpora BIBREF7 is likely a good starting point for a shared LID task dataset for the South African languages BIBREF8. The NCHLT text corpora contains enough data to have 3500 training samples and 600 testing samples of 300+ character sentences per language. Researchers have recently started applying existing algorithms for tasks like neural machine translation in earnest to such South African language datasets BIBREF9.
Existing NLP datasets, models and services BIBREF10 are available for South African languages. These include an LID algorithm BIBREF11 that uses a character level n-gram language model. Multiple papers have shown that 'shallow' naive Bayes classifiers BIBREF12, BIBREF8, BIBREF13, BIBREF14, SVMs BIBREF15 and similar models work very well for doing LID. The DSL 2017 paper BIBREF1, for example, gives an overview of the solutions of all of the teams that competed on the shared task and the winning approach BIBREF16 used an SVM with character n-gram, parts of speech tag features and some other engineered features. The winning approach for DSL 2015 used an ensemble naive Bayes classifier. The fasttext classifier BIBREF17 is perhaps one of the best known efficient 'shallow' text classifiers that have been used for LID .
Multiple papers have proposed hierarchical stacked classifiers (including lexicons) that would for example first classify a piece of text by language group and then by exact language BIBREF18, BIBREF19, BIBREF8, BIBREF0. Some work has also been done on classifying surnames between Tshivenda, Xitsonga and Sepedi BIBREF20. Additionally, data augmentation BIBREF21 and adversarial training BIBREF22 approaches are potentially very useful to reduce the requirement for data.
Researchers have investigated deeper LID models like bidirectional recurrent neural networks BIBREF23 or ensembles of recurrent neural networks BIBREF24. The latter is reported to achieve 95.12% in the DSL 2015 shared task. In these models text features can include character and word n-grams as well as informative character and word-level features learnt BIBREF25 from the training data. The neural methods seem to work well in tasks where more training data is available.
In summary, LID of short texts, informal styles and similar languages remains a difficult problem which is actively being researched. Increased confusion can in general be expected between shorter pieces of text and languages that are more closely related. Shallow methods still seem to work well compared to deeper models for LID. Other remaining research opportunities seem to be data harvesting, building standardised datasets and creating shared tasks for South Africa and Africa. Support for language codes that include more languages seems to be growing and discoverability of research is improving with more survey papers coming out. Paywalls also seem to no longer be a problem; the references used in this paper was either openly published or available as preprint papers.
Methodology
The proposed LID algorithm builds on the work in BIBREF8 and BIBREF26. We apply a naive Bayesian classifier with character (2, 4 & 6)-grams, word unigram and word bigram features with a hierarchical lexicon based classifier.
The naive Bayesian classifier is trained to predict the specific language label of a piece of text, but used to first classify text as belonging to either the Nguni family, the Sotho family, English, Afrikaans, Xitsonga or Tshivenda. The scikit-learn multinomial naive Bayes classifier is used for the implementation with an alpha smoothing value of 0.01 and hashed text features.
The lexicon based classifier is then used to predict the specific language within a language group. For the South African languages this is done for the Nguni and Sotho groups. If the lexicon prediction of the specific language has high confidence then its result is used as the final label else the naive Bayesian classifier's specific language prediction is used as the final result. The lexicon is built over all the data and therefore includes the vocabulary from both the training and testing sets.
The lexicon based classifier is designed to trade higher precision for lower recall. The proposed implementation is considered confident if the number of words from the winning language is at least one more than the number of words considered to be from the language scored in second place.
The stacked classifier is tested against three public LID implementations BIBREF17, BIBREF23, BIBREF8. The LID implementation described in BIBREF17 is available on GitHub and is trained and tested according to a post on the fasttext blog. Character (5-6)-gram features with 16 dimensional vectors worked the best. The implementation discussed in BIBREF23 is available from https://github.com/tomkocmi/LanideNN. Following the instructions for an OSX pip install of an old r0.8 release of TensorFlow, the LanideNN code could be executed in Python 3.7.4. Settings were left at their defaults and a learning rate of 0.001 was used followed by a refinement with learning rate of 0.0001. Only one code modification was applied to return the results from a method that previously just printed to screen. The LID algorithm described in BIBREF8 is also available on GitHub.
The stacked classifier is also tested against the results reported for four other algorithms BIBREF16, BIBREF26, BIBREF24, BIBREF15. All the comparisons are done using the NCHLT BIBREF7, DSL 2015 BIBREF19 and DSL 2017 BIBREF1 datasets discussed in Section SECREF2.
Results and Analysis
The average classification accuracy results are summarised in Table TABREF9. The accuracies reported are for classifying a piece of text by its specific language label. Classifying text only by language group or family is a much easier task as reported in BIBREF8.
Different variations of the proposed classifier were evaluated. A single NB classifier (NB), a stack of two NB classifiers (NB+NB), a stack of a NB classifier and lexicon (NB+Lex) and a lexicon (Lex) by itself. A lexicon with a 50% training token dropout is also listed to show the impact of the lexicon support on the accuracy.
From the results it seems that the DSL 2017 task might be harder than the DSL 2015 and NCHLT tasks. Also, the results for the implementation discussed in BIBREF23 might seem low, but the results reported in that paper is generated on longer pieces of text so lower scores on the shorter pieces of text derived from the NCHLT corpora is expected.
The accuracy of the proposed algorithm seems to be dependent on the support of the lexicon. Without a good lexicon a non-stacked naive Bayesian classifier might even perform better.
The execution performance of some of the LID implementations are shown in Table TABREF10. Results were generated on an early 2015 13-inch Retina MacBook Pro with a 2.9 GHz CPU (Turbo Boosted to 3.4 GHz) and 8GB RAM. The C++ implementation in BIBREF17 is the fastest. The implementation in BIBREF8 makes use of un-hashed feature representations which causes it to be slower than the proposed sklearn implementation. The execution performance of BIBREF23 might improve by a factor of five to ten when executed on a GPU.
Conclusion
LID of short texts, informal styles and similar languages remains a difficult problem which is actively being researched. The proposed algorithm was evaluated on three existing datasets and compared to the implementations of three public LID implementations as well as to reported results of four other algorithms. It performed well relative to the other methods beating their results. However, the performance is dependent on the support of the lexicon.
We would like to investigate the value of a lexicon in a production system and how to possibly maintain it using self-supervised learning. We are investigating the application of deeper language models some of which have been used in more recent DSL shared tasks. We would also like to investigate data augmentation strategies to reduce the amount of training data that is required.
Further research opportunities include data harvesting, building standardised datasets and shared tasks for South Africa as well as the rest of Africa. In general, the support for language codes that include more languages seems to be growing, discoverability of research is improving and paywalls seem to no longer be a big problem in getting access to published research.
| [
"Nguni languages (zul, xho, nbl, ssw), Sotho languages (nso, sot, tsn)",
"The Nguni languages are similar to each other, The same is true of the Sotho languages"
] | 1,877 | qasper | en | null | 93004cdb0e6d24f5de0568ed952fd50655e42900dc9dbfdb | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Which languages are similar to each other?
Answer: | [
"Nguni languages (zul, xho, nbl, ssw), Sotho languages (nso, sot, tsn)",
"The Nguni languages are similar to each other, The same is true of the Sotho languages"
] | qasper | 128 | 37 | |
which lstm models did they compare with? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Recently, deep neural network has been widely employed in various recognition tasks. Increasing the depth of neural network is a effective way to improve the performance, and convolutional neural network (CNN) has benefited from it in visual recognition task BIBREF0 . Deeper long short-term memory (LSTM) recurrent neural networks (RNNs) are also applied in large vocabulary continuous speech recognition (LVCSR) task, because LSTM networks have shown better performance than Fully-connected feed-forward deep neural network BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 .
Training neural network becomes more challenge when it goes deep. A conceptual tool called linear classifier probe is introduced to better understand the dynamics inside a neural network BIBREF5 . The discriminating features of linear classifier is the hidden units of a intermediate layer. For deep neural networks, it is observed that deeper layer's accuracy is lower than that of shallower layers. Therefore, the tool shows the difficulty of deep neural model training visually.
Layer-wise pre-training is a successful method to train very deep neural networks BIBREF6 . The convergence becomes harder with increasing the number of layers, even though the model is initialized with Xavier or its variants BIBREF7 , BIBREF8 . But the deeper network which is initialized with a shallower trained network could converge well.
The size of LVCSR training dataset goes larger and training with only one GPU becomes high time consumption inevitably. Therefore, parallel training with multi-GPUs is more suitable for LVCSR system. Mini-batch based stochastic gradient descent (SGD) is the most popular method in neural network training procedure. Asynchronous SGD is a successful effort for parallel training based on it BIBREF9 , BIBREF10 . It can many times speed up the training time without decreasing the accuracy. Besides, synchronous SGD is another effective effort, where the parameter server waits for every works to finish their computation and sent their local models to it, and then it sends updated model back to all workers BIBREF11 . Synchronous SGD converges well in parallel training with data parallelism, and is also easy to implement.
In order to further improve the performance of deep neural network with parallel training, several methods are proposed. Model averaging method achieves linear speedup, as the final model is averaged from all parameters of local models in different workers BIBREF12 , BIBREF13 , but the accuracy decreases compared with single GPU training. Moreover, blockwise model-updating filter (BMUF) provides another almost linear speedup approach with multi-GPUs on the basis of model averaging. It can achieve improvement or no-degradation of recognition performance compared with mini-batch SGD on single GPU BIBREF14 .
Moving averaged (MA) approaches are also proposed for parallel training. It is demonstrated that the moving average of the parameters obtained by SGD performs as well as the parameters that minimize the empirical cost, and moving average parameters can be used as the estimator of them, if the size of training data is large enough BIBREF15 . One pass learning is then proposed, which is the combination of learning rate schedule and averaged SGD using moving average BIBREF16 . Exponential moving average (EMA) is proposed as a non-interference method BIBREF17 . EMA model is not broadcasted to workers to update their local models, and it is applied as the final model of entire training process. EMA method is utilized with model averaging and BMUF to further decrease the character error rate (CER). It is also easy to implement in existing parallel training systems.
Frame stacking can also speed up the training time BIBREF18 . The super frame is stacked by several regular frames, and it contains the information of them. Thus, the network can see multiple frames at a time, as the super frame is new input. Frame stacking can also lead to faster decoding.
For streaming voice search service, it needs to display intermediate recognition results while users are still speaking. As a result, the system needs to fulfill high real-time requirement, and we prefer unidirectional LSTM network rather than bidirectional one. High real-time requirement means low real time factor (RTF), but the RTF of deep LSTM model is higher inevitably. The dilemma of recognition accuracy and real-time requirement is an obstacle to the employment of deep LSTM network. Deep model outperforms because it contains more knowledge, but it is also cumbersome. As a result, the knowledge of deep model can be distilled to a shallow model BIBREF19 . It provided a effective way to employ the deep model to the real-time system.
In this paper, we explore a entire deep LSTM RNN training framework, and employ it to real-time application. The deep learning systems benefit highly from a large quantity of labeled training data. Our first and basic speech recognition system is trained on 17000 hours of Shenma voice search dataset. It is a generic dataset sampled from diverse aspects of search queries. The requirement of speech recognition system also addressed by specific scenario, such as map and navigation task. The labeled dataset is too expensive, and training a new model with new large dataset from the beginning costs lots of time. Thus, it is natural to think of transferring the knowledge from basic model to new scenario's model. Transfer learning expends less data and less training time than full training. In this paper, we also introduce a novel transfer learning strategy with segmental Minimum Bayes-Risk (sMBR). As a result, transfer training with only 1000 hours data can match equivalent performance for full training with 7300 hours data.
Our deep LSTM training framework for LVCSR is presented in Section 2. Section 3 describes how the very deep models does apply in real world applications, and how to transfer the model to another task. The framework is analyzed and discussed in Section 4, and followed by the conclusion in Section 5.
Layer-wise Training with Soft Target and Hard Target
Gradient-based optimization of deep LSTM network with random initialization get stuck in poor solution easily. Xavier initialization can partially solve this problem BIBREF7 , so this method is the regular initialization method of all training procedure. However, it does not work well when it is utilized to initialize very deep model directly, because of vanishing or exploding gradients. Instead, layer-wise pre-training method is a effective way to train the weights of very deep architecture BIBREF6 , BIBREF20 . In layer-wise pre-training procedure, a one-layer LSTM model is firstly trained with normalized initialization. Sequentially, two-layers LSTM model's first layer is initialized by trained one-layer model, and its second layer is regularly initialized. In this way, a deep architecture is layer-by-layer trained, and it can converge well.
In conventional layer-wise pre-training, only parameters of shallower network are transfered to deeper one, and the learning targets are still the alignments generated by HMM-GMM system. The targets are vectors that only one state's probability is one, and the others' are zeros. They are known as hard targets, and they carry limited knowledge as only one state is active. In contrast, the knowledge of shallower network should be also transfered to deeper one. It is obtained by the softmax layer of existing model typically, so each state has a probability rather than only zero or one, and called as soft target. As a result, the deeper network which is student network learns the parameters and knowledge from shallower one which is called teacher network. When training the student network from the teacher network, the final alignment is the combination of hard target and soft target in our layer-wise training phase. The final alignment provides various knowledge which transfered from teacher network and extracted from true labels. If only soft target is learned, student network perform no better than teacher network, but it could outperform teacher network as it also learns true labels.
The deeper network spends less time to getting the same level of original network than the network trained from the beginning, as a period of low performance is skipped. Therefore, training with hard and soft target is a time saving method. For large training dataset, training with the whole dataset still spends too much time. A network firstly trained with only a small part of dataset could go deeper as well, and so the training time reducing rapidly. When the network is deep enough, it then trained on the entire dataset to get further improvement. There is no gap of accuracy between these two approaches, but latter one saves much time.
Differential Saturation Check
The objects of conventional saturation check are gradients and the cell activations BIBREF4 . Gradients are clipped to range [-5, 5], while the cell activations clipped to range [-50, 50]. Apart from them, the differentials of recurrent layers is also limited. If the differentials go beyond the range, corresponding back propagation is skipped, while if the gradients and cell activations go beyond the bound, values are set as the boundary values. The differentials which are too large or too small will lead to the gradients easily vanishing, and it demonstrates the failure of this propagation. As a result, the parameters are not updated, and next propagation .
Sequence Discriminative Training
Cross-entropy (CE) is widely used in speech recognition training system as a frame-wise discriminative training criterion. However, it is not well suited to speech recognition, because speech recognition training is a sequential learning problem. In contrast, sequence discriminative training criterion has shown to further improve performance of neural network first trained with cross-entropy BIBREF21 , BIBREF22 , BIBREF23 . We choose state-level minimum bayes risk (sMBR) BIBREF21 among a number of sequence discriminative criterion is proposed, such as maximum mutual information (MMI) BIBREF24 and minimum phone error (MPE) BIBREF25 . MPE and sMBR are designed to minimize the expected error of different granularity of labels, while CE aims to minimizes expected frame error, and MMI aims to minimizes expected sentence error. State-level information is focused on by sMBR.
a frame-level accurate model is firstly trained by CE loss function, and then sMBR loss function is utilized for further training to get sequence-level accuracy. Only a part of training dataset is needed in sMBR training phase on the basis of whole dataset CE training.
Parallel Training
It is demonstrated that training with larger dataset can improve recognition accuracy. However, larger dataset means more training samples and more model parameters. Therefore, parallel training with multiple GPUs is essential, and it makes use of data parallelism BIBREF9 . The entire training data is partitioned into several split without overlapping and they are distributed to different GPUs. Each GPU trains with one split of training dataset locally. All GPUs synchronize their local models with model average method after a mini-batch optimization BIBREF12 , BIBREF13 .
Model average method achieves linear speedup in training phase, but the recognition accuracy decreases compared with single GPU training. Block-wise model updating filter (BMUF) is another successful effort in parallel training with linear speedup as well. It can achieve no-degradation of recognition accuracy with multi-GPUs BIBREF14 . In the model average method, aggregated model INLINEFORM0 is computed and broadcasted to GPUs. On the basis of it, BMUF proposed a novel model updating strategy: INLINEFORM1 INLINEFORM2
Where INLINEFORM0 denotes model update, and INLINEFORM1 is the global-model update. There are two parameters in BMUF, block momentum INLINEFORM2 , and block learning rate INLINEFORM3 . Then, the global model is updated as INLINEFORM4
Consequently, INLINEFORM0 is broadcasted to all GPUs to initial their local models, instead of INLINEFORM1 in model average method.
Averaged SGD is proposed to further accelerate the convergence speed of SGD. Averaged SGD leverages the moving average (MA) INLINEFORM0 as the estimator of INLINEFORM1 BIBREF15 : INLINEFORM2
Where INLINEFORM0 is computed by model averaging or BMUF. It is shown that INLINEFORM1 can well converge to INLINEFORM2 , with the large enough training dataset in single GPU training. It can be considered as a non-interference strategy that INLINEFORM3 does not participate the main optimization process, and only takes effect after the end of entire optimization. However, for the parallel training implementation, each INLINEFORM4 is computed by model averaging and BMUF with multiple models, and moving average model INLINEFORM5 does not well converge, compared with single GPU training.
Model averaging based methods are employed in parallel training of large scale dataset, because of their faster convergence, and especially no-degradation implementation of BMUF. But combination of model averaged based methods and moving average does not match the expectation of further enhance performance and it is presented as INLINEFORM0
The weight of each INLINEFORM0 is equal in moving average method regardless the effect of temporal order. But INLINEFORM1 closer to the end of training achieve higher accuracy in the model averaging based approach, and thus it should be with more proportion in final INLINEFORM2 . As a result, exponential moving average(EMA) is appropriate, which the weight for each older parameters decrease exponentially, and never reaching zero. After moving average based methods, the EMA parameters are updated recursively as INLINEFORM3
Here INLINEFORM0 represents the degree of weight decrease, and called exponential updating rate. EMA is also a non-interference training strategy that is implemented easily, as the updated model is not broadcasted. Therefore, there is no need to add extra learning rate updating approach, as it can be appended to existing training procedure directly.
Deployment
There is a high real time requirement in real world application, especially in online voice search system. Shenma voice search is one of the most popular mobile search engines in China, and it is a streaming service that intermediate recognition results displayed while users are still speaking. Unidirectional LSTM network is applied, rather than bidirectional one, because it is well suited to real-time streaming speech recognition.
Distillation
It is demonstrated that deep neural network architecture can achieve improvement in LVCSR. However, it also leads to much more computation and higher RTF, so that the recognition result can not be displayed in real time. It should be noted that deeper neural network contains more knowledge, but it is also cumbersome. the knowledge is key to improve the performance. If it can be transfered from cumbersome model to a small model, the recognition ability can also be transfered to the small model. Knowledge transferring to small model is called distillation BIBREF19 . The small model can perform as well as cumbersome model, after distilling. It provide a way to utilize high-performance but high RTF model in real time system. The class probability produced by the cumbersome model is regarded as soft target, and the generalization ability of cumbersome model is transfered to small model with it. Distillation is model's knowledge transferring approach, so there is no need to use the hard target, which is different with the layer-wise training method.
9-layers unidirectional LSTM model achieves outstanding performance, but meanwhile it is too computationally expensive to allow deployment in real time recognition system. In order to ensure real-time of the system, the number of layers needs to be reduced. The shallower network can learn the knowledge of deeper network with distillation. It is found that RTF of 2-layers network is acceptable, so the knowledge is distilled from 9-layers well-trained model to 2-layers model. Table TABREF16 shows that distillation from 9-layers to 2-layers brings RTF decrease of relative 53%, while CER only increases 5%. The knowledge of deep network is almost transfered with distillation, Distillation brings promising RTF reduction, but only little knowledge of deep network is lost. Moreover, CER of 2-layers distilled LSTM decreases relative 14%, compared with 2-layers regular-trained LSTM.
Transfer Learning with sMBR
For a certain specific scenario, the model trained with the data recorded from it has better adaptation than the model trained with generic scenario. But it spends too much time training a model from the beginning, if there is a well-trained model for generic scenarios. Moreover, labeling a large quantity of training data in new scenario is both costly and time consuming. If a model transfer trained with smaller dataset can obtained the similar recognition accuracy compared with the model directly trained with larger dataset, it is no doubt that transfer learning is more practical. Since specific scenario is a subset of generic scenario, some knowledge can be shared between them. Besides, generic scenario consists of various conditions, so its model has greater robustness. As a result, not only shared knowledge but also robustness can be transfered from the model of generic scenario to the model of specific one.
As the model well trained from generic scenario achieves good performance in frame level classification, sequence discriminative training is required to adapt new model to specific scenario additionally. Moreover, it does not need alignment from HMM-GMM system, and it also saves amount of time to prepare alignment.
Training Data
A large quantity of labeled data is needed for training a more accurate acoustic model. We collect the 17000 hours labeled data from Shenma voice search, which is one of the most popular mobile search engines in China. The dataset is created from anonymous online users' search queries in Mandarin, and all audio file's sampling rate is 16kHz, recorded by mobile phones. This dataset consists of many different conditions, such as diverse noise even low signal-to-noise, babble, dialects, accents, hesitation and so on.
In the Amap, which is one of the most popular web mapping and navigation services in China, users can search locations and navigate to locations they want though voice search. To present the performance of transfer learning with sequence discriminative training, the model trained from Shenma voice search which is greneric scenario transfer its knowledge to the model of Amap voice search. 7300 hours labeled data is collected in the similar way of Shenma voice search data collection.
Two dataset is divided into training set, validation set and test set separately, and the quantity of them is shown in Table TABREF10 . The three sets are split according to speakers, in order to avoid utterances of same speaker appearing in three sets simultaneously. The test sets of Shenma and Amap voice search are called Shenma Test and Amap Test.
Experimental setup
LSTM RNNs outperform conventional RNNs for speech recognition system, especially deep LSTM RNNs, because of its long-range dependencies more accurately for temporal sequence conditions BIBREF26 , BIBREF23 . Shenma and Amap voice search is a streaming service that intermediate recognition results displayed while users are still speaking. So as for online recognition in real time, we prefer unidirectional LSTM model rather than bidirectional one. Thus, the training system is unidirectional LSTM-based.
A 26-dimensional filter bank and 2-dimensional pitch feature is extracted for each frame, and is concatenated with first and second order difference as the final input of the network. The super frame are stacked by 3 frames without overlapping. The architecture we trained consists of two LSTM layers with sigmoid activation function, followed by a full-connection layer. The out layer is a softmax layer with 11088 hidden markov model (HMM) tied-states as output classes, the loss function is cross-entropy (CE). The performance metric of the system in Mandarin is reported with character error rate (CER). The alignment of frame-level ground truth is obtained by GMM-HMM system. Mini-batched SGD is utilized with momentum trick and the network is trained for a total of 4 epochs. The block learning rate and block momentum of BMUF are set as 1 and 0.9. 5-gram language model is leveraged in decoder, and the vocabulary size is as large as 760000. Differentials of recurrent layers is limited to range [-10000,10000], while gradients are clipped to range [-5, 5] and cell activations clipped to range [-50, 50]. After training with CE loss, sMBR loss is employed to further improve the performance.
It has shown that BMUF outperforms traditional model averaging method, and it is utilized at the synchronization phase. After synchronizing with BMUF, EMA method further updates the model in non-interference way. The training system is deployed on the MPI-based HPC cluster where 8 GPUs. Each GPU processes non-overlap subset split from the entire large scale dataset in parallel.
Local models from distributed workers synchronize with each other in decentralized way. In the traditional model averaging and BMUF method, a parameter server waits for all workers to send their local models, aggregate them, and send the updated model to all workers. Computing resource of workers is wasted until aggregation of the parameter server done. Decentralized method makes full use of computing resource, and we employ the MPI-based Mesh AllReduce method. It is mesh topology as shown in Figure FIGREF12 . There is no centralized parameter server, and peer to peer communication is used to transmit local models between workers. Local model INLINEFORM0 of INLINEFORM1 -th worker in INLINEFORM2 workers cluster is split to INLINEFORM3 pieces INLINEFORM4 , and send to corresponding worker. In the aggregation phase, INLINEFORM5 -th worker computed INLINEFORM6 splits of model INLINEFORM7 and send updated model INLINEFORM8 back to workers. As a result, all workers participate in aggregation and no computing resource is dissipated. It is significant to promote training efficiency, when the size of neural network model is too large. The EMA model is also updated additionally, but not broadcasting it.
Results
In order to evaluate our system, several sets of experiments are performed. The Shenma test set including about 9000 samples and Amap test set including about 7000 samples contain various real world conditions. It simulates the majority of user scenarios, and can well evaluates the performance of a trained model. Firstly, we show the results of models trained with EMA method. Secondly, for real world applications, very deep LSTM is distilled to a shallow one, so as for lower RTF. The model of Amap is also needed to train for map and navigation scenarios. The performance of transfer learning from Shenma voice search to Amap voice search is also presented.
Layer-wise Training
In layer-wise training, the deeper model learns both parameters and knowledge from the shallower model. The deeper model is initialized by the shallower one, and its alignment is the combination of hard target and soft target of shallower one. Two targets have the same weights in our framework. The teacher model is trained with CE. For each layer-wise trained CE model, corresponding sMBR model is also trained, as sMBR could achieve additional improvement. In our framework, 1000 hours data is randomly selected from the total dataset for sMBR training. There is no obvious performance enhancement when the size of sMBR training dataset increases.
For very deep unidirectional LSTM initialized with Xavier initialization algorithm, 6-layers model converges well, but there is no further improvement with increasing the number of layers. Therefore, the first 6 layers of 7-layers model is initialized by 6-layers model, and soft target is provided by 6-layers model. Consequently, deeper LSTM is also trained in the same way. It should be noticed that the teacher model of 9-layers model is the 8-layers model trained by sMBR, while the other teacher model is CE model. As shown in Table TABREF15 , the layer-wise trained models always performs better than the models with Xavier initialization, as the model is deep. Therefore, for the last layer training, we choose 8-layers sMBR model as the teacher model instead of CE model. A comparison between 6-layers and 9-layers sMBR models shows that 3 additional layers of layer-wise training brings relative 12.6% decreasing of CER. It is also significant that the averaged CER of sMBR models with different layers decreases absolute 0.73% approximately compared with CE models, so the improvement of sequence discriminative learning is promising.
Transfer Learning
2-layers distilled model of Shenma voice search has shown a impressive performance on Shenma Test, and we call it Shenma model. It is trained for generic search scenario, but it has less adaptation for specific scenario like Amap voice search. Training with very large dataset using CE loss is regarded as improvement of frame level recognition accuracy, and sMBR with less dataset further improves accuracy as sequence discriminative training. If robust model of generic scenario is trained, there is no need to train a model with very large dataset, and sequence discriminative training with less dataset is enough. Therefore, on the basis of Shenma model, it is sufficient to train a new Amap model with small dataset using sMBR. As shown in Table TABREF18 , Shenma model presents the worst performance among three methods, since it does not trained for Amap scenario. 2-layers Shenma model further trained with sMBR achieves about 8.1% relative reduction, compared with 2-layers regular-trained Amap model. Both training sMBR datasets contain the same 1000 hours data. As a result, with the Shenma model, only about 14% data usage achieves lower CER, and it leads to great time and cost saving with less labeled data. Besides, transfer learning with sMBR does not use the alignment from the HMM-GMM system, so it also saves huge amount of time.
Conclusion
We have presented a whole deep unidirectional LSTM parallel training system for LVCSR. The recognition performance improves when the network goes deep. Distillation makes it possible that deep LSTM model transfer its knowledge to shallow model with little loss. The model could be distilled to 2-layers model with very low RTF, so that it can display the immediate recognition results. As a result, its CER decrease relatively 14%, compared with the 2-layers regular trained model. In addition, transfer learning with sMBR is also proposed. If a great model has well trained from generic scenario, only 14% of the size of training dataset is needed to train a more accuracy acoustic model for specific scenario. Our future work includes 1) finding more effective methods to reduce the CER by increasing the number of layers; 2) applying this training framework to Connectionist Temporal Classification (CTC) and attention-based neural networks.
| [
"Unidirectional LSTM networks with 2, 6, 7, 8, and 9 layers."
] | 4,286 | qasper | en | null | 3598040772b4950e6241b50125f7ab7038a8149060e3c381 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: which lstm models did they compare with?
Answer: | [
"Unidirectional LSTM networks with 2, 6, 7, 8, and 9 layers."
] | qasper | 128 | 38 | |
How large is their data set? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The task of document quality assessment is to automatically assess a document according to some predefined inventory of quality labels. This can take many forms, including essay scoring (quality = language quality, coherence, and relevance to a topic), job application filtering (quality = suitability for role + visual/presentational quality of the application), or answer selection in community question answering (quality = actionability + relevance of the answer to the question). In the case of this paper, we focus on document quality assessment in two contexts: Wikipedia document quality classification, and whether a paper submitted to a conference was accepted or not.
Automatic quality assessment has obvious benefits in terms of time savings and tractability in contexts where the volume of documents is large. In the case of dynamic documents (possibly with multiple authors), such as in the case of Wikipedia, it is particularly pertinent, as any edit potentially has implications for the quality label of that document (and around 10 English Wikipedia documents are edited per second). Furthermore, when the quality assessment task is decentralized (as in the case of Wikipedia and academic paper assessment), quality criteria are often applied inconsistently by different people, where an automatic document quality assessment system could potentially reduce inconsistencies and enable immediate author feedback.
Current studies on document quality assessment mainly focus on textual features. For example, BIBREF0 examine features such as the article length and the number of headings to predict the quality class of a Wikipedia article. In contrast to these studies, in this paper, we propose to combine text features with visual features, based on a visual rendering of the document. Figure 1 illustrates our intuition, relative to Wikipedia articles. Without being able to read the text, we can tell that the article in Figure 1 has higher quality than Figure 1 , as it has a detailed infobox, extensive references, and a variety of images. Based on this intuition, we aim to answer the following question: can we achieve better accuracy on document quality assessment by complementing textual features with visual features?
Our visual model is based on fine-tuning an Inception V3 model BIBREF1 over visual renderings of documents, while our textual model is based on a hierarchical biLSTM. We further combine the two into a joint model. We perform experiments on two datasets: a Wikipedia dataset novel to this paper, and an arXiv dataset provided by BIBREF2 split into three sub-parts based on subject category. Experimental results on the visual renderings of documents show that implicit quality indicators, such as images and visual layout, can be captured by an image classifier, at a level comparable to a text classifier. When we combine the two models, we achieve state-of-the-art results over 3/4 of our datasets.
This paper makes the following contributions:
All code and data associated with this research will be released on publication.
Related Work
A variety of approaches have been proposed for document quality assessment across different domains: Wikipedia article quality assessment, academic paper rating, content quality assessment in community question answering (cQA), and essay scoring. Among these approaches, some use hand-crafted features while others use neural networks to learn features from documents. For each domain, we first briefly describe feature-based approaches and then review neural network-based approaches. Wikipedia article quality assessment: Quality assessment of Wikipedia articles is a task that assigns a quality class label to a given Wikipedia article, mirroring the quality assessment process that the Wikipedia community carries out manually. Many approaches have been proposed that use features from the article itself, meta-data features (e.g., the editors, and Wikipedia article revision history), or a combination of the two. Article-internal features capture information such as whether an article is properly organized, with supporting evidence, and with appropriate terminology. For example, BIBREF3 use writing styles represented by binarized character trigram features to identify featured articles. BIBREF4 and BIBREF0 explore the number of headings, images, and references in the article. BIBREF5 use nine readability scores, such as the percentage of difficult words in the document, to measure the quality of the article. Meta-data features, which are indirect indicators of article quality, are usually extracted from revision history, and the interaction between editors and articles. For example, one heuristic that has been proposed is that higher-quality articles have more edits BIBREF6 , BIBREF7 . BIBREF8 use the percentage of registered editors and the total number of editors of an article. Article–editor dependencies have also been explored. For example, BIBREF9 use the authority of editors to measure the quality of Wikipedia articles, where the authority of editors is determined by the articles they edit. Deep learning approaches to predicting Wikipedia article quality have also been proposed. For example, BIBREF10 use a version of doc2vec BIBREF11 to represent articles, and feed the document embeddings into a four hidden layer neural network. BIBREF12 first obtain sentence representations by averaging words within a sentence, and then apply a biLSTM BIBREF13 to learn a document-level representation, which is combined with hand-crafted features as side information. BIBREF14 exploit two stacked biLSTMs to learn document representations.
Academic paper rating: Academic paper rating is a relatively new task in NLP/AI, with the basic formulation being to automatically predict whether to accept or reject a paper. BIBREF2 explore hand-crafted features, such as the length of the title, whether specific words (such as outperform, state-of-the-art, and novel) appear in the abstract, and an embedded representation of the abstract as input to different downstream learners, such as logistic regression, decision tree, and random forest. BIBREF15 exploit a modularized hierarchical convolutional neural network (CNN), where each paper section is treated as a module. For each paper section, they train an attention-based CNN, and an attentive pooling layer is applied to the concatenated representation of each section, which is then fed into a softmax layer.
Content quality assessment in cQA: Automatic quality assessment in cQA is the task of determining whether an answer is of high quality, selected as the best answer, or ranked higher than other answers. To measure answer content quality in cQA, researchers have exploited various features from different sources, such as the answer content itself, the answerer's profile, interactions among users, and usage of the content. The most common feature used is the answer length BIBREF16 , BIBREF17 , with other features including: syntactic and semantic features, such as readability scores. BIBREF18 ; similarity between the question and the answer at lexical, syntactic, and semantic levels BIBREF18 , BIBREF19 , BIBREF20 ; or user data (e.g., a user's status points or the number of answers written by the user). There have also been approaches using neural networks. For example, BIBREF21 combine CNN-learned representations with hand-crafted features to predict answer quality. BIBREF22 use a 2-dimensional CNN to learn the semantic relevance of an answer to the question, and apply an LSTM to the answer sequence to model thread context. BIBREF23 and BIBREF24 model the problem similarly to machine translation quality estimation, treating answers as competing translation hypotheses and the question as the reference translation, and apply neural machine translation to the problem. Essay scoring: Automated essay scoring is the task of assigning a score to an essay, usually in the context of assessing the language ability of a language learner. The quality of an essay is affected by the following four primary dimensions: topic relevance, organization and coherence, word usage and sentence complexity, and grammar and mechanics. To measure whether an essay is relevant to its “prompt” (the description of the essay topic), lexical and semantic overlap is commonly used BIBREF25 , BIBREF26 . BIBREF27 explore word features, such as the number of verb formation errors, average word frequency, and average word length, to measure word usage and lexical complexity. BIBREF28 use sentence structure features to measure sentence variety. The effects of grammatical and mechanic errors on the quality of an essay are measured via word and part-of-speech $n$ -gram features and “mechanics” features BIBREF29 (e.g., spelling, capitalization, and punctuation), respectively. BIBREF30 , BIBREF31 , and BIBREF32 use an LSTM to obtain an essay representation, which is used as the basis for classification. Similarly, BIBREF33 utilize a CNN to obtain sentence representation and an LSTM to obtain essay representation, with an attention layer at both the sentence and essay levels.
The Proposed Joint Model
We treat document quality assessment as a classification problem, i.e., given a document, we predict its quality class (e.g., whether an academic paper should be accepted or rejected). The proposed model is a joint model that integrates visual features learned through Inception V3 with textual features learned through a biLSTM. In this section, we present the details of the visual and textual embeddings, and finally describe how we combine the two. We return to discuss hyper-parameter settings and the experimental configuration in the Experiments section.
Visual Embedding Learning
A wide range of models have been proposed to tackle the image classification task, such as VGG BIBREF34 , ResNet BIBREF35 , Inception V3 BIBREF1 , and Xception BIBREF36 . However, to the best of our knowledge, there is no existing work that has proposed to use visual renderings of documents to assess document quality. In this paper, we use Inception V3 pretrained on ImageNet (“Inception” hereafter) to obtain visual embeddings of documents, noting that any image classifier could be applied to our task. The input to Inception is a visual rendering (screenshot) of a document, and the output is a visual embedding, which we will later integrate with our textual embedding.
Based on the observation that it is difficult to decide what types of convolution to apply to each layer (such as 3 $\times $ 3 or 5 $\times $ 5), the basic Inception model applies multiple convolution filters in parallel and concatenates the resulting features, which are fed into the next layer. This has the benefit of capturing both local features through smaller convolutions and abstracted features through larger convolutions. Inception is a hybrid of multiple Inception models of different architectures. To reduce computational cost, Inception also modifies the basic model by applying a 1 $\times $ 1 convolution to the input and factorizing larger convolutions into smaller ones.
Textual Embedding Learning
We adopt a bi-directional LSTM model to generate textual embeddings for document quality assessment, following the method of BIBREF12 (“biLSTM” hereafter). The input to biLSTM is a textual document, and the output is a textual embedding, which will later integrate with the visual embedding.
For biLSTM, each word is represented as a word embedding BIBREF37 , and an average-pooling layer is applied to the word embeddings to obtain the sentence embedding, which is fed into a bi-directional LSTM to generate the document embedding from the sentence embeddings. Then a max-pooling layer is applied to select the most salient features from the component sentences.
The Joint Model
The proposed joint model (“Joint” hereafter) combines the visual and textual embeddings (output of Inception and biLSTM) via a simple feed-forward layer and softmax over the document label set, as shown in Figure 2 . We optimize our model based on cross-entropy loss.
Experiments
In this section, we first describe the two datasets used in our experiments: (1) Wikipedia, and (2) arXiv. Then, we report the experimental details and results.
Datasets
The Wikipedia dataset consists of articles from English Wikipedia, with quality class labels assigned by the Wikipedia community. Wikipedia articles are labelled with one of six quality classes, in descending order of quality: Featured Article (“FA”), Good Article (“GA”), B-class Article (“B”), C-class Article (“C”), Start Article (“Start”), and Stub Article (“Stub”). A description of the criteria associated with the different classes can be found in the Wikipedia grading scheme page. The quality class of a Wikipedia article is assigned by Wikipedia reviewers or any registered user, who can discuss through the article's talk page to reach consensus. We constructed the dataset by first crawling all articles from each quality class repository, e.g., we get FA articles by crawling pages from the FA repository: https://en.wikipedia.org/wiki/Category:Featured_articles. This resulted in around 5K FA, 28K GA, 212K B, 533K C, 2.6M Start, and 3.2M Stub articles.
We randomly sampled 5,000 articles from each quality class and removed all redirect pages, resulting in a dataset of 29,794 articles. As the wikitext contained in each document contains markup relating to the document category such as {Featured Article} or {geo-stub}, which reveals the label, we remove such information. We additionally randomly partitioned this dataset into training, development, and test splits based on a ratio of 8:1:1. Details of the dataset are summarized in Table 1 .
We generate a visual representation of each document via a 1,000 $\times $ 2,000-pixel screenshot of the article via a PhantomJS script over the rendered version of the article, ensuring that the screenshot and wikitext versions of the article are the same version. Any direct indicators of document quality (such as the FA indicator, which is a bronze star icon in the top right corner of the webpage) are removed from the screenshot.
The arXiv dataset BIBREF2 consists of three subsets of academic articles under the arXiv repository of Computer Science (cs), from the three subject areas of: Artificial Intelligence (cs.ai), Computation and Language (cs.cl), and Machine Learning (cs.lg). In line with the original dataset formulation BIBREF2 , a paper is considered to have been accepted (i.e. is positively labeled) if it matches a paper in the DBLP database or is otherwise accepted by any of the following conferences: ACL, EMNLP, NAACL, EACL, TACL, NIPS, ICML, ICLR, or AAAI. Failing this, it is considered to be rejected (noting that some of the papers may not have been submitted to one of these conferences). The median numbers of pages for papers in cs.ai, cs.cl, and cs.lg are 11, 10, and 12, respectively. To make sure each page in the PDF file has the same size in the screenshot, we crop the PDF file of a paper to the first 12; we pad the PDF file with blank pages if a PDF file has less than 12 pages, using the PyPDF2 Python package. We then use ImageMagick to convert the 12-page PDF file to a single 1,000 $\times $ 2,000 pixel screenshot. Table 2 details this dataset, where the “Accepted” column denotes the percentage of positive instances (accepted papers) in each subset.
Experimental Setting
As discussed above, our model has two main components — biLSTM and Inception— which generate textual and visual representations, respectively. For the biLSTM component, the documents are preprocessed as described in BIBREF12 , where an article is divided into sentences and tokenized using NLTK BIBREF38 . Words appearing more than 20 times are retained when building the vocabulary. All other words are replaced by the special UNK token. We use the pre-trained GloVe BIBREF39 50-dimensional word embeddings to represent words. For words not in GloVe, word embeddings are randomly initialized based on sampling from a uniform distribution $U(-1, 1)$ . All word embeddings are updated in the training process. We set the LSTM hidden layer size to 256. The concatenation of the forward and backward LSTMs thus gives us 512 dimensions for the document embedding. A dropout layer is applied at the sentence and document level, respectively, with a probability of 0.5.
For Inception, we adopt data augmentation techniques in the training with a “nearest” filling mode, a zoom range of 0.1, a width shift range of 0.1, and a height shift range of 0.1. As the original screenshots have the size of 1,000 $\times 2$ ,000 pixels, they are resized to 500 $\times $ 500 to feed into Inception, where the input shape is (500, 500, 3). A dropout layer is applied with a probability of 0.5. Then, a GlobalAveragePooling2D layer is applied, which produces a 2,048 dimensional representation.
For the Joint model, we get a representation of 2,560 dimensions by concatenating the 512 dimensional representation from the biLSTM with the 2,048 dimensional representation from Inception. The dropout layer is applied to the two components with a probability of 0.5. For biLSTM, we use a mini-batch size of 128 and a learning rate of 0.001. For both Inception and joint model, we use a mini-batch size of 16 and a learning rate of 0.0001. All hyper-parameters were set empirically over the development data, and the models were optimized using the Adam optimizer BIBREF40 .
In the training phase, the weights in Inception are initialized by parameters pretrained on ImageNet, and the weights in biLSTM are randomly initialized (except for the word embeddings). We train each model for 50 epochs. However, to prevent overfitting, we adopt early stopping, where we stop training the model if the performance on the development set does not improve for 20 epochs. For evaluation, we use (micro-)accuracy, following previous studies BIBREF5 , BIBREF2 .
Baseline Approaches
We compare our models against the following five baselines:
Majority: the model labels all test samples with the majority class of the training data.
Benchmark: a benchmark method from the literature. In the case of Wikipedia, this is BIBREF5 , who use structural features and readability scores as features to build a random forest classifier; for arXiv, this is BIBREF2 , who use hand-crafted features, such as the number of references and TF-IDF weighted bag-of-words in abstract, to build a classifier based on the best of logistic regression, multi-layer perception, and AdaBoost.
Doc2Vec: doc2vec BIBREF11 to learn document embeddings with a dimension of 500, and a 4-layer feed-forward classification model on top of this, with 2000, 1000, 500, and 200 dimensions, respectively.
biLSTM: first derive a sentence representation by averaging across words in a sentence, then feed the sentence representation into a biLSTM and a maxpooling layer over output sequence to learn a document level representation with a dimension of 512, which is used to predict document quality.
Inception $_{\text{fixed}}$ : the frozen Inception model, where only parameters in the last layer are fine-tuned during training.
The hyper-parameters of Benchmark, Doc2Vec, and biLSTM are based on the corresponding papers except that: (1) we fine-tune the feed forward layer of Doc2Vec on the development set and train the model 300 epochs on Wikipedia and 50 epochs on arXiv; (2) we do not use hand-crafted features for biLSTM as we want the baselines to be comparable to our models, and the main focus of this paper is not to explore the effects of hand-crafted features (e.g., see BIBREF12 ).
Experimental Results
Table 3 shows the performance of the different models over our two datasets, in the form of the average accuracy on the test set (along with the standard deviation) over 10 runs, with different random initializations.
On Wikipedia, we observe that the performance of biLSTM, Inception, and Joint is much better than that of all four baselines. Inception achieves 2.9% higher accuracy than biLSTM. The performance of Joint achieves an accuracy of 59.4%, which is 5.3% higher than using textual features alone (biLSTM) and 2.4% higher than using visual features alone (Inception). Based on a one-tailed Wilcoxon signed-rank test, the performance of Joint is statistically significant ( $p<0.05$ ). This shows that the textual and visual features complement each other, achieving state-of-the-art results in combination.
For arXiv, baseline methods Majority, Benchmark, and Inception $_{\text{fixed}}$ outperform biLSTM over cs.ai, in large part because of the class imbalance in this dataset (90% of papers are rejected). Surprisingly, Inception $_{\text{fixed}}$ is better than Majority and Benchmark over the arXiv cs.lg subset, which verifies the usefulness of visual features, even when only the last layer is fine-tuned. Table 3 also shows that Inception and biLSTM achieve similar performance on arXiv, showing that textual and visual representations are equally discriminative: Inception and biLSTM are indistinguishable over cs.cl; biLSTM achieves 1.8% higher accuracy over cs.lg, while Inception achieves 1.3% higher accuracy over cs.ai. Once again, the Joint model achieves the highest accuracy on cs.ai and cs.cl by combining textual and visual representations (at a level of statistical significance for cs.ai). This, again, confirms that textual and visual features complement each other, and together they achieve state-of-the-art results. On arXiv cs.lg, Joint achieves a 0.6% higher accuracy than Inception by combining visual features and textual features, but biLSTM achieves the highest accuracy. One characteristic of cs.lg documents is that they tend to contain more equations than the other two arXiv datasets, and preliminary analysis suggests that the biLSTM is picking up on a correlation between the volume/style of mathematical presentation and the quality of the document.
Analysis
In this section, we first analyze the performance of Inception and Joint. We also analyze the performance of different models on different quality classes. The high-level representations learned by different models are also visualized and discussed. As the Wikipedia test set is larger and more balanced than that of arXiv, our analysis will focus on Wikipedia.
Inception
To better understand the performance of Inception, we generated the gradient-based class activation map BIBREF41 , by maximizing the outputs of each class in the penultimate layer, as shown in Figure 3 . From Figure 3 and Figure 3 , we can see that Inception identifies the two most important regions (one at the top corresponding to the table of contents, and the other at the bottom, capturing both document length and references) that contribute to the FA class prediction, and a region in the upper half of the image that contributes to the GA class prediction (capturing the length of the article body). From Figure 3 and Figure 3 , we can see that the most important regions in terms of B and C class prediction capture images (down the left and right of the page, in the case of B and C), and document length/references. From Figure 3 and Figure 3 , we can see that Inception finds that images in the top right corner are the strongest predictor of Start class prediction, and (the lack of) images/the link bar down the left side of the document are the most important for Stub class prediction.
Joint
Table 4 shows the confusion matrix of Joint on Wikipedia. We can see that more than 50% of documents for each quality class are correctly classified, except for the C class where more documents are misclassified into B. Analysis shows that when misclassified, documents are usually misclassified into adjacent quality classes, which can be explained by the Wikipedia grading scheme, where the criteria for adjacent quality classes are more similar.
We also provide a breakdown of precision (“ $\mathcal {P}$ ”), recall (“ $\mathcal {R}$ ”), and F1 score (“ $\mathcal {F}_{\beta =1}$ ”) for biLSTM, Inception, and Joint across the quality classes in Table 5 . We can see that Joint achieves the highest accuracy in 11 out of 18 cases. It is also worth noting that all models achieve higher scores for FA, GA, and Stub articles than B, C and Start articles. This can be explained in part by the fact that FA and GA articles must pass an official review based on structured criteria, and in part by the fact that Stub articles are usually very short, which is discriminative for Inception, and Joint. All models perform worst on the B and C quality classes. It is difficult to differentiate B articles from C articles even for Wikipedia contributors. As evidence of this, when we crawled a new dataset including talk pages with quality class votes from Wikipedia contributors, we found that among articles with three or more quality labels, over 20% percent of B and C articles have inconsistent votes from Wikipedia contributors, whereas for FA and GA articles the number is only 0.7%.
We further visualize the learned document representations of biLSTM, Inception, and Joint in the form of a t-SNE plot BIBREF42 in Figure 4 . The degree of separation between Start and Stub achieved by Inception is much greater than for biLSTM, with the separation between Start and Stub achieved by Joint being the clearest among the three models. Inception and Joint are better than biLSTM at separating Start and C. Joint achieves slightly better performance than Inception in separating GA and FA. We can also see that it is difficult for all models to separate B and C, which is consistent with the findings of Tables 4 and 5 .
Conclusions
We proposed to use visual renderings of documents to capture implicit document quality indicators, such as font choices, images, and visual layout, which are not captured in textual content. We applied neural network models to capture visual features given visual renderings of documents. Experimental results show that we achieve a 2.9% higher accuracy than state-of-the-art approaches based on textual features over Wikipedia, and performance competitive with or surpassing state-of-the-art approaches over arXiv. We further proposed a joint model, combining textual and visual representations, to predict the quality of a document. Experimental results show that our joint model outperforms the visual-only model in all cases, and the text-only model on Wikipedia and two subsets of arXiv. These results underline the feasibility of assessing document quality via visual features, and the complementarity of visual and textual document representations for quality assessment.
| [
"a sample of 29,794 wikipedia articles and 2,794 arXiv papers "
] | 4,187 | qasper | en | null | ecf37304ab84334d46491b0bd3d8bdc7cda33d54a15870f0 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How large is their data set?
Answer: | [
"a sample of 29,794 wikipedia articles and 2,794 arXiv papers "
] | qasper | 128 | 39 | |
How were the human judgements assembled? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The use of RNNs in the field of Statistical Machine Translation (SMT) has revolutionised the approaches to automated translation. As opposed to traditional shallow SMT models, which require a lot of memory to run, these neural translation models require only a small fraction of memory used, about 5% BIBREF0 . Also, neural translation models are optimized such that every module is trained to jointly improve translation quality. With that being said, one of the main downsides of neural translation models is the heavy corpus requirement in order to ensure learning of deeper contexts. This is where the application of these encoder decoder architectures in translation to and/or from morphologically rich languages takes a severe hit.
For any language pair, the efficiency of an MT system depends on two major factors: the availability and size of parallel corpus used for training and the syntactic divergence between the two languages i.e morphological richness, word order differences, grammatical structure etc. BIBREF0 . The main differences between the languages stem from the fact that languages similar to English are predominantly fusional languages whereas many of the morphologically rich languages are agglutinative in nature. The nature of morphologically rich languages being structurally and semantically discordant from languages like English adds to the difficulty of SMT involving such languages.
In morphologically rich languages, any suffix can be added to any verb or noun to simply mean one specific thing about that particular word that the suffix commonly represents (agglutination). This means that there exists a lot of inflectional forms of the same noun and verb base words, conveying similar notions. For example, in Tamil, there are at least 30,000 inflectional forms of any given verb and about 5,000 forms of inflectional forms for any noun. The merged words carry information about part of speech (POS) tags, tense, plurality and so forth that are important for analyzing text for Machine Translation (MT). Not only are these hidden meanings not captured, the corresponding root words are trained as different units, thereby increasing the complexity of developing such MT systems BIBREF1 .
To add to the complexities of being a morphologically rich language, there are several factors unique to Tamil that make translation very difficult. The availability of parallel corpus for Tamil is very scarce. Most of the other models in the field of English–Tamil MT have made use of their own translation corpora that were manually created for the purposes of research. Most of these corpora are not available online for use.
Another issue specific to Tamil is the addition of suffix characters included to the words in the language for smoothness in pronunciation. These characters are of so many different types; there is a unique suffix for each and every consonant in the language. These suffixes degrade performance of MT because the same words with different such pronounciation-based suffixes will be taken as different words in training.
Also to take into consideration is the existence of two different forms of the language being used. Traditionally defined Tamil and its pronunciations aren't acoustically pleasing to use. There's no linguistic flow between syllables and its usage in verbal communication is time consuming. Therefore, there exists two forms of the language, the written form, rigid in structure and syntax, and the spoken form, in which the flow and pace of the language is given priority over syntax and correctness of spelling. This divide leads to the corpus having 2 different versions of the language that increase the vocabulary even with the same words. This can be evidently seen in the corpus between the sentences used in the Bible, which is in traditional Tamil and sentences from movie subtitles, being in spoken Tamil format.
To account for such difficulties, a trade-off between domain specificity and size of the corpus is integral in building an English–Tamil neural MT system.
Corpus
The corpus selected for this experiment was a combination of different corpora from various domains. The major part of the corpus was made up by the EnTam v2 corpus BIBREF2 . This corpus contained sentences taken from parallel news articles, English and Tamil bible corpus and movie subtitles. It also comprised of a tourism corpus that was obtained from TDIL (Technology Development for Indian Languages) and a corpus created from Tamil novels and short stories from AU-KBC, Anna university. The complete corpus consisted of 197,792 sentences. Fig. FIGREF20 shows the skinny shift and heatmap representations of the relativity between the sentences in terms of their sentence lengths.
An extra monolingual Tamil corpus, collated from various online sources was used for the word2vec embedding of the Tamil target language to enhance the richness of context of the word vectors. It was also used to create the language model for the phrase-based SMT model. This corpus contained 567,772 sentences and was self-collected by combining hundreds of ancient Tamil scriptures, novels and poems by accessing the websites of popular online ebook libraries in Python using the urllib package. Since the sources had Tamil text in different encodings, the encoding scheme was standardized to be UTF-8 for the entirety of the monolingual and parallel corpora using the chardet package. The corpora were cleaned for any stray special characters, unnecessary html tags and website URLs.
Word2Vec
The word embeddings of the source and target language sentences are used as initial vectors of the model to improve contextualization. The skip gram model of the word2vec algorithm optimizes the vectors by accounting for the average log probability of context words given a source word. DISPLAYFORM0
where k is the context window taken for the vectorization, INLINEFORM0 refers to the INLINEFORM1 word of the corpus and INLINEFORM2 is the size of the training corpus in terms of the number of words. Here, the probabily INLINEFORM3 is computed as a hierarchical softmax of the product of the transpose of the output vector of INLINEFORM4 and the input vector of INLINEFORM5 for each and every pair over the entire vocabulary. The processes of negative sampling and subsampling of frequent words that were used in the original model aren't used in this experiment BIBREF3 .
For the process of creating semantically meaningful word embeddings, a monolingual corpus of 569,772 Tamil sentences was used. This gave the vectors more contextual richness due to the increased size of the corpus as opposed to using just the bilingual corpus' target side sentences BIBREF3 .
In the experiment, the word2vec model was trained using a vector size of 100 to ensure that the bulk of the limited memory of the GPU will be used for the neural attention translation model. It has been shown that any size over that of 150 used for word vectorization gives similar results and that a size of 100 performs close to the model with 150-sized word vectors BIBREF7 . A standard size of 5 was used as window size and the model was trained over 7 worker threads simultaneously. A batch size of 50 words was used for training. The negative sampling was set at 1 as it is the nature of morphologically rich languages to have a lot of important words that don't occur more than once in the corpus. The gensim word2vec toolkit was used to implement this word embedding process BIBREF8 .
Neural Translation Model
The model used for translation is the one implemented by Bahdanau et al. Bahdanau2014. A bidirectional LSTM encoder first takes the source sentence and encodes it into a context vector which acts as input for the decoder. The decoder is attention-based where the hidden states of the decoder get as input the weighted sum of all the hidden layer outputs of the encoder alongwith the output of the previous hidden layer and the previously decoded word. This provides a contextual reference into the source language sentence BIBREF4 .
Neural Machine Translation models directly compute the probability of the target language sentence given the source language sentence, word by word for every time step. The model with a basic decoder without the attention module computes the log probability of target sentence given source sentence as the sum of log probabilities of every word given every word before that. The attention-based model, on the other hand, calculates: DISPLAYFORM0
where INLINEFORM0 is the number of words in the target sentence, INLINEFORM1 is the target sentence, INLINEFORM2 is the source sentence, INLINEFORM3 is the fixed length output vector of the encoder and INLINEFORM4 is the weighted sum of all the hidden layer outputs of the encoder at every time step. Both the encoder's output context vector and the weighted sum (known as attention vector) help to improve the quality of translation by enabling selective source sentence lookup.
The decoder LSTM computes: DISPLAYFORM0
where the probability is computed as a function of the decoder's output in the previous time step INLINEFORM0 , the hidden layer vector of the decoder in the current timestep INLINEFORM1 and the context vector from the attention mechanism INLINEFORM2 . The context vector INLINEFORM3 for time step INLINEFORM4 is computed as a weighted sum of the output of the entire sentence using a weight parameter INLINEFORM5 : DISPLAYFORM0
where INLINEFORM0 is the number of tokens in the source sentence, INLINEFORM1 refers to the value of the hidden layer of the encoder at time step INLINEFORM2 , and INLINEFORM3 is the alignment parameter. This parameter is calculated by means of a feed forward neural network to ensure that the alignment model is free from the difficulties of contextualization of long sentences into a single vector. The feed forward network is trained along with the neural translation model to jointly improve the performance of the translation. Mathematically, DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 is the softmax output of the result of the feedforward network, INLINEFORM1 is the hidden state value of the decoder at timestep INLINEFORM2 and INLINEFORM3 is the encoder's hidden layer annotation at timestep INLINEFORM4 . A concatenation of the forward and the reverse hidden layer parameters of the encoder is used at each step to compute the weights INLINEFORM5 for the attention mechanism. This is done to enable an overall context of the sentence, as opposed to a context of only all the previous words of the sentence for every word in consideration. Fig. FIGREF12 is the general architecture of the neural translation model without the Bidirectional LSTM encoder.
A global attention mechanism is preferred over local attention because the differences in the structures of the languages cannot be mapped efficiently to enable lookup into the right parts of the source sentence. Using local attention mechanism with a monotonic context lookup, where the region around INLINEFORM0 source word is looked up for the prediction of the INLINEFORM1 target word, is impractical because of the structural discordance between the English and Tamil sentences (see Figs. FIGREF37 and FIGREF44 ). The use of gaussian and other such distributions to facilitate local attention would also be inefficient because the existence of various forms of translations for the same source sentence involving morphological and structural variations that don't stay uniform through the entire corpus BIBREF5 .
The No Peepholes (NP) variant of the LSTM cell, formulated in Greff et al. greff2015lstm is used in this experiment as it proved to give the best results amongst all the variants of an LSTM cell. It is specified by means of a gated mechanism designed to ensure that the vanishing gradient problem is prevented. LSTM maintains its hidden layer in two components, the cell vector INLINEFORM0 and the actual hidden layer output vector INLINEFORM1 . The cell vector is ensured to never reach zero by means of a weighted sum of the previous layer's cell vector INLINEFORM2 regulated by the forget gate INLINEFORM3 and an activation of the weighted sum of the input INLINEFORM4 in the current timestep INLINEFORM5 and the previous timestep's hidden layer output vector INLINEFORM6 . The combination is similarly regulated by the input gate INLINEFORM7 . The hidden layer output is determined as an activation of the cell gate, regulated by the output gate INLINEFORM8 . The interplay between these two vectors ( INLINEFORM9 and INLINEFORM10 ) at every timestep ensures that the problem of vanishing gradients doesn't occur. The three gates are also formed as a sigmoid of the weighted sum of the previous hidden layer output INLINEFORM11 and the input in the current timestep INLINEFORM12 . The output generated out of the LSTM's hidden layer is specified as a weighted softmax over the hidden layer output INLINEFORM13 . The learnable parameters of an LSTM cell are all the weights INLINEFORM14 and the biases INLINEFORM15 . DISPLAYFORM0
The LSTM specified by equations 7 through 11 is the one used for the decoder of the model. The encoder uses a bidirectional RNN LSTM cell in which there are two hidden layer components INLINEFORM0 and INLINEFORM1 that contribute to the output INLINEFORM2 of each time step INLINEFORM3 . Both the components have their own sets of LSTM equations in such a way that INLINEFORM4 for every timestep is computed from the first timestep till the INLINEFORM5 token is reached and INLINEFORM6 is computed from the INLINEFORM7 timestep backwards until the first token is reached. All the five vectors of the two components are all exactly the same as the LSTM equations specified with one variation in the computation of the result. DISPLAYFORM0
Morphological Segmentation
The morphological segmentation used is a semi-supervised extension to the generative probabilistic model of maximizing the probability of a INLINEFORM0 prefix,root,postfix INLINEFORM1 recursive split up of words based on an exhaustive combination of all possible morphemes. The details of this model are specified and extensively studied in Kohonen et al. kohonen2010semi. The model parameters INLINEFORM2 include the morph type count, morph token count of training data, the morph strings and their counts. The model is trained by maximizing the Maximum A Posteriori (MAP) probability using Bayes' rule: DISPLAYFORM0
where INLINEFORM0 refers to every word in the training lexicon. The prior INLINEFORM1 is estimated using the Minimum Description Length(MDL) principle. The likelihood INLINEFORM2 is estimated as: DISPLAYFORM0
where INLINEFORM0 refers to the intermediate analyses and INLINEFORM1 refers to the INLINEFORM2 morpheme of word INLINEFORM3 .
An extension to the Viterbi algorithm is used for the decoding step based on exhaustive mapping of morphemes. To account for over-segmentation and under-segmentation issues associated with unsupervised morphological segmentation, extra parameters ( INLINEFORM0 ) and ( INLINEFORM1 ) are used with the cost function INLINEFORM2 DISPLAYFORM0
where INLINEFORM0 is the likelihood of the cost function, INLINEFORM1 describes the likelihood of contribution of the annotated dataset to the cost function and INLINEFORM2 is the likelihood of the labeled data. A decrease in the value of INLINEFORM3 will cause smaller segments and vice versa. INLINEFORM4 takes care of size discrepancies due to reduced availability of annotated corpus as compared to the training corpus BIBREF2 , BIBREF6 .
The Python extension to the morphological segmentation tool morfessor 2.0 was used for this experiment to perform the segmentation. The annotation data for Tamil language collated and released by Anoop Kunchukkutan in the Indic NLP Library was used as the semi-supervised input to the model BIBREF9 , BIBREF6 .
Experiment
The complexities of neural machine translation of morphologically rich languages were studied with respect to English to Tamil machine translation using the RNN LSTM Bi-directional encoder attention decoder architecture. To compare with a baseline system, a phrase based SMT system was implemented using the same corpus. The Factored SMT model with source-side preprocessing by Kumar et al. kumar2014improving was used as a reference for the translation between these language pairs. Also, an additional 569,772 monolingual Tamil sentences were used for the language model of the SMT system. The model used could be split up into various modules as expanded in Fig. FIGREF17 .
Bucketing
The input source and target language sentences used for training were taken and divided into bucketed pairs of sentences of a fixed number of sizes. This relationship was determined by examining the distribution of words in the corpus primarily to minimize the number of PAD tokens in the sentence. The heat map of the number of words in the English–Tamil sentence pairs of the corpus revealed that the distribution is centered around the 10–20 words region. Therefore, more buckets in that region were applied as there would be enough number of examples in each of these bucket pairs for the model to learn about the sentences in each and every bucket. The exact scheme used for the RNNSearch models is specified by Fig. FIGREF21 . The bucketing scheme for the RNNMorph model, involving morphs instead of words, was a simple shifted scheme of the one used in Fig. FIGREF21 , where every target sentence bucket count was increased uniformly by 5.
Model Details
Due to various computational constraints and lack of availability of comprehensive corpora, the vocabularies for English and Tamil languages for the RNNSearch model were restricted to 60,000 out of 67,768 and 150,000 out of 340,325 respectively. The vocabulary of the languages for the RNNMorph didn't have to be restricted and the actual number of words in the corpus i.e. 67,768 words for English and 41,906 words for Tamil could be accommodated into the training. Words not in the vocabulary from the test set input and output were replaced with the universal INLINEFORM0 UNK INLINEFORM1 token, symbolizing an unknown word. The LSTM hidden layer size, the training batch size, and the vocabulary sizes of the languages, together, acted as a bottleneck. The model was run on a 2GB NVIDIA GeForce GT 650M card with 384 cores and the memory allotment was constrained to the limits of the GPU. Therefore, after repeated experimentation, it was determined that with a batch size of 16, the maximum hidden layer size possible was 500, which was the size used. Attempts to reduce the batch size resulted in poor convergence, and so the parameters were set to center around the batch size of 16. The models used were of 4 layers of LSTM hidden units in the bidirectional encoder and attention decoder.
The model used a Stochastic Gradient Descent (SGD) optimization algorithm with a sampled softmax loss of 512 per sample to handle large vocabulary size of the target language BIBREF10 . The model was trained with a learning rate 1.0 and a decay of rate 0.5 enforced manually. Gradient clipping based on the global norm of 5.0 was carried out to prevent gradients exploding and going to unrecoverable values tending towards infinity. The model described is the one used in the Tensorflow BIBREF11 seq2seq library.
Results and Discussion
The BLEU metric parameters (modified 1-gram, 2-gram, 3-gram and 4-gram precision values) and human evaluation metrics of adequacy, fluency and relative ranking values were used to evaluate the performance of the models.
BLEU Evaluation
The BLEU scores obtained using the various models used in the experiment are tabulated in Table TABREF25 .
The BLEU metric computes the BLEU unigram, bigram, trigram and BLEU-4 modified precision values, each micro-averaged over the test set sentences BIBREF7 . It was observed, as expected, that the performance of the phrase-based SMT model was inferior to that of the RNNSearch model. The baseline RNNSearch system was further refined by using word2vec vectors to embed semantic understanding, as observed with the slight increase in the BLEU scores. Fig. FIGREF26 plots the BLEU scores as a line graph for visualization of the improvement in performance. Also, the 4-gram BLEU scores for the various models were plotted as a bar graph in Fig. FIGREF26
Due to the agglutinative and morphologically rich nature of the target language i.e. Tamil, the use of morphological segmentation to split the words into morphemes further improved the BLEU precision values in the RNNMorph model. One of the reasons for the large extent of increase in the BLEU score could be attributed to the overall increase in the number of word units per sentence. Since the BLEU score computes micro-average precision scores, an increase in both the numerator and denominator of the precision scores is apparent with an increase in the number of tokens due to morphological segmentation of the target language. Thus, the numeric extent of the increase of accuracy might not efficiently describe the improvement in performance of the translation.
Human Evaluation
To ensure that the increase in BLEU score correlated to actual increase in performance of translation, human evaluation metrics like adequacy, precision and ranking values (between RNNSearch and RNNMorph outputs) were estimated in Table TABREF30 . A group of 50 native people who were well-versed in both English and Tamil languages acted as annotators for the evaluation. A collection of samples of about 100 sentences were taken from the test set results for comparison. This set included a randomized selection of the translation results to ensure the objectivity of evaluation. Fluency and adequacy results for the RNNMorph results are tabulated. Adequacy rating was calculated on a 5-point scale of how much of the meaning is conveyed by the translation (All, Most, Much, Little, None). The fluency rating was calculated based on grammatical correctness on a 5-point scale of (Flawless, Good, Non-native, Disfluent, Incomprehensive). For the comparison process, the RNNMorph and the RNNSearch + Word2Vec models’ sentence level translations were individually ranked between each other, permitting the two translations to have ties in the ranking. The intra-annotator values were computed for these metrics and the scores are shown in Table TABREF32 BIBREF12 , BIBREF13 .
The human evaluation Kappa co-efficient results are calculated with respect to: DISPLAYFORM0
It was observed that the ranking Kappa co-efficient for intra-annotator ranking of the RNNMorph model was at 0.573, higher that the 0.410 of the RNNSearch+Word2Vec model, implying that the annotators found the RNNMorph model to produce better results when compared to the RNNSearch + Word2Vec model.
Model Parameters
The learning rate decay through the training process of the RNNMorph model is showcased in the graph in Fig. FIGREF34 . This process was done manually where the learning rate was decayed after the end of specific epochs based on an observed stagnation in perplexity.The RNNMorph model achieved saturation of perplexities much earlier through the epochs than the RNNSearch + Word2Vec model. This conforms to the expected outcome as the morphological segmentation has reduced the vocabulary size of the target language from 340,325 words to a mere 41,906 morphs.
The error function used was the sampled SoftMax loss to ensure a large target vocabulary could be accommodated BIBREF10 . A zoomed inset graph (Fig. FIGREF35 ) has been used to visualize the values of the error function for the RNNSearch + Word2Vec and RNNMorph models with 4 hidden layers. It can be seen that the RNNMorph model is consistently better in terms of the perplexity values through the time steps.
Attention Vectors
In order to further demonstrate the quality of the RNNMorph model, the attention vectors of both the RNNSearch with Word2Vec embedding and RNNMorph models are compared for several good translations in Figs. FIGREF37 and FIGREF44 . It is observed that the reduction in vocabulary size has improved the source sentence lookup by quite an extent. Each cell in the heatmap displays the magnitude of the attention layer weight INLINEFORM0 for the INLINEFORM1 Tamil word and the INLINEFORM2 English word in the respective sentences. The intensity of black corresponds to the magnitude of the cell INLINEFORM3 . Also, the attention vectors of the RNNSearch model with Word2Vec embeddings tend to attend to INLINEFORM4 EOS INLINEFORM5 token in the middle of the sentence leading to incomplete translations. This could be due to the fact that only 44% of the Tamil vocabulary and 74% of the English vocabulary is taken for training in this model, as opposed to 100% of English and Tamil words in the RNNMorph model.
Target vocabulary size
A very large target vocabulary is an inadvertent consequence of the morphological richness of the Tamil language. This creates a potential restriction on the accuracy of the model as many inflectional forms of the same word are trained as independent units. One of the advantages of morphological segmentation of Tamil text is that the target vocabulary size decreased from 340,325 to a mere 41,906. This reduction helps improve the performance of the translation as the occurrence of unknown tokens was reduced compared to the RNNSearch model. This morphologically segmented vocabulary is divided into a collection of morphological roots and inflections as individual units.
Repetitions
Some of the translations of the RNNMorph model have repetitions of the same phrases (Fig. FIGREF53 ), whereas such repetitions occur much less frequently in the RNNSearch predictions. Such translations would make for good results if the repetitions weren't present and all parts of the sentence occur just once. These repetitions might be due to the increase in the general sequence length of the target sentences because of the morphological segmentation. While it is true the target vocabulary size has decreased due to morphological segmentation, the RNNMorph has more input units (morphs) per sentence, which makes it more demanding of the LSTM's memory units and the feed forward network of the attention model. Additionally, this behavior could also be attributed to the errors in the semi-supervised morphological segmentation due to the complexities of the Tamil language and the extent of the corpus.
Model Outputs
The translation outputs of the RNNSearch + Word2Vec and Morph2Vec models for the same input sentences from the test set demonstrate the effectiveness of using a morphological segmentation tool and how the morphemes have changed the sentence to be more grammatically sound. It is also observed (from Fig. FIGREF55 ) that most of the translation sentences of the Morph2Vec model have no INLINEFORM0 UNK INLINEFORM1 tokens. They exist in the predictions mostly only due to a word in the English test sentence not present in the source vocabulary.
Related Work
Professors CN Krishnan, Sobha et al developed a machine-aided-translation (MAT) system similar to the Anusaakara English Hindi MT system, using a small corpus and very few transfer rules, available at AU-KBC website BIBREF14 . Balajapally et al. balajapally2006multilingual developed an example based machine translation (EBMT) system with 700000 sentences for English to INLINEFORM0 Tamil, Kannada, Hindi INLINEFORM1 transliterated text BIBREF15 , BIBREF16 . Renganathan renganathan2002interactive developed a rule based MT system for English and Tamil using grammar rules for the language pair. Vetrivel et al. vetrivel2010english used HMMs to align and translate English and Tamil parallel sentences to build an SMT system. Irvine et al. irvine2013combining tried to combine parallel and similar corpora to improve the performance of English to Tamil SMT amongst other languages. Kasthuri et al. kasthuri2014rule used a rule based MT system using transfer lexicon and morphological analysis tools. Anglabharathi was developed at IIT Kanpur, a system translating English to a collection of Indian languages including Tamil using CFG like structures to create a pseudo target to convert to Indian languages BIBREF17 , BIBREF18 . A variety of hybrid approaches have also been used for English–Tamil MT in combinations of rule based (transfer methods), interlingua representations BIBREF19 , BIBREF20 , BIBREF21 . The use of Statistical Machine Translation took over the English–Tamil MT system research because of its desirable properties of language independence, better generalization features and a reduced requirement of linguistic expertise BIBREF1 , BIBREF22 , BIBREF23 . Various enhancement techniques external to the MT system have also been proposed to improve the performance of translation using morphological pre and post processing techniques BIBREF24 , BIBREF25 , BIBREF26 .
The use of RNN Encoder Decoder models in machine translation has shown good results in languages with similar grammatical structure. Deep MT systems have been performing better than the other shallow SMT models recently, with the availability of computational resources and hardware making it feasible to train such models. The first of these models came in 2014, with Cho et al SecondOneByCho. The model used was the RNN LSTM encoder decoder model with the context vector output of the encoder (run for every word in the sentence) is fed to every decoder unit along with the previous word output until INLINEFORM0 EOS INLINEFORM1 is reached. This model was used to score translation results of another MT system. Sutskever et al. sutskever2014sequence created a similar encoder decoder model with the decoder getting the context vector only for the first word of the target language sentence. After that, only the decoded target outputs act as inputs to the various time steps of the decoder. One major drawback of these models is the size of the context vector of the encoder being static in nature. The same sized vector was expected to to represent sentences of arbitrary length, which was impractical when it came to very long sentences.
The next breakthrough came from Bahdanau et al. Bahdanau2014 where variable length word vectors were used and instead of just the context vector, a weighted sum of the inputs is given for the decoder. This enabled selective lookup to the source sentence during decoding and is known as the attention mechanism BIBREF27 . The attention mechanism was further analysed by Luong et al. luong2015effective where they made a distinction between global and local attention by means of AER scores of the attention vectors. A Gaussian distribution and a monotonic lookup were used to facilitate the corresponding local source sentence look-up.
Conclusion
Thus, it is seen that the use of morphological segmentation on a morphologically rich language before translation helps with the performance of the translation in multiple ways. Thus, machine translation involving morphologically rich languages should ideally be carried out only after morphological segmentation. If the translation has to be carried out between two morphologically rich languages, then both the languages' sentences should be individually segmented based on morphology. This is because while it is true that they are both morphologically rich languages, the schemes that the languages use for the process of agglutination might be different, in which case a mapping between the units would be difficult without the segmentation.
One drawback of morphological segmentation is the increase in complexity of the model due to an increase in the average sentence lengths. This cannot be avoided as it is essential to enable a correspondence between the sentences of the two languages when one of them is a simple fusional language. Even with the increase in the average sentence length, the attention models that have been developed to ensure correctness of translation of long sequences can be put to good use when involving morphologically rich languages. Another point to note here is that morphologically rich languages like Tamil generally have lesser number of words per sentence than languages like English due to the inherent property of agglutination.
Future Work
The model implemented in this paper only includes source-side morphological segmentation and does not include a target side morphological agglutination to give back the output in words rather than morphemes. In order to implement an end-to-end translation system for morphologically rich languages, a morphological generator is essential because the output units of the translation cannot be morphemes.
The same model implemented can be further enhanced by means of a better corpus that can generalize over more than just domain specific source sentences. Also, the use of a better GPU would result in a better allocation of the hidden layer sizes and the batch sizes thereby possibly increasing the scope and accuracy of learning of the translation model.
Although not directly related to Machine Translation, the novel encoder– decoder architecture proposed in by Rocktaschel et al. rocktaschel2015reasoning for Natural Language Inference (NLI) can be used for the same. Their model fuses inferences from each and every individual word, summarizing information at each step, thereby linking the hidden state of the encoder with that of the decoder by means of a weighted sum, trained for optimization.
Acknowledgements
I would like to thank Dr. M. Anand Kumar, Assistant Professor, Amrita Vishwa Vidyapeetham for his continuous support and guidance. I would also like to thank Dr. Arvindan, Professor, SSN College Of Engineering for his inputs and suggestions.
| [
"50 human annotators ranked a random sample of 100 translations by Adequacy, Fluency and overall ranking on a 5-point scale.",
"adequacy, precision and ranking values"
] | 5,344 | qasper | en | null | 1a1daeb5809cde7c7131d972f42ee43654a378c78d5182c8 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How were the human judgements assembled?
Answer: | [
"50 human annotators ranked a random sample of 100 translations by Adequacy, Fluency and overall ranking on a 5-point scale.",
"adequacy, precision and ranking values"
] | qasper | 128 | 40 | |
Do they test their framework performance on commonly used language pairs, such as English-to-German? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Neural Machine Translation (NMT) has shown its effectiveness in translation tasks when NMT systems perform best in recent machine translation campaigns BIBREF0 , BIBREF1 . Compared to phrase-based Statistical Machine Translation (SMT) which is basically an ensemble of different features trained and tuned separately, NMT directly modeling the translation relationship between source and target sentences. Unlike SMT, NMT does not require much linguistic information and large monolingual data to achieve good performances.
An NMT consists of an encoder which recursively reads and represents the whole source sentence into a context vector and a recurrent decoder which takes the context vector and its previous state to predict the next target word. It is then trained in an end-to-end fashion to learn parameters which maximizes the likelihood between the outputs and the references. Recently, attention-based NMT has been featured in most state-of-the-art systems. First introduced by BIBREF2 , attention mechanism is integrated in decoder side as feedforward layers. It allows the NMT to decide which source words should take part in the predicting process of the next target words. It helps to improve NMTs significantly. Nevertheless, since the attention mechanism is specific to a particular source sentence and the considering target word, it is also specific to particular language pairs.
Some recent work has focused on extending the NMT framework to multilingual scenarios. By training such network using parallel corpora in number of different languages, NMT could benefit from additional information embedded in a common semantic space across languages. Basically, the proposed NMT are required to employ multiple encoders or multiple decoders to deal with multilinguality. Furthermore, in order to avoid the tight dependency of the attention mechanism to specific language pairs, they also need to modify their architecture to combine either the encoders or the attention layers. These modifications are specific to the purpose of the tasks as well. Thus, those multilingual NMTs are more complicated, much more free parameters to learn and more difficult to perform standard trainings compared to the original NMT.
In this paper, we introduce a unified approach to seamlessly extend the original NMT to multilingual settings. Our approach allows us to integrate any language in any side of the encoder-decoder architecture with only one encoder and one decoder for all the languages involved. Moreover, it is not necessary to do any network modification to enable attention mechanism in our NMT systems. We then apply our proprosed framework in two demanding scenarios: under-resourced translation and zero-resourced translation. The results show that bringing multilinguality to NMT helps to improve individual translations. With some insightful analyses of the results, we set our goal toward a fully multilingual NMT framework.
The paper starts with a detailed introduction to attention-based NMT. In Section SECREF3 , related work about multi-task NMT is reviewed. Section SECREF5 describes our proposed approach and thorough comparisons to the related work. It is followed by a section of evaluating our systems in two aforementioned scenarios, in which different strategies have been employed under a unified approach (Section SECREF4 ). Finally, the paper ends with conclusion and future work.
This work is licenced under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/
Neural Machine Translation: Background
An NMT system consists of an encoder which automatically learns the characteristics of a source sentence into fix-length context vectors and a decoder that recursively combines the produced context vectors with the previous target word to generate the most probable word from a target vocabulary.
More specifically, a bidirectional recurrent encoder reads every words INLINEFORM0 of a source sentence INLINEFORM1 and encodes a representation INLINEFORM2 of the sentence into a fixed-length vector INLINEFORM3 concatinated from those of the forward and backward directions: INLINEFORM4
Here INLINEFORM0 is the one-hot vector of the word INLINEFORM1 and INLINEFORM2 is the word embedding matrix which is shared across the source words. INLINEFORM3 is the recurrent unit computing the current hidden state of the encoder based on the previous hidden state. INLINEFORM4 is then called an annotation vector, which encodes the source sentence up to the time INLINEFORM5 from both forward and backward directions. Recurrent units in NMT can be a simple recurrent neural network unit (RNN), a Long Short-Term Memory unit (LSTM) BIBREF3 or a Gated Recurrent Unit (GRU) BIBREF4
Similar to the encoder, the recurrent decoder generates one target word INLINEFORM0 to form a translated target sentence INLINEFORM1 in the end. At the time INLINEFORM2 , it takes the previous hidden state of the decoder INLINEFORM3 , the previous embedded word representation INLINEFORM4 and a time-specific context vector INLINEFORM5 as inputs to calculate the current hidden state INLINEFORM6 : INLINEFORM7
Again, INLINEFORM0 is the recurrent activation function of the decoder and INLINEFORM1 is the shared word embedding matrix of the target sentences. The context vector INLINEFORM2 is calculated based on the annotation vectors from the encoder. Before feeding the annotation vectors into the decoder, an attention mechanism is set up in between, in order to choose which annotation vectors should contribute to the predicting decision of the next target word. Intuitively, a relevance between the previous target word and the annotation vectors can be used to form some attention scenario. There exists several ways to calculate the relevance as shown in BIBREF5 , but what we describe here follows the proposed method of BIBREF2 DISPLAYFORM0
In BIBREF2 , this attention mechanism, originally called alignment model, has been employed as a simple feedforward network with the first layer is a learnable layer via INLINEFORM0 , INLINEFORM1 and INLINEFORM2 . The relevance scores INLINEFORM3 are then normalized into attention weights INLINEFORM4 and the context vector INLINEFORM5 is calculated as the weighted sum of all annotation vectors INLINEFORM6 . Depending on how much attention the target word at time INLINEFORM7 put on the source states INLINEFORM8 , a soft alignment is learned. By being employed this way, word alignment is not a latent variable but a parametrized function, making the alignment model differentiable. Thus, it could be trained together with the whole architecture using backpropagation.
One of the most severe problems of NMT is handling of the rare words, which are not in the short lists of the vocabularies, i.e. out-of-vocabulary (OOV) words, or do not appear in the training set at all. In BIBREF6 , the rare target words are copied from their aligned source words after the translation. This heuristic works well with OOV words and named entities but unable to translate unseen words. In BIBREF7 , their proposed NMT models have been shown to not only be effective on reducing vocabulary sizes but also have the ability to generate unseen words. This is achieved by segmenting the rare words into subword units and translating them. The state-of-the-art translation systems essentially employ subword NMT BIBREF7 .
Universal Encoder and Decoder for Multilingual Neural Machine Translation
While the majority of previous research has focused on improving the performance of NMT on individual language pairs with individual NMT systems, recent work has started investigating potential ways to conduct the translation involved in multiple languages using a single NMT system. The possible reason explaining these efforts lies on the unique architecture of NMT. Unlike SMT, NMT consists of separated neural networks for the source and target sides, or the encoder and decoder, respectively. This allows these components to map a sentence in any language to a representation in an embedding space which is believed to share common semantics among the source languages involved. From that shared space, the decoder, with some implicit or explicit relevant constraints, could transform the representation into a concrete sentence in any desired language. In this section, we review some related work on this matter. We then describe a unified approach toward an universal attention-based NMT scheme. Our approach does not require any architecture modification and it can be trained to learn a minimal number of parameters compared to the other work.
Related Work
By extending the solution of sequence-to-sequence modeling using encoder-decoder architectures to multi-task learning, Luong2016 managed to achieve better performance on some INLINEFORM0 tasks such as translation, parsing and image captioning compared to individual tasks. Specifically in translation, the work utilizes multiple encoders to translate from multiple languages, and multiple decoders to translate to multiple languages. In this view of multilingual translation, each language in source or target side is modeled by one encoder or decoder, depending on the side of the translation. Due to the natural diversity between two tasks in that multi-task learning scenario, e.g. translation and parsing, it could not feature the attention mechanism although it has proven its effectiveness in NMT. There exists two directions which proposed for multilingual translation scenarios where they leverage the attention mechanism. The first one is indicated in the work from BIBREF8 , where it introduce an one-to-many multilingual NMT system to translates from one source language into multiple target languages. Having one source language, the attention mechanism is then handed over to the corresponding decoder. The objective function is changed to adapt to multilingual settings. In testing time, the parameters specific to a desired language pair are used to perform the translation.
Firat2016 proposed another approach which genuinely delivers attention-based NMT to multilingual translation. As in BIBREF9 , their approach utilizes one encoder per source language and one decoder per target language for many-to-many translation tasks. Instead of a quadratic number of independent attention layers, however, one single attention mechanism is integrated into their NMT, performing an affine transformation between the hidden layer of INLINEFORM0 source languages and that one of INLINEFORM1 target languages. It is required to change their architecture to accomodate such a complicated shared attention mechanism.
In a separate effort to achieve multilingual NMT, the work of Zoph2016 leverages available parallel data from other language pairs to help reducing possible ambiguities in the translation process into a single target language. They employed the multi-source attention-based NMT in a way that only one attention mechanism is required despite having multiple encoders. To achieve this, the outputs of the encoders were combined before feeding to the attention layer. They implemented two types of encoder combination; One is adding a non-linear layer on the concatenation of the encoders' hidden states. The other is using a variant of LSTM taking the respective gate values from the individual LSTM units of the encoders. As a result, the combined hidden states contain information from both encoders , thus encode the common semantic of the two source languages.
Universal Encoder and Decoder
Inspired by the multi-source NMT as additional parallel data in several languages are expected to benefit single translations, we aim to develop a NMT-based approach toward an universal framework to perform multilingual translation. Our solution features two treatments: 1) Coding the words in different languages as different words in the language-mixed vocabularies and 2) Forcing the NMT to translating a representation of source sentences into the sentences in a desired target language.
Language-specific Coding. When the encoder of a NMT system considers words across languages as different words, with a well-chosen architecture, it is expected to be able to learn a good representation of the source words in an embedding space in which words carrying similar meaning would have a closer distance to each others than those are semantically different. This should hold true when the words have the same or similar surface form, such as (@de@Obama; @en@Obama) or (@de@Projektion; @en@projection). This should also hold true when the words have the same or similar meaning across languages, such as (@en@car; @en@automobile) or (@de@Flussufer; @en@bank). Our encoder then acts similarly to the one of multi-source approach BIBREF10 , collecting additional information from other sources for better translations, but with a much simpler embedding function. Unlike them, we need only one encoder, so we could reduce the number of parameters to learn. Furthermore, we neither need to change the network architecture nor depend on which recurrent unit (GRU, LSTM or simple RNN) is currently using in the encoder.
We could apply the same trick to the target sentences and thus enable many-to-many translation capability of our NMT system. Similar to the multi-target translation BIBREF8 , we exploit further the correlation in semantics of those target sentences across different languages. The main difference between our approach and the work of BIBREF8 is that we need only one decoder for all target languages. Given one encoder for multiple source languages and one decoder for multiple target languages, it is trivial to incorporate the attention mechanism as in the case of a regular NMT for single language translation. In training, the attention layers were directed to learn relevant alignments between words in specific language pair and forward the produced context vector to the decoder. Now we rely totally on the network to learn good alignments between source and target sides. In fact, giving more information, our system are able to form nice alignments.
In comparison to other research that could perform complete multi-task learning, e.g. the work from BIBREF9 or the approach proposed by BIBREF11 , our method is able to accommodate the attention layers seemlessly and easily. It also draws a clear distinction from those works in term of the complexity of the whole network: considerably less parameters to learn, thus reduces overfitting, with a conventional attention mechanism and a standard training procedure.
Target Forcing. While language-specific coding allows us to implement a multilingual attention-based NMT, there are two issues we have to consider before training the network. The first is that the number of rare words would increase in proportion with the number of languages involved. This might be solved by applying a rare word treatment method with appropriate awareness of the vocabularies' size. The second one is more problematic: Ambiguity level in the translation process definitely increases due to the additional introduction of words having the same or similar meaning across languages at both source and target sides. We deal with the problem by explicitly forcing the attention and translation to the direction that we prefer, expecting the information would limit the ambiguity to the scope of one language instead of all target languages. We realize this idea by adding at the beginning and at the end of every source sentences a special symbol indicating the language they would be translated into. For example, in a multilingual NMT, when a source sentence is German and the target language is English, the original sentence (already language-specific coded) is:
@de@darum @de@geht @de@es @de@in @de@meinem @de@Vortrag
Now when we force it to be translated into English, the target-forced sentence becomes:
<E> @de@darum @de@geht @de@es @de@in @de@meinem @de@Vortrag <E>
Due to the nature of recurrent units used in the encoder and decoder, in training, those starting symbols encourage the network learning the translation of following target words in a particular language pair. In testing time, information of the target language we provided help to limit the translated candidates, hence forming the translation in the desired language.
Figure FIGREF8 illustrates the essence of our approach. With two steps in the preprocessing phase, namely language-specific coding and target forcing, we are able to employ multilingual attention-based NMT without any special treatment in training such a standard architecture. Our encoder and attention-enable decoder can be seen as a shared encoder and decoder across languages, or an universal encoder and decoder. The flexibitily of our approach allow us to integrate any language into source or target side. As we will see in Section SECREF4 , it has proven to be extremely helpful not only in low-resourced scenarios but also in translation of well-resourced language pairs as it provides a novel way to make use of large monolingual corpora in NMT.
Evaluation
In this section, we describe the evaluation of our proposed approach in comparisons with the strong baselines using NMT in two scenarios: the translation of an under-resource language pair and the translation of a language pair that does not exist any paralled data at all.
Experimental Settings
Training Data. We choose WIT3's TED corpus BIBREF12 as the basis of our experiments since it might be the only high-quality parallel data of many low-resourced language pairs. TED is also multilingual in a sense that it includes numbers of talks which are commonly translated into many languages. In addition, we use a much larger corpus provided freely by WMT organizers when we evaluate the impact of our approach in a real machine translation campaign. It includes the paralled corpus extracted from the digital corpus of European Parliament (EPPS), the News Commentary (NC) and the web-crawled parallel data (CommonCrawl). While the number of sentences in popular TED corpora varies from 13 thousands to 17 thousands, the total number of sentences in those larger corpus is approximately 3 million sentences.
Neural Machine Translation Setup. All experiments have been conducted using NMT framework Nematus, Following the work of Sennrich2016a, subword segmentation is handled in the prepocessing phase using Byte-Pair Encoding (BPE). Excepts stated clearly in some experiments, we set the number of BPE merging operations at 39500 on the joint of source and target data. When training all NMT systems, we take out the sentence pairs exceeding 50-word length and shuffle them inside every minibatch. Our short-list vocabularies contain 40,000 most frequent words while the others are considered as rare words and applied the subword translation. We use an 1024-cell GRU layer and 1000-dimensional embeddings with dropout at every layer with the probability of 0.2 in the embedding and hidden layers and 0.1 in the input and ourput layers. We trained our systems using gradient descent optimization with Adadelta BIBREF13 on minibatches of size 80 and the gradient is rescaled whenever its norm exceed 1.0. All the trainings last approximately seven days if the early-stopping condition could not be reached. At a certain time, an external evaluation script on BLEU BIBREF14 is conducted on a development set to decide the early-stopping condition. This evaluation script has also being used to choose the model archiving the best BLEU on the development set instead of the maximal loglikelihood between the translations and target sentences while training. In translation, the framework produces INLINEFORM0 -best candidates and we then use a beam search with the beam size of 12 to get the best translation.
Under-resourced Translation
First, we consider the translation for an under-resourced pair of languages. Here a small portion of the large parallel corpus for English-German is used as a simulation for the scenario where we do not have much parallel data: Translating texts in English to German. We perform language-specific coding in both source and target sides. By accommodating the German monolingual data as an additional input (German INLINEFORM0 German), which we called the mix-source approach, we could enrich the training data in a simple, natural way. Given this under-resourced situation, it could help our NMT obtain a better representation of the source side, hence, able to learn the translation relationship better. Including monolingual data in this way might also improve the translation of some rare word types such as named entities. Furthermore, as the ultimate goal of our work, we would like to investigate the advantages of multilinguality in NMT. We incorporate a similar portion of French-German parallel corpus into the English-German one. As discussed in Section SECREF5 , it is expected to help reducing the ambiguity in translation between one language pair since it utilizes the semantic context provided by the other source language. We name this mix-multi-source.
Table TABREF16 summarizes the performance of our systems measured in BLEU on two test sets, tst2013 and tst2014. Compared to the baseline NMT system which is solely trained on TED English-German data, our mix-source system achieves a considerable improvement of 2.6 BLEU points on tst2013 and 2.1 BLEU points on and tst2014 . Adding French data to the source side and their corresponding German data to the target side in our mix-multi-source system also help to gain 2.2 and 1.6 BLEU points more on tst2013 tst2014, respectively. We observe a better improvement from our mix-source system compared to our mix-multi-source system. We speculate the reason that the mix-source encoder utilize the same information shared in two languages while the mix-multi-source receives and processes similar information in the other language but not necessarily the same. We might validate this hypothesis by comparing two systems trained on a common English-German-French corpus of TED. We put it in our future work's plan.
As we expected Figure FIGREF19 shows how different words in different languages can be close in the shared space after being learned to translate into a common language. We extract the word embeddings from the encoder of the mix-multi-source (En,Fr INLINEFORM0 De,De) after training, remove the language-specific codes (@en@ and @fr@)and project the word vectors to the 2D space using t-SNE BIBREF15 .
Using large monolingual data in NMT.
A standard NMT system employs parallel data only. While good parallel corpora are limited in number, getting monolingual data of an arbitrary language is trivial. To make use of German monolingual corpus in an English INLINEFORM0 German NMT system, sennrich2016b built a separate German INLINEFORM1 English NMT using the same parallel corpus, then they used that system to translate the German monolingual corpus back to English, forming a synthesis parallel data. gulcehre2015 trained another RNN-based language model to score the monolingual corpus and integrate it to the NMT system through shallow or deep fusion. Both methods requires to train separate systems with possibly different hyperparameters for each. Conversely, by applying mix-source method to the big monolingual data, we need to train only one network. We mix the TED parallel corpus and the substantial monolingual corpus (EPPS+NC+CommonCrawl) and train a mix-source NMT system from those data.
The first result is not encouraging when its performance is even worse than the baseline NMT which is trained on the small parallel data only. Not using the same information in the source side, as we discussed in case of mix-multi-source strategy, could explain the degrading in performance of such a system. But we believe that the magnitude and unbalancing of the corpus are the main reasons. The data contains nearly four millions sentences but only around twenty thousands of them (0.5%) are the genuine parallel data. As a quick attempt, after we get the model with that big data, we continue training on the real parallel corpus for some more epochs. When this adaptation is applied, our system brings an improvement of +1.52 BLEU on tst2013 and +1.06 BLEU on tst2014 (Table TABREF21 ).
Zero-resourced Translation
Among low-resourced scenarios, zero-resourced translation task stands in an extreme level. A zero-resourced translation task is one of the most difficult situation when there is no parallel data between the translating language pair. To the best of our knowledge, there have been yet existed a published work about using NMT for zero-resourced translation tasks up to now. In this section, we extend our strategies using the proposed multilingual NMT approach as first attempts to this extreme situation.
We employ language-specific coding and target forcing in a strategy called bridge. Unlike the strategies used in under-resourced translation task, bridge is an entire many-to-many multilingual NMT. Simulating a zero-resourced German INLINEFORM0 French translation task given the available German-English and English-French parallel corpora, after applying language-specific coding and target forcing for each corpus, we mix those data with an English-English data as a “bridge” creating some connection between German and French. We also propose a variant of this strategy that we incorporate French-French data. And we call it universal.
We evaluate bridge and universal systems on two German INLINEFORM0 French test sets. They are compared to a direct system, which is an NMT trained on German INLINEFORM1 French data, and to a pivot system, which essentially consists of two separate NMTs trained to translate from German to English and English to French. The direct system should not exist in a real zero-resourced situation. We refer it as the perfect system for comparison purpose only. In case of the pivot system, to generate a translated text in French from a German sentence, we first translate it to English, then the output sentence is fed to the English INLINEFORM2 German NMT system to obtain the French translation. Since there are more than two languages involved in those systems, we increase the number of BPE merging operations proportionally in order to reduce the number of rare words in such systems. We do not expect our proposed systems to perform well with this primitive way of building direct translating connections since this is essentially a difficult task. We report the performance of those systems in Table TABREF23 .
Unsupprisingly, both bridge and universal systems perform worse than the pivot one. We consider two possible reasons:
Our target forcing mechanism is moderately primitive. Since the process is applied after language-specific coding, the target forcing symbol is the same for all source sentences in every languages. Thus, the forcing strength might not be enough to guide the decision of the next words. Once the very first word is translated into a word in wrong language, the following words tend to be translated into that wrong language again. Table TABREF24 shows some statistics of the translated words and sentences in wrong language.
Balancing of the training corpus. Although it is not severe as in the case of mix-source system for large monolingual data, the limited number of sentences in target language can affect the training. The difference of 1.07 BLEU points between bridge and universal might explain this assumption as we added more target data (French) in universal strategy, thus reducing the unbalance in training.
Those issues would be addressed in our following future work toward the multilingual attention-based NMT.
Conclusion and Future Work
In this paper, we present our first attempts in building a multilingual Neural Machine Translation framework. By treating words in different languages as different words and force the attention and translation to the direction of desired target language, we are able to employ attention-enable NMT toward a multilingual translation system. Our proposed approach alleviates the need of complicated architecture re-designing when accommodating attention mechanism. In addition, the number of free parameters to learn in our network does not go beyond that magnitute of a single NMT system. With its universality, our approach has shown its effectiveness in an under-resourced translation task with considerable improvements. In addition, the approach has achieved interesting and promising results when applied in the translation task that there is no direct parallel corpus between source and target languages.
Nevertheless, there are issues that we can continue working on to address in future work. A more balancing data would be helpful for this framework. The mechanism of forcing the NMT system to the right target language could be improved. We could conduct more detailed analyses of the various strategies under the framework to show its universarity.
| [
"Yes",
"Yes"
] | 4,472 | qasper | en | null | d1d93cefe6e2c643ecf128643f9362e3fa137ff54253a3ec | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Do they test their framework performance on commonly used language pairs, such as English-to-German?
Answer: | [
"Yes",
"Yes"
] | qasper | 128 | 41 | |
How are models evaluated in this human-machine communication game? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Suppose a user wants to write a sentence “I will be 10 minutes late.” Ideally, she would type just a few keywords such as “10 minutes late” and an autocomplete system would be able to infer the intended sentence (Figure FIGREF1). Existing left-to-right autocomplete systems BIBREF0, BIBREF1 can often be inefficient, as the prefix of a sentence (e.g. “I will be”) fails to capture the core meaning of the sentence. Besides the practical goal of building a better autocomplete system, we are interested in exploring the tradeoffs inherent to such communication schemes between the efficiency of typing keywords, accuracy of reconstruction, and interpretability of keywords.
One approach to learn such schemes is to collect a supervised dataset of keywords-sentence pairs as a training set, but (i) it would be expensive to collect such data from users, and (ii) a static dataset would not capture a real user's natural predilection to adapt to the system BIBREF2. Another approach is to avoid supervision and jointly learn a user-system communication scheme to directly optimize the combination of efficiency and accuracy. However, learning in this way can lead to communication schemes that are uninterpretable to humans BIBREF3, BIBREF4 (see Appendix for additional related work).
In this work, we propose a simple, unsupervised approach to an autocomplete system that is efficient, accurate, and interpretable. For interpretability, we restrict keywords to be subsequences of their source sentences based on the intuition that humans can infer most of the original meaning from a few keywords. We then apply multi-objective optimization approaches to directly control and achieve desirable tradeoffs between efficiency and accuracy.
We observe that naively optimizing a linear combination of efficiency and accuracy terms is unstable and leads to suboptimal schemes. Thus, we propose a new objective which optimizes for communication efficiency under an accuracy constraint. We show this new objective is more stable and efficient than the linear objective at all accuracy levels.
As a proof-of-concept, we build an autocomplete system within this framework which allows a user to write sentences by specifying keywords. We empirically show that our framework produces communication schemes that are 52.16% more accurate than rule-based baselines when specifying 77.37% of sentences, and 11.73% more accurate than a naive, weighted optimization approach when specifying 53.38% of sentences. Finally, we demonstrate that humans can easily adapt to the keyword-based autocomplete system and save nearly 50% of time compared to typing a full sentence in our user study.
Approach
Consider a communication game in which the goal is for a user to communicate a target sequence $x= (x_1, ..., x_m)$ to a system by passing a sequence of keywords $z= (z_1, ..., z_n)$. The user generates keywords $z$ using an encoding strategy $q_{\alpha }(z\mid x)$, and the system attempts to guess the target sequence $x$ via a decoding strategy $p_{\beta }(x\mid z)$.
A good communication scheme $(q_{\alpha }, p_{\beta })$ should be both efficient and accurate. Specifically, we prefer schemes that use fewer keywords (cost), and the target sentence $x$ to be reconstructed with high probability (loss) where
Based on our assumption that humans have an intuitive sense of retaining important keywords, we restrict the set of schemes to be a (potentially noncontiguous) subsequence of the target sentence. Our hypothesis is that such subsequence schemes naturally ensure interpretability, as efficient human and machine communication schemes are both likely to involve keeping important content words.
Approach ::: Modeling with autoencoders.
To learn communication schemes without supervision, we model the cooperative communication between a user and system through an encoder-decoder framework. Concretely, we model the user's encoding strategy $q_{\alpha }(z\mid x)$ with an encoder which encodes the target sentence $x$ into the keywords $z$ by keeping a subset of the tokens. This stochastic encoder $q_{\alpha }(z\mid x)$ is defined by a model which returns the probability of each token retained in the final subsequence $z$. Then, we sample from Bernoulli distributions according to these probabilities to either keep or drop the tokens independently (see Appendix for an example).
We model the autocomplete system's decoding strategy $p_{\beta }(x\mid z)$ as a probabilistic model which conditions on the keywords $z$ and returns a distribution over predictions $x$. We use a standard sequence-to-sequence model with attention and copying for the decoder, but any model architecture can be used (see Appendix for details).
Approach ::: Multi-objective optimization.
Our goal now is to learn encoder-decoder pairs which optimally balance the communication cost and reconstruction loss. The simplest approach to balancing efficiency and accuracy is to weight $\mathrm {cost}(x, \alpha )$ and $\mathrm {loss}(x, \alpha , \beta )$ linearly using a weight $\lambda $ as follows,
where the expectation is taken over the population distribution of source sentences $x$, which is omitted to simplify notation. However, we observe that naively weighting and searching over $\lambda $ is suboptimal and highly unstable—even slight changes to the weighting results in degenerate schemes which keep all or none of its tokens. This instability motivates us to develop a new stable objective.
Our main technical contribution is to draw inspiration from the multi-objective optimization literature and view the tradeoff as a sequence of constrained optimization problems, where we minimize the expected cost subject to varying expected reconstruction error constraints $\epsilon $,
This greatly improves the stability of the training procedure. We empirically observe that the model initially keeps most of the tokens to meet the constraints, and slowly learns to drop uninformative words from the keywords to minimize the cost. Furthermore, $\epsilon $ in Eq (DISPLAY_FORM6) allows us to directly control the maximum reconstruction error of resulting schemes, whereas $\lambda $ in Eq (DISPLAY_FORM5) is not directly related to any of our desiderata.
To optimize the constrained objective, we consider the Lagrangian of Eq (DISPLAY_FORM6),
Much like the objective in Eq (DISPLAY_FORM5) we can compute unbiased gradients by replacing the expectations with their averages over random minibatches. Although gradient descent guarantees convergence on Eq (DISPLAY_FORM7) only when the objective is convex, we find that not only is the optimization stable, the resulting solution achieves better performance than the weighting approach in Eq (DISPLAY_FORM5).
Approach ::: Optimization.
Optimization with respect to $q_{\alpha }(z\mid x)$ is challenging as $z$ is discrete, and thus, we cannot differentiate $\alpha $ through $z$ via the chain rule. Because of this, we use the stochastic REINFORCE estimate BIBREF5 as follows:
We perform joint updates on $(\alpha , \beta , \lambda )$, where $\beta $ and $\lambda $ are updated via standard gradient computations, while $\alpha $ uses an unbiased, stochastic gradient estimate where we approximate the expectation in Eq (DISPLAY_FORM9). We use a single sample from $q_{\alpha }(z\mid x)$ and moving-average of rewards as a baseline to reduce variance.
Experiments
We evaluate our approach by training an autocomplete system on 500K randomly sampled sentences from Yelp reviews BIBREF6 (see Appendix for details). We quantify the efficiency of a communication scheme $(q_{\alpha },p_{\beta })$ by the retention rate of tokens, which is measured as the fraction of tokens that are kept in the keywords. The accuracy of a scheme is measured as the fraction of sentences generated by greedily decoding the model that exactly matches the target sentence.
Experiments ::: Effectiveness of constrained objective.
We first show that the linear objective in Eq (DISPLAY_FORM5) is suboptimal compared to the constrained objective in Eq (DISPLAY_FORM6). Figure FIGREF10 compares the achievable accuracy and efficiency tradeoffs for the two objectives, which shows that the constrained objective results in more efficient schemes than the linear objective at every accuracy level (e.g. 11.73% more accurate at a 53.38% retention rate).
We also observe that the linear objective is highly unstable as a function of the tradeoff parameter $\lambda $ and requires careful tuning. Even slight changes to $\lambda $ results in degenerate schemes that keep all or none of the tokens (e.g. $\lambda \le 4.2$ and $\lambda \ge 4.4$). On the other hand, the constrained objective is substantially more stable as a function of $\epsilon $ (e.g. points for $\epsilon $ are more evenly spaced than $\lambda $).
Experiments ::: Efficiency-accuracy tradeoff.
We quantify the efficiency-accuracy tradeoff compared to two rule-based baselines: Unif and Stopword. The Unif encoder randomly keeps tokens to generate keywords with the probability $\delta $. The Stopword encoder keeps all tokens but drops stop words (e.g. `the', `a', `or') all the time ($\delta =0$) or half of the time ($\delta =0.5$). The corresponding decoders for these encoders are optimized using gradient descent to minimize the reconstruction error (i.e. $\mathrm {loss}(x, \alpha , \beta )$).
Figure FIGREF10 shows that two baselines achieve similar tradeoff curves, while the constrained model achieves a substantial 52.16% improvement in accuracy at a 77.37% retention rate compared to Unif, thereby showing the benefits of jointly training the encoder and decoder.
Experiments ::: Robustness and analysis.
We provide additional experimental results on the robustness of learned communication schemes as well as in-depth analysis on the correlation between the retention rates of tokens and their properties, which we defer to Appendix and for space.
Experiments ::: User study.
We recruited 100 crowdworkers on Amazon Mechanical Turk (AMT) and measured completion times and accuracies for typing randomly sampled sentences from the Yelp corpus. Each user was shown alternating autocomplete and writing tasks across 50 sentences (see Appendix for user interface). For the autocomplete task, we gave users a target sentence and asked them to type a set of keywords into the system. The users were shown the top three suggestions from the autocomplete system, and were asked to mark whether each of these three suggestions was semantically equivalent to the target sentence. For the writing task, we gave users a target sentence and asked them to either type the sentence verbatim or a sentence that preserves the meaning of the target sentence.
Table TABREF13 shows two examples of the autocomplete task and actual user-provided keywords. Each column contains a set of keywords and its corresponding top three suggestions generated by the autocomplete system with beam search. We observe that the system is likely to propose generic sentences for under-specified keywords (left column) and almost the same sentences for over-specified keywords (right column). For properly specified keywords (middle column), the system completes sentences accordingly by adding a verb, adverb, adjective, preposition, capitalization, and punctuation.
Overall, the autocomplete system achieved high accuracy in reconstructing the keywords. Users marked the top suggestion from the autocomplete system to be semantically equivalent to the target $80.6$% of the time, and one of the top 3 was semantically equivalent $90.11$% of the time. The model also achieved a high exact match accuracy of 18.39%. Furthermore, the system was efficient, as users spent $3.86$ seconds typing keywords compared to $5.76$ seconds for full sentences on average. The variance of the typing time was $0.08$ second for keywords and $0.12$ second for full sentences, indicating that choosing and typing keywords for the system did not incur much overhead.
Experiments ::: Acknowledgments
We thank the reviewers and Yunseok Jang for their insightful comments. This work was supported by NSF CAREER Award IIS-1552635 and an Intuit Research Award.
Experiments ::: Reproducibility
All code, data and experiments are available on CodaLab at https://bit.ly/353fbyn.
| [
"by training an autocomplete system on 500K randomly sampled sentences from Yelp reviews",
"efficiency of a communication scheme $(q_{\\alpha },p_{\\beta })$ by the retention rate of tokens, which is measured as the fraction of tokens that are kept in the keywords, accuracy of a scheme is measured as the fraction... | 1,873 | qasper | en | null | 92da01e7242f30f5266e431b4269fee1b0ca5fcc23aee095 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How are models evaluated in this human-machine communication game?
Answer: | [
"by training an autocomplete system on 500K randomly sampled sentences from Yelp reviews",
"efficiency of a communication scheme $(q_{\\alpha },p_{\\beta })$ by the retention rate of tokens, which is measured as the fraction of tokens that are kept in the keywords, accuracy of a scheme is measured as the fraction... | qasper | 128 | 42 | |
What evaluation metrics are looked at for classification tasks? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Performance appraisal (PA) is an important HR process, particularly for modern organizations that crucially depend on the skills and expertise of their workforce. The PA process enables an organization to periodically measure and evaluate every employee's performance. It also provides a mechanism to link the goals established by the organization to its each employee's day-to-day activities and performance. Design and analysis of PA processes is a lively area of research within the HR community BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 .
The PA process in any modern organization is nowadays implemented and tracked through an IT system (the PA system) that records the interactions that happen in various steps. Availability of this data in a computer-readable database opens up opportunities to analyze it using automated statistical, data-mining and text-mining techniques, to generate novel and actionable insights / patterns and to help in improving the quality and effectiveness of the PA process BIBREF4 , BIBREF5 , BIBREF6 . Automated analysis of large-scale PA data is now facilitated by technological and algorithmic advances, and is becoming essential for large organizations containing thousands of geographically distributed employees handling a wide variety of roles and tasks.
A typical PA process involves purposeful multi-step multi-modal communication between employees, their supervisors and their peers. In most PA processes, the communication includes the following steps: (i) in self-appraisal, an employee records his/her achievements, activities, tasks handled etc.; (ii) in supervisor assessment, the supervisor provides the criticism, evaluation and suggestions for improvement of performance etc.; and (iii) in peer feedback (aka INLINEFORM0 view), the peers of the employee provide their feedback. There are several business questions that managers are interested in. Examples:
In this paper, we develop text mining techniques that can automatically produce answers to these questions. Since the intended users are HR executives, ideally, the techniques should work with minimum training data and experimentation with parameter setting. These techniques have been implemented and are being used in a PA system in a large multi-national IT company.
The rest of the paper is organized as follows. Section SECREF2 summarizes related work. Section SECREF3 summarizes the PA dataset used in this paper. Section SECREF4 applies sentence classification algorithms to automatically discover three important classes of sentences in the PA corpus viz., sentences that discuss strengths, weaknesses of employees and contain suggestions for improving her performance. Section SECREF5 considers the problem of mapping the actual targets mentioned in strengths, weaknesses and suggestions to a fixed set of attributes. In Section SECREF6 , we discuss how the feedback from peers for a particular employee can be summarized. In Section SECREF7 we draw conclusions and identify some further work.
Related Work
We first review some work related to sentence classification. Semantically classifying sentences (based on the sentence's purpose) is a much harder task, and is gaining increasing attention from linguists and NLP researchers. McKnight and Srinivasan BIBREF7 and Yamamoto and Takagi BIBREF8 used SVM to classify sentences in biomedical abstracts into classes such as INTRODUCTION, BACKGROUND, PURPOSE, METHOD, RESULT, CONCLUSION. Cohen et al. BIBREF9 applied SVM and other techniques to learn classifiers for sentences in emails into classes, which are speech acts defined by a verb-noun pair, with verbs such as request, propose, amend, commit, deliver and nouns such as meeting, document, committee; see also BIBREF10 . Khoo et al. BIBREF11 uses various classifiers to classify sentences in emails into classes such as APOLOGY, INSTRUCTION, QUESTION, REQUEST, SALUTATION, STATEMENT, SUGGESTION, THANKING etc. Qadir and Riloff BIBREF12 proposes several filters and classifiers to classify sentences on message boards (community QA systems) into 4 speech acts: COMMISSIVE (speaker commits to a future action), DIRECTIVE (speaker expects listener to take some action), EXPRESSIVE (speaker expresses his or her psychological state to the listener), REPRESENTATIVE (represents the speaker's belief of something). Hachey and Grover BIBREF13 used SVM and maximum entropy classifiers to classify sentences in legal documents into classes such as FACT, PROCEEDINGS, BACKGROUND, FRAMING, DISPOSAL; see also BIBREF14 . Deshpande et al. BIBREF15 proposes unsupervised linguistic patterns to classify sentences into classes SUGGESTION, COMPLAINT.
There is much work on a closely related problem viz., classifying sentences in dialogues through dialogue-specific categories called dialogue acts BIBREF16 , which we will not review here. Just as one example, Cotterill BIBREF17 classifies questions in emails into the dialogue acts of YES_NO_QUESTION, WH_QUESTION, ACTION_REQUEST, RHETORICAL, MULTIPLE_CHOICE etc.
We could not find much work related to mining of performance appraisals data. Pawar et al. BIBREF18 uses kernel-based classification to classify sentences in both performance appraisal text and product reviews into classes SUGGESTION, APPRECIATION, COMPLAINT. Apte et al. BIBREF6 provides two algorithms for matching the descriptions of goals or tasks assigned to employees to a standard template of model goals. One algorithm is based on the co-training framework and uses goal descriptions and self-appraisal comments as two separate perspectives. The second approach uses semantic similarity under a weak supervision framework. Ramrakhiyani et al. BIBREF5 proposes label propagation algorithms to discover aspects in supervisor assessments in performance appraisals, where an aspect is modelled as a verb-noun pair (e.g. conduct training, improve coding).
Dataset
In this paper, we used the supervisor assessment and peer feedback text produced during the performance appraisal of 4528 employees in a large multi-national IT company. The corpus of supervisor assessment has 26972 sentences. The summary statistics about the number of words in a sentence is: min:4 max:217 average:15.5 STDEV:9.2 Q1:9 Q2:14 Q3:19.
Sentence Classification
The PA corpus contains several classes of sentences that are of interest. In this paper, we focus on three important classes of sentences viz., sentences that discuss strengths (class STRENGTH), weaknesses of employees (class WEAKNESS) and suggestions for improving her performance (class SUGGESTION). The strengths or weaknesses are mostly about the performance in work carried out, but sometimes they can be about the working style or other personal qualities. The classes WEAKNESS and SUGGESTION are somewhat overlapping; e.g., a suggestion may address a perceived weakness. Following are two example sentences in each class.
STRENGTH:
WEAKNESS:
SUGGESTION:
Several linguistic aspects of these classes of sentences are apparent. The subject is implicit in many sentences. The strengths are often mentioned as either noun phrases (NP) with positive adjectives (Excellent technology leadership) or positive nouns (engineering strength) or through verbs with positive polarity (dedicated) or as verb phrases containing positive adjectives (delivers innovative solutions). Similarly for weaknesses, where negation is more frequently used (presentations are not his forte), or alternatively, the polarities of verbs (avoid) or adjectives (poor) tend to be negative. However, sometimes the form of both the strengths and weaknesses is the same, typically a stand-alone sentiment-neutral NP, making it difficult to distinguish between them; e.g., adherence to timing or timely closure. Suggestions often have an imperative mood and contain secondary verbs such as need to, should, has to. Suggestions are sometimes expressed using comparatives (better process compliance). We built a simple set of patterns for each of the 3 classes on the POS-tagged form of the sentences. We use each set of these patterns as an unsupervised sentence classifier for that class. If a particular sentence matched with patterns for multiple classes, then we have simple tie-breaking rules for picking the final class. The pattern for the STRENGTH class looks for the presence of positive words / phrases like takes ownership, excellent, hard working, commitment, etc. Similarly, the pattern for the WEAKNESS class looks for the presence of negative words / phrases like lacking, diffident, slow learner, less focused, etc. The SUGGESTION pattern not only looks for keywords like should, needs to but also for POS based pattern like “a verb in the base form (VB) in the beginning of a sentence”.
We randomly selected 2000 sentences from the supervisor assessment corpus and manually tagged them (dataset D1). This labelled dataset contained 705, 103, 822 and 370 sentences having the class labels STRENGTH, WEAKNESS, SUGGESTION or OTHER respectively. We trained several multi-class classifiers on this dataset. Table TABREF10 shows the results of 5-fold cross-validation experiments on dataset D1. For the first 5 classifiers, we used their implementation from the SciKit Learn library in Python (scikit-learn.org). The features used for these classifiers were simply the sentence words along with their frequencies. For the last 2 classifiers (in Table TABREF10 ), we used our own implementation. The overall accuracy for a classifier is defined as INLINEFORM0 , where the denominator is 2000 for dataset D1. Note that the pattern-based approach is unsupervised i.e., it did not use any training data. Hence, the results shown for it are for the entire dataset and not based on cross-validation.
Comparison with Sentiment Analyzer
We also explored whether a sentiment analyzer can be used as a baseline for identifying the class labels STRENGTH and WEAKNESS. We used an implementation of sentiment analyzer from TextBlob to get a polarity score for each sentence. Table TABREF13 shows the distribution of positive, negative and neutral sentiments across the 3 class labels STRENGTH, WEAKNESS and SUGGESTION. It can be observed that distribution of positive and negative sentiments is almost similar in STRENGTH as well as SUGGESTION sentences, hence we can conclude that the information about sentiments is not much useful for our classification problem.
Discovering Clusters within Sentence Classes
After identifying sentences in each class, we can now answer question (1) in Section SECREF1 . From 12742 sentences predicted to have label STRENGTH, we extract nouns that indicate the actual strength, and cluster them using a simple clustering algorithm which uses the cosine similarity between word embeddings of these nouns. We repeat this for the 9160 sentences with predicted label WEAKNESS or SUGGESTION as a single class. Tables TABREF15 and TABREF16 show a few representative clusters in strengths and in weaknesses, respectively. We also explored clustering 12742 STRENGTH sentences directly using CLUTO BIBREF19 and Carrot2 Lingo BIBREF20 clustering algorithms. Carrot2 Lingo discovered 167 clusters and also assigned labels to these clusters. We then generated 167 clusters using CLUTO as well. CLUTO does not generate cluster labels automatically, hence we used 5 most frequent words within the cluster as its labels. Table TABREF19 shows the largest 5 clusters by both the algorithms. It was observed that the clusters created by CLUTO were more meaningful and informative as compared to those by Carrot2 Lingo. Also, it was observed that there is some correspondence between noun clusters and sentence clusters. E.g. the nouns cluster motivation expertise knowledge talent skill (Table TABREF15 ) corresponds to the CLUTO sentence cluster skill customer management knowledge team (Table TABREF19 ). But overall, users found the nouns clusters to be more meaningful than the sentence clusters.
PA along Attributes
In many organizations, PA is done from a predefined set of perspectives, which we call attributes. Each attribute covers one specific aspect of the work done by the employees. This has the advantage that we can easily compare the performance of any two employees (or groups of employees) along any given attribute. We can correlate various performance attributes and find dependencies among them. We can also cluster employees in the workforce using their supervisor ratings for each attribute to discover interesting insights into the workforce. The HR managers in the organization considered in this paper have defined 15 attributes (Table TABREF20 ). Each attribute is essentially a work item or work category described at an abstract level. For example, FUNCTIONAL_EXCELLENCE covers any tasks, goals or activities related to the software engineering life-cycle (e.g., requirements analysis, design, coding, testing etc.) as well as technologies such as databases, web services and GUI.
In the example in Section SECREF4 , the first sentence (which has class STRENGTH) can be mapped to two attributes: FUNCTIONAL_EXCELLENCE and BUILDING_EFFECTIVE_TEAMS. Similarly, the third sentence (which has class WEAKNESS) can be mapped to the attribute INTERPERSONAL_EFFECTIVENESS and so forth. Thus, in order to answer the second question in Section SECREF1 , we need to map each sentence in each of the 3 classes to zero, one, two or more attributes, which is a multi-class multi-label classification problem.
We manually tagged the same 2000 sentences in Dataset D1 with attributes, where each sentence may get 0, 1, 2, etc. up to 15 class labels (this is dataset D2). This labelled dataset contained 749, 206, 289, 207, 91, 223, 191, 144, 103, 80, 82, 42, 29, 15, 24 sentences having the class labels listed in Table TABREF20 in the same order. The number of sentences having 0, 1, 2, or more than 2 attributes are: 321, 1070, 470 and 139 respectively. We trained several multi-class multi-label classifiers on this dataset. Table TABREF21 shows the results of 5-fold cross-validation experiments on dataset D2.
Precision, Recall and F-measure for this multi-label classification are computed using a strategy similar to the one described in BIBREF21 . Let INLINEFORM0 be the set of predicted labels and INLINEFORM1 be the set of actual labels for the INLINEFORM2 instance. Precision and recall for this instance are computed as follows: INLINEFORM3
It can be observed that INLINEFORM0 would be undefined if INLINEFORM1 is empty and similarly INLINEFORM2 would be undefined when INLINEFORM3 is empty. Hence, overall precision and recall are computed by averaging over all the instances except where they are undefined. Instance-level F-measure can not be computed for instances where either precision or recall are undefined. Therefore, overall F-measure is computed using the overall precision and recall.
Summarization of Peer Feedback using ILP
The PA system includes a set of peer feedback comments for each employee. To answer the third question in Section SECREF1 , we need to create a summary of all the peer feedback comments about a given employee. As an example, following are the feedback comments from 5 peers of an employee.
The individual sentences in the comments written by each peer are first identified and then POS tags are assigned to each sentence. We hypothesize that a good summary of these multiple comments can be constructed by identifying a set of important text fragments or phrases. Initially, a set of candidate phrases is extracted from these comments and a subset of these candidate phrases is chosen as the final summary, using Integer Linear Programming (ILP). The details of the ILP formulation are shown in Table TABREF36 . As an example, following is the summary generated for the above 5 peer comments.
humble nature, effective communication, technical expertise, always supportive, vast knowledge
Following rules are used to identify candidate phrases:
Various parameters are used to evaluate a candidate phrase for its importance. A candidate phrase is more important:
A complete list of parameters is described in detail in Table TABREF36 .
There is a trivial constraint INLINEFORM0 which makes sure that only INLINEFORM1 out of INLINEFORM2 candidate phrases are chosen. A suitable value of INLINEFORM3 is used for each employee depending on number of candidate phrases identified across all peers (see Algorithm SECREF6 ). Another set of constraints ( INLINEFORM4 to INLINEFORM5 ) make sure that at least one phrase is selected for each of the leadership attributes. The constraint INLINEFORM6 makes sure that multiple phrases sharing the same headword are not chosen at a time. Also, single word candidate phrases are chosen only if they are adjectives or nouns with lexical category noun.attribute. This is imposed by the constraint INLINEFORM7 . It is important to note that all the constraints except INLINEFORM8 are soft constraints, i.e. there may be feasible solutions which do not satisfy some of these constraints. But each constraint which is not satisfied, results in a penalty through the use of slack variables. These constraints are described in detail in Table TABREF36 .
The objective function maximizes the total importance score of the selected candidate phrases. At the same time, it also minimizes the sum of all slack variables so that the minimum number of constraints are broken.
INLINEFORM0 : No. of candidate phrases INLINEFORM1 : No. of phrases to select as part of summary
INLINEFORM0 INLINEFORM1 INLINEFORM2 INLINEFORM3 INLINEFORM4 INLINEFORM5 INLINEFORM6 INLINEFORM7 INLINEFORM8
INLINEFORM0 and INLINEFORM1 INLINEFORM2 INLINEFORM3 INLINEFORM4 INLINEFORM5 INLINEFORM6
INLINEFORM0 (For determining number of phrases to select to include in summary)
Evaluation of auto-generated summaries
We considered a dataset of 100 employees, where for each employee multiple peer comments were recorded. Also, for each employee, a manual summary was generated by an HR personnel. The summaries generated by our ILP-based approach were compared with the corresponding manual summaries using the ROUGE BIBREF22 unigram score. For comparing performance of our ILP-based summarization algorithm, we explored a few summarization algorithms provided by the Sumy package. A common parameter which is required by all these algorithms is number of sentences keep in the final summary. ILP-based summarization requires a similar parameter K, which is automatically decided based on number of total candidate phrases. Assuming a sentence is equivalent to roughly 3 phrases, for Sumy algorithms, we set number of sentences parameter to the ceiling of K/3. Table TABREF51 shows average and standard deviation of ROUGE unigram f1 scores for each algorithm, over the 100 summaries. The performance of ILP-based summarization is comparable with the other algorithms, as the two sample t-test does not show statistically significant difference. Also, human evaluators preferred phrase-based summary generated by our approach to the other sentence-based summaries.
Conclusions and Further Work
In this paper, we presented an analysis of the text generated in Performance Appraisal (PA) process in a large multi-national IT company. We performed sentence classification to identify strengths, weaknesses and suggestions for improvements found in the supervisor assessments and then used clustering to discover broad categories among them. As this is non-topical classification, we found that SVM with ADWS kernel BIBREF18 produced the best results. We also used multi-class multi-label classification techniques to match supervisor assessments to predefined broad perspectives on performance. Logistic Regression classifier was observed to produce the best results for this topical classification. Finally, we proposed an ILP-based summarization technique to produce a summary of peer feedback comments for a given employee and compared it with manual summaries.
The PA process also generates much structured data, such as supervisor ratings. It is an interesting problem to compare and combine the insights from discovered from structured data and unstructured text. Also, we are planning to automatically discover any additional performance attributes to the list of 15 attributes currently used by HR.
| [
"Precision, Recall, F-measure, accuracy",
"Precision, Recall and F-measure"
] | 3,044 | qasper | en | null | b5f62b7acf975c5454df077adf727c1bcc62985be5e5a660 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What evaluation metrics are looked at for classification tasks?
Answer: | [
"Precision, Recall, F-measure, accuracy",
"Precision, Recall and F-measure"
] | qasper | 128 | 43 | |
What are the source and target domains? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
In practice, it is often difficult and costly to annotate sufficient training data for diverse application domains on-the-fly. We may have sufficient labeled data in an existing domain (called the source domain), but very few or no labeled data in a new domain (called the target domain). This issue has motivated research on cross-domain sentiment classification, where knowledge in the source domain is transferred to the target domain in order to alleviate the required labeling effort.
One key challenge of domain adaptation is that data in the source and target domains are drawn from different distributions. Thus, adaptation performance will decline with an increase in distribution difference. Specifically, in sentiment analysis, reviews of different products have different vocabulary. For instance, restaurants reviews would contain opinion words such as “tender”, “tasty”, or “undercooked” and movie reviews would contain “thrilling”, “horrific”, or “hilarious”. The intersection between these two sets of opinion words could be small which makes domain adaptation difficult.
Several techniques have been proposed for addressing the problem of domain shifting. The aim is to bridge the source and target domains by learning domain-invariant feature representations so that a classifier trained on a source domain can be adapted to another target domain. In cross-domain sentiment classification, many works BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 utilize a key intuition that domain-specific features could be aligned with the help of domain-invariant features (pivot features). For instance, “hilarious” and “tasty” could be aligned as both of them are relevant to “good”.
Despite their promising results, these works share two major limitations. First, they highly depend on the heuristic selection of pivot features, which may be sensitive to different applications. Thus the learned new representations may not effectively reduce the domain difference. Furthermore, these works only utilize the unlabeled target data for representation learning while the sentiment classifier was solely trained on the source domain. There have not been many studies on exploiting unlabeled target data for refining the classifier, even though it may contain beneficial information. How to effectively leverage unlabeled target data still remains an important challenge for domain adaptation.
In this work, we argue that the information from unlabeled target data is beneficial for domain adaptation and we propose a novel Domain Adaptive Semi-supervised learning framework (DAS) to better exploit it. Our main intuition is to treat the problem as a semi-supervised learning task by considering target instances as unlabeled data, assuming the domain distance can be effectively reduced through domain-invariant representation learning. Specifically, the proposed approach jointly performs feature adaptation and semi-supervised learning in a multi-task learning setting. For feature adaptation, it explicitly minimizes the distance between the encoded representations of the two domains. On this basis, two semi-supervised regularizations – entropy minimization and self-ensemble bootstrapping – are jointly employed to exploit unlabeled target data for classifier refinement.
We evaluate our method rigorously under multiple experimental settings by taking label distribution and corpus size into consideration. The results show that our model is able to obtain significant improvements over strong baselines. We also demonstrate through a series of analysis that the proposed method benefits greatly from incorporating unlabeled target data via semi-supervised learning, which is consistent with our motivation. Our datasets and source code can be obtained from https://github.com/ruidan/DAS.
Related Work
Domain Adaptation: The majority of feature adaptation methods for sentiment analysis rely on a key intuition that even though certain opinion words are completely distinct for each domain, they can be aligned if they have high correlation with some domain-invariant opinion words (pivot words) such as “excellent” or “terrible”. Blitzer et al. ( BIBREF0 ) proposed a method based on structural correspondence learning (SCL), which uses pivot feature prediction to induce a projected feature space that works well for both the source and the target domains. The pivot words are selected in a way to cover common domain-invariant opinion words. Subsequent research aims to better align the domain-specific words BIBREF1 , BIBREF5 , BIBREF3 such that the domain discrepancy could be reduced. More recently, Yu and Jiang ( BIBREF4 ) borrow the idea of pivot feature prediction from SCL and extend it to a neural network-based solution with auxiliary tasks. In their experiment, substantial improvement over SCL has been observed due to the use of real-valued word embeddings. Unsupervised representation learning with deep neural networks (DNN) such as denoising autoencoders has also been explored for feature adaptation BIBREF6 , BIBREF7 , BIBREF8 . It has been shown that DNNs could learn transferable representations that disentangle the underlying factors of variation behind data samples.
Although the aforementioned methods aim to reduce the domain discrepancy, they do not explicitly minimize the distance between distributions, and some of them highly rely on the selection of pivot features. In our method, we formally construct an objective for this purpose. Similar ideas have been explored in many computer vision problems, where the representations of the underlying domains are encouraged to be similar through explicit objectives BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 such as maximum mean discrepancy (MMD) BIBREF14 . In NLP tasks, Li et al. ( BIBREF15 ) and Chen et al. ( BIBREF16 ) both proposed using adversarial training framework for reducing domain difference. In their model, a sub-network is added as a domain discriminator while deep features are learned to confuse the discriminator. The feature adaptation component in our model shares similar intuition with MMD and adversary training. We will show a detailed comparison with them in our experiments.
Semi-supervised Learning: We attempt to treat domain adaptation as a semi-supervised learning task by considering the target instances as unlabeled data. Some efforts have been initiated on transfer learning from unlabeled data BIBREF17 , BIBREF18 , BIBREF19 . In our model, we reduce the domain discrepancy by feature adaptation, and thereafter adopt semi-supervised learning techniques to learn from unlabeled data. Primarily motivated by BIBREF20 and BIBREF21 , we employed entropy minimization and self-ensemble bootstrapping as regularizations to incorporate unlabeled data. Our experimental results show that both methods are effective when jointly trained with the feature adaptation objective, which confirms to our motivation.
Notations and Model Overview
We conduct most of our experiments under an unsupervised domain adaptation setting, where we have no labeled data from the target domain. Consider two sets INLINEFORM0 and INLINEFORM1 . INLINEFORM2 is from the source domain with INLINEFORM3 labeled examples, where INLINEFORM4 is a one-hot vector representation of sentiment label and INLINEFORM5 denotes the number of classes. INLINEFORM6 is from the target domain with INLINEFORM7 unlabeled examples. INLINEFORM8 denotes the total number of training documents including both labeled and unlabeled. We aim to learn a sentiment classifier from INLINEFORM13 and INLINEFORM14 such that the classifier would work well on the target domain. We also present some results under a setting where we assume that a small number of labeled target examples are available (see Figure FIGREF27 ).
For the proposed model, we denote INLINEFORM0 parameterized by INLINEFORM1 as a neural-based feature encoder that maps documents from both domains to a shared feature space, and INLINEFORM2 parameterized by INLINEFORM3 as a fully connected layer with softmax activation serving as the sentiment classifier. We aim to learn feature representations that are domain-invariant and at the same time discriminative on both domains, thus we simultaneously consider three factors in our objective: (1) minimize the classification error on the labeled source examples; (2) minimize the domain discrepancy; and (3) leverage unlabeled data via semi-supervised learning.
Suppose we already have the encoded features of documents INLINEFORM0 (see Section SECREF10 ), the objective function for purpose (1) is thus the cross entropy loss on the labeled source examples DISPLAYFORM0
where INLINEFORM0 denotes the predicted label distribution. In the following subsections, we will explain how to perform feature adaptation and domain adaptive semi-supervised learning in details for purpose (2) and (3) respectively.
Feature Adaptation
Unlike prior works BIBREF0 , BIBREF4 , our method does not attempt to align domain-specific words through pivot words. In our preliminary experiments, we found that word embeddings pre-trained on a large corpus are able to adequately capture this information. As we will later show in our experiments, even without adaptation, a naive neural network classifier with pre-trained word embeddings can already achieve reasonably good results.
We attempt to explicitly minimize the distance between the source and target feature representations ( INLINEFORM0 and INLINEFORM1 ). A few methods from literature can be applied such as Maximum Mean Discrepancy (MMD) BIBREF14 or adversary training BIBREF15 , BIBREF16 . The main idea of MMD is to estimate the distance between two distributions as the distance between sample means of the projected embeddings in Hilbert space. MMD is implicitly computed through a characteristic kernel, which is used to ensure that the sample mean is injective, leading to the MMD being zero if and only if the distributions are identical. In our implementation, we skip the mapping procedure induced by a characteristic kernel for simplifying the computation and learning. We simply estimate the distribution distance as the distance between the sample means in the current embedding space. Although this approximation cannot preserve all statistical features of the underlying distributions, we find it performs comparably to MMD on our problem. The following equations formally describe the feature adaptation loss INLINEFORM2 : DISPLAYFORM0
INLINEFORM0 normalization is applied on the mean representations INLINEFORM1 and INLINEFORM2 , rescaling the vectors such that all entries sum to 1. We adopt a symmetric version of KL divergence BIBREF12 as the distance function. Given two distribution vectors INLINEFORM3 , INLINEFORM4 .
Domain Adaptive Semi-supervised Learning (DAS)
We attempt to exploit the information in target data through semi-supervised learning objectives, which are jointly trained with INLINEFORM0 and INLINEFORM1 . Normally, to incorporate target data, we can minimize the cross entropy loss between the true label distributions INLINEFORM2 and the predicted label distributions INLINEFORM3 over target samples. The challenge here is that INLINEFORM4 is unknown, and thus we attempt to estimate it via semi-supervised learning. We use entropy minimization and bootstrapping for this purpose. We will later show in our experiments that both methods are effective, and jointly employing them overall yields the best results.
Entropy Minimization: In this method, INLINEFORM0 is estimated as the predicted label distribution INLINEFORM1 , which is a function of INLINEFORM2 and INLINEFORM3 . The loss can thus be written as DISPLAYFORM0
Assume the domain discrepancy can be effectively reduced through feature adaptation, by minimizing the entropy penalty, training of the classifier is influenced by the unlabeled target data and will generally maximize the margins between the target examples and the decision boundaries, increasing the prediction confidence on the target domain.
Self-ensemble Bootstrapping: Another way to estimate INLINEFORM0 corresponds to bootstrapping. The idea is to estimate the unknown labels as the predictions of the model learned from the previous round of training. Bootstrapping has been explored for domain adaptation in previous works BIBREF18 , BIBREF19 . However, in their methods, domain discrepancy was not explicitly minimized via feature adaptation. Applying bootstrapping or other semi-supervised learning techniques in this case may worsen the results as the classifier can perform quite bad on the target data.
[t] Pseudocode for training DAS INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3 INLINEFORM4 = ensembling momentum, INLINEFORM5 INLINEFORM6 = weight ramp-up function INLINEFORM7 INLINEFORM8 INLINEFORM9 each minibatch INLINEFORM10 , INLINEFORM11 , INLINEFORM12 in
INLINEFORM0 , INLINEFORM1 , INLINEFORM2 compute loss INLINEFORM3 on INLINEFORM4 compute loss INLINEFORM5 on INLINEFORM6 compute loss INLINEFORM7 on INLINEFORM8 compute loss INLINEFORM9 on INLINEFORM10 INLINEFORM11
update network parameters INLINEFORM0 , for INLINEFORM1 INLINEFORM2 INLINEFORM3
Inspired by the ensembling method proposed in BIBREF21 , we estimate INLINEFORM0 by forming ensemble predictions of labels during training, using the outputs on different training epochs. The loss is formulated as follows: DISPLAYFORM0
where INLINEFORM0 denotes the estimated labels computed on the ensemble predictions from different epochs. The loss is applied on all documents. It serves for bootstrapping on the unlabeled target data, and it also serves as a regularization that encourages the network predictions to be consistent in different training epochs. INLINEFORM1 is jointly trained with INLINEFORM2 , INLINEFORM3 , and INLINEFORM4 . Algorithm SECREF6 illustrates the overall training process of the proposed domain adaptive semi-supervised learning (DAS) framework.
In Algorithm SECREF6 , INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 are weights to balance the effects of INLINEFORM3 , INLINEFORM4 , and INLINEFORM5 respectively. INLINEFORM6 and INLINEFORM7 are constant hyper-parameters. We set INLINEFORM8 as a Gaussian curve to ramp up the weight from 0 to INLINEFORM9 . This is to ensure the ramp-up of the bootstrapping loss component is slow enough in the beginning of the training. After each training epoch, we compute INLINEFORM10 which denotes the predictions made by the network in current epoch, and then the ensemble prediction INLINEFORM11 is updated as a weighted average of the outputs from previous epochs and the current epoch, with recent epochs having larger weight. For generating estimated labels INLINEFORM12 , INLINEFORM13 is converted to a one-hot vector where the entry with the maximum value is set to one and other entries are set to zeros. The self-ensemble bootstrapping is a generalized version of bootstrappings that only use the outputs from the previous round of training BIBREF18 , BIBREF19 . The ensemble prediction is likely to be closer to the correct, unknown labels of the target data.
CNN Encoder Implementation
We have left the feature encoder INLINEFORM0 unspecified, for which, a few options can be considered. In our implementation, we adopt a one-layer CNN structure from previous works BIBREF22 , BIBREF4 , as it has been demonstrated to work well for sentiment classification tasks. Given a review document INLINEFORM1 consisting of INLINEFORM2 words, we begin by associating each word with a continuous word embedding BIBREF23 INLINEFORM3 from an embedding matrix INLINEFORM4 , where INLINEFORM5 is the vocabulary size and INLINEFORM6 is the embedding dimension. INLINEFORM7 is jointly updated with other network parameters during training. Given a window of dense word embeddings INLINEFORM8 , the convolution layer first concatenates these vectors to form a vector INLINEFORM9 of length INLINEFORM10 and then the output vector is computed by Equation ( EQREF11 ): DISPLAYFORM0
INLINEFORM0 , INLINEFORM1 is the parameter set of the encoder INLINEFORM2 and is shared across all windows of the sequence. INLINEFORM3 is an element-wise non-linear activation function. The convolution operation can capture local contextual dependencies of the input sequence and the extracted feature vectors are similar to INLINEFORM4 -grams. After the convolution operation is applied to the whole sequence, we obtain a list of hidden vectors INLINEFORM5 . A max-over-time pooling layer is applied to obtain the final vector representation INLINEFORM6 of the input document.
Datasets and Experimental Settings
Existing benchmark datasets such as the Amazon benchmark BIBREF0 typically remove reviews with neutral labels in both domains. This is problematic as the label information of the target domain is not accessible in an unsupervised domain adaptation setting. Furthermore, removing neutral instances may bias the dataset favorably for max-margin-based algorithms like ours, since the resulting dataset has all uncertain labels removed, leaving only high confidence examples. Therefore, we construct new datasets by ourselves. The results on the original Amazon benchmark is qualitatively similar, and we present them in Appendix SECREF6 for completeness since most of previous works reported results on it.
Small-scale datasets: Our new dataset was derived from the large-scale Amazon datasets released by McAuley et al. ( BIBREF24 ). It contains four domains: Book (BK), Electronics (E), Beauty (BT), and Music (M). Each domain contains two datasets. Set 1 contains 6000 instances with exactly balanced class labels, and set 2 contains 6000 instances that are randomly sampled from the large dataset, preserving the original label distribution, which we believe better reflects the label distribution in real life. The examples in these two sets do not overlap. Detailed statistics of the generated datasets are given in Table TABREF9 .
In all our experiments on the small-scale datasets, we use set 1 of the source domain as the only source with sentiment label information during training, and we evaluate the trained model on set 1 of the target domain. Since we cannot control the label distribution of unlabeled data during training, we consider two different settings:
Setting (1): Only set 1 of the target domain is used as the unlabeled set. This tells us how the method performs in a condition when the target domain has a close-to-balanced label distribution. As we also evaluate on set 1 of the target domain, this is also considered as a transductive setting.
Setting (2): Set 2 from both the source and target domains are used as unlabeled sets. Since set 2 is directly sampled from millions of reviews, it better reflects real-life sentiment distribution.
Large-scale datasets: We further conduct experiments on four much larger datasets: IMDB (I), Yelp2014 (Y), Cell Phone (C), and Baby (B). IMDB and Yelp2014 were previously used in BIBREF25 , BIBREF26 . Cell phone and Baby are from the large-scale Amazon dataset BIBREF24 , BIBREF27 . Detailed statistics are summarized in Table TABREF9 . We keep all reviews in the original datasets and consider a transductive setting where all target examples are used for both training (without label information) and evaluation. We perform sampling to balance the classes of labeled source data in each minibatch INLINEFORM3 during training.
Selection of Development Set
Ideally, the development set should be drawn from the same distribution as the test set. However, under the unsupervised domain adaptation setting, we do not have any labeled target data at training phase which could be used as development set. In all of our experiments, for each pair of domains, we instead sample 1000 examples from the training set of the source domain as development set. We train the network for a fixed number of epochs, and the model with the minimum classification error on this development set is saved for evaluation. This approach works well on most of the problems since the target domain is supposed to behave like the source domain if the domain difference is effectively reduced.
Another problem is how to select the values for hyper-parameters. If we tune INLINEFORM0 and INLINEFORM1 directly on the development set from the source domain, most likely both of them will be set to 0, as unlabeled target data is not helpful for improving in-domain accuracy of the source domain. Other neural network models also have the same problem for hyper-parameter tuning. Therefore, our strategy is to use the development set from the target domain to optimize INLINEFORM2 and INLINEFORM3 for one problem (e.g., we only do this on E INLINEFORM4 BK), and fix their values on the other problems. This setting assumes that we have at least two labeled domains such that we can optimize the hyper-parameters, and then we fix them for other new unlabeled domains to transfer to.
Training Details and Hyper-parameters
We initialize word embeddings using the 300-dimension GloVe vectors supplied by Pennington et al., ( BIBREF28 ), which were trained on 840 billion tokens from the Common Crawl. For each pair of domains, the vocabulary consists of the top 10000 most frequent words. For words in the vocabulary but not present in the pre-trained embeddings, we randomly initialize them.
We set hyper-parameters of the CNN encoder following previous works BIBREF22 , BIBREF4 without specific tuning on our datasets. The window size is set to 3 and the size of the hidden layer is set to 300. The nonlinear activation function is Relu. For regularization, we also follow their settings and employ dropout with probability set to 0.5 on INLINEFORM0 before feeding it to the output layer INLINEFORM1 , and constrain the INLINEFORM2 -norm of the weight vector INLINEFORM3 , setting its max norm to 3.
On the small-scale datasets and the Aamzon benchmark, INLINEFORM0 and INLINEFORM1 are set to 200 and 1, respectively, tuned on the development set of task E INLINEFORM2 BK under setting 1. On the large-scale datasets, INLINEFORM3 and INLINEFORM4 are set to 500 and 0.2, respectively, tuned on I INLINEFORM5 Y. We use a Gaussian curve INLINEFORM6 to ramp up the weight of the bootstrapping loss INLINEFORM7 from 0 to INLINEFORM8 , where INLINEFORM9 denotes the maximum number of training epochs. We train 30 epochs for all experiments. We set INLINEFORM10 to 3 and INLINEFORM11 to 0.5 for all experiments.
The batch size is set to 50 on the small-scale datasets and the Amazon benchmark. We increase the batch size to 250 on the large-scale datasets to reduce the number of iterations. RMSProp optimizer with learning rate set to 0.0005 is used for all experiments.
Models for Comparison
We compare with the following baselines:
(1) Naive: A non-domain-adaptive baseline with bag-of-words representations and SVM classifier trained on the source domain.
(2) mSDA BIBREF7 : This is the state-of-the-art method based on discrete input features. Top 1000 bag-of-words features are kept as pivot features. We set the number of stacked layers to 3 and the corruption probability to 0.5.
(3) NaiveNN: This is a non-domain-adaptive CNN trained on source domain, which is a variant of our model by setting INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 to zeros.
(4) AuxNN BIBREF4 : This is a neural model that exploits auxiliary tasks, which has achieved state-of-the-art results on cross-domain sentiment classification. The sentence encoder used in this model is the same as ours.
(5) ADAN BIBREF16 : This method exploits adversarial training to reduce representation difference between domains. The original paper uses a simple feedforward network as encoder. For fair comparison, we replace it with our CNN-based encoder. We train 5 iterations on the discriminator per iteration on the encoder and sentiment classifier as suggested in their paper.
(6) MMD: MMD has been widely used for minimizing domain discrepancy on images. In those works BIBREF9 , BIBREF13 , variants of deep CNNs are used for encoding images and the MMDs of multiple layers are jointly minimized. In NLP, adding more layers of CNNs may not be very helpful and thus those models from image-related tasks can not be directly applied to our problem. To compare with MMD-based method, we train a model that jointly minimize the classification loss INLINEFORM0 on the source domain and MMD between INLINEFORM1 and INLINEFORM2 . For computing MMD, we use a Gaussian RBF which is a common choice for characteristic kernel.
In addition to the above baselines, we also show results of different variants of our model. DAS as shown in Algorithm SECREF6 denotes our full model. DAS-EM denotes the model with only entropy minimization for semi-supervised learning (set INLINEFORM0 ). DAS-SE denotes the model with only self-ensemble bootstrapping for semi-supervised learning (set INLINEFORM1 ). FANN (feature-adaptation neural network) denotes the model without semi-supervised learning performed (set both INLINEFORM2 and INLINEFORM3 to zeros).
Main Results
Figure FIGREF17 shows the comparison of adaptation results (see Appendix SECREF7 for the exact numerical numbers). We report classification accuracy on the small-scale dataset. For the large-scale dataset, macro-F1 is instead used since the label distribution in the test set is extremely unbalanced. Key observations are summarized as follows. (1) Both DAS-EM and DAS-SE perform better in most cases compared with ADAN, MDD, and FANN, in which only feature adaptation is performed. This demonstrates the effectiveness of the proposed domain adaptive semi-supervised learning framework. DAS-EM is more effective than DAS-SE in most cases, and the full model DAS with both techniques jointly employed overall has the best performance. (2) When comparing the two settings on the small-scale dataset, all domain-adaptive methods generally perform better under setting 1. In setting 1, the target examples are balanced in classes, which can provide more diverse opinion-related features. However, when considering unsupervised domain adaptation, we should not presume the label distribution of the unlabeled data. Thus, it is necessary to conduct experiments using datasets that reflect real-life sentiment distribution as what we did on setting2 and the large-scale dataset. Unfortunately, this is ignored by most of previous works. (3) Word-embeddings are very helpful, as we can see even NaiveNN can substantially outperform mSDA on most tasks.
To see the effect of semi-supervised learning alone, we also conduct experiments by setting INLINEFORM0 to eliminate the effect of feature adaptation. Both entropy minimization and bootstrapping perform very badly in this setting. Entropy minimization gives almost random predictions with accuracy below 0.4, and the results of bootstrapping are also much lower compared to NaiveNN. This suggests that the feature adaptation component is essential. Without it, the learned target representations are less meaningful and discriminative. Applying semi-supervised learning in this case is likely to worsen the results.
Further Analysis
In Figure FIGREF23 , we show the change of accuracy with respect to the percentage of unlabeled data used for training on three particular problems under setting 1. The value at INLINEFORM0 denotes the accuracies of NaiveNN which does not utilize any target data. For DAS, we observe a nonlinear increasing trend where the accuracy quickly improves at the beginning, and then gradually stabilizes. For other methods, this trend is less obvious, and adding more unlabeled data sometimes even worsen the results. This finding again suggests that the proposed approach can better exploit the information from unlabeled data.
We also conduct experiments under a setting with a small number of labeled target examples available. Figure FIGREF27 shows the change of accuracy with respect to the number of labeled target examples added for training. We can observe that DAS is still more effective under this setting, while the performance differences to other methods gradually decrease with the increasing number of labeled target examples.
CNN Filter Analysis
In this subsection, we aim to better understand DAS by analyzing sentiment-related CNN filters. To do that, 1) we first select a list of the most related CNN filters for predicting each sentiment label (positive, negative neutral). Those filters can be identified according to the learned weights INLINEFORM0 of the output layer INLINEFORM1 . Higher weight indicates stronger relatedness. 2) Recall that in our implementation, each CNN filter has a window size of 3 with Relu activation. We can thus represent each selected filter as a ranked list of trigrams with highest activation values.
We analyze the CNN filters learned by NaiveNN, FANN and DAS respectively on task E INLINEFORM0 BT under setting 1. We focus on E INLINEFORM1 BT for study because electronics and beauty are very different domains and each of them has a diverse set of domain-specific sentiment expressions. For each method, we identify the top 10 most related filters for each sentiment label, and extract the top trigrams of each selected filter on both source and target domains. Since labeled source examples are used for training, we find the filters learned by the three methods capture similar expressions on the source domain, containing both domain-invariant and domain-specific trigrams. On the target domain, DAS captures more target-specific expressions compared to the other two methods. Due to space limitation, we only present a small subset of positive-sentiment-related filters in Table TABREF34 . The complete results are provided in Appendix SECREF8 . From Table TABREF34 , we can observe that the filters learned by NaiveNN are almost unable to capture target-specific sentiment expressions, while FANN is able to capture limited target-specific words such as “clean” and “scent”. The filters learned by DAS are more domain-adaptive, capturing diverse sentiment expressions in the target domain.
Conclusion
In this work, we propose DAS, a novel framework that jointly performs feature adaptation and semi-supervised learning. We have demonstrated through multiple experiments that DAS can better leverage unlabeled data, and achieve substantial improvements over baseline methods. We have also shown that feature adaptation is an essential component, without which, semi-supervised learning is not able to function properly. The proposed framework could be potentially adapted to other domain adaptation tasks, which is the focus of our future studies.
Results on Amazon Benchmark
Most previous works BIBREF0 , BIBREF1 , BIBREF6 , BIBREF7 , BIBREF29 carried out experiments on the Amazon benchmark released by Blitzer et al. ( BIBREF0 ). The dataset contains 4 different domains: Book (B), DVDs (D), Electronics (E), and Kitchen (K). Following their experimental settings, we consider the binary classification task to predict whether a review is positive or negative on the target domain. Each domain consists of 1000 positive and 1000 negative reviews respectively. We also allow 4000 unlabeled reviews to be used for both the source and the target domains, of which the positive and negative reviews are balanced as well, following the settings in previous works. We construct 12 cross-domain sentiment classification tasks and split the labeled data in each domain into a training set of 1600 reviews and a test set of 400 reviews. The classifier is trained on the training set of the source domain and is evaluated on the test set of the target domain. The comparison results are shown in Table TABREF37 .
Numerical Results of Figure
Due to space limitation, we only show results in figures in the paper. All numerical numbers used for plotting Figure FIGREF17 are presented in Table TABREF38 . We can observe that DAS-EM, DAS-SE, and DAS all achieve substantial improvements over baseline methods under different settings.
CNN Filter Analysis Full Results
As mentioned in Section SECREF36 , we conduct CNN filter analysis on NaiveNN, FANN, and DAS. For each method, we identify the top 10 most related filters for positive, negative, neutral sentiment labels respectively, and then represent each selected filter as a ranked list of trigrams with the highest activation values on it. Table TABREF39 , TABREF40 , TABREF41 in the following pages illustrate the trigrams from the target domain (beauty) captured by the selected filters learned on E INLINEFORM0 BT for each method.
We can observe that compared to NaiveNN and FANN, DAS is able to capture a more diverse set of relevant sentiment expressions on the target domain for each sentiment label. This observation is consistent with our motivation. Since NaiveNN, FANN and other baseline methods solely train the sentiment classifier on the source domain, the learned encoder is not able to produce discriminative features on the target domain. DAS addresses this problem by refining the classifier on the target domain with semi-supervised learning, and the overall objective forces the encoder to learn feature representations that are not only domain-invariant but also discriminative on both domains.
| [
"Book, electronics, beauty, music, IMDB, Yelp, cell phone, baby, DVDs, kitchen",
"we use set 1 of the source domain as the only source with sentiment label information during training, and we evaluate the trained model on set 1 of the target domain, Book (BK), Electronics (E), Beauty (BT), and Music (M)"
] | 5,061 | qasper | en | null | c92d96ed55bc5dcd92f963d3c5d26e52661b74e71090f24a | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What are the source and target domains?
Answer: | [
"Book, electronics, beauty, music, IMDB, Yelp, cell phone, baby, DVDs, kitchen",
"we use set 1 of the source domain as the only source with sentiment label information during training, and we evaluate the trained model on set 1 of the target domain, Book (BK), Electronics (E), Beauty (BT), and Music (M)"
] | qasper | 128 | 44 | |
what previous RNN models do they compare with? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Long short term memory (LSTM) units BIBREF1 are popular for many sequence modeling tasks and are used extensively in language modeling. A key to their success is their articulated gating structure, which allows for more control over the information passed along the recurrence. However, despite the sophistication of the gating mechanisms employed in LSTMs and similar recurrent units, the input and context vectors are treated with simple linear transformations prior to gating. Non-linear transformations such as convolutions BIBREF2 have been used, but these have not achieved the performance of well regularized LSTMs for language modeling BIBREF3 .
A natural way to improve the expressiveness of linear transformations is to increase the number of dimensions of the input and context vectors, but this comes with a significant increase in the number of parameters which may limit generalizability. An example is shown in Figure FIGREF1 , where LSTMs performance decreases with the increase in dimensions of the input and context vectors. Moreover, the semantics of the input and context vectors are different, suggesting that each may benefit from specialized treatment.
Guided by these insights, we introduce a new recurrent unit, the Pyramidal Recurrent Unit (PRU), which is based on the LSTM gating structure. Figure FIGREF2 provides an overview of the PRU. At the heart of the PRU is the pyramidal transformation (PT), which uses subsampling to effect multiple views of the input vector. The subsampled representations are combined in a pyramidal fusion structure, resulting in richer interactions between the individual dimensions of the input vector than is possible with a linear transformation. Context vectors, which have already undergone this transformation in the previous cell, are modified with a grouped linear transformation (GLT) which allows the network to learn latent representations in high dimensional space with fewer parameters and better generalizability (see Figure FIGREF1 ).
We show that PRUs can better model contextual information and demonstrate performance gains on the task of language modeling. The PRU improves the perplexity of the current state-of-the-art language model BIBREF0 by up to 1.3 points, reaching perplexities of 56.56 and 64.53 on the Penn Treebank and WikiText2 datasets while learning 15-20% fewer parameters. Replacing an LSTM with a PRU results in improvements in perplexity across a variety of experimental settings. We provide detailed ablations which motivate the design of the PRU architecture, as well as detailed analysis of the effect of the PRU on other components of the language model.
Related work
Multiple methods, including a variety of gating structures and transformations, have been proposed to improve the performance of recurrent neural networks (RNNs). We first describe these approaches and then provide an overview of recent work in language modeling.
Pyramidal Recurrent Units
We introduce Pyramidal Recurrent Units (PRUs), a new RNN architecture which improves modeling of context by allowing for higher dimensional vector representations while learning fewer parameters. Figure FIGREF2 provides an overview of PRU. We first elaborate on the details of the pyramidal transformation and the grouped linear transformation. We then describe our recurrent unit, PRU.
Pyramidal transformation for input
The basic transformation in many recurrent units is a linear transformation INLINEFORM0 defined as: DISPLAYFORM0
where INLINEFORM0 are learned weights that linearly map INLINEFORM1 to INLINEFORM2 . To simplify notation, we omit the biases.
Motivated by successful applications of sub-sampling in computer vision (e.g., BIBREF22 , BIBREF23 , BIBREF9 , BIBREF24 ), we subsample input vector INLINEFORM0 into INLINEFORM1 pyramidal levels to achieve representation of the input vector at multiple scales. This sub-sampling operation produces INLINEFORM2 vectors, represented as INLINEFORM3 , where INLINEFORM4 is the sampling rate and INLINEFORM5 . We learn scale-specific transformations INLINEFORM6 for each INLINEFORM7 . The transformed subsamples are concatenated to produce the pyramidal analog to INLINEFORM8 , here denoted as INLINEFORM9 : DISPLAYFORM0
where INLINEFORM0 indicates concatenation. We note that pyramidal transformation with INLINEFORM1 is the same as the linear transformation.
To improve gradient flow inside the recurrent unit, we combine the input and output using an element-wise sum (when dimension matches) to produce residual analog of pyramidal transformation, as shown in Figure FIGREF2 BIBREF25 .
We sub-sample the input vector INLINEFORM0 into INLINEFORM1 pyramidal levels using the kernel-based approach BIBREF8 , BIBREF9 . Let us assume that we have a kernel INLINEFORM2 with INLINEFORM3 elements. Then, the input vector INLINEFORM4 can be sub-sampled as: DISPLAYFORM0
where INLINEFORM0 represents the stride and INLINEFORM1 .
The number of parameters learned by the linear transformation and the pyramidal transformation with INLINEFORM0 pyramidal levels to map INLINEFORM1 to INLINEFORM2 are INLINEFORM3 and INLINEFORM4 respectively. Thus, pyramidal transformation reduces the parameters of a linear transformation by a factor of INLINEFORM5 . For example, the pyramidal transformation (with INLINEFORM6 and INLINEFORM7 ) learns INLINEFORM8 fewer parameters than the linear transformation.
Grouped linear transformation for context
Many RNN architectures apply linear transformations to both the input and context vector. However, this may not be ideal due to the differing semantics of each vector. In many NLP applications including language modeling, the input vector is a dense word embedding which is shared across all contexts for a given word in a dataset. In contrast, the context vector is highly contextualized by the current sequence. The differences between the input and context vector motivate their separate treatment in the PRU architecture.
The weights learned using the linear transformation (Eq. EQREF9 ) are reused over multiple time steps, which makes them prone to over-fitting BIBREF26 . To combat over-fitting, various methods, such as variational dropout BIBREF26 and weight dropout BIBREF0 , have been proposed to regularize these recurrent connections. To further improve generalization abilities while simultaneously enabling the recurrent unit to learn representations at very high dimensional space, we propose to use grouped linear transformation (GLT) instead of standard linear transformation for recurrent connections BIBREF27 . While pyramidal and linear transformations can be applied to transform context vectors, our experimental results in Section SECREF39 suggests that GLTs are more effective.
The linear transformation INLINEFORM0 maps INLINEFORM1 linearly to INLINEFORM2 . Grouped linear transformations break the linear interactions by factoring the linear transformation into two steps. First, a GLT splits the input vector INLINEFORM3 into INLINEFORM4 smaller groups such that INLINEFORM5 . Second, a linear transformation INLINEFORM6 is applied to map INLINEFORM7 linearly to INLINEFORM8 , for each INLINEFORM9 . The INLINEFORM10 resultant output vectors INLINEFORM11 are concatenated to produce the final output vector INLINEFORM12 . DISPLAYFORM0
GLTs learn representations at low dimensionality. Therefore, a GLT requires INLINEFORM0 fewer parameters than the linear transformation. We note that GLTs are subset of linear transformations. In a linear transformation, each neuron receives an input from each element in the input vector while in a GLT, each neuron receives an input from a subset of the input vector. Therefore, GLT is the same as a linear transformation when INLINEFORM1 .
Pyramidal Recurrent Unit
We extend the basic gating architecture of LSTM with the pyramidal and grouped linear transformations outlined above to produce the Pyramidal Recurrent Unit (PRU), whose improved sequence modeling capacity is evidenced in Section SECREF4 .
At time INLINEFORM0 , the PRU combines the input vector INLINEFORM1 and the previous context vector (or previous hidden state vector) INLINEFORM2 using the following transformation function as: DISPLAYFORM0
where INLINEFORM0 indexes the various gates in the LSTM model, and INLINEFORM1 and INLINEFORM2 represents the pyramidal and grouped linear transformations defined in Eqns. EQREF10 and EQREF15 , respectively.
We will now incorporate INLINEFORM0 into LSTM gating architecture to produce PRU. At time INLINEFORM1 , a PRU cell takes INLINEFORM2 , INLINEFORM3 , and INLINEFORM4 as inputs to produce forget INLINEFORM5 , input INLINEFORM6 , output INLINEFORM7 , and content INLINEFORM8 gate signals. The inputs are combined with these gate signals to produce context vector INLINEFORM9 and cell state INLINEFORM10 . Mathematically, the PRU with the LSTM gating architecture can be defined as: DISPLAYFORM0
where INLINEFORM0 represents the element-wise multiplication operation, and INLINEFORM1 and INLINEFORM2 are the sigmoid and hyperbolic tangent activation functions. We note that LSTM is a special case of PRU when INLINEFORM3 = INLINEFORM4 =1.
Experiments
To showcase the effectiveness of the PRU, we evaluate the performance on two standard datasets for word-level language modeling and compare with state-of-the-art methods. Additionally, we provide a detailed examination of the PRU and its behavior on the language modeling tasks.
Set-up
Following recent works, we compare on two widely used datasets, the Penn Treebank (PTB) BIBREF28 as prepared by BIBREF29 and WikiText2 (WT-2) BIBREF20 . For both datasets, we follow the same training, validation, and test splits as in BIBREF0 .
We extend the language model, AWD-LSTM BIBREF0 , by replacing LSTM layers with PRU. Our model uses 3-layers of PRU with an embedding size of 400. The number of parameters learned by state-of-the-art methods vary from 18M to 66M with majority of the methods learning about 22M to 24M parameters on the PTB dataset. For a fair comparison with state-of-the-art methods, we fix the model size to 19M and vary the value of INLINEFORM0 and hidden layer sizes so that total number of learned parameters is similar across different configurations. We use 1000, 1200, and 1400 as hidden layer sizes for values of INLINEFORM1 =1,2, and 4, respectively. We use the same settings for the WT-2 dataset. We set the number of pyramidal levels INLINEFORM2 to two in our experiments and use average pooling for sub-sampling. These values are selected based on our ablation experiments on the validation set (Section SECREF39 ). We measure the performance of our models in terms of word-level perplexity. We follow the same training strategy as in BIBREF0 .
To understand the effect of regularization methods on the performance of PRUs, we perform experiments under two different settings: (1) Standard dropout: We use a standard dropout BIBREF12 with probability of 0.5 after embedding layer, the output between LSTM layers, and the output of final LSTM layer. (2) Advanced dropout: We use the same dropout techniques with the same dropout values as in BIBREF0 . We call this model as AWD-PRU.
Results
Table TABREF23 compares the performance of the PRU with state-of-the-art methods. We can see that the PRU achieves the best performance with fewer parameters.
PRUs achieve either the same or better performance than LSTMs. In particular, the performance of PRUs improves with the increasing value of INLINEFORM0 . At INLINEFORM1 , PRUs outperform LSTMs by about 4 points on the PTB dataset and by about 3 points on the WT-2 dataset. This is explained in part by the regularization effect of the grouped linear transformation (Figure FIGREF1 ). With grouped linear and pyramidal transformations, PRUs learn rich representations at very high dimensional space while learning fewer parameters. On the other hand, LSTMs overfit to the training data at such high dimensions and learn INLINEFORM2 to INLINEFORM3 more parameters than PRUs.
With the advanced dropouts, the performance of PRUs improves by about 4 points on the PTB dataset and 7 points on the WT-2 dataset. This further improves with finetuning on the PTB (about 2 points) and WT-2 (about 1 point) datasets.
For similar number of parameters, the PRU with standard dropout outperforms most of the state-of-the-art methods by large margin on the PTB dataset (e.g. RAN BIBREF7 by 16 points with 4M less parameters, QRNN BIBREF33 by 16 points with 1M more parameters, and NAS BIBREF31 by 1.58 points with 6M less parameters). With advanced dropouts, the PRU delivers the best performance. On both datasets, the PRU improves the perplexity by about 1 point while learning 15-20% fewer parameters.
PRU is a drop-in replacement for LSTM, therefore, it can improve language models with modern inference techniques such as dynamic evaluation BIBREF21 . When we evaluate PRU-based language models (only with standard dropout) with dynamic evaluation on the PTB test set, the perplexity of PRU ( INLINEFORM0 ) improves from 62.42 to 55.23 while the perplexity of an LSTM ( INLINEFORM1 ) with similar settings improves from 66.29 to 58.79; suggesting that modern inference techniques are equally applicable to PRU-based language models.
Analysis
It is shown above that the PRU can learn representations at higher dimensionality with more generalization power, resulting in performance gains for language modeling. A closer analysis of the impact of the PRU in a language modeling system reveals several factors that help explain how the PRU achieves these gains.
As exemplified in Table TABREF34 , the PRU tends toward more confident decisions, placing more of the probability mass on the top next-word prediction than the LSTM. To quantify this effect, we calculate the entropy of the next-token distribution for both the PRU and the LSTM using 3687 contexts from the PTB validation set. Figure FIGREF32 shows a histogram of the entropies of the distribution, where bins of size 0.23 are used to effect categories. We see that the PRU more often produces lower entropy distributions corresponding to higher confidences for next-token choices. This is evidenced by the mass of the red PRU curve lying in the lower entropy ranges compared to the blue LSTM's curve. The PRU can produce confident decisions in part because more information is encoded in the higher dimensional context vectors.
The PRU has the ability to model individual words at different resolutions through the pyramidal transform; which provides multiple paths for the gradient to the embedding layer (similar to multi-task learning) and improves the flow of information. When considering the embeddings by part of speech, we find that the pyramid level 1 embeddings exhibit higher variance than the LSTM across all POS categories (Figure FIGREF33 ), and that pyramid level 2 embeddings show extremely low variance. We hypothesize that the LSTM must encode both coarse group similarities and individual word differences into the same vector space, reducing the space between individual words of the same category. The PRU can rely on the subsampled embeddings to account for coarse-grained group similarities, allowing for finer individual word distinctions in the embedding layer. This hypothesis is strengthened by the entropy results described above: a model which can make finer distinctions between individual words can more confidently assign probability mass. A model that cannot make these distinctions, such as the LSTM, must spread its probability mass across a larger class of similar words.
Saliency analysis using gradients help identify relevant words in a test sequence that contribute to the prediction BIBREF34 , BIBREF35 , BIBREF36 . These approaches compute the relevance as the squared norm of the gradients obtained through back-propagation. Table TABREF34 visualizes the heatmaps for different sequences. PRUs, in general, give more relevance to contextual words than LSTMs, such as southeast (sample 1), cost (sample 2), face (sample 4), and introduced (sample 5), which help in making more confident decisions. Furthermore, when gradients during back-propagation are visualized BIBREF37 (Table TABREF34 ), we find that PRUs have better gradient coverage than LSTMs, suggesting PRUs use more features than LSTMs that contributes to the decision. This also suggests that PRUs update more parameters at each iteration which results in faster training. Language model in BIBREF0 takes 500 and 750 epochs to converge with PRU and LSTM as a recurrent unit, respectively.
Ablation studies
In this section, we provide a systematic analysis of our design choices. Our training methodology is the same as described in Section SECREF19 with the standard dropouts. For a thorough understanding of our design choices, we use a language model with a single layer of PRU and fix the size of embedding and hidden layers to 600. The word-level perplexities are reported on the validation sets of the PTB and the WT-2 datasets.
The two hyper-parameters that control the trade-off between performance and number of parameters in PRUs are the number of pyramidal levels INLINEFORM0 and groups INLINEFORM1 . Figure FIGREF35 provides a trade-off between perplexity and recurrent unit (RU) parameters.
Variable INLINEFORM0 and fixed INLINEFORM1 : When we increase the number of pyramidal levels INLINEFORM2 at a fixed value of INLINEFORM3 , the performance of the PRU drops by about 1 to 4 points while reducing the total number of recurrent unit parameters by up to 15%. We note that the PRU with INLINEFORM4 at INLINEFORM5 delivers similar performance as the LSTM while learning about 15% fewer recurrent unit parameters.
Fixed INLINEFORM0 and variable INLINEFORM1 : When we vary the value of INLINEFORM2 at fixed number of pyramidal levels INLINEFORM3 , the total number of recurrent unit parameters decreases significantly with a minimal impact on the perplexity. For example, PRUs with INLINEFORM4 and INLINEFORM5 learns 77% fewer recurrent unit parameters while its perplexity (lower is better) increases by about 12% in comparison to LSTMs. Moreover, the decrease in number of parameters at higher value of INLINEFORM6 enables PRUs to learn the representations in high dimensional space with better generalizability (Table TABREF23 ).
Table TABREF43 shows the impact of different transformations of the input vector INLINEFORM0 and the context vector INLINEFORM1 . We make following observations: (1) Using the pyramidal transformation for the input vectors improves the perplexity by about 1 point on both the PTB and WT-2 datasets while reducing the number of recurrent unit parameters by about 14% (see R1 and R4). We note that the performance of the PRU drops by up to 1 point when residual connections are not used (R4 and R6). (2) Using the grouped linear transformation for context vectors reduces the total number of recurrent unit parameters by about 75% while the performance drops by about 11% (see R3 and R4). When we use the pyramidal transformation instead of the linear transformation, the performance drops by up to 2% while there is no significant drop in the number of parameters (R4 and R5).
We set sub-sampling kernel INLINEFORM0 (Eq. EQREF12 ) with stride INLINEFORM1 and size of 3 ( INLINEFORM2 ) in four different ways: (1) Skip: We skip every other element in the input vector. (2) Convolution: We initialize the elements of INLINEFORM3 randomly from normal distribution and learn them during training the model. We limit the output values between -1 and 1 using INLINEFORM4 activation function to make training stable. (3) Avg. pool: We initialize the elements of INLINEFORM5 to INLINEFORM6 . (4) Max pool: We select the maximum value in the kernel window INLINEFORM7 .
Table TABREF45 compares the performance of the PRU with different sampling methods. Average pooling performs the best while skipping give comparable performance. Both of these methods enable the network to learn richer word representations while representing the input vector in different forms, thus delivering higher performance. Surprisingly, a convolution-based sub-sampling method does not perform as well as the averaging method. The INLINEFORM0 function used after convolution limits the range of output values which are further limited by the LSTM gating structure, thereby impeding in the flow of information inside the cell. Max pooling forces the network to learn representations from high magnitude elements, thus distinguishing features between elements vanishes, resulting in poor performance.
Conclusion
We introduce the Pyramidal Recurrent Unit, which better model contextual information by admitting higher dimensional representations with good generalizability. When applied to the task of language modeling, PRUs improve perplexity across several settings, including recent state-of-the-art systems. Our analysis shows that the PRU improves the flow of gradient and expand the word embedding subspace, resulting in more confident decisions. Here we have shown improvements for language modeling. In future, we plan to study the performance of PRUs on different tasks, including machine translation and question answering. In addition, we will study the performance of the PRU on language modeling with more recent inference techniques, such as dynamic evaluation and mixture of softmax.
Acknowledgments
This research was supported by NSF (IIS 1616112, III 1703166), Allen Distinguished Investigator Award, and gifts from Allen Institute for AI, Google, Amazon, and Bloomberg. We are grateful to Aaron Jaech, Hannah Rashkin, Mandar Joshi, Aniruddha Kembhavi, and anonymous reviewers for their helpful comments.
| [
"Variational LSTM, CharCNN, Pointer Sentinel-LSTM, RHN, NAS Cell, SRU, QRNN, RAN, 4-layer skip-connection LSTM, AWD-LSTM, Quantized LSTM"
] | 3,319 | qasper | en | null | eeab2b9167f294f68d3058752acc01e1145eb7d89437d5c4 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: what previous RNN models do they compare with?
Answer: | [
"Variational LSTM, CharCNN, Pointer Sentinel-LSTM, RHN, NAS Cell, SRU, QRNN, RAN, 4-layer skip-connection LSTM, AWD-LSTM, Quantized LSTM"
] | qasper | 128 | 45 | |
What neural network modules are included in NeuronBlocks? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Deep Neural Networks (DNN) have been widely employed in industry for solving various Natural Language Processing (NLP) tasks, such as text classification, sequence labeling, question answering, etc. However, when engineers apply DNN models to address specific NLP tasks, they often face the following challenges.
The above challenges often hinder the productivity of engineers, and result in less optimal solutions to their given tasks. This motivates us to develop an NLP toolkit for DNN models, which facilitates engineers to develop DNN approaches. Before designing this NLP toolkit, we conducted a survey among engineers and identified a spectrum of three typical personas.
To satisfy the requirements of all the above three personas, the NLP toolkit has to be generic enough to cover as many tasks as possible. At the same time, it also needs to be flexible enough to allow alternative network architectures as well as customized modules. Therefore, we analyzed the NLP jobs submitted to a commercial centralized GPU cluster. Table TABREF11 showed that about 87.5% NLP related jobs belong to a few common tasks, including sentence classification, text matching, sequence labeling, MRC, etc. It further suggested that more than 90% of the networks were composed of several common components, such as embedding, CNN/RNN, Transformer and so on.
Based on the above observations, we developed NeuronBlocks, a DNN toolkit for NLP tasks. The basic idea is to provide two layers of support to the engineers. The upper layer targets common NLP tasks. For each task, the toolkit contains several end-to-end network templates, which can be immediately instantiated with simple configuration. The bottom layer consists of a suite of reusable and standard components, which can be adopted as building blocks to construct networks with complex architecture. By following the interface guidelines, users can also contribute to this gallery of components with their own modules.
The technical contributions of NeuronBlocks are summarized into the following three aspects.
Related Work
There are several general-purpose deep learning frameworks, such as TensorFlow, PyTorch and Keras, which have gained popularity in NLP community. These frameworks offer huge flexibility in DNN model design and support various NLP tasks. However, building models under these frameworks requires a large overhead of mastering these framework details. Therefore, higher level abstraction to hide the framework details is favored by many engineers.
There are also several popular deep learning toolkits in NLP, including OpenNMT BIBREF0 , AllenNLP BIBREF1 etc. OpenNMT is an open-source toolkit mainly targeting neural machine translation or other natural language generation tasks. AllenNLP provides several pre-built models for NLP tasks, such as semantic role labeling, machine comprehension, textual entailment, etc. Although these toolkits reduce the development cost, they are limited to certain tasks, and thus not flexible enough to support new network architectures or new components.
Design
The Neuronblocks is built on PyTorch. The overall framework is illustrated in Figure FIGREF16 . It consists of two layers: the Block Zoo and the Model Zoo. In Block Zoo, the most commonly used components of deep neural networks are categorized into several groups according to their functions. Within each category, several alternative components are encapsulated into standard and reusable blocks with a consistent interface. These blocks serve as basic and exchangeable units to construct complex network architectures for different NLP tasks. In Model Zoo, the most popular NLP tasks are identified. For each task, several end-to-end network templates are provided in the form of JSON configuration files. Users can simply browse these configurations and choose one or more to instantiate. The whole task can be completed without any coding efforts.
Block Zoo
We recognize the following major functional categories of neural network components. Each category covers as many commonly used modules as possible. The Block Zoo is an open framework, and more modules can be added in the future.
[itemsep= -0.4em,topsep = 0.3em, align=left, labelsep=-0.6em, leftmargin=1.2em]
Embedding Layer: Word/character embedding and extra handcrafted feature embedding such as pos-tagging are supported.
Neural Network Layers: Block zoo provides common layers like RNN, CNN, QRNN BIBREF2 , Transformer BIBREF3 , Highway network, Encoder Decoder architecture, etc. Furthermore, attention mechanisms are widely used in neural networks. Thus we also support multiple attention layers, such as Linear/Bi-linear Attention, Full Attention BIBREF4 , Bidirectional attention flow BIBREF5 , etc. Meanwhile, regularization layers such as Dropout, Layer Norm, Batch Norm, etc are also supported for improving generalization ability.
Loss Function: Besides of the loss functions built in PyTorch, we offer more options such as Focal Loss BIBREF6 .
Metrics: For classification task, AUC, Accuracy, Precision/Recall, F1 metrics are supported. For sequence labeling task, F1/Accuracy are supported. For knowledge distillation task, MSE/RMSE are supported. For MRC task, ExactMatch/F1 are supported.
Model Zoo
In NeuronBlocks, we identify four types of most popular NLP tasks. For each task, we provide various end-to-end network templates.
[itemsep= -0.4em,topsep = 0.3em, align=left, labelsep=-0.6em, leftmargin=1.2em]
Text Classification and Matching. Tasks such as domain/intent classification, question answer matching are supported.
Sequence Labeling. Predict each token in a sequence into predefined types. Common tasks include NER, POS tagging, Slot tagging, etc.
Knowledge Distillation BIBREF7 . Teacher-Student based knowledge distillation is one common approach for model compression. NeuronBlocks provides knowledge distillation template to improve the inference speed of heavy DNN models like BERT/GPT.
Extractive Machine Reading Comprehension. Given a pair of question and passage, predict the start and end positions of the answer spans in the passage.
User Interface
NeuronBlocks provides convenient user interface for users to build, train, and test DNN models. The details are described in the following.
[itemsep= -0.4em,topsep = 0.3em, align=left, labelsep=-0.6em, leftmargin=1.2em]
I/O interface. This part defines model input/output, such as training data, pre-trained models/embeddings, model saving path, etc.
Model Architecture interface. This is the key part of the configuration file, which defines the whole model architecture. Figure FIGREF19 shows an example of how to specify a model architecture using the blocks in NeuronBlocks. To be more specific, it consists of a list of layers/blocks to construct the architecture, where the blocks are supplied in the gallery of Block Zoo.
Training Parameters interface. In this part, the model optimizer as well as all other training hyper parameters are indicated.
Workflow
Figure FIGREF34 shows the workflow of building DNN models in NeuronBlocks. Users only need to write a JSON configuration file. They can either instantiate an existing template from Model Zoo, or construct a new architecture based on the blocks from Block Zoo. This configuration file is shared across training, test, and prediction. For model hyper-parameter tuning or architecture modification, users just need to change the JSON configuration file. Advanced users can also contribute novel customized blocks into Block Zoo, as long as they follow the same interface guidelines with the existing blocks. These new blocks can be further shared across all users for model architecture design. Moreover, NeuronBlocks has flexible platform support, such as GPU/CPU, GPU management platforms like PAI.
Experiments
To verify the performance of NeuronBlocks, we conducted extensive experiments for common NLP tasks on public data sets including CoNLL-2003 BIBREF14 , GLUE benchmark BIBREF13 , and WikiQA corpus BIBREF15 . The experimental results showed that the models built with NeuronBlocks can achieve reliable and competitive results on various tasks, with productivity greatly improved.
Sequence Labeling
For sequence labeling task, we evaluated NeuronBlocks on CoNLL-2003 BIBREF14 English NER dataset, following most works on the same task. This dataset includes four types of named entities, namely, PERSON, LOCATION, ORGANIZATION, and MISC. We adopted the BIOES tagging scheme instead of IOB, as many previous works indicated meaningful improvement with BIOES scheme BIBREF16 , BIBREF17 . Table TABREF28 shows the results on CoNLL-2003 Englist testb dataset, with 12 different combinations of network layers/blocks, such as word/character embedding, CNN/LSTM and CRF. The results suggest that the flexible combination of layers/blocks in NeuronBlocks can easily reproduce the performance of original models, with comparative or slightly better performance.
GLUE Benchmark
The General Language Understanding Evaluation (GLUE) benchmark BIBREF13 is a collection of natural language understanding tasks. We experimented on the GLUE benchmark tasks using BiLSTM and Attention based models. As shown in Table TABREF29 , the models built by NeuronBlocks can achieve competitive or even better results on GLUE tasks with minimal coding efforts.
Knowledge Distillation
We evaluated Knowledge Distillation task in NeuronBlocks on a dataset collected from one commercial search engine. We refer to this dataset as Domain Classification Dataset. Each sample in this dataset consists of two parts, i.e., a question and a binary label indicating whether the question belongs to a specific domain. Table TABREF36 shows the results, where Area Under Curve (AUC) metric is used as the performance evaluation criteria and Queries per Second (QPS) is used to measure inference speed. By knowledge distillation training approach, the student model by NeuronBlocks managed to get 23-27 times inference speedup with only small performance regression compared with BERTbase fine-tuned classifier.
WikiQA
The WikiQA corpus BIBREF15 is a publicly available dataset for open-domain question answering. This dataset contains 3,047 questions from Bing query logs, each associated with some candidate answer sentences from Wikipedia. We conducted experiments on WikiQA dataset using CNN, BiLSTM, and Attention based models. The results are shown in Table TABREF41 . The models built in NeuronBlocks achieved competitive or even better results with simple model configurations.
Conclusion and Future Work
In this paper, we introduce NeuronBlocks, a DNN toolkit for NLP tasks built on PyTorch. NeuronBlocks targets three types of engineers, and provides a two-layer solution to satisfy the requirements from all three types of users. To be more specific, the Model Zoo consists of various templates for the most common NLP tasks, while the Block Zoo supplies a gallery of alternative layers/modules for the networks. Such design achieves a balance between generality and flexibility. Extensive experiments have verified the effectiveness of this approach. NeuronBlocks has been widely used in a product team of a commercial search engine, and significantly improved the productivity for developing NLP DNN approaches.
As an open-source toolkit, we will further extend it in various directions. The following names a few examples.
| [
"Embedding Layer, Neural Network Layers, Loss Function, Metrics",
"Embedding Layer, Neural Network Layers, Loss Function, Metrics"
] | 1,678 | qasper | en | null | d5280384a4496ef3358dc45dc0199b9198bd0d927be302ef | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What neural network modules are included in NeuronBlocks?
Answer: | [
"Embedding Layer, Neural Network Layers, Loss Function, Metrics",
"Embedding Layer, Neural Network Layers, Loss Function, Metrics"
] | qasper | 128 | 46 | |
what datasets did they use? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Accurate grapheme-to-phoneme conversion (g2p) is important for any application that depends on the sometimes inconsistent relationship between spoken and written language. Most prominently, this includes text-to-speech and automatic speech recognition. Most work on g2p has focused on a few languages for which extensive pronunciation data is available BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 . Most languages lack these resources. However, a low resource language's writing system is likely to be similar to the writing systems of languages that do have sufficient pronunciation data. Therefore g2p may be possible for low resource languages if this high resource data can be properly utilized.
We attempt to leverage high resource data by treating g2p as a multisource neural machine translation (NMT) problem. The source sequences for our system are words in the standard orthography in any language. The target sequences are the corresponding representation in the International Phonetic Alphabet (IPA). Our results show that the parameters learned by the shared encoder–decoder are able to exploit the orthographic and phonemic similarities between the various languages in our data.
Low Resource g2p
Our approach is similar in goal to deri2016grapheme's model for adapting high resource g2p models for low resource languages. They trained weighted finite state transducer (wFST) models on a variety of high resource languages, then transferred those models to low resource languages, using a language distance metric to choose which high resource models to use and a phoneme distance metric to map the high resource language's phonemes to the low resource language's phoneme inventory. These distance metrics are computed based on data from Phoible BIBREF4 and URIEL BIBREF5 .
Other low resource g2p systems have used a strategy of combining multiple models. schlippe2014combining trained several data-driven g2p systems on varying quantities of monolingual data and combined their outputs with a phoneme-level voting scheme. This led to improvements over the best-performing single system for small quantities of data in some languages. jyothilow trained recurrent neural networks for small data sets and found that a version of their system that combined the neural network output with the output of the wFST-based Phonetisaurus system BIBREF1 did better than either system alone.
A different approach came from kim2012universal, who used supervised learning with an undirected graphical model to induce the grapheme–phoneme mappings for languages written in the Latin alphabet. Given a short text in a language, the model predicts the language's orthographic rules. To create phonemic context features from the short text, the model naïvely maps graphemes to IPA symbols written with the same character, and uses the features of these symbols to learn an approximation of the phonotactic constraints of the language. In their experiments, these phonotactic features proved to be more valuable than geographical and genetic features drawn from WALS BIBREF6 .
Multilingual Neural NLP
In recent years, neural networks have emerged as a common way to use data from several languages in a single system. Google's zero-shot neural machine translation system BIBREF7 shares an encoder and decoder across all language pairs. In order to facilitate this multi-way translation, they prepend an artificial token to the beginning of each source sentence at both training and translation time. The token identifies what language the sentence should be translated to. This approach has three benefits: it is far more efficient than building a separate model for each language pair; it allows for translation between languages that share no parallel data; and it improves results on low-resource languages by allowing them to implicitly share parameters with high-resource languages. Our g2p system is inspired by this approach, although it differs in that there is only one target “language”, IPA, and the artificial tokens identify the language of the source instead of the language of the target.
Other work has also made use of multilingually-trained neural networks. Phoneme-level polyglot language models BIBREF8 train a single model on multiple languages and additionally condition on externally constructed typological data about the language. ostling2017continuous used a similar approach, in which a character-level neural language model is trained on a massively multilingual corpus. A language embedding vector is concatenated to the input at each time step. The language embeddings their system learned correlate closely to the genetic relationships between languages. However, neither of these models was applied to g2p.
Grapheme-to-Phoneme
g2p is the problem of converting the orthographic representation of a word into a phonemic representation. A phoneme is an abstract unit of sound which may have different realizations in different contexts. For example, the English phoneme has two phonetic realizations (or allophones):
English speakers without linguistic training often struggle to perceive any difference between these sounds. Writing systems usually do not distinguish between allophones: and are both written as INLINEFORM0 p INLINEFORM1 in English. The sounds are written differently in languages where they contrast, such as Hindi and Eastern Armenian.
Most writing systems in use today are glottographic, meaning that their symbols encode solely phonological information. But despite being glottographic, in few writing systems do graphemes correspond one-to-one with phonemes. There are cases in which multiple graphemes represent a single phoneme, as in the word the in English:
There are cases in which a single grapheme represents multiple phonemes, such as syllabaries, in which each symbol represents a syllable.
In many languages, there are silent letters, as in the word hora in Spanish:
There are more complicated correspondences, such as the silent e in English that affects the pronunciation of the previous vowel, as seen in the pair of words cape and cap.
It is possible for an orthographic system to have any or all of the above phenomena while remaining unambiguous. However, some orthographic systems contain ambiguities. English is well-known for its spelling ambiguities. Abjads, used for Arabic and Hebrew, do not give full representation to vowels.
Consequently, g2p is harder than simply replacing each grapheme symbol with a corresponding phoneme symbol. It is the problem of replacing a grapheme sequence INLINEFORM0
with a phoneme sequence INLINEFORM0
where the sequences are not necessarily of the same length. Data-driven g2p is therefore the problem of finding the phoneme sequence that maximizes the likelihood of the grapheme sequence: INLINEFORM0
Data-driven approaches are especially useful for problems in which the rules that govern them are complex and difficult to engineer by hand. g2p for languages with ambiguous orthographies is such a problem. Multilingual g2p, in which the various languages have similar but different and possibly contradictory spelling rules, can be seen as an extreme case of that. Therefore, a data-driven sequence-to-sequence model is a natural choice.
Encoder–Decoder Models
In order to find the best phoneme sequence, we use a neural encoder–decoder model with attention BIBREF9 . The model consists of two main parts: the encoder compresses each source grapheme sequence INLINEFORM0 into a fixed-length vector. The decoder, conditioned on this fixed-length vector, generates the output phoneme sequence INLINEFORM1 .
The encoder and decoder are both implemented as recurrent neural networks, which have the advantage of being able to process sequences of arbitrary length and use long histories efficiently. They are trained jointly to minimize cross-entropy on the training data. We had our best results when using a bidirectional encoder, which consists of two separate encoders which process the input in forward and reverse directions. We used long short-term memory units BIBREF10 for both the encoder and decoder. For the attention mechanism, we used the general global attention architecture described by luong2015effective.
We implemented all models with OpenNMT BIBREF11 . Our hyperparameters, which we determined by experimentation, are listed in Table TABREF8 .
Training Multilingual Models
Presenting pronunciation data in several languages to the network might create problems because different languages have different pronunciation patterns. For example, the string `real' is pronounced differently in English, German, Spanish, and Portuguese. We solve this problem by prepending each grapheme sequence with an artificial token consisting of the language's ISO 639-3 code enclosed in angle brackets. The English word `real', for example, would be presented to the system as
INLINEFORM0 eng INLINEFORM1 r e a l
The artificial token is treated simply as an element of the grapheme sequence. This is similar to the approach taken by johnson2016google in their zero-shot NMT system. However, their source-side artificial tokens identify the target language, whereas ours identify the source language. An alternative approach, used by ostling2017continuous, would be to concatenate a language embedding to the input at each time step. They do not evaluate their approach on grapheme-to-phoneme conversion.
Data
In order to train a neural g2p system, one needs a large quantity of pronunciation data. A standard dataset for g2p is the Carnegie Mellon Pronouncing Dictionary BIBREF12 . However, that is a monolingual English resource, so it is unsuitable for our multilingual task. Instead, we use the multilingual pronunciation corpus collected by deri2016grapheme for all experiments. This corpus consists of spelling–pronunciation pairs extracted from Wiktionary. It is already partitioned into training and test sets. Corpus statistics are presented in Table TABREF10 .
In addition to the raw IPA transcriptions extracted from Wiktionary, the corpus provides an automatically cleaned version of transcriptions. Cleaning is a necessary step because web-scraped data is often noisy and may be transcribed at an inconsistent level of detail. The data cleaning used here attempts to make the transcriptions consistent with the phonemic inventories used in Phoible BIBREF4 . When a transcription contains a phoneme that is not in its language's inventory in Phoible, that phoneme is replaced by the phoneme with the most similar articulatory features that is in the language's inventory. Sometimes this cleaning algorithm works well: in the German examples in Table TABREF11 , the raw German symbols and are both converted to . This is useful because the in Ansbach and the in Kaninchen are instances of the same phoneme, so their phonemic representations should use the same symbol. However, the cleaning algorithm can also have negative effects on the data quality. For example, the phoneme is not present in the Phoible inventory for German, but it is used in several German transcriptions in the corpus. The cleaning algorithm converts to in all German transcriptions, whereas would be a more reasonable guess. The cleaning algorithm also removes most suprasegmentals, even though these are often an important part of a language's phonology. Developing a more sophisticated procedure for cleaning pronunciation data is a direction for future work, but in this paper we use the corpus's provided cleaned transcriptions in order to ease comparison to previous results.
Experiments
We present experiments with two versions of our sequence-to-sequence model. LangID prepends each training, validation, and test sample with an artificial token identifying the language of the sample. NoLangID omits this token. LangID and NoLangID have identical structure otherwise. To translate the test corpus, we used a beam width of 100. Although this is an unusually wide beam and had negligible performance effects, it was necessary to compute our error metrics.
Evaluation
We use the following three evaluation metrics:
Phoneme Error Rate (PER) is the Levenshtein distance between the predicted phoneme sequences and the gold standard phoneme sequences, divided by the length of the gold standard phoneme sequences.
Word Error Rate (WER) is the percentage of words in which the predicted phoneme sequence does not exactly match the gold standard phoneme sequence.
Word Error Rate 100 (WER 100) is the percentage of words in the test set for which the correct guess is not in the first 100 guesses of the system.
In system evaluations, WER, WER 100, and PER numbers presented for multiple languages are averaged, weighting each language equally BIBREF13 .
It would be interesting to compute error metrics that incorporate phoneme similarity, such as those proposed by hixon2011phonemic. PER weights all phoneme errors the same, even though some errors are more harmful than others: and are usually contrastive, whereas and almost never are. Such statistics would be especially interesting for evaluating a multilingual system, because different languages often map the same grapheme to phonemes that are only subtly different from each other. However, these statistics have not been widely reported for other g2p systems, so we omit them here.
Baseline
Results on LangID and NoLangID are compared to the system presented by deri2016grapheme, which is identified in our results as wFST. Their results can be divided into two parts:
High resource results, computed with wFSTs trained on a combination of Wiktionary pronunciation data and g2p rules extracted from Wikipedia IPA Help pages. They report high resource results for 85 languages.
Adapted results, where they apply various mapping strategies in order to adapt high resource models to other languages. The final adapted results they reported include most of the 85 languages with high resource results, as well as the various languages they were able to adapt them for, for a total of 229 languages. This test set omits 23 of the high resource languages that are written in unique scripts or for which language distance metrics could not be computed.
Training
We train the LangID and NoLangID versions of our model each on three subsets of the Wiktionary data:
LangID-High and NoLangID-High: Trained on data from the 85 languages for which BIBREF13 used non-adapted wFST models.
LangID-Adapted and NoLangID-Adapted: Trained on data from any of the 229 languages for which they built adapted models. Because many of these languages had no training data at all, the model is actually only trained on data in 157 languages. As is noted above, the Adapted set omits 23 languages which are in the High test set.
LangID-All and NoLangID-All: Trained on data in all 311 languages in the Wiktionary training corpus.
In order to ease comparison to Deri and Knight's system, we limited our use of the training corpus to 10,000 words per language. We set aside 10 percent of the data in each language for validation, so the maximum number of training words for any language is 9000 for our systems.
Adapted Results
On the 229 languages for which deri2016grapheme presented their final results, the LangID version of our system outperforms the baseline by a wide margin. The best performance came with the version of our model that was trained on data in all available languages, not just the languages it was tested on. Using a language ID token improves results considerably, but even NoLangID beats the baseline in WER and WER 100. Full results are presented in Table TABREF24 .
High Resource Results
Having shown that our model exceeds the performance of the wFST-adaptation approach, we next compare it to the baseline models for just high resource languages. The wFST models here are purely monolingual – they do not use data adaptation because there is sufficient training data for each of them. Full results are presented in Table TABREF26 . We omit models trained on the Adapted languages because they were not trained on high resource languages with unique writing systems, such as Georgian and Greek, and consequently performed very poorly on them.
In contrast to the larger-scale Adapted results, in the High Resource experiments none of the sequence-to-sequence approaches equal the performance of the wFST model in WER and PER, although LangID-High does come close. The LangID models do beat wFST in WER 100. A possible explanation is that a monolingual wFST model will never generate phonemes that are not part of the language's inventory. A multilingual model, on the other hand, could potentially generate phonemes from the inventories of any language it has been trained on.
Even if LangID-High does not present a more accurate result, it does present a more compact one: LangID-High is 15.4 MB, while the combined wFST high resource models are 197.5 MB.
Results on Unseen Languages
Finally, we report our models' results on unseen languages in Table TABREF28 . The unseen languages are any that are present in the test corpus but absent from the training data. Deri and Knight did not report results specifically on these languages. Although the NoLangID models sometimes do better on WER 100, even here the LangID models have a slight advantage in WER and PER. This is somewhat surprising because the LangID models have not learned embeddings for the language ID tokens of unseen languages. Perhaps negative associations are also being learned, driving the model towards predicting more common pronunciations for unseen languages.
Language ID Tokens
Adding a language ID token always improves results in cases where an embedding has been learned for that token. The power of these embeddings is demonstrated by what happens when one feeds the same input word to the model with different language tokens, as is seen in Table TABREF30 . Impressively, this even works when the source sequence is in the wrong script for the language, as is seen in the entry for Arabic.
Language Embeddings
Because these language ID tokens are so useful, it would be good if they could be effectively estimated for unseen languages. ostling2017continuous found that the language vectors their models learned correlated well to genetic relationships, so it would be interesting to see if the embeddings our source encoder learned for the language ID tokens showed anything similar. In a few cases they do (the languages closest to German in the vector space are Luxembourgish, Bavarian, and Yiddish, all close relatives). However, for the most part the structure of these vectors is not interpretable. Therefore, it would be difficult to estimate the embedding for an unseen language, or to “borrow” the language ID token of a similar language. A more promising way forward is to find a model that uses an externally constructed typological representation of the language.
Phoneme Embeddings
In contrast to the language embeddings, the phoneme embeddings appear to show many regularities (see Table TABREF33 ). This is a sign that our multilingual model learns similar embeddings for phonemes that are written with the same grapheme in different languages. These phonemes tend to be phonetically similar to each other.
Perhaps the structure of the phoneme embedding space is what leads to our models' very good performance on WER 100. Even when the model's first predicted pronunciation is not correct, it tends to assign more probability mass to guesses that are more similar to the correct one. Applying some sort of filtering or reranking of the system output might therefore lead to better performance.
Future Work
Because the language ID token is so beneficial to performance, it would be very interesting to find ways to extend a similar benefit to unseen languages. One possible way to do so is with tokens that identify something other than the language, such as typological features about the language's phonemic inventory. This could enable better sharing of resources among languages. Such typological knowledge is readily available in databases like Phoible and WALS for a wide variety of languages. It would be interesting to explore if any of these features is a good predictor of a language's orthographic rules.
It would also be interesting to apply the artificial token approach to other problems besides multilingual g2p. One closely related application is monolingual English g2p. Some of the ambiguity of English spelling is due to the wide variety of loanwords in the language, many of which have unassimilated spellings. Knowing the origins of these loanwords could provide a useful hint for figuring out their pronunciations. The etymology of a word could be tagged in an analogous way to how language ID is tagged in multilingual g2p.
| [
"the Carnegie Mellon Pronouncing Dictionary BIBREF12, the multilingual pronunciation corpus collected by deri2016grapheme , ranscriptions extracted from Wiktionary",
"multilingual pronunciation corpus collected by deri2016grapheme"
] | 3,244 | qasper | en | null | 0b90a0a4b2cdceda62a9e2b165f0cfa1a34d34d800a46086 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: what datasets did they use?
Answer: | [
"the Carnegie Mellon Pronouncing Dictionary BIBREF12, the multilingual pronunciation corpus collected by deri2016grapheme , ranscriptions extracted from Wiktionary",
"multilingual pronunciation corpus collected by deri2016grapheme"
] | qasper | 128 | 47 | |
What were the baselines? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The task of speculation detection and scope resolution is critical in distinguishing factual information from speculative information. This has multiple use-cases, like systems that determine the veracity of information, and those that involve requirement analysis. This task is particularly important to the biomedical domain, where patient reports and medical articles often use this feature of natural language. This task is commonly broken down into two subtasks: the first subtask, speculation cue detection, is to identify the uncertainty cue in a sentence, while the second subtask: scope resolution, is to identify the scope of that cue. For instance, consider the example:
It might rain tomorrow.
The speculation cue in the sentence above is ‘might’ and the scope of the cue ‘might’ is ‘rain tomorrow’. Thus, the speculation cue is the word that expresses the speculation, while the words affected by the speculation are in the scope of that cue.
This task was the CoNLL-2010 Shared Task (BIBREF0), which had 3 different subtasks. Task 1B was speculation cue detection on the BioScope Corpus, Task 1W was weasel identification from Wikipedia articles, and Task 2 was speculation scope resolution from the BioScope Corpus. For each task, the participants were provided the train and test set, which is henceforth referred to as Task 1B CoNLL and Task 2 CoNLL throughout this paper.
For our experimentation, we use the sub corpora of the BioScope Corpus (BIBREF1), namely the BioScope Abstracts sub corpora, which is referred to as BA, and the BioScope Full Papers sub corpora, which is referred to as BF. We also use the SFU Review Corpus (BIBREF2), which is referred to as SFU.
This subtask of natural language processing, along with another similar subtask, negation detection and scope resolution, have been the subject of a body of work over the years. The approaches used to solve them have evolved from simple rule-based systems (BIBREF3) based on linguistic information extracted from the sentences, to modern deep-learning based methods. The Machine Learning techniques used varied from Maximum Entropy Classifiers (BIBREF4) to Support Vector Machines (BIBREF5,BIBREF6,BIBREF7,BIBREF8), while the deep learning approaches included Recursive Neural Networks (BIBREF9,BIBREF10), Convolutional Neural Networks (BIBREF11) and most recently transfer learning-based architectures like Bidirectional Encoder Representation from Transformers (BERT) (BIBREF12). Figures FIGREF1 and FIGREF1 contain a summary of the papers addressing speculation detection and scope resolution (BIBREF13, BIBREF5, BIBREF9, BIBREF3, BIBREF14, BIBREF15, BIBREF16, BIBREF17, BIBREF6, BIBREF11, BIBREF18, BIBREF10, BIBREF19, BIBREF7, BIBREF4, BIBREF8).
Inspired by the most recent approach of applying BERT to negation detection and scope resolution (BIBREF12), we take this approach one step further by performing a comparative analysis of three popular transformer-based architectures: BERT (BIBREF20), XLNet (BIBREF21) and RoBERTa (BIBREF22), applied to speculation detection and scope resolution. We evaluate the performance of each model across all datasets via the single dataset training approach, and report all scores including inter-dataset scores (i.e. train on one dataset, evaluate on another) to test the generalizability of the models. This approach outperforms all existing systems on the task of speculation detection and scope resolution. Further, we jointly train on multiple datasets and obtain improvements over the single dataset training approach on most datasets.
Contrary to results observed on benchmark GLUE tasks, we observe XLNet consistently outperforming RoBERTa. To confirm this observation, we apply these models to the negation detection and scope resolution task, and observe a continuity in this trend, reporting state-of-the-art results on three of four datasets on the negation scope resolution task.
The rest of the paper is organized as follows: In Section 2, we provide a detailed description of our methodology and elaborate on the experimentation details. In Section 3, we present our results and analysis on the speculation detection and scope resolution task, using the single dataset and the multiple dataset training approach. In Section 4, we show the results of applying XLNet and RoBERTa on negation detection and scope resolution and propose a few reasons to explain why XLNet performs better than RoBERTa. Finally, the future scope and conclusion is mentioned in Section 5.
Methodology and Experimental Setup
We use the methodology by Khandelwal and Sawant (BIBREF12), and modify it to support experimentation with multiple models.
For Speculation Cue Detection:
Input Sentence: It might rain tomorrow.
True Labels: Not-A-Cue, Cue, Not-A-Cue, Not-A-Cue.
First, this sentence is preprocessed to get the target labels as per the following annotation schema:
1 – Normal Cue 2 – Multiword Cue 3 – Not a cue 4 – Pad token
Thus, the preprocessed sequence is as follows:
Input Sentence: [It, might, rain, tomorrow]
True Labels: [3,1,3,3]
Then, we preprocess the input using the tokenizer for the model being used (BERT, XLNet or RoBERTa): splitting each word into one or more tokens, and converting each token to it’s corresponding tokenID, and padding it to the maximum input length of the model. Thus,
Input Sentence: [wtt(It), wtt(might), wtt(rain), wtt(tom), wtt(## or), wtt(## row), wtt(〈pad 〉),wtt(〈pad 〉)...]
True Labels: [3,1,3,3,3,3,4,4,4,4,...]
The word ‘tomorrow' has been split into 3 tokens, ‘tom', ‘##or' and ‘##row'. The function to convert the word to tokenID is represented by wtt.
For Speculation Scope Resolution:
If a sentence has multiple cues, each cue's scope will be resolved individually.
Input Sentence: It might rain tomorrow.
True Labels: Out-Of-Scope, Out-Of-Scope, In-Scope, In-Scope.
First, this sentence is preprocessed to get the target labels as per the following annotation schema:
0 – Out-Of-Scope 1 – In-Scope
Thus, the preprocessed sequence is as follows:
True Scope Labels: [0,0,1,1]
As for cue detection, we preprocess the input using the tokenizer for the model being used. Additionally, we need to indicate which cue's scope we want to find in the input sentence. We do this by inserting a special token representing the token type (according to the cue detection annotation schema) before the cue word whose scope is being resolved. Here, we want to find the scope of the cue ‘might'. Thus,
Input Sentence: [wtt(It), wtt(〈token[1]〉), wtt(might), wtt(rain), wtt(tom), wtt(##or), wtt(##row), wtt(〈pad〉), wtt(〈pad〉)...]
True Scope Labels: [0,0,1,1,1,1,0,0,0,0,...]
Now, the preprocessed input for cue detection and similarly for scope detection is fed as input to our model as follows:
X = Model (Input)
Y = W*X + b
The W matrix is a matrix of size n_hidden x num_classes (n_hidden is the size of the representation of a token within the model). These logits are fed to the loss function.
We use the following variants of each model:
BERT: bert-base-uncaseds3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased.tar.gz (The model used by BIBREF12)
RoBERTa: roberta-bases3.amazonaws.com/models.huggingface.co/bert/roberta-base-pytorch_model.bin (RoBERTa-base does not have an uncased variant)
XLNet: xlnet-base-caseds3.amazonaws.com/models.huggingface.co/bert/xlnet-base-cased-pytorch_model.bin (XLNet-base does not have an uncased variant)
The output of the model is a vector of probabilities per token. The loss is calculated for each token, by using the output vector and the true label for that token. We use class weights for the loss function, setting the weight for label 4 to 0 and all other labels to 1 (for cue detection only) to avoid training on padding token’s output.
We calculate the scores (Precision, Recall, F1) for the model per word of the input sentence, not per token that was fed to the model, as the tokens could be different for different models leading to inaccurate scores. For the above example, we calculate the output label for the word ‘tomorrow', not for each token it was split into (‘tom', ‘##or' and ‘##row'). To find the label for each word from the tokens it was split into, we experiment with 2 methods:
Average: We average the output vectors (softmax probabilities) for each token that the word was split into by the model's tokenizer. In the example above, we average the output of ‘tom', ‘##or' and ‘##row' to get the output for ‘tomorrow'. Then, we take an argmax over the resultant vector. This is then compared with the true label for the original word.
First Token: Here, we only consider the first token's probability vector (among all tokens the word was split into) as the output for that word, and get the label by an argmax over this vector. In the example above, we would consider the output vector corresponding to the token ‘tom' as the output for the word ‘tomorrow'.
For cue detection, the results are reported for the Average method only, while we report the scores for both Average and First Token for Scope Resolution.
For fair comparison, we use the same hyperparameters for the entire architecture for all 3 models. Only the tokenizer and the model are changed for each model. All other hyperparameters are kept same. We finetune the models for 60 epochs, using early stopping with a patience of 6 on the F1 score (word level) on the validation dataset. We use an initial learning rate of 3e-5, with a batch size of 8. We use the Categorical Cross Entropy loss function.
We use the Huggingface’s Pytorch Transformer library (BIBREF23) for the models and train all our models on Google Colaboratory.
Results: Speculation Cue Detection and Scope Resolution
We use a default train-validation-test split of 70-15-15 for each dataset. For the speculation detection and scope resolution subtasks using single-dataset training, we report the results as an average of 5 runs of the model. For training the model on multiple datasets, we perform a 70-15-15 split of each training dataset, after which the train and validation part of the individual datasets are merged while the scores are reported for the test part of the individual datasets, which is not used for training or validation. We report the results as an average of 3 runs of the model. Figure FIGREF8 contains results for speculation cue detection and scope resolution when trained on a single dataset. All models perform the best when trained on the same dataset as they are evaluated on, except for BF, which gets the best results when trained on BA. This is because of the transfer learning capabilities of the models and the fact that BF is a smaller dataset than BA (BF: 2670 sentences, BA: 11871 sentences). For speculation cue detection, there is lesser generalizability for models trained on BF or BA, while there is more generalizability for models trained on SFU. This could be because of the different nature of the biomedical domain.
Figure FIGREF11 contains the results for speculation detection and scope resolution for models trained jointly on multiple datasets. We observe that training on multiple datasets helps the performance of all models on each dataset, as the quantity of data available to train the model increases. We also observe that XLNet consistently outperforms BERT and RoBERTa. To confirm this observation, we apply the 2 models to the related task of negation detection and scope resolution
Negation Cue Detection and Scope Resolution
We use a default train-validation-test split of 70-15-15 for each dataset, and use all 4 datasets (BF, BA, SFU and Sherlock). The results for BERT are taken from BIBREF12. The results for XLNet and RoBERTa are averaged across 5 runs for statistical significance. Figure FIGREF14 contains results for negation cue detection and scope resolution. We report state-of-the-art results on negation scope resolution on BF, BA and SFU datasets. Contrary to popular opinion, we observe that XLNet is better than RoBERTa for the cue detection and scope resolution tasks. A few possible reasons for this trend are:
Domain specificity, as both negation and speculation are closely related subtasks. Further experimentation on different tasks is needed to verify this.
Most benchmark tasks are sentence classification tasks, whereas the subtasks we experiment on are sequence labelling tasks. Given the pre-training objective of XLNet (training on permutations of the input), it may be able to capture long-term dependencies better, essential for sequence labelling tasks.
We work with the base variants of the models, while most results are reported with the large variants of the models.
Conclusion and Future Scope
In this paper, we expanded on the work of Khandelwal and Sawant (BIBREF12) by looking at alternative transfer-learning models and experimented with training on multiple datasets. On the speculation detection task, we obtained a gain of 0.42 F1 points on BF, 1.98 F1 points on BA and 0.29 F1 points on SFU, while on the scope resolution task, we obtained a gain of 8.06 F1 points on BF, 4.27 F1 points on BA and 11.87 F1 points on SFU, when trained on a single dataset. While training on multiple datasets, we observed a gain of 10.6 F1 points on BF and 1.94 F1 points on BA on the speculation detection task and 2.16 F1 points on BF and 0.25 F1 points on SFU on the scope resolution task over the single dataset training approach. We thus significantly advance the state-of-the-art for speculation detection and scope resolution. On the negation scope resolution task, we applied the XLNet and RoBERTa and obtained a gain of 3.16 F1 points on BF, 0.06 F1 points on BA and 0.3 F1 points on SFU. Thus, we demonstrated the usefulness of transformer-based architectures in the field of negation and speculation detection and scope resolution. We believe that a larger and more general dataset would go a long way in bolstering future research and would help create better systems that are not domain-specific.
| [
"varied from Maximum Entropy Classifiers (BIBREF4) to Support Vector Machines (BIBREF5,BIBREF6,BIBREF7,BIBREF8), Recursive Neural Networks (BIBREF9,BIBREF10), Convolutional Neural Networks (BIBREF11) and most recently transfer learning-based architectures like Bidirectional Encoder Representation from Transformers ... | 2,215 | qasper | en | null | c9d48df129798d2aed0161fa805a289e5f75306af3e22336 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What were the baselines?
Answer: | [
"varied from Maximum Entropy Classifiers (BIBREF4) to Support Vector Machines (BIBREF5,BIBREF6,BIBREF7,BIBREF8), Recursive Neural Networks (BIBREF9,BIBREF10), Convolutional Neural Networks (BIBREF11) and most recently transfer learning-based architectures like Bidirectional Encoder Representation from Transformers ... | qasper | 128 | 48 | |
What are the languages they use in their experiment? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
While most NLP resources are English-specific, there have been several recent efforts to build multilingual benchmarks. One possibility is to collect and annotate data in multiple languages separately BIBREF0, but most existing datasets have been created through translation BIBREF1, BIBREF2. This approach has two desirable properties: it relies on existing professional translation services rather than requiring expertise in multiple languages, and it results in parallel evaluation sets that offer a meaningful measure of the cross-lingual transfer gap of different models. The resulting multilingual datasets are generally used for evaluation only, relying on existing English datasets for training.
Closely related to that, cross-lingual transfer learning aims to leverage large datasets available in one language—typically English—to build multilingual models that can generalize to other languages. Previous work has explored 3 main approaches to that end: machine translating the test set into English and using a monolingual English model (Translate-Test), machine translating the training set into each target language and training the models on their respective languages (Translate-Train), or using English data to fine-tune a multilingual model that is then transferred to the rest of languages (Zero-Shot).
The dataset creation and transfer procedures described above result in a mixture of original, human translated and machine translated data when dealing with cross-lingual models. In fact, the type of text a system is trained on does not typically match the type of text it is exposed to at test time: Translate-Test systems are trained on original data and evaluated on machine translated test sets, Zero-Shot systems are trained on original data and evaluated on human translated test sets, and Translate-Train systems are trained on machine translated data and evaluated on human translated test sets.
Despite overlooked to date, we show that such mismatch has a notable impact in the performance of existing cross-lingual models. By using back-translation BIBREF3 to paraphrase each training instance, we obtain another English version of the training set that better resembles the test set, obtaining substantial improvements for the Translate-Test and Zero-Shot approaches in cross-lingual Natural Language Inference (NLI). While improvements brought by machine translation have previously been attributed to data augmentation BIBREF4, we reject this hypothesis and show that the phenomenon is only present in translated test sets, but not in original ones. Instead, our analysis reveals that this behavior is caused by subtle artifacts arising from the translation process itself. In particular, we show that translating different parts of each instance separately (e.g. the premise and the hypothesis in NLI) can alter superficial patterns in the data (e.g. the degree of lexical overlap between them), which severely affects the generalization ability of current models. Based on the gained insights, we improve the state-of-the-art in XNLI, and show that some previous findings need to be reconsidered in the light of this phenomenon.
Related work ::: Cross-lingual transfer learning.
Current cross-lingual models work by pre-training multilingual representations using some form of language modeling, which are then fine-tuned on the relevant task and transferred to different languages. Some authors leverage parallel data to that end BIBREF5, BIBREF6, but training a model akin to BERT BIBREF7 on the combination of monolingual corpora in multiple languages is also effective BIBREF8. Closely related to our work, BIBREF4 showed that replacing segments of the training data with their translation during fine-tuning is helpful. However, they attribute this behavior to a data augmentation effect, which we believe should be reconsidered given the new evidence we provide.
Related work ::: Multilingual benchmarks.
Most benchmarks covering a wide set of languages have been created through translation, as it is the case of XNLI BIBREF1 for NLI, PAWS-X BIBREF9 for adversarial paraphrase identification, and XQuAD BIBREF2 and MLQA BIBREF10 for Question Answering (QA). A notable exception is TyDi QA BIBREF0, a contemporaneous QA dataset that was separately annotated in 11 languages. Other cross-lingual datasets leverage existing multilingual resources, as it is the case of MLDoc BIBREF11 for document classification and Wikiann BIBREF12 for named entity recognition. Concurrent to our work, BIBREF13 combine some of these datasets into a single multilingual benchmark, and evaluate some well-known methods on it.
Related work ::: Annotation artifacts.
Several studies have shown that NLI datasets like SNLI BIBREF14 and MultiNLI BIBREF15 contain spurious patterns that can be exploited to obtain strong results without making real inferential decisions. For instance, BIBREF16 and BIBREF17 showed that a hypothesis-only baseline performs better than chance due to cues on their lexical choice and sentence length. Similarly, BIBREF18 showed that NLI models tend to predict entailment for sentence pairs with a high lexical overlap. Several authors have worked on adversarial datasets to diagnose these issues and provide a more challenging benchmark BIBREF19, BIBREF20, BIBREF21. Besides NLI, other tasks like QA have also been found to be susceptible to annotation artifacts BIBREF22, BIBREF23. While previous work has focused on the monolingual scenario, we show that translation can interfere with these artifacts in multilingual settings.
Related work ::: Translationese.
Translated texts are known to have unique features like simplification, explicitation, normalization and interference, which are refer to as translationese BIBREF24. This phenomenon has been reported to have a notable impact in machine translation evaluation BIBREF25, BIBREF26. For instance, back-translation brings large BLEU gains for reversed test sets (i.e. when translationese is on the source side and original text is used as reference), but its effect diminishes in the natural direction BIBREF27. While connected, the phenomenon we analyze is different in that it arises from translation inconsistencies due to the lack of context, and affects cross-lingual transfer learning rather than machine translation.
Experimental design
Our goal is to analyze the effect of both human and machine translation in cross-lingual models. For that purpose, the core idea of our work is to (i) use machine translation to either translate the training set into other languages, or generate English paraphrases of it through back-translation, and (ii) evaluate the resulting systems on original, human translated and machine translated test sets in comparison with systems trained on original data. We next describe the models used in our experiments (§SECREF6), the specific training variants explored (§SECREF8), and the evaluation procedure followed (§SECREF10).
Experimental design ::: Models and transfer methods
We experiment with two models that are representative of the state-of-the-art in monolingual and cross-lingual pre-training: (i) Roberta BIBREF28, which is an improved version of BERT that uses masked language modeling to pre-train an English Transformer model, and (ii) XLM-R BIBREF8, which is a multilingual extension of the former pre-trained on 100 languages. In both cases, we use the large models released by the authors under the fairseq repository. As discussed next, we explore different variants of the training set to fine-tune each model on different tasks. At test time, we try both machine translating the test set into English (Translate-Test) and, in the case of XLM-R, using the actual test set in the target language (Zero-Shot).
Experimental design ::: Training variants
We try 3 variants of each training set to fine-tune our models: (i) the original one in English (Orig), (ii) an English paraphrase of it generated through back-translation using Spanish or Finnish as pivot (BT-ES and BT-FI), and (iii) a machine translated version in Spanish or Finnish (MT-ES and MT-FI). For sentences occurring multiple times in the training set (e.g. premises repeated for multiple hypotheses), we use the exact same translation for all occurrences, as our goal is to understand the inherent effect of translation rather than its potential application as a data augmentation method.
In order to train the machine translation systems for MT-XX and BT-XX, we use the big Transformer model BIBREF29 with the same settings as BIBREF30 and SentencePiece tokenization BIBREF31 with a joint vocabulary of 32k subwords. For English-Spanish, we train for 10 epochs on all parallel data from WMT 2013 BIBREF32 and ParaCrawl v5.0 BIBREF33. For English-Finnish, we train for 40 epochs on Europarl and Wiki Titles from WMT 2019 BIBREF34, ParaCrawl v5.0, and DGT, EUbookshop and TildeMODEL from OPUS BIBREF35. In both cases, we remove sentences longer than 250 tokens, with a source/target ratio exceeding 1.5, or for which langid.py BIBREF36 predicts a different language, resulting in a final corpus size of 48M and 7M sentence pairs, respectively. We use sampling decoding with a temperature of 0.5 for inference, which produces more diverse translations than beam search BIBREF37 and performed better in our preliminary experiments.
Experimental design ::: Tasks and evaluation procedure
We use the following tasks for our experiments:
Experimental design ::: Tasks and evaluation procedure ::: Natural Language Inference (NLI).
Given a premise and a hypothesis, the task is to determine whether there is an entailment, neutral or contradiction relation between them. We fine-tune our models on MultiNLI BIBREF15 for 10 epochs using the same settings as BIBREF28. In most of our experiments, we evaluate on XNLI BIBREF1, which comprises 2490 development and 5010 test instances in 15 languages. These were originally annotated in English, and the resulting premises and hypotheses were independently translated into the rest of the languages by professional translators. For the Translate-Test approach, we use the machine translated versions from the authors. Following BIBREF8, we select the best epoch checkpoint according to the average accuracy in the development set.
Experimental design ::: Tasks and evaluation procedure ::: Question Answering (QA).
Given a context paragraph and a question, the task is to identify the span answering the question in the context. We fine-tune our models on SQuAD v1.1 BIBREF38 for 2 epochs using the same settings as BIBREF28, and report test results for the last epoch. We use two datasets for evaluation: XQuAD BIBREF2, a subset of the SQuAD development set translated into 10 other languages, and MLQA BIBREF10 a dataset consisting of parallel context paragraphs plus the corresponding questions annotated in English and translated into 6 other languages. In both cases, the translation was done by professional translators at the document level (i.e. when translating a question, the text answering it was also shown). For our BT-XX and MT-XX variants, we translate the context paragraph and the questions independently, and map the answer spans using the same procedure as BIBREF39. For the Translate-Test approach, we use the official machine translated versions of MLQA, run inference over them, and map the predicted answer spans back to the target language.
Both for NLI and QA, we run each system 5 times with different random seeds and report the average results. Space permitting, we also report the standard deviation across the 5 runs.
NLI experiments
We next discuss our main results in the XNLI development set (§SECREF15, §SECREF16), run additional experiments to better understand the behavior of our different variants (§SECREF17, §SECREF22, §SECREF25), and compare our results to previous work in the XNLI test set (§SECREF30).
NLI experiments ::: Translate-Test results
We start by analyzing XNLI development results for Translate-Test. Recall that, in this approach, the test set is machine translated into English, but training is typically done on original English data. Our BT-ES and BT-FI variants close this gap by training on a machine translated English version of the training set generated through back-translation. As shown in Table TABREF9, this brings substantial gains for both Roberta and XLM-R, with an average improvement of 4.6 points in the best case. Quite remarkably, MT-ES and MT-FI also outperform Orig by a substantial margin, and are only 0.8 points below their BT-ES and BT-FI counterparts. Recall that, for these two systems, training is done in machine translated Spanish or Finnish, while inference is done in machine translated English. This shows that the loss of performance when generalizing from original data to machine translated data is substantially larger than the loss of performance when generalizing from one language to another.
NLI experiments ::: Zero-Shot results
We next analyze the results for the Zero-Shot approach. In this case, inference is done in the test set in each target language which, in the case of XNLI, was human translated from English. As such, different from the Translate-Test approach, neither training on original data (Orig) nor training on machine translated data (BT-XX and MT-XX) makes use of the exact same type of text that the system is exposed to at test time. However, as shown in Table TABREF9, both BT-XX and MT-XX outperform Orig by approximately 2 points, which suggests that our (back-)translated versions of the training set are more similar to the human translated test sets than the original one. This also provides a new perspective on the Translate-Train approach, which was reported to outperform Orig in previous work BIBREF5: while the original motivation was to train the model on the same language that it is tested on, our results show that machine translating the training set is beneficial even when the target language is different.
NLI experiments ::: Original vs. translated test sets
So as to understand whether the improvements observed so far are limited to translated test sets or apply more generally, we conduct additional experiments comparing translated test sets to original ones. However, to the best of our knowledge, all existing non-English NLI benchmarks were created through translation. For that reason, we build a new test set that mimics XNLI, but is annotated in Spanish rather than English. We first collect the premises from a filtered version of CommonCrawl BIBREF42, taking a subset of 5 websites that represent a diverse set of genres: a newspaper, an economy forum, a celebrity magazine, a literature blog, and a consumer magazine. We then ask native Spanish annotators to generate an entailment, a neutral and a contradiction hypothesis for each premise. We collect a total of 2490 examples using this procedure, which is the same size as the XNLI development set. Finally, we create a human translated and a machine translated English version of the dataset using professional translators from Gengo and our machine translation system described in §SECREF8, respectively. We report results for the best epoch checkpoint on each set.
As shown in Table TABREF18, both BT-XX and MT-XX clearly outperform Orig in all test sets created through translation, which is consistent with our previous results. In contrast, the best results on the original English set are obtained by Orig, and neither BT-XX nor MT-XX obtain any clear improvement on the one in Spanish either. This confirms that the underlying phenomenon is limited to translated test sets. In addition, it is worth mentioning that the results for the machine translated test set in English are slightly better than those for the human translated one, which suggests that the difficulty of the task does not only depend on the translation quality. Finally, it is also interesting that MT-ES is only marginally better than MT-FI in both Spanish test sets, even if it corresponds to the Translate-Train approach, whereas MT-FI needs to Zero-Shot transfer from Finnish into Spanish. This reinforces the idea that it is training on translated data rather than training on the target language that is key in Translate-Train.
NLI experiments ::: Stress tests
In order to better understand how systems trained on original and translated data differ, we run additional experiments on the NLI Stress Tests BIBREF19, which were designed to test the robustness of NLI models to specific linguistic phenomena in English. The benchmark consists of a competence test, which evaluates the ability to understand antonymy relation and perform numerical reasoning, a distraction test, which evaluates the robustness to shallow patterns like lexical overlap and the presence of negation words, and a noise test, which evaluates robustness to spelling errors. Just as with previous experiments, we report results for the best epoch checkpoint in each test set.
As shown in Table TABREF23, Orig outperforms BT-FI and MT-FI on the competence test by a large margin, but the opposite is true on the distraction test. In particular, our results show that BT-FI and MT-FI are less reliant on lexical overlap and the presence of negative words. This feels intuitive, as translating the premise and hypothesis independently—as BT-FI and MT-FI do—is likely to reduce the lexical overlap between them. More generally, the translation process can alter similar superficial patterns in the data, which NLI models are sensitive to (§SECREF2). This would explain why the resulting models have a different behavior on different stress tests.
NLI experiments ::: Output class distribution
With the aim to understand the effect of the previous phenomenon in cross-lingual settings, we look at the output class distribution of our different models in the XNLI development set. As shown in Table TABREF28, the predictions of all systems are close to the true class distribution in the case of English. Nevertheless, Orig is strongly biased for the rest of languages, and tends to underpredict entailment and overpredict neutral. This can again be attributed to the fact that the English test set is original, whereas the rest are human translated. In particular, it is well-known that NLI models tend to predict entailment when there is a high lexical overlap between the premise and the hypothesis (§SECREF2). However, the degree of overlap will be smaller in the human translated test sets given that the premise and the hypothesis were translated independently, which explains why entailment is underpredicted. In contrast, BT-FI and MT-FI are exposed to the exact same phenomenon during training, which explains why they are not that heavily affected.
So as to measure the impact of this phenomenon, we explore a simple approach to correct this bias: having fine-tuned each model, we adjust the bias term added to the logit of each class so the model predictions match the true class distribution for each language. As shown in Table TABREF29, this brings large improvements for Orig, but is less effective for BT-FI and MT-FI. This shows that the performance of Orig was considerably hindered by this bias, which BT-FI and MT-FI effectively mitigate.
NLI experiments ::: Comparison with the state-of-the-art
So as to put our results into perspective, we compare our best variant to previous work on the XNLI test set. As shown in Table TABREF31, our method improves the state-of-the-art for both the Translate-Test and the Zero-Shot approaches by 4.3 and 2.8 points, respectively. It also obtains the best overall results published to date, with the additional advantage that the previous state-of-the-art required a machine translation system between English and each of the 14 target languages, whereas our method uses a single machine translation system between English and Finnish (which is not one of the target languages). While the main goal of our work is not to design better cross-lingual models, but to analyze their behavior in connection to translation, this shows that the phenomenon under study is highly relevant, to the extent that it can be exploited to improve the state-of-the-art.
QA experiments
So as to understand whether our previous findings apply to other tasks besides NLI, we run additional experiments on QA. As shown in Table TABREF32, BT-FI and BT-ES do indeed outperform Orig for the Translate-Test approach on MLQA. The improvement is modest, but very consistent across different languages, models and runs. The results for MT-ES and MT-FI are less conclusive, presumably because mapping the answer spans across languages might introduce some noise. In contrast, we do not observe any clear improvement for the Zero-Shot approach on this dataset. Our XQuAD results in Table TABREF33 are more positive, but still inconclusive.
These results can partly be explained by the translation procedure used to create the different benchmarks: the premises and hypotheses of XNLI were translated independently, whereas the questions and context paragraphs of XQuAD were translated together. Similarly, MLQA made use of parallel contexts, and translators were shown the sentence containing each answer when translating the corresponding question. As a result, one can expect both QA benchmarks to have more consistent translations than XNLI, which would in turn diminish this phenomenon. In contrast, the questions and context paragraphs are independently translated when using machine translation, which explains why BT-ES and BT-FI outperform Orig for the Translate-Test approach. We conclude that the translation artifacts revealed by our analysis are not exclusive to NLI, as they also show up on QA for the Translate-Test approach, but their actual impact can be highly dependent on the translation procedure used and the nature of the task.
Discussion
Our analysis prompts to reconsider previous findings in cross-lingual transfer learning as follows:
Discussion ::: The cross-lingual transfer gap on XNLI was overestimated.
Given the parallel nature of XNLI, accuracy differences across languages are commonly interpreted as the loss of performance when generalizing from English to the rest of languages. However, our work shows that there is another factor that can have a much larger impact: the loss of performance when generalizing from original to translated data. Our results suggest that the real cross-lingual generalization ability of XLM-R is considerably better than what the accuracy numbers in XNLI reflect.
Discussion ::: Overcoming the cross-lingual gap is not what makes Translate-Train work.
The original motivation for Translate-Train was to train the model on the same language it is tested on. However, we show that it is training on translated data, rather than training on the target language, that is key for this approach to outperform Zero-Shot as reported by previous authors.
Discussion ::: Improvements previously attributed to data augmentation should be reconsidered.
The method by BIBREF4 combines machine translated premises and hypotheses in different languages (§SECREF2), resulting in an effect similar to BT-XX and MT-XX. As such, we believe that this method should be analyzed from the point of view of dataset artifacts rather than data augmentation, as the authors do. From this perspective, having the premise and the hypotheses in different languages can reduce the superficial patterns between them, which would explain why this approach is better than using examples in a single language.
Discussion ::: The potential of Translate-Test was underestimated.
The previous best results for Translate-Test on XNLI lagged behind the state-of-the-art by 4.6 points. Our work reduces this gap to only 0.8 points by addressing the underlying translation artifacts. The reason why Translate-Test is more severely affected by this phenomenon is twofold: (i) the effect is doubled by first using human translation to create the test set and then machine translation to translate it back to English, and (ii) Translate-Train was inadvertently mitigating this issue (see above), but equivalent techniques were never applied to Translate-Test.
Discussion ::: Future evaluation should better account for translation artifacts.
The evaluation issues raised by our analysis do not have a simple solution. In fact, while we use the term translation artifacts to highlight that they are an unintended effect of translation that impacts final evaluation, one could also argue that it is the original datasets that contain the artifacts, which translation simply alters or even mitigates. In any case, this is a more general issue that falls beyond the scope of cross-lingual transfer learning, so we argue that it should be carefully controlled when evaluating cross-lingual models. In the absence of more robust datasets, we recommend that future multilingual benchmarks should at least provide consistent test sets for English and the rest of languages. This can be achieved by (i) using original annotations in all languages, (ii) using original annotations in a non-English language and translating them into English and other languages, or (iii) if translating from English, doing so at the document level to minimize translation inconsistencies.
Conclusions
In this paper, we have shown that both human and machine translation can alter superficial patterns in data, which requires reconsidering previous findings in cross-lingual transfer learning. Based on the gained insights, we have improved the state-of-the-art in XNLI for the Translate-Test and Zero-Shot approaches by a substantial margin. Finally, we have shown that the phenomenon is not specific to NLI but also affects QA, although it is less pronounced there thanks to the translation procedure used in the corresponding benchmarks. So as to facilitate similar studies in the future, we release our NLI dataset, which, unlike previous benchmarks, was annotated in a non-English language and human translated into English.
Acknowledgments
We thank Nora Aranberri and Uxoa Iñurrieta for helpful discussion during the development of this work, as well as the rest of our colleagues from the IXA group that worked as annotators for our NLI dataset.
This research was partially funded by a Facebook Fellowship, the Basque Government excellence research group (IT1343-19), the Spanish MINECO (UnsupMT TIN2017‐91692‐EXP MCIU/AEI/FEDER, UE), Project BigKnowledge (Ayudas Fundación BBVA a equipos de investigación científica 2018), and the NVIDIA GPU grant program.
| [
"English\nFrench\nSpanish\nGerman\nGreek\nBulgarian\nRussian\nTurkish\nArabic\nVietnamese\nThai\nChinese\nHindi\nSwahili\nUrdu\nFinnish",
"English, Spanish, Finnish"
] | 4,086 | qasper | en | null | 443d051a54d96c296d9135dad3794d09f89ff91d6433d092 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What are the languages they use in their experiment?
Answer: | [
"English\nFrench\nSpanish\nGerman\nGreek\nBulgarian\nRussian\nTurkish\nArabic\nVietnamese\nThai\nChinese\nHindi\nSwahili\nUrdu\nFinnish",
"English, Spanish, Finnish"
] | qasper | 128 | 49 | |
What other tasks do they test their method on? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
We understand from Zipf's Law that in any natural language corpus a majority of the vocabulary word types will either be absent or occur in low frequency. Estimating the statistical properties of these rare word types is naturally a difficult task. This is analogous to the curse of dimensionality when we deal with sequences of tokens - most sequences will occur only once in the training data. Neural network architectures overcome this problem by defining non-linear compositional models over vector space representations of tokens and hence assign non-zero probability even to sequences not seen during training BIBREF0 , BIBREF1 . In this work, we explore a similar approach to learning distributed representations of social media posts by composing them from their constituent characters, with the goal of generalizing to out-of-vocabulary words as well as sequences at test time.
Traditional Neural Network Language Models (NNLMs) treat words as the basic units of language and assign independent vectors to each word type. To constrain memory requirements, the vocabulary size is fixed before-hand; therefore, rare and out-of-vocabulary words are all grouped together under a common type `UNKNOWN'. This choice is motivated by the assumption of arbitrariness in language, which means that surface forms of words have little to do with their semantic roles. Recently, BIBREF2 challenge this assumption and present a bidirectional Long Short Term Memory (LSTM) BIBREF3 for composing word vectors from their constituent characters which can memorize the arbitrary aspects of word orthography as well as generalize to rare and out-of-vocabulary words.
Encouraged by their findings, we extend their approach to a much larger unicode character set, and model long sequences of text as functions of their constituent characters (including white-space). We focus on social media posts from the website Twitter, which are an excellent testing ground for character based models due to the noisy nature of text. Heavy use of slang and abundant misspellings means that there are many orthographically and semantically similar tokens, and special characters such as emojis are also immensely popular and carry useful semantic information. In our moderately sized training dataset of 2 million tweets, there were about 0.92 million unique word types. It would be expensive to capture all these phenomena in a word based model in terms of both the memory requirement (for the increased vocabulary) and the amount of training data required for effective learning. Additional benefits of the character based approach include language independence of the methods, and no requirement of NLP preprocessing such as word-segmentation.
A crucial step in learning good text representations is to choose an appropriate objective function to optimize. Unsupervised approaches attempt to reconstruct the original text from its latent representation BIBREF4 , BIBREF0 . Social media posts however, come with their own form of supervision annotated by millions of users, in the form of hashtags which link posts about the same topic together. A natural assumption is that the posts with the same hashtags should have embeddings which are close to each other. Hence, we formulate our training objective to maximize cross-entropy loss at the task of predicting hashtags for a post from its latent representation.
We propose a Bi-directional Gated Recurrent Unit (Bi-GRU) BIBREF5 neural network for learning tweet representations. Treating white-space as a special character itself, the model does a forward and backward pass over the entire sequence, and the final GRU states are linearly combined to get the tweet embedding. Posterior probabilities over hashtags are computed by projecting this embedding to a softmax output layer. Compared to a word-level baseline this model shows improved performance at predicting hashtags for a held-out set of posts. Inspired by recent work in learning vector space text representations, we name our model tweet2vec.
Related Work
Using neural networks to learn distributed representations of words dates back to BIBREF0 . More recently, BIBREF4 released word2vec - a collection of word vectors trained using a recurrent neural network. These word vectors are in widespread use in the NLP community, and the original work has since been extended to sentences BIBREF1 , documents and paragraphs BIBREF6 , topics BIBREF7 and queries BIBREF8 . All these methods require storing an extremely large table of vectors for all word types and cannot be easily generalized to unseen words at test time BIBREF2 . They also require preprocessing to find word boundaries which is non-trivial for a social network domain like Twitter.
In BIBREF2 , the authors present a compositional character model based on bidirectional LSTMs as a potential solution to these problems. A major benefit of this approach is that large word lookup tables can be compacted into character lookup tables and the compositional model scales to large data sets better than other state-of-the-art approaches. While BIBREF2 generate word embeddings from character representations, we propose to generate vector representations of entire tweets from characters in our tweet2vec model.
Our work adds to the growing body of work showing the applicability of character models for a variety of NLP tasks such as Named Entity Recognition BIBREF9 , POS tagging BIBREF10 , text classification BIBREF11 and language modeling BIBREF12 , BIBREF13 .
Previously, BIBREF14 dealt with the problem of estimating rare word representations by building them from their constituent morphemes. While they show improved performance over word-based models, their approach requires a morpheme parser for preprocessing which may not perform well on noisy text like Twitter. Also the space of all morphemes, though smaller than the space of all words, is still large enough that modelling all morphemes is impractical.
Hashtag prediction for social media has been addressed earlier, for example in BIBREF15 , BIBREF16 . BIBREF15 also use a neural architecture, but compose text embeddings from a lookup table of words. They also show that the learned embeddings can generalize to an unrelated task of document recommendation, justifying the use of hashtags as supervision for learning text representations.
Tweet2Vec
Bi-GRU Encoder: Figure 1 shows our model for encoding tweets. It uses a similar structure to the C2W model in BIBREF2 , with LSTM units replaced with GRU units.
The input to the network is defined by an alphabet of characters $C$ (this may include the entire unicode character set). The input tweet is broken into a stream of characters $c_1, c_2, ... c_m$ each of which is represented by a 1-by- $|C|$ encoding. These one-hot vectors are then projected to a character space by multiplying with the matrix $P_C \in \mathbb {R}^{|C| \times d_c}$ , where $d_c$ is the dimension of the character vector space. Let $x_1, x_2, ... x_m$ be the sequence of character vectors for the input tweet after the lookup. The encoder consists of a forward-GRU and a backward-GRU. Both have the same architecture, except the backward-GRU processes the sequence in reverse order. Each of the GRU units process these vectors sequentially, and starting with the initial state $h_0$ compute the sequence $h_1, h_2, ... h_m$ as follows: $ r_t &= \sigma (W_r x_t + U_r h_{t-1} + b_r), \\ z_t &= \sigma (W_z x_t + U_z h_{t-1} + b_z), \\ \tilde{h}_t &= tanh(W_h x_t + U_h (r_t \odot h_{t-1}) + b_h), \\ h_t &= (1-z_t) \odot h_{t-1} + z_t \odot \tilde{h}_t. $
Here $r_t$ , $z_t$ are called the reset and update gates respectively, and $\tilde{h}_t$ is the candidate output state which is converted to the actual output state $h_t$ . $W_r, W_z, W_h$ are $d_h \times d_c$ matrices and $U_r, U_z, U_h$ are $d_h \times d_h$ matrices, where $d_h$ is the hidden state dimension of the GRU. The final states $h_m^f$ from the forward-GRU, and $z_t$0 from the backward GRU are combined using a fully-connected layer to the give the final tweet embedding $z_t$1 :
$$e_t = W^f h_m^f + W^b h_0^b$$ (Eq. 3)
Here $W^f, W^b$ are $d_t \times d_h$ and $b$ is $d_t \times 1$ bias term, where $d_t$ is the dimension of the final tweet embedding. In our experiments we set $d_t=d_h$ . All parameters are learned using gradient descent.
Softmax: Finally, the tweet embedding is passed through a linear layer whose output is the same size as the number of hashtags $L$ in the data set. We use a softmax layer to compute the posterior hashtag probabilities:
$$P(y=j |e) = \frac{exp(w_j^Te + b_j)}{\sum _{i=1}^L exp(w_i^Te + b_j)}.$$ (Eq. 4)
Objective Function: We optimize the categorical cross-entropy loss between predicted and true hashtags:
$$J = \frac{1}{B} \sum _{i=1}^{B} \sum _{j=1}^{L} -t_{i,j}log(p_{i,j}) + \lambda \Vert \Theta \Vert ^2.$$ (Eq. 5)
Here $B$ is the batch size, $L$ is the number of classes, $p_{i,j}$ is the predicted probability that the $i$ -th tweet has hashtag $j$ , and $t_{i,j} \in \lbrace 0,1\rbrace $ denotes the ground truth of whether the $j$ -th hashtag is in the $i$ -th tweet. We use L2-regularization weighted by $\lambda $ .
Word Level Baseline
Since our objective is to compare character-based and word-based approaches, we have also implemented a simple word-level encoder for tweets. The input tweet is first split into tokens along white-spaces. A more sophisticated tokenizer may be used, but for a fair comparison we wanted to keep language specific preprocessing to a minimum. The encoder is essentially the same as tweet2vec, with the input as words instead of characters. A lookup table stores word vectors for the $V$ (20K here) most common words, and the rest are grouped together under the `UNK' token.
Data
Our dataset consists of a large collection of global posts from Twitter between the dates of June 1, 2013 to June 5, 2013. Only English language posts (as detected by the lang field in Twitter API) and posts with at least one hashtag are retained. We removed infrequent hashtags ( $<500$ posts) since they do not have enough data for good generalization. We also removed very frequent tags ( $>19K$ posts) which were almost always from automatically generated posts (ex: #androidgame) which are trivial to predict. The final dataset contains 2 million tweets for training, 10K for validation and 50K for testing, with a total of 2039 distinct hashtags. We use simple regex to preprocess the post text and remove hashtags (since these are to be predicted) and HTML tags, and replace user-names and URLs with special tokens. We also removed retweets and convert the text to lower-case.
Implementation Details
Word vectors and character vectors are both set to size $d_L=150$ for their respective models. There were 2829 unique characters in the training set and we model each of these independently in a character look-up table. Embedding sizes were chosen such that each model had roughly the same number of parameters (Table 2 ). Training is performed using mini-batch gradient descent with Nesterov's momentum. We use a batch size $B=64$ , initial learning rate $\eta _0=0.01$ and momentum parameter $\mu _0=0.9$ . L2-regularization with $\lambda =0.001$ was applied to all models. Initial weights were drawn from 0-mean gaussians with $\sigma =0.1$ and initial biases were set to 0. The hyperparameters were tuned one at a time keeping others fixed, and values with the lowest validation cost were chosen. The resultant combination was used to train the models until performance on validation set stopped increasing. During training, the learning rate is halved everytime the validation set precision increases by less than 0.01 % from one epoch to the next. The models converge in about 20 epochs. Code for training both the models is publicly available on github.
Results
We test the character and word-level variants by predicting hashtags for a held-out test set of posts. Since there may be more than one correct hashtag per post, we generate a ranked list of tags for each post from the output posteriors, and report average precision@1, recall@10 and mean rank of the correct hashtags. These are listed in Table 3 .
To see the performance of each model on posts containing rare words (RW) and frequent words (FW) we selected two test sets each containing 2,000 posts. We populated these sets with posts which had the maximum and minimum number of out-of-vocabulary words respectively, where vocabulary is defined by the 20K most frequent words. Overall, tweet2vec outperforms the word model, doing significantly better on RW test set and comparably on FW set. This improved performance comes at the cost of increased training time (see Table 2 ), since moving from words to characters results in longer input sequences to the GRU.
We also study the effect of model size on the performance of these models. For the word model we set vocabulary size $V$ to 8K, 15K and 20K respectively. For tweet2vec we set the GRU hidden state size to 300, 400 and 500 respectively. Figure 2 shows precision 1 of the two models as the number of parameters is increased, for each test set described above. There is not much variation in the performance, and moreover tweet2vec always outperforms the word based model for the same number of parameters.
Table 4 compares the models as complexity of the task is increased. We created 3 datasets (small, medium and large) with an increasing number of hashtags to be predicted. This was done by varying the lower threshold of the minimum number of tags per post for it to be included in the dataset. Once again we observe that tweet2vec outperforms its word-based counterpart for each of the three settings.
Finally, table 1 shows some predictions from the word level model and tweet2vec. We selected these to highlight some strengths of the character based approach - it is robust to word segmentation errors and spelling mistakes, effectively interprets emojis and other special characters to make predictions, and also performs comparably to the word-based approach for in-vocabulary tokens.
Conclusion
We have presented tweet2vec - a character level encoder for social media posts trained using supervision from associated hashtags. Our result shows that tweet2vec outperforms the word based approach, doing significantly better when the input post contains many rare words. We have focused only on English language posts, but the character model requires no language specific preprocessing and can be extended to other languages. For future work, one natural extension would be to use a character-level decoder for predicting the hashtags. This will allow generation of hashtags not seen in the training dataset. Also, it will be interesting to see how our tweet2vec embeddings can be used in domains where there is a need for semantic understanding of social media, such as tracking infectious diseases BIBREF17 . Hence, we provide an off-the-shelf encoder trained on medium dataset described above to compute vector-space representations of tweets along with our code on github.
Acknowledgments
We would like to thank Alex Smola, Yun Fu, Hsiao-Yu Fish Tung, Ruslan Salakhutdinov, and Barnabas Poczos for useful discussions. We would also like to thank Juergen Pfeffer for providing access to the Twitter data, and the reviewers for their comments.
| [
"None"
] | 2,473 | qasper | en | null | c2df5ec4b6b7dd1c56e05d6b3c0e244d0c44d68f117caf0e | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What other tasks do they test their method on?
Answer: | [
"None"
] | qasper | 128 | 50 | |
Do they use pretrained embeddings? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Rendering natural language descriptions from structured data is required in a wide variety of commercial applications such as generating descriptions of products, hotels, furniture, etc., from a corresponding table of facts about the entity. Such a table typically contains {field, value} pairs where the field is a property of the entity (e.g., color) and the value is a set of possible assignments to this property (e.g., color = red). Another example of this is the recently introduced task of generating one line biography descriptions from a given Wikipedia infobox BIBREF0 . The Wikipedia infobox serves as a table of facts about a person and the first sentence from the corresponding article serves as a one line description of the person. Figure FIGREF2 illustrates an example input infobox which contains fields such as Born, Residence, Nationality, Fields, Institutions and Alma Mater. Each field further contains some words (e.g., particle physics, many-body theory, etc.). The corresponding description is coherent with the information contained in the infobox.
Note that the number of fields in the infobox and the ordering of the fields within the infobox varies from person to person. Given the large size (700K examples) and heterogeneous nature of the dataset which contains biographies of people from different backgrounds (sports, politics, arts, etc.), it is hard to come up with simple rule-based templates for generating natural language descriptions from infoboxes, thereby making a case for data-driven models. Based on the recent success of data-driven neural models for various other NLG tasks BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , one simple choice is to treat the infobox as a sequence of {field, value} pairs and use a standard seq2seq model for this task. However, such a model is too generic and does not exploit the specific characteristics of this task as explained below. First, note that while generating such descriptions from structured data, a human keeps track of information at two levels. Specifically, at a macro level, she would first decide which field to mention next and then at a micro level decide which of the values in the field needs to be mentioned next. For example, she first decides that at the current step, the field occupation needs attention and then decides which is the next appropriate occupation to attend to from the set of occupations (actor, director, producer, etc.). To enable this, we use a bifocal attention mechanism which computes an attention over fields at a macro level and over values at a micro level. We then fuse these attention weights such that the attention weight for a field also influences the attention over the values within it. Finally, we feed a fused context vector to the decoder which contains both field level and word level information. Note that such two-level attention mechanisms BIBREF6 , BIBREF7 , BIBREF8 have been used in the context of unstructured data (as opposed to structured data in our case), where at a macro level one needs to pay attention to sentences and at a micro level to words in the sentences.
Next, we observe that while rendering the output, once the model pays attention to a field (say, occupation) it needs to stay on this field for a few timesteps (till all the occupations are produced in the output). We refer to this as the stay on behavior. Further, we note that once the tokens of a field are referred to, they are usually not referred to later. For example, once all the occupations have been listed in the output we will never visit the occupation field again because there is nothing left to say about it. We refer to this as the never look back behavior. To model the stay on behaviour, we introduce a forget (or remember) gate which acts as a signal to decide when to forget the current field (or equivalently to decide till when to remember the current field). To model the never look back behaviour we introduce a gated orthogonalization mechanism which ensures that once a field is forgotten, subsequent field context vectors fed to the decoder are orthogonal to (or different from) the previous field context vectors.
We experiment with the WikiBio dataset BIBREF0 which contains around 700K {infobox, description} pairs and has a vocabulary of around 400K words. We show that the proposed model gives a relative improvement of 21% and 20% as compared to current state of the art models BIBREF0 , BIBREF9 on this dataset. The proposed model also gives a relative improvement of 10% as compared to the basic seq2seq model. Further, we introduce new datasets for French and German on the same lines as the English WikiBio dataset. Even on these two datasets, our model outperforms the state of the art methods mentioned above.
Related work
Natural Language Generation has always been of interest to the research community and has received a lot of attention in the past. The approaches for NLG range from (i) rule based approaches (e.g., BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 ) (ii) modular statistical approaches which divide the process into three phases (planning, selection and surface realization) and use data driven approaches for one or more of these phases BIBREF15 , BIBREF16 , BIBREF17 , BIBREF18 , BIBREF19 (iii) hybrid approaches which rely on a combination of handcrafted rules and corpus statistics BIBREF20 , BIBREF21 , BIBREF22 and (iv) the more recent neural network based models BIBREF1 .
Neural models for NLG have been proposed in the context of various tasks such as machine translation BIBREF1 , document summarization BIBREF2 , BIBREF4 , paraphrase generation BIBREF23 , image captioning BIBREF24 , video summarization BIBREF25 , query based document summarization BIBREF5 and so on. Most of these models are data hungry and are trained on large amounts of data. On the other hand, NLG from structured data has largely been studied in the context of small datasets such as WeatherGov BIBREF26 , RoboCup BIBREF27 , NFL Recaps BIBREF15 , Prodigy-Meteo BIBREF28 and TUNA Challenge BIBREF29 . Recently weather16 proposed RNN/LSTM based neural encoder-decoder models with attention for WeatherGov and RoboCup datasets.
Unlike the datasets mentioned above, the biography dataset introduced by lebret2016neural is larger (700K {table, descriptions} pairs) and has a much larger vocabulary (400K words as opposed to around 350 or fewer words in the above datasets). Further, unlike the feed-forward neural network based model proposed by BIBREF0 we use a sequence to sequence model and introduce components to address the peculiar characteristics of the task. Specifically, we introduce neural components to address the need for attention at two levels and to address the stay on and never look back behaviour required by the decoder. KiddonZC16 have explored the use of checklists to track previously visited ingredients while generating recipes from ingredients. Note that two-level attention mechanisms have also been used in the context of summarization BIBREF6 , document classification BIBREF7 , dialog systems BIBREF8 , etc. However, these works deal with unstructured data (sentences at the higher level and words at a lower level) as opposed to structured data in our case.
Proposed model
As input we are given an infobox INLINEFORM0 , which is a set of pairs INLINEFORM1 where INLINEFORM2 corresponds to field names and INLINEFORM3 is the sequence of corresponding values and INLINEFORM4 is the total number of fields in INLINEFORM5 . For example, INLINEFORM6 could be one such pair in this set. Given such an input, the task is to generate a description INLINEFORM7 containing INLINEFORM8 words. A simple solution is to treat the infobox as a sequence of fields followed by the values corresponding to the field in the order of their appearance in the infobox. For example, the infobox could be flattened to produce the following input sequence (the words in bold are field names which act as delimiters)
[Name] John Doe [Birth_Date] 19 March 1981 [Nationality] Indian .....
The problem can then be cast as a seq2seq generation problem and can be modeled using a standard neural architecture comprising of three components (i) an input encoder (using GRU/LSTM cells), (ii) an attention mechanism to attend to important values in the input sequence at each time step and (iii) a decoder to decode the output one word at a time (again, using GRU/LSTM cells). However, this standard model is too generic and does not exploit the specific characteristics of this task. We propose additional components, viz., (i) a fused bifocal attention mechanism which operates on fields (macro) and values (micro) and (ii) a gated orthogonalization mechanism to model stay on and never look back behavior.
Fused Bifocal Attention Mechanism
Intuitively, when a human writes a description from a table she keeps track of information at two levels. At the macro level, it is important to decide which is the appropriate field to attend to next and at a micro level (i.e., within a field) it is important to know which values to attend to next. To capture this behavior, we use a bifocal attention mechanism as described below.
Macro Attention: Consider the INLINEFORM0 -th field INLINEFORM1 which has values INLINEFORM2 . Let INLINEFORM3 be the representation of this field in the infobox. This representation can either be (i) the word embedding of the field name or (ii) some function INLINEFORM4 of the values in the field or (iii) a concatenation of (i) and (ii). The function INLINEFORM5 could simply be the sum or average of the embeddings of the values in the field. Alternately, this function could be a GRU (or LSTM) which treats these values within a field as a sequence and computes the field representation as the final representation of this sequence (i.e., the representation of the last time-step). We found that bidirectional GRU is a better choice for INLINEFORM6 and concatenating the embedding of the field name with this GRU representation works best. Further, using a bidirectional GRU cell to take contextual information from neighboring fields also helps (these are the orange colored cells in the top-left block in Figure FIGREF3 with macro attention). Given these representations INLINEFORM7 for all the INLINEFORM8 fields we compute an attention over the fields (macro level). DISPLAYFORM0
where INLINEFORM0 is the state of the decoder at time step INLINEFORM1 . INLINEFORM2 and INLINEFORM3 are parameters, INLINEFORM4 is the total number of fields in the input, INLINEFORM5 is the macro (field level) context vector at the INLINEFORM6 -th time step of the decoder.
Micro Attention: Let INLINEFORM0 be the representation of the INLINEFORM1 -th value in a given field. This representation could again either be (i) simply the embedding of this value (ii) or a contextual representation computed using a function INLINEFORM2 which also considers the other values in the field. For example, if INLINEFORM3 are the values in a field then these values can be treated as a sequence and the representation of the INLINEFORM4 -th value can be computed using a bidirectional GRU over this sequence. Once again, we found that using a bi-GRU works better then simply using the embedding of the value. Once we have such a representation computed for all values across all the fields, we compute the attention over these values (micro level) as shown below : DISPLAYFORM0
where INLINEFORM0 is the state of the decoder at time step INLINEFORM1 . INLINEFORM2 and INLINEFORM3 are parameters, INLINEFORM4 is the total number of values across all the fields.
Fused Attention: Intuitively, the attention weights assigned to a field should have an influence on all the values belonging to the particular field. To ensure this, we reweigh the micro level attention weights based on the corresponding macro level attention weights. In other words, we fuse the attention weights at the two levels as: DISPLAYFORM0
where INLINEFORM0 is the field corresponding to the INLINEFORM1 -th value, INLINEFORM2 is the macro level context vector.
Gated Orthogonalization for Modeling Stay-On and Never Look Back behaviour
We now describe a series of choices made to model stay-on and never look back behavior. We first begin with the stay-on property which essentially implies that if we have paid attention to the field INLINEFORM0 at timestep INLINEFORM1 then we are likely to pay attention to the same field for a few more time steps. For example, if we are focusing on the occupation field at this timestep then we are likely to focus on it for the next few timesteps till all relevant values in this field have been included in the generated description. In other words, we want to remember the field context vector INLINEFORM2 for a few timesteps. One way of ensuring this is to use a remember (or forget) gate as given below which remembers the previous context vector when required and forgets it when it is time to move on from that field. DISPLAYFORM0
where INLINEFORM0 are parameters to be learned. The job of the forget gate is to ensure that INLINEFORM1 is similar to INLINEFORM2 when required (i.e., by learning INLINEFORM3 when we want to continue focusing on the same field) and different when it is time to move on (by learning that INLINEFORM4 ).
Next, the never look back property implies that once we have moved away from a field we are unlikely to pay attention to it again. For example, once we have rendered all the occupations in the generated description there is no need to return back to the occupation field. In other words, once we have moved on ( INLINEFORM0 ), we want the successive field context vectors INLINEFORM1 to be very different from the previous field vectors INLINEFORM2 . One way of ensuring this is to orthogonalize successive field vectors using DISPLAYFORM0
where INLINEFORM0 is the dot product between vectors INLINEFORM1 and INLINEFORM2 . The above equation essentially subtracts the component of INLINEFORM3 along INLINEFORM4 . INLINEFORM5 is a learned parameter which controls the degree of orthogonalization thereby allowing a soft orthogonalization (i.e., the entire component along INLINEFORM6 is not subtracted but only a fraction of it). The above equation only ensures that INLINEFORM7 is soft-orthogonal to INLINEFORM8 . Alternately, we could pass the sequence of context vectors, INLINEFORM9 generated so far through a GRU cell. The state of this GRU cell at each time step would thus be aware of the history of the field vectors till that timestep. Now instead of orthogonalizing INLINEFORM10 to INLINEFORM11 we could orthogonalize INLINEFORM12 to the hidden state of this GRU at time-step INLINEFORM13 . In practice, we found this to work better as it accounts for all the field vectors in the history instead of only the previous field vector.
In summary, Equation provides a mechanism for remembering the current field vector when appropriate (thus capturing stay-on behavior) using a remember gate. On the other hand, Equation EQREF10 explicitly ensures that the field vector is very different (soft-orthogonal) from the previous field vectors once it is time to move on (thus capturing never look back behavior). The value of INLINEFORM0 computed in Equation EQREF10 is then used in Equation . The INLINEFORM1 (macro) thus obtained is then concatenated with INLINEFORM2 (micro) and fed to the decoder (see Fig. FIGREF3 )
Experimental setup
We now describe our experimental setup:
Datasets
We use the WikiBio dataset introduced by lebret2016neural. It consists of INLINEFORM0 biography articles from English Wikipedia. A biography article corresponds to a person (sportsman, politician, historical figure, actor, etc.). Each Wikipedia article has an accompanying infobox which serves as the structured input and the task is to generate the first sentence of the article (which typically is a one-line description of the person). We used the same train, valid and test sets which were made publicly available by lebret2016neural.
We also introduce two new biography datasets, one in French and one in German. These datasets were created and pre-processed using the same procedure as outlined in lebret2016neural. Specifically, we extracted the infoboxes and the first sentence from the corresponding Wikipedia article. As with the English dataset, we split the French and German datasets randomly into train (80%), test (10%) and valid (10%). The French and German datasets extracted by us has been made publicly available. The number of examples was 170K and 50K and the vocabulary size was 297K and 143K for French and German respectively. Although in this work we focus only on generating descriptions in one language, we hope that this dataset will also be useful for developing models which jointly learn to generate descriptions from structured data in multiple languages.
Models compared
We compare with the following models:
1. BIBREF0 : This is a conditional language model which uses a feed-forward neural network to predict the next word in the description conditioned on local characteristics (i.e., words within a field) and global characteristics (i.e., overall structure of the infobox).
2. BIBREF9 : This model was proposed in the context of the WeatherGov and RoboCup datasets which have a much smaller vocabulary. They use an improved attention model with additional regularizer terms which influence the weights assigned to the fields.
3. Basic Seq2Seq: This is the vanilla encode-attend-decode model BIBREF1 . Further, to deal with the large vocabulary ( INLINEFORM0 400K words) we use a copying mechanism as a post-processing step. Specifically, we identify the time steps at which the decoder produces unknown words (denoted by the special symbol UNK). For each such time step, we look at the attention weights on the input words and replace the UNK word by that input word which has received maximum attention at this timestep. This process is similar to the one described in BIBREF30 . Even lebret2016neural have a copying mechanism tightly integrated with their model.
Hyperparameter tuning
We tuned the hyperparameters of all the models using a validation set. As mentioned earlier, we used a bidirectional GRU cell as the function INLINEFORM0 for computing the representation of the fields and the values (see Section SECREF4 ). For all the models, we experimented with GRU state sizes of 128, 256 and 512. The total number of unique words in the corpus is around 400K (this includes the words in the infobox and the descriptions). Of these, we retained only the top 20K words in our vocabulary (same as BIBREF0 ). We initialized the embeddings of these words with 300 dimensional Glove embeddings BIBREF31 . We used Adam BIBREF32 with a learning rate of INLINEFORM1 , INLINEFORM2 and INLINEFORM3 . We trained the model for a maximum of 20 epochs and used early stopping with the patience set to 5 epochs.
Results and Discussions
We now discuss the results of our experiments.
Comparison of different models
Following lebret2016neural, we used BLEU-4, NIST-4 and ROUGE-4 as the evaluation metrics. We first make a few observations based on the results on the English dataset (Table TABREF15 ). The basic seq2seq model, as well as the model proposed by weather16, perform better than the model proposed by lebret2016neural. Our final model with bifocal attention and gated orthogonalization gives the best performance and does 10% (relative) better than the closest baseline (basic seq2seq) and 21% (relative) better than the current state of the art method BIBREF0 . In Table TABREF16 , we show some qualitative examples of the output generated by different models.
Human Evaluations
To make a qualitative assessment of the generated sentences, we conducted a human study on a sample of 500 Infoboxes which were sampled from English dataset. The annotators for this task were undergraduate and graduate students. For each of these infoboxes, we generated summaries using the basic seq2seq model and our final model with bifocal attention and gated orthogonalization. For each description and for each model, we asked three annotators to rank the output of the systems based on i) adequacy (i.e. does it capture relevant information from the infobox), (ii) fluency (i.e. grammar) and (iii) relative preference (i.e., which of the two outputs would be preferred). Overall the average fluency/adequacy (on a scale of 5) for basic seq2seq model was INLINEFORM0 and INLINEFORM1 for our model respectively.
The results from Table TABREF17 suggest that in general gated orthogonalization model performs better than the basic seq2seq model. Additionally, annotators were asked to verify if the generated summaries look natural (i.e, as if they were generated by humans). In 423 out of 500 cases, the annotators said “Yes” suggesting that gated orthogonalization model indeed produces good descriptions.
Performance on different languages
The results on the French and German datasets are summarized in Tables TABREF20 and TABREF20 respectively. Note that the code of BIBREF0 is not publicly available, hence we could not report numbers for French and German using their model. We observe that our final model gives the best performance - though the bifocal attention model performs poorly as compared to the basic seq2seq model on French. However, the overall performance for French and German are much smaller than those for English. There could be multiple reasons for this. First, the amount of training data in these two languages is smaller than that in English. Specifically, the amount of training data available in French (German) is only INLINEFORM0 ( INLINEFORM1 )% of that available for English. Second, on average the descriptions in French and German are longer than that in English (EN: INLINEFORM2 words, FR: INLINEFORM3 words and DE: INLINEFORM4 words). Finally, a manual inspection across the three languages suggests that the English descriptions have a more consistent structure than the French descriptions. For example, most English descriptions start with name followed by date of birth but this is not the case in French. However, this is only a qualitative observation and it is hard to quantify this characteristic of the French and German datasets.
Visualizing Attention Weights
If the proposed model indeed works well then we should see attention weights that are consistent with the stay on and never look back behavior. To verify this, we plotted the attention weights in cases where the model with gated orthogonalization does better than the model with only bifocal attention. Figure FIGREF21 shows the attention weights corresponding to infobox in Figure FIGREF25 . Notice that the model without gated orthogonalization has attention on both name field and article title while rendering the name. The model with gated orthogonalization, on the other hand, stays on the name field for as long as it is required but then moves and never returns to it (as expected).
Due to lack of space, we do not show similar plots for French and German but we would like to mention that, in general, the differences between the attention weights learned by the model with and without gated orthogonalization were more pronounced for the French/German dataset than the English dataset. This is in agreement with the results reported in Table TABREF20 and TABREF20 where the improvements given by gated orthogonalization are more for French/German than for English.
Out of domain results
What if the model sees a different INLINEFORM0 of person at test time? For example, what if the training data does not contain any sportspersons but at test time we encounter the infobox of a sportsperson. This is the same as seeing out-of-domain data at test time. Such a situation is quite expected in the products domain where new products with new features (fields) get frequently added to the catalog. We were interested in three questions here. First, we wanted to see if testing the model on out-of-domain data indeed leads to a drop in the performance. For this, we compared the performance of our best model in two scenarios (i) trained on data from all domains (including the target domain) and tested on the target domain (sports, arts) and (ii) trained on data from all domains except the target domain and tested on the target domain. Comparing rows 1 and 2 of Table TABREF32 we observed a significant drop in the performance. Note that the numbers for sports domain in row 1 are much better than the Arts domain because roughly 40% of the WikiBio training data contains sportspersons.
Next, we wanted to see if we can use a small amount of data from the target domain to fine tune a model trained on the out of domain data. We observe that even with very small amounts of target domain data the performance starts improving significantly (see rows 3 and 4 of Table TABREF32 ). Note that if we train a model from scratch with only limited data from the target domain instead of fine-tuning a model trained on a different source domain then the performance is very poor. In particular, training a model from scratch with 10K training instances we get a BLEU score of INLINEFORM0 and INLINEFORM1 for arts and sports respectively. Finally, even though the actual words used for describing a sportsperson (footballer, cricketer, etc.) would be very different from the words used to describe an artist (actor, musician, etc.) they might share many fields (for example, date of birth, occupation, etc.). As seen in Figure FIGREF28 (attention weights corresponding to the infobox in Figure FIGREF27 ), the model predicts the attention weights correctly for common fields (such as occupation) but it is unable to use the right vocabulary to describe the occupation (since it has not seen such words frequently in the training data). However, once we fine tune the model with limited data from the target domain we see that it picks up the new vocabulary and produces a correct description of the occupation.
Conclusion
We present a model for generating natural language descriptions from structured data. To address specific characteristics of the problem we propose neural components for fused bifocal attention and gated orthogonalization to address stay on and never look back behavior while decoding. Our final model outperforms an existing state of the art model on a large scale WikiBio dataset by 21%. We also introduce datasets for French and German and demonstrate that our model gives state of the art results on these datasets. Finally, we perform experiments with an out-of-domain model and show that if such a model is fine-tuned with small amounts of in domain data then it can give an improved performance on the target domain.
Given the multilingual nature of the new datasets, as future work, we would like to build models which can jointly learn to generate natural language descriptions from structured data in multiple languages. One idea is to replace the concepts in the input infobox by Wikidata concept ids which are language agnostic. A large amount of input vocabulary could thus be shared across languages thereby facilitating joint learning.
Acknowledgements
We thank Google for supporting Preksha Nema through their Google India Ph.D. Fellowship program. We also thank Microsoft Research India for supporting Shreyas Shetty through their generous travel grant for attending the conference.
| [
"Yes",
"Yes"
] | 4,457 | qasper | en | null | e0a92574e7c31adb9146851dcc95729b585305d15cf491d4 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Do they use pretrained embeddings?
Answer: | [
"Yes",
"Yes"
] | qasper | 128 | 51 | |
Was PolyReponse evaluated against some baseline? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction and Background
Task-oriented dialogue systems are primarily designed to search and interact with large databases which contain information pertaining to a certain dialogue domain: the main purpose of such systems is to assist the users in accomplishing a well-defined task such as flight booking BIBREF0, tourist information BIBREF1, restaurant search BIBREF2, or booking a taxi BIBREF3. These systems are typically constructed around rigid task-specific ontologies BIBREF1, BIBREF4 which enumerate the constraints the users can express using a collection of slots (e.g., price range for restaurant search) and their slot values (e.g., cheap, expensive for the aforementioned slots). Conversations are then modelled as a sequence of actions that constrain slots to particular values. This explicit semantic space is manually engineered by the system designer. It serves as the output of the natural language understanding component as well as the input to the language generation component both in traditional modular systems BIBREF5, BIBREF6 and in more recent end-to-end task-oriented dialogue systems BIBREF7, BIBREF8, BIBREF9, BIBREF3.
Working with such explicit semantics for task-oriented dialogue systems poses several critical challenges on top of the manual time-consuming domain ontology design. First, it is difficult to collect domain-specific data labelled with explicit semantic representations. As a consequence, despite recent data collection efforts to enable training of task-oriented systems across multiple domains BIBREF0, BIBREF3, annotated datasets are still few and far between, as well as limited in size and the number of domains covered. Second, the current approach constrains the types of dialogue the system can support, resulting in artificial conversations, and breakdowns when the user does not understand what the system can and cannot support. In other words, training a task-based dialogue system for voice-controlled search in a new domain always implies the complex, expensive, and time-consuming process of collecting and annotating sufficient amounts of in-domain dialogue data.
In this paper, we present a demo system based on an alternative approach to task-oriented dialogue. Relying on non-generative response retrieval we describe the PolyResponse conversational search engine and its application in the task of restaurant search and booking. The engine is trained on hundreds of millions of real conversations from a general domain (i.e., Reddit), using an implicit representation of semantics that directly optimizes the task at hand. It learns what responses are appropriate in different conversational contexts, and consequently ranks a large pool of responses according to their relevance to the current user utterance and previous dialogue history (i.e., dialogue context).
The technical aspects of the underlying conversational search engine are explained in detail in our recent work BIBREF11, while the details concerning the Reddit training data are also available in another recent publication BIBREF12. In this demo, we put focus on the actual practical usefulness of the search engine by demonstrating its potential in the task of restaurant search, and extending it to deal with multi-modal data. We describe a PolyReponse system that assists the users in finding a relevant restaurant according to their preference, and then additionally helps them to make a booking in the selected restaurant. Due to its retrieval-based design, with the PolyResponse engine there is no need to engineer a structured ontology, or to solve the difficult task of general language generation. This design also bypasses the construction of dedicated decision-making policy modules. The large ranking model already encapsulates a lot of knowledge about natural language and conversational flow.
Since retrieved system responses are presented visually to the user, the PolyResponse restaurant search engine is able to combine text responses with relevant visual information (e.g., photos from social media associated with the current restaurant and related to the user utterance), effectively yielding a multi-modal response. This setup of using voice as input, and responding visually is becoming more and more prevalent with the rise of smart screens like Echo Show and even mixed reality. Finally, the PolyResponse restaurant search engine is multilingual: it is currently deployed in 8 languages enabling search over restaurants in 8 cities around the world. System snapshots in four different languages are presented in Figure FIGREF16, while screencast videos that illustrate the dialogue flow with the PolyResponse engine are available at: https://tinyurl.com/y3evkcfz.
PolyResponse: Conversational Search
The PolyResponse system is powered by a single large conversational search engine, trained on a large amount of conversational and image data, as shown in Figure FIGREF2. In simple words, it is a ranking model that learns to score conversational replies and images in a given conversational context. The highest-scoring responses are then retrieved as system outputs. The system computes two sets of similarity scores: 1) $S(r,c)$ is the score of a candidate reply $r$ given a conversational context $c$, and 2) $S(p,c)$ is the score of a candidate photo $p$ given a conversational context $c$. These scores are computed as a scaled cosine similarity of a vector that represents the context ($h_c$), and a vector that represents the candidate response: a text reply ($h_r$) or a photo ($h_p$). For instance, $S(r,c)$ is computed as $S(r,c)=C cos(h_r,h_c)$, where $C$ is a learned constant. The part of the model dealing with text input (i.e., obtaining the encodings $h_c$ and $h_r$) follows the architecture introduced recently by Henderson:2019acl. We provide only a brief recap here; see the original paper for further details.
PolyResponse: Conversational Search ::: Text Representation.
The model, implemented as a deep neural network, learns to respond by training on hundreds of millions context-reply $(c,r)$ pairs. First, similar to Henderson:2017arxiv, raw text from both $c$ and $r$ is converted to unigrams and bigrams. All input text is first lower-cased and tokenised, numbers with 5 or more digits get their digits replaced by a wildcard symbol #, while words longer than 16 characters are replaced by a wildcard token LONGWORD. Sentence boundary tokens are added to each sentence. The vocabulary consists of the unigrams that occur at least 10 times in a random 10M subset of the Reddit training set (see Figure FIGREF2) plus the 200K most frequent bigrams in the same random subset.
During training, we obtain $d$-dimensional feature representations ($d=320$) shared between contexts and replies for each unigram and bigram jointly with other neural net parameters. A state-of-the-art architecture based on transformers BIBREF13 is then applied to unigram and bigram vectors separately, which are then averaged to form the final 320-dimensional encoding. That encoding is then passed through three fully-connected non-linear hidden layers of dimensionality $1,024$. The final layer is linear and maps the text into the final $l$-dimensional ($l=512$) representation: $h_c$ and $h_r$. Other standard and more sophisticated encoder models can also be used to provide final encodings $h_c$ and $h_r$, but the current architecture shows a good trade-off between speed and efficacy with strong and robust performance in our empirical evaluations on the response retrieval task using Reddit BIBREF14, OpenSubtitles BIBREF15, and AmazonQA BIBREF16 conversational test data, see BIBREF12 for further details.
In training the constant $C$ is constrained to lie between 0 and $\sqrt{l}$. Following Henderson:2017arxiv, the scoring function in the training objective aims to maximise the similarity score of context-reply pairs that go together, while minimising the score of random pairings: negative examples. Training proceeds via SGD with batches comprising 500 pairs (1 positive and 499 negatives).
PolyResponse: Conversational Search ::: Photo Representation.
Photos are represented using convolutional neural net (CNN) models pretrained on ImageNet BIBREF17. We use a MobileNet model with a depth multiplier of 1.4, and an input dimension of $224 \times 224$ pixels as in BIBREF18. This provides a $1,280 \times 1.4 = 1,792$-dimensional representation of a photo, which is then passed through a single hidden layer of dimensionality $1,024$ with ReLU activation, before being passed to a hidden layer of dimensionality 512 with no activation to provide the final representation $h_p$.
PolyResponse: Conversational Search ::: Data Source 1: Reddit.
For training text representations we use a Reddit dataset similar to AlRfou:2016arxiv. Our dataset is large and provides natural conversational structure: all Reddit data from January 2015 to December 2018, available as a public BigQuery dataset, span almost 3.7B comments BIBREF12. We preprocess the dataset to remove uninformative and long comments by retaining only sentences containing more than 8 and less than 128 word tokens. After pairing all comments/contexts $c$ with their replies $r$, we obtain more than 727M context-reply $(c,r)$ pairs for training, see Figure FIGREF2.
PolyResponse: Conversational Search ::: Data Source 2: Yelp.
Once the text encoding sub-networks are trained, a photo encoder is learned on top of a pretrained MobileNet CNN, using data taken from the Yelp Open dataset: it contains around 200K photos and their captions. Training of the multi-modal sub-network then maximises the similarity of captions encoded with the response encoder $h_r$ to the photo representation $h_p$. As a result, we can compute the score of a photo given a context using the cosine similarity of the respective vectors. A photo will be scored highly if it looks like its caption would be a good response to the current context.
PolyResponse: Conversational Search ::: Index of Responses.
The Yelp dataset is used at inference time to provide text and photo candidates to display to the user at each step in the conversation. Our restaurant search is currently deployed separately for each city, and we limit the responses to a given city. For instance, for our English system for Edinburgh we work with 396 restaurants, 4,225 photos (these include additional photos obtained using the Google Places API without captions), 6,725 responses created from the structured information about restaurants that Yelp provides, converted using simple templates to sentences of the form such as “Restaurant X accepts credit cards.”, 125,830 sentences extracted from online reviews.
PolyResponse: Conversational Search ::: PolyResponse in a Nutshell.
The system jointly trains two encoding functions (with shared word embeddings) $f(context)$ and $g(reply)$ which produce encodings $h_c$ and $h_r$, so that the similarity $S(c,r)$ is high for all $(c,r)$ pairs from the Reddit training data and low for random pairs. The encoding function $g()$ is then frozen, and an encoding function $t(photo)$ is learnt which makes the similarity between a photo and its associated caption high for all (photo, caption) pairs from the Yelp dataset, and low for random pairs. $t$ is a CNN pretrained on ImageNet, with a shallow one-layer DNN on top. Given a new context/query, we then provide its encoding $h_c$ by applying $f()$, and find plausible text replies and photo responses according to functions $g()$ and $t()$, respectively. These should be responses that look like answers to the query, and photos that look like they would have captions that would be answers to the provided query.
At inference, finding relevant candidates given a context reduces to computing $h_c$ for the context $c$ , and finding nearby $h_r$ and $h_p$ vectors. The response vectors can all be pre-computed, and the nearest neighbour search can be further optimised using standard libraries such as Faiss BIBREF19 or approximate nearest neighbour retrieval BIBREF20, giving an efficient search that scales to billions of candidate responses.
The system offers both voice and text input and output. Speech-to-text and text-to-speech conversion in the PolyResponse system is currently supported by the off-the-shelf Google Cloud tools.
Dialogue Flow
The ranking model lends itself to the one-shot task of finding the most relevant responses in a given context. However, a restaurant-browsing system needs to support a dialogue flow where the user finds a restaurant, and then asks questions about it. The dialogue state for each search scenario is represented as the set of restaurants that are considered relevant. This starts off as all the restaurants in the given city, and is assumed to monotonically decrease in size as the conversation progresses until the user converges to a single restaurant. A restaurant is only considered valid in the context of a new user input if it has relevant responses corresponding to it. This flow is summarised here:
S1. Initialise $R$ as the set of all restaurants in the city. Given the user's input, rank all the responses in the response pool pertaining to restaurants in $R$.
S2. Retrieve the top $N$ responses $r_1, r_2, \ldots , r_N$ with corresponding (sorted) cosine similarity scores: $s_1 \ge s_2 \ge \ldots \ge s_N$.
S3. Compute probability scores $p_i \propto \exp (a \cdot s_i)$ with $\sum _{i=1}^N p_i$, where $a>0$ is a tunable constant.
S4. Compute a score $q_e$ for each restaurant/entity $e \in R$, $q_e = \sum _{i: r_i \in e} p_i$.
S5. Update $R$ to the smallest set of restaurants with highest $q$ whose $q$-values sum up to more than a predefined threshold $t$.
S6. Display the most relevant responses associated with the updated $R$, and return to S2.
If there are multiple relevant restaurants, one response is shown from each. When only one restaurant is relevant, the top $N$ responses are all shown, and relevant photos are also displayed. The system does not require dedicated understanding, decision-making, and generation modules, and this dialogue flow does not rely on explicit task-tailored semantics. The set of relevant restaurants is kept internally while the system narrows it down across multiple dialogue turns. A simple set of predefined rules is used to provide a templatic spoken system response: e.g., an example rule is “One review of $e$ said $r$”, where $e$ refers to the restaurant, and $r$ to a relevant response associated with $e$. Note that while the demo is currently focused on the restaurant search task, the described “narrowing down” dialogue flow is generic and applicable to a variety of applications dealing with similar entity search.
The system can use a set of intent classifiers to allow resetting the dialogue state, or to activate the separate restaurant booking dialogue flow. These classifiers are briefly discussed in §SECREF4.
Other Functionality ::: Multilinguality.
The PolyResponse restaurant search is currently available in 8 languages and for 8 cities around the world: English (Edinburgh), German (Berlin), Spanish (Madrid), Mandarin (Taipei), Polish (Warsaw), Russian (Moscow), Korean (Seoul), and Serbian (Belgrade). Selected snapshots are shown in Figure FIGREF16, while we also provide videos demonstrating the use and behaviour of the systems at: https://tinyurl.com/y3evkcfz. A simple MT-based translate-to-source approach at inference time is currently used to enable the deployment of the system in other languages: 1) the pool of responses in each language is translated to English by Google Translate beforehand, and pre-computed encodings of their English translations are used as representations of each foreign language response; 2) a provided user utterance (i.e., context) is translated to English on-the-fly and its encoding $h_c$ is then learned. We plan to experiment with more sophisticated multilingual models in future work.
Other Functionality ::: Voice-Controlled Menu Search.
An additional functionality enables the user to get parts of the restaurant menu relevant to the current user utterance as responses. This is achieved by performing an additional ranking step of available menu items and retrieving the ones that are semantically relevant to the user utterance using exactly the same methodology as with ranking other responses. An example of this functionality is shown in Figure FIGREF21.
Other Functionality ::: Resetting and Switching to Booking.
The restaurant search system needs to support the discrete actions of restarting the conversation (i.e., resetting the set $R$), and should enable transferring to the slot-based table booking flow. This is achieved using two binary intent classifiers, that are run at each step in the dialogue. These classifiers make use of the already-computed $h_c$ vector that represents the user's latest text. A single-layer neural net is learned on top of the 512-dimensional encoding, with a ReLU activation and 100 hidden nodes. To train the classifiers, sets of 20 relevant paraphrases (e.g., “Start again”) are provided as positive examples. Finally, when the system successfully switches to the booking scenario, it proceeds to the slot filling task: it aims to extract all the relevant booking information from the user (e.g., date, time, number of people to dine). The entire flow of the system illustrating both the search phase and the booking phase is provided as the supplemental video material.
Conclusion and Future Work
This paper has presented a general approach to search-based dialogue that does not rely on explicit semantic representations such as dialogue acts or slot-value ontologies, and allows for multi-modal responses. In future work, we will extend the current demo system to more tasks and languages, and work with more sophisticated encoders and ranking functions. Besides the initial dialogue flow from this work (§SECREF3), we will also work with more complex flows dealing, e.g., with user intent shifts.
| [
"No",
"No"
] | 2,738 | qasper | en | null | f545e80cf01375e891406755e35019032cb4b7621338b707 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Was PolyReponse evaluated against some baseline?
Answer: | [
"No",
"No"
] | qasper | 128 | 52 | |
How do they obtain psychological dimensions of people? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Blogging gained momentum in 1999 and became especially popular after the launch of freely available, hosted platforms such as blogger.com or livejournal.com. Blogging has progressively been used by individuals to share news, ideas, and information, but it has also developed a mainstream role to the extent that it is being used by political consultants and news services as a tool for outreach and opinion forming as well as by businesses as a marketing tool to promote products and services BIBREF0 .
For this paper, we compiled a very large geolocated collection of blogs, written by individuals located in the U.S., with the purpose of creating insightful mappings of the blogging community. In particular, during May-July 2015, we gathered the profile information for all the users that have self-reported their location in the U.S., along with a number of posts for all their associated blogs. We utilize this blog collection to generate maps of the U.S. that reflect user demographics, language use, and distributions of psycholinguistic and semantic word classes. We believe that these maps can provide valuable insights and partial verification of previous claims in support of research in linguistic geography BIBREF1 , regional personality BIBREF2 , and language analysis BIBREF3 , BIBREF4 , as well as psychology and its relation to human geography BIBREF5 .
Data Collection
Our premise is that we can generate informative maps using geolocated information available on social media; therefore, we guide the blog collection process with the constraint that we only accept blogs that have specific location information. Moreover, we aim to find blogs belonging to writers from all 50 U.S. states, which will allow us to build U.S. maps for various dimensions of interest.
We first started by collecting a set of profiles of bloggers that met our location specifications by searching individual states on the profile finder on http://www.blogger.com. Starting with this list, we can locate the profile page for a user, and subsequently extract additional information, which includes fields such as name, email, occupation, industry, and so forth. It is important to note that the profile finder only identifies users that have an exact match to the location specified in the query; we thus built and ran queries that used both state abbreviations (e.g., TX, AL), as well as the states' full names (e.g., Texas, Alabama).
After completing all the processing steps, we identified 197,527 bloggers with state location information. For each of these bloggers, we found their blogs (note that a blogger can have multiple blogs), for a total of 335,698 blogs. For each of these blogs, we downloaded the 21 most recent blog postings, which were cleaned of HTML tags and tokenized, resulting in a collection of 4,600,465 blog posts.
Maps from Blogs
Our dataset provides mappings between location, profile information, and language use, which we can leverage to generate maps that reflect demographic, linguistic, and psycholinguistic properties of the population represented in the dataset.
People Maps
The first map we generate depicts the distribution of the bloggers in our dataset across the U.S. Figure FIGREF1 shows the density of users in our dataset in each of the 50 states. For instance, the densest state was found to be California with 11,701 users. The second densest is Texas, with 9,252 users, followed by New York, with 9,136. The state with the fewest bloggers is Delaware with 1,217 users. Not surprisingly, this distribution correlates well with the population of these states, with a Spearman's rank correlation INLINEFORM0 of 0.91 and a p-value INLINEFORM1 0.0001, and is very similar to the one reported in Lin and Halavais Lin04.
Figure FIGREF1 also shows the cities mentioned most often in our dataset. In particular, it illustrates all 227 cities that have at least 100 bloggers. The bigger the dot on the map, the larger the number of users found in that city. The five top blogger-dense cities, in order, are: Chicago, New York, Portland, Seattle, and Atlanta.
We also generate two maps that delineate the gender distribution in the dataset. Overall, the blogging world seems to be dominated by females: out of 153,209 users who self-reported their gender, only 52,725 are men and 100,484 are women. Figures FIGREF1 and FIGREF1 show the percentage of male and female bloggers in each of the 50 states. As seen in this figure, there are more than the average number of male bloggers in states such as California and New York, whereas Utah and Idaho have a higher percentage of women bloggers.
Another profile element that can lead to interesting maps is the Industry field BIBREF6 . Using this field, we created different maps that plot the geographical distribution of industries across the country. As an example, Figure FIGREF2 shows the percentage of the users in each state working in the automotive and tourism industries respectively.
Linguistic Maps
Another use of the information found in our dataset is to build linguistic maps, which reflect the geographic lexical variation across the 50 states BIBREF7 . We generate maps that represent the relative frequency by which a word occurs in the different states. Figure FIGREF3 shows sample maps created for two different words. The figure shows the map generated for one location specific word, Maui, which unsurprisingly is found predominantly in Hawaii, and a map for a more common word, lake, which has a high occurrence rate in Minnesota (Land of 10,000 Lakes) and Utah (home of the Great Salt Lake). Our demo described in Section SECREF4 , can also be used to generate maps for function words, which can be very telling regarding people's personality BIBREF8 .
Psycholinguistic and Semantic Maps
LIWC. In addition to individual words, we can also create maps for word categories that reflect a certain psycholinguistic or semantic property. Several lexical resources, such as Roget or Linguistic Inquiry and Word Count BIBREF9 , group words into categories. Examples of such categories are Money, which includes words such as remuneration, dollar, and payment; or Positive feelings with words such as happy, cheerful, and celebration. Using the distribution of the individual words in a category, we can compile distributions for the entire category, and therefore generate maps for these word categories. For instance, figure FIGREF8 shows the maps created for two categories: Positive Feelings and Money. The maps are not surprising, and interestingly they also reflect an inverse correlation between Money and Positive Feelings .
Values. We also measure the usage of words related to people's core values as reported by Boyd et al. boyd2015. The sets of words, or themes, were excavated using the Meaning Extraction Method (MEM) BIBREF10 . MEM is a topic modeling approach applied to a corpus of texts created by hundreds of survey respondents from the U.S. who were asked to freely write about their personal values. To illustrate, Figure FIGREF9 shows the geographical distributions of two of these value themes: Religion and Hard Work. Southeastern states often considered as the nation's “Bible Belt” BIBREF11 were found to have generally higher usage of Religion words such as God, bible, and church. Another broad trend was that western-central states (e.g., Wyoming, Nebraska, Iowa) commonly blogged about Hard Work, using words such as hard, work, and job more often than bloggers in other regions.
Web Demonstration
A prototype, interactive charting demo is available at http://lit.eecs.umich.edu/~geoliwc/. In addition to drawing maps of the geographical distributions on the different LIWC categories, the tool can report the three most and least correlated LIWC categories in the U.S. and compare the distributions of any two categories.
Conclusions
In this paper, we showed how we can effectively leverage a prodigious blog dataset. Not only does the dataset bring out the extensive linguistic content reflected in the blog posts, but also includes location information and rich metadata. These data allow for the generation of maps that reflect the demographics of the population, variations in language use, and differences in psycholinguistic and semantic categories. These mappings can be valuable to both psychologists and linguists, as well as lexicographers. A prototype demo has been made available together with the code used to collect our dataset.
Acknowledgments
This material is based in part upon work supported by the National Science Foundation (#1344257) and by the John Templeton Foundation (#48503). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation or the John Templeton Foundation. We would like to thank our colleagues Hengjing Wang, Jiatao Fan, Xinghai Zhang, and Po-Jung Huang who provided technical help with the implementation of the demo.
| [
"using the Meaning Extraction Method",
"Unanswerable"
] | 1,440 | qasper | en | null | 43ffd7775c3b4a541e227c120bffcc7c7b31fb184ba94d69 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How do they obtain psychological dimensions of people?
Answer: | [
"using the Meaning Extraction Method",
"Unanswerable"
] | qasper | 128 | 53 | |
What argument components do the ML methods aim to identify? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The art of argumentation has been studied since the early work of Aristotle, dating back to the 4th century BC BIBREF0 . It has been exhaustively examined from different perspectives, such as philosophy, psychology, communication studies, cognitive science, formal and informal logic, linguistics, computer science, educational research, and many others. In a recent and critically well-acclaimed study, Mercier.Sperber.2011 even claim that argumentation is what drives humans to perform reasoning. From the pragmatic perspective, argumentation can be seen as a verbal activity oriented towards the realization of a goal BIBREF1 or more in detail as a verbal, social, and rational activity aimed at convincing a reasonable critic of the acceptability of a standpoint by putting forward a constellation of one or more propositions to justify this standpoint BIBREF2 .
Analyzing argumentation from the computational linguistics point of view has very recently led to a new field called argumentation mining BIBREF3 . Despite the lack of an exact definition, researchers within this field usually focus on analyzing discourse on the pragmatics level and applying a certain argumentation theory to model and analyze textual data at hand.
Our motivation for argumentation mining stems from a practical information seeking perspective from the user-generated content on the Web. For example, when users search for information in user-generated Web content to facilitate their personal decision making related to controversial topics, they lack tools to overcome the current information overload. One particular use-case example dealing with a forum post discussing private versus public schools is shown in Figure FIGREF4 . Here, the lengthy text on the left-hand side is transformed into an argument gist on the right-hand side by (i) analyzing argument components and (ii) summarizing their content. Figure FIGREF5 shows another use-case example, in which users search for reasons that underpin certain standpoint in a given controversy (which is homeschooling in this case). In general, the output of automatic argument analysis performed on the large scale in Web data can provide users with analyzed arguments to a given topic of interest, find the evidence for the given controversial standpoint, or help to reveal flaws in argumentation of others.
Satisfying the above-mentioned information needs cannot be directly tackled by current methods for, e.g., opinion mining, questions answering, or summarization and requires novel approaches within the argumentation mining field. Although user-generated Web content has already been considered in argumentation mining, many limitations and research gaps can be identified in the existing works. First, the scope of the current approaches is restricted to a particular domain or register, e.g., hotel reviews BIBREF5 , Tweets related to local riot events BIBREF6 , student essays BIBREF7 , airline passenger rights and consumer protection BIBREF8 , or renewable energy sources BIBREF9 . Second, not all the related works are tightly connected to argumentation theories, resulting into a gap between the substantial research in argumentation itself and its adaptation in NLP applications. Third, as an emerging research area, argumentation mining still suffers from a lack of labeled corpora, which is crucial for designing, training, and evaluating the algorithms. Although some works have dealt with creating new data sets, the reliability (in terms of inter-annotator agreement) of the annotated resources is often unknown BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 .
Annotating and automatically analyzing arguments in unconstrained user-generated Web discourse represent challenging tasks. So far, the research in argumentation mining “has been conducted on domains like news articles, parliamentary records and legal documents, where the documents contain well-formed explicit arguments, i.e., propositions with supporting reasons and evidence present in the text” BIBREF8 . [p. 50]Boltuzic.Snajder.2014 point out that “unlike in debates or other more formal argumentation sources, the arguments provided by the users, if any, are less formal, ambiguous, vague, implicit, or often simply poorly worded.” Another challenge stems from the different nature of argumentation theories and computational linguistics. Whereas computational linguistics is mainly descriptive, the empirical research that is carried out in argumentation theories does not constitute a test of the theoretical model that is favored, because the model of argumentation is a normative instrument for assessing the argumentation BIBREF15 . So far, no fully fledged descriptive argumentation theory based on empirical research has been developed, thus feasibility of adapting argumentation models to the Web discourse represents an open issue.
These challenges can be formulated into the following research questions:
In this article, we push the boundaries of the argumentation mining field by focusing on several novel aspects. We tackle the above-mentioned research questions as well as the previously discussed challenges and issues. First, we target user-generated Web discourse from several domains across various registers, to examine how argumentation is communicated in different contexts. Second, we bridge the gap between argumentation theories and argumentation mining through selecting the argumenation model based on research into argumentation theories and related fields in communication studies or psychology. In particular, we adapt normative models from argumentation theory to perform empirical research in NLP and support our application of argumentation theories with an in-depth reliability study. Finally, we use state-of-the-art NLP techniques in order to build robust computational models for analyzing arguments that are capable of dealing with a variety of genres on the Web.
Our contributions
We create a new corpus which is, to the best of our knowledge, the largest corpus that has been annotated within the argumentation mining field to date. We choose several target domains from educational controversies, such as homeschooling, single-sex education, or mainstreaming. A novel aspect of the corpus is its coverage of different registers of user-generated Web content, such as comments to articles, discussion forum posts, blog posts, as well as professional newswire articles.
Since the data come from a variety of sources and no assumptions about its actual content with respect to argumentation can be drawn, we conduct two extensive annotation studies. In the first study, we tackle the problem of relatively high “noise” in the retrieved data. In particular, not all of the documents are related to the given topics in a way that makes them candidates for further deep analysis of argumentation (this study results into 990 annotated documents). In the second study, we discuss the selection of an appropriate argumentation model based on evidence in argumentation research and propose a model that is suitable for analyzing micro-level argumention in user-generated Web content. Using this model, we annotate 340 documents (approx. 90,000 tokens), reaching a substantial inter-annotator agreement. We provide a hand-analysis of all the phenomena typical to argumentation that are prevalent in our data. These findings may also serve as empirical evidence to issues that are on the spot of current argumentation research.
From the computational perspective, we experiment on the annotated data using various machine learning methods in order to extract argument structure from documents. We propose several novel feature sets and identify configurations that run best in in-domain and cross-domain scenarios. To foster research in the community, we provide the annotated data as well as all the experimental software under free license.
The rest of the article is structured as follows. First, we provide an essential background in argumentation theory in section SECREF2 . Section SECREF3 surveys related work in several areas. Then we introduce the dataset and two annotation studies in section SECREF4 . Section SECREF5 presents our experimental work and discusses the results and errors and section SECREF6 concludes this article.
Theoretical background
Let us first present some definitions of the term argumentation itself. [p. 3]Ketcham.1917 defines argumentation as “the art of persuading others to think or act in a definite way. It includes all writing and speaking which is persuasive in form.” According to MacEwan.1898, “argumentation is the process of proving or disproving a proposition. Its purpose is to induce a new belief, to establish truth or combat error in the mind of another.” [p. 2]Freeley.Steinberg.2008 narrow the scope of argumentation to “reason giving in communicative situations by people whose purpose is the justification of acts, beliefs, attitudes, and values.” Although these definitions vary, the purpose of argumentation remains the same – to persuade others.
We would like to stress that our perception of argumentation goes beyond somehow limited giving reasons BIBREF17 , BIBREF18 . Rather, we see the goal of argumentation as to persuade BIBREF19 , BIBREF20 , BIBREF21 . Persuasion can be defined as a successful intentional effort at influencing another's mental state through communication in a circumstance in which the persuadee has some measure of freedom BIBREF22 , although, as OKeefe2011 points out, there is no correct or universally-endorsed definition of either `persuasion' or `argumentation'. However, broader understanding of argumentation as a means of persuasion allows us to take into account not only reasoned discourse, but also non-reasoned mechanisms of influence, such as emotional appeals BIBREF23 .
Having an argument as a product within the argumentation process, we should now define it. One typical definition is that an argument is a claim supported by reasons BIBREF24 . The term claim has been used since 1950's, introduced by Toulmin.1958, and in argumentation theory it is a synonym for standpoint or point of view. It refers to what is an issue in the sense what is being argued about. The presence of a standpoint is thus crucial for argumentation analysis. However, the claim as well as other parts of the argument might be implicit; this is known as enthymematic argumentation, which is rather usual in ordinary argumentative discourse BIBREF25 .
One fundamental problem with the definition and formal description of arguments and argumentation is that there is no agreement even among argumentation theorists. As [p. 29]vanEmeren.et.al.2014 admit in their very recent and exhaustive survey of the field, ”as yet, there is no unitary theory of argumentation that encompasses the logical, dialectical, and rhetorical dimensions of argumentation and is universally accepted. The current state of the art in argumentation theory is characterized by the coexistence of a variety of theoretical perspectives and approaches, which differ considerably from each other in conceptualization, scope, and theoretical refinement.”
Argumentation models
Despite the missing consensus on the ultimate argumentation theory, various argumentation models have been proposed that capture argumentation on different levels. Argumentation models abstract from the language level to a concept level that stresses the links between the different components of an argument or how arguments relate to each other BIBREF26 . Bentahar.et.al.2010 propose a taxonomy of argumentation models, that is horizontally divided into three categories – micro-level models, macro-level models, and rhetorical models.
In this article, we deal with argumentation on the micro-level (also called argumentation as a product or monological models). Micro-level argumentation focuses on the structure of a single argument. By contrast, macro-level models (also called dialogical models) and rhetorical models highlight the process of argumentation in a dialogue BIBREF27 . In other words, we examine the structure of a single argument produced by a single author in term of its components, not the relations that can exist among arguments and their authors in time. A detailed discussion of these different perspectives can be found, e.g., in BIBREF28 , BIBREF29 , BIBREF30 , BIBREF1 , BIBREF31 , BIBREF32 .
Dimensions of argument
The above-mentioned models focus basically only on one dimension of the argument, namely the logos dimension. According to the classical Aristotle's theory BIBREF0 , argument can exist in three dimensions, which are logos, pathos, and ethos. Logos dimension represents a proof by reason, an attempt to persuade by establishing a logical argument. For example, syllogism belongs to this argumentation dimension BIBREF34 , BIBREF25 . Pathos dimension makes use of appealing to emotions of the receiver and impacts its cognition BIBREF35 . Ethos dimension of argument relies on the credibility of the arguer. This distinction will have practical impact later in section SECREF51 which deals with argumentation on the Web.
Original Toulmin's model
We conclude the theoretical section by presenting one (micro-level) argumentation model in detail – a widely used conceptual model of argumentation introduced by Toulmin.1958, which we will henceforth denote as the Toulmin's original model. This model will play an important role later in the annotation studies (section SECREF51 ) and experimental work (section SECREF108 ). The model consists of six parts, referred as argument components, where each component plays a distinct role.
is an assertion put forward publicly for general acceptance BIBREF38 or the conclusion we seek to establish by our arguments BIBREF17 .
It is the evidence to establish the foundation of the claim BIBREF24 or, as simply put by Toulmin, “the data represent what we have to go on.” BIBREF37 . The name of this concept was later changed to grounds in BIBREF38 .
The role of warrant is to justify a logical inference from the grounds to the claim.
is a set of information that stands behind the warrant, it assures its trustworthiness.
limits the degree of certainty under which the argument should be accepted. It is the degree of force which the grounds confer on the claim in virtue of the warrant BIBREF37 .
presents a situation in which the claim might be defeated.
A schema of the Toulmin's original model is shown in Figure FIGREF29 . The lines and arrows symbolize implicit relations between the components. An example of an argument rendered using the Toulmin's scheme can be seen in Figure FIGREF30 .
We believe that this theoretical overview should provide sufficient background for the argumentation mining research covered in this article; for further references, we recommend for example BIBREF15 .
Related work in computational linguistics
We structure the related work into three sub-categories, namely argumentation mining, stance detection, and persuasion and on-line dialogs, as these areas are closest to this article's focus. For a recent overview of general discourse analysis see BIBREF39 . Apart from these, research on computer-supported argumentation has been also very active; see, e.g., BIBREF40 for a survey of various models and argumentation formalisms from the educational perspective or BIBREF41 which examines argumentation in the Semantic Web.
Argumentation Mining
The argumentation mining field has been evolving very rapidly in the recent years, resulting into several workshops co-located with major NLP conferences. We first present related works with a focus on annotations and then review experiments with classifying argument components, schemes, or relations.
One of the first papers dealing with annotating argumentative discourse was Argumentative Zoning for scientific publications BIBREF42 . Later, Teufel.et.al.2009 extended the original 7 categories to 15 and annotated 39 articles from two domains, where each sentence is assigned a category. The obtained Fleiss' INLINEFORM0 was 0.71 and 0.65. In their approach, they tried to deliberately ignore the domain knowledge and rely only on general, rhetorical and logical aspect of the annotated texts. By contrast to our work, argumentative zoning is specific to scientific publications and has been developed solely for that task.
Reed.Rowe.2004 presented Araucaria, a tool for argumentation diagramming which supports both convergent and linked arguments, missing premises (enthymemes), and refutations. They also released the AracuariaDB corpus which has later been used for experiments in the argumentation mining field. However, the creation of the dataset in terms of annotation guidelines and reliability is not reported – these limitations as well as its rather small size have been identified BIBREF10 .
Biran.Rambow.2011 identified justifications for subjective claims in blog threads and Wikipedia talk pages. The data were annotated with claims and their justifications reaching INLINEFORM0 0.69, but a detailed description of the annotation approach was missing.
[p. 1078]Schneider.et.al.2013b annotated Wikipedia talk pages about deletion using 17 Walton's schemes BIBREF43 , reaching a moderate agreement (Cohen's INLINEFORM0 0.48) and concluded that their analysis technique can be reused, although “it is intensive and difficult to apply.”
Stab.Gurevych.2014 annotated 90 argumentative essays (about 30k tokens), annotating claims, major claims, and premises and their relations (support, attack). They reached Krippendorff's INLINEFORM0 0.72 for argument components and Krippendorff's INLINEFORM1 0.81 for relations between components.
Rosenthal2012 annotated sentences that are opinionated claims, in which the author expresses a belief that should be adopted by others. Two annotators labeled sentences as claims without any context and achieved Cohen's INLINEFORM0 0.50 (2,000 sentences from LiveJournal) and 0.56 (2,000 sentences from Wikipedia).
Aharoni.et.al.2014 performed an annotation study in order to find context-dependent claims and three types of context-dependent evidence in Wikipedia, that were related to 33 controversial topics. The claim and evidence were annotated in 104 articles. The average Cohen's INLINEFORM0 between a group of 20 expert annotators was 0.40. Compared to our work, the linguistic properties of Wikipedia are qualitatively different from other user-generated content, such as blogs or user comments BIBREF44 .
Wacholder.et.al.2014 annotated “argument discourse units” in blog posts and criticized the Krippendorff's INLINEFORM0 measure. They proposed a new inter-annotator metric by taking the most overlapping part of one annotation as the “core” and all annotations as a “cluster”. The data were extended by Ghosh2014, who annotated “targets” and “callouts” on the top of the units.
Park.Cardie.2014 annotated about 10k sentences from 1,047 documents into four types of argument propositions with Cohen's INLINEFORM0 0.73 on 30% of the dataset. Only 7% of the sentences were found to be non-argumentative.
Faulkner2014 used Amazon Mechanical Turk to annotate 8,179 sentences from student essays. Three annotators decided whether the given sentence offered reasons for or against the main prompt of the essay (or no reason at all; 66% of the sentences were found to be neutral and easy to identify). The achieved Cohen's INLINEFORM0 was 0.70.
The research has also been active on non-English datasets. Goudas.et.al.2014 focused on user-generated Greek texts. They selected 204 documents and manually annotated sentences that contained an argument (760 out of 16,000). They distinguished claims and premises, but the claims were always implicit. However, the annotation agreement was not reported, neither was the number of annotators or the guidelines. A study on annotation of arguments was conducted by Peldszus.Stede.2013, who evaluate agreement among 26 “naive" annotators (annotators with very little training). They manually constructed 23 German short texts, each of them contains exactly one central claim, two premises, and one objection (rebuttal or undercut) and analyzed annotator agreement on this artificial data set. Peldszus.2014 later achieved higher inter-rater agreement with expert annotators on an extended version of the same data. Kluge.2014 built a corpus of argumentative German Web documents, containing 79 documents from 7 educational topics, which were annotated by 3 annotators according to the claim-premise argumentation model. The corpus comprises 70,000 tokens and the inter-annotator agreement was 0.40 (Krippendorff's INLINEFORM0 ). Houy.et.al.2013 targeted argumentation mining of German legal cases.
Table TABREF33 gives an overview of annotation studies with their respective argumentation model, domain, size, and agreement. It also contains other studies outside of computational linguistics and few proposals and position papers.
Arguments in the legal domain were targeted in BIBREF11 . Using argumentation formalism inspired by Walton.2012, they employed multinomial Naive Bayes classifier and maximum entropy model for classifying argumentative sentences on the AraucariaDB corpus BIBREF45 . The same test dataset was used by Feng.Hirst.2011, who utilized the C4.5 decision classifier. Rooney.et.al.2012 investigated the use of convolution kernel methods for classifying whether a sentence belongs to an argumentative element or not using the same corpus.
Stab.Gurevych.2014b classified sentences to four categories (none, major claim, claim, premise) using their previously annotated corpus BIBREF7 and reached 0.72 macro- INLINEFORM0 score. In contrast to our work, their documents are expected to comply with a certain structure of argumentative essays and are assumed to always contain argumentation.
Biran.Rambow.2011 identified justifications on the sentence level using a naive Bayes classifier over a feature set based on statistics from the RST Treebank, namely n-grams which were manually processed by deleting n-grams that “seemed irrelevant, ambiguous or domain-specific.”
Llewellyn2014 experimented with classifying tweets into several argumentative categories, namely claims and counter-claims (with and without evidence) and verification inquiries previously annotated by Procter.et.al.2013. They used unigrams, punctuations, and POS as features in three classifiers.
Park.Cardie.2014 classified propositions into three classes (unverifiable, verifiable non-experimental, and verifiable experimental) and ignored non-argumentative texts. Using multi-class SVM and a wide range of features (n-grams, POS, sentiment clue words, tense, person) they achieved Macro INLINEFORM0 0.69.
Peldszus.2014 experimented with a rather complex labeling schema of argument segments, but their data were artificially created for their task and manually cleaned, such as removing segments that did not meet the criteria or non-argumentative segments.
In the first step of their two-phase approach, Goudas.et.al.2014 sampled the dataset to be balanced and identified argumentative sentences with INLINEFORM0 0.77 using the maximum entropy classifier. For identifying premises, they used BIO encoding of tokens and achieved INLINEFORM1 score 0.42 using CRFs.
Saint-Dizier.2012 developed a Prolog engine using a lexicon of 1300 words and a set of 78 hand-crafted rules with the focus on a particular argument structure “reasons supporting conclusions” in French.
Taking the dialogical perspective, Cabrio.Villata.2012 built upon an argumentation framework proposed by Dung.1995 which models arguments within a graph structure and provides a reasoning mechanism for resolving accepted arguments. For identifying support and attack, they relied on existing research on textual entailment BIBREF46 , namely using the off-the-shelf EDITS system. The test data were taken from a debate portal Debatepedia and covered 19 topics. Evaluation was performed in terms of measuring the acceptance of the “main argument" using the automatically recognized entailments, yielding INLINEFORM0 score about 0.75. By contrast to our work which deals with micro-level argumentation, the Dung's model is an abstract framework intended to model dialogical argumentation.
Finding a bridge between existing discourse research and argumentation has been targeted by several researchers. Peldszus2013a surveyed literature on argumentation and proposed utilization of Rhetorical Structure Theory (RST) BIBREF47 . They claimed that RST is by its design well-suited for studying argumentative texts, but an empirical evidence has not yet been provided. Penn Discourse Tree Bank (PDTB) BIBREF48 relations have been under examination by argumentation mining researchers too. Cabrio2013b examined a connection between five Walton's schemes and discourse markers in PDTB, however an empirical evaluation is missing.
Stance detection
Research related to argumentation mining also involves stance detection. In this case, the whole document (discussion post, article) is assumed to represent the writer's standpoint to the discussed topic. Since the topic is stated as a controversial question, the author is either for or against it.
Somasundaran.Wiebe.2009 built a computational model for recognizing stances in dual-topic debates about named entities in the electronic products domain by combining preferences learned from the Web data and discourse markers from PDTB BIBREF48 . Hasan.Ng.2013 determined stance in on-line ideological debates on four topics using data from createdebate.com, employing supervised machine learning and features ranging from n-grams to semantic frames. Predicting stance of posts in Debatepedia as well as external articles using a probabilistic graphical model was presented in BIBREF49 . This approach also employed sentiment lexicons and Named Entity Recognition as a preprocessing step and achieved accuracy about 0.80 in binary prediction of stances in debate posts.
Recent research has involved joint modeling, taking into account information about the users, the dialog sequences, and others. Hasan.Ng.2012 proposed machine learning approach to debate stance classification by leveraging contextual information and author's stances towards the topic. Qiu.et.al.2013 introduced a computational debate side model to cluster posts or users by sides for general threaded discussions using a generative graphical model employing words from various subjectivity lexicons as well as all adjectives and adverbs in the posts. Qiu.Jiang.2013 proposed a graphical model for viewpoint discovery in discussion threads. Burfoot.et.al.2011 exploited the informal citation structure in U.S. Congressional floor-debate transcripts and use a collective classification which outperforms methods that consider documents in isolation.
Some works also utilize argumentation-motivated features. Park.et.al.2011 dealt with contentious issues in Korean newswire discourse. Although they annotate the documents with “argument frames”, the formalism remains unexplained and does not refer to any existing research in argumentation. Walker.et.al.2012b incorporated features with some limited aspects of the argument structure, such as cue words signaling rhetorical relations between posts, POS generalized dependencies, and a representation of the parent post (context) to improve stance classification over 14 topics from convinceme.net.
Online persuasion
Another stream of research has been devoted to persuasion in online media, which we consider as a more general research topic than argumentation.
Schlosser.2011 investigated persuasiveness of online reviews and concluded that presenting two sides is not always more helpful and can even be less persuasive than presenting one side. Mohammadi.et.al.2013 explored persuasiveness of speakers in YouTube videos and concluded that people are perceived more persuasive in video than in audio and text. Miceli.et.al.2006 proposed a computational model that attempts to integrate emotional and non-emotional persuasion. In the study of Murphy.2001, persuasiveness was assigned to 21 articles (out of 100 manually preselected) and four of them are later analyzed in detail for comparing the perception of persuasion between expert and students. Bernard.et.al.2012 experimented with children's perception of discourse connectives (namely with “because”) to link statements in arguments and found out that 4- and 5-years-old and adults are sensitive to the connectives. Le.2004 presented a study of persuasive texts and argumentation in newspaper editorials in French.
A coarse-grained view on dialogs in social media was examined by Bracewell.et.al.2013, who proposed a set of 15 social acts (such as agreement, disagreement, or supportive behavior) to infer the social goals of dialog participants and presented a semi-supervised model for their classification. Their social act types were inspired by research in psychology and organizational behavior and were motivated by work in dialog understanding. They annotated a corpus in three languages using in-house annotators and achieved INLINEFORM0 in the range from 0.13 to 0.53.
Georgila.et.al.2011 focused on cross-cultural aspects of persuasion or argumentation dialogs. They developed a novel annotation scheme stemming from different literature sources on negotiation and argumentation as well as from their original analysis of the phenomena. The annotation scheme is claimed to cover three dimensions of an utterance, namely speech act, topic, and response or reference to a previous utterance. They annotated 21 dialogs and reached Krippendorff's INLINEFORM0 between 0.38 and 0.57.
Given the broad landscape of various approaches to argument analysis and persuasion studies presented in this section, we would like to stress some novel aspects of the current article. First, we aim at adapting a model of argument based on research by argumentation scholars, both theoretical and empirical. We pose several pragmatical constraints, such as register independence (generalization over several registers). Second, our emphasis is put on reliable annotations and sufficient data size (about 90k tokens). Third, we deal with fairly unrestricted Web-based sources, so additional steps of distinguishing whether the texts are argumentative are required. Argumentation mining has been a rapidly evolving field with several major venues in 2015. We encourage readers to consult an upcoming survey article by Lippi.Torroni.2016 or the proceedings of the 2nd Argumentation Mining workshop BIBREF50 to keep up with recent developments. However, to the best of our knowledge, the main findings of this article have not yet been made obsolete by any related work.
Annotation studies and corpus creation
This section describes the process of data selection, annotation, curation, and evaluation with the goal of creating a new corpus suitable for argumentation mining research in the area of computational linguistics. As argumentation mining is an evolving discipline without established and widely-accepted annotation schemes, procedures, and evaluation, we want to keep this overview detailed to ensure full reproducibility of our approach. Given the wide range of perspectives on argumentation itself BIBREF15 , variety of argumentation models BIBREF27 , and high costs of discourse or pragmatic annotations BIBREF48 , creating a new, reliable corpus for argumentation mining represents a substantial effort.
A motivation for creating a new corpus stems from the various use-cases discussed in the introduction, as well as some research gaps pointed in section SECREF1 and further discussed in the survey in section SECREF31 (e.g., domain restrictions, missing connection to argumentation theories, non-reported reliability or detailed schemes).
Topics and registers
As a main field of interest in the current study, we chose controversies in education. One distinguishing feature of educational topics is their breadth of sub-topics and points of view, as they attract researchers, practitioners, parents, students, or policy-makers. We assume that this diversity leads to the linguistic variability of the education topics and thus represents a challenge for NLP. In a cooperation with researchers from the German Institute for International Educational Research we identified the following current controversial topics in education in English-speaking countries: (1) homeschooling, (2) public versus private schools, (3) redshirting — intentionally delaying the entry of an age-eligible child into kindergarten, allowing their child more time to mature emotionally and physically BIBREF51 , (4) prayer in schools — whether prayer in schools should be allowed and taken as a part of education or banned completely, (5) single-sex education — single-sex classes (males and females separate) versus mixed-sex classes (“co-ed”), and (6) mainstreaming — including children with special needs into regular classes.
Since we were also interested in whether argumentation differs across registers, we included four different registers — namely (1) user comments to newswire articles or to blog posts, (2) posts in discussion forums (forum posts), (3) blog posts, and (4) newswire articles. Throughout this work, we will refer to each article, blog post, comment, or forum posts as a document. This variety of sources covers mainly user-generated content except newswire articles which are written by professionals and undergo an editing procedure by the publisher. Since many publishers also host blog-like sections on their portals, we consider as blog posts all content that is hosted on personal blogs or clearly belong to a blog category within a newswire portal.
Raw corpus statistics
Given the six controversial topics and four different registers, we compiled a collection of plain-text documents, which we call the raw corpus. It contains 694,110 tokens in 5,444 documents. As a coarse-grained analysis of the data, we examined the lengths and the number of paragraphs (see Figure FIGREF43 ). Comments and forum posts follow a similar distribution, being shorter than 300 tokens on average. By contrast, articles and blogs are longer than 400 tokens and have 9.2 paragraphs on average. The process of compiling the raw corpus and its further statistics are described in detail in Appendix UID158 .
Annotation study 1: Identifying persuasive documents in forums and comments
The goal of this study was to select documents suitable for a fine-grained analysis of arguments. In a preliminary study on annotating argumentation using a small sample (50 random documents) of forum posts and comments from the raw corpus, we found that many documents convey no argumentation at all, even in discussions about controversies. We observed that such contributions do not intend to persuade; these documents typically contain story-sharing, personal worries, user interaction (asking questions, expressing agreement), off-topic comments, and others. Such characteristics are typical to on-line discussions in general, but they have not been examined with respect to argumentation or persuasion. Indeed, we observed that there are (1) documents that are completely unrelated and (2) documents that are related to the topic, but do not contain any argumentation. This issue has been identified among argumentation theorist; for example as external relevance by Paglieri.Castelfranchia.2014. Similar findings were also confirmed in related literature in argumentation mining, however never tackled empirically BIBREF53 , BIBREF8 These documents are thus not suitable for analyzing argumentation.
In order to filter documents that are suitable for argumentation annotation, we defined a binary document-level classification task. The distinction is made between either persuasive documents or non-persuasive (which includes all other sorts of texts, such as off-topic, story sharing, unrelated dialog acts, etc.).
The two annotated categories were on-topic persuasive and non-persuasive. Three annotators with near-native English proficiency annotated a set of 990 documents (a random subset of comments and forum posts) reaching 0.59 Fleiss' INLINEFORM0 . The final label was selected by majority voting. The annotation study took on average of 15 hours per annotator with approximately 55 annotated documents per hour. The resulting labels were derived by majority voting. Out of 990 documents, 524 (53%) were labeled as on-topic persuasive. We will refer to this corpus as gold data persuasive.
We examined all disagreements between annotators and discovered some typical problems, such as implicitness or topic relevance. First, the authors often express their stance towards the topic implicitly, so it must be inferred by the reader. To do so, certain common-ground knowledge is required. However, such knowledge heavily depends on many aspects, such as the reader's familiarity with the topic or her cultural background, as well as the context of the source website or the discussion forum thread. This also applies for sarcasm and irony. Second, the decision whether a particular topic is persuasive was always made with respect to the controversial topic under examination. Some authors shift the focus to a particular aspect of the given controversy or a related issue, making the document less relevant.
We achieved moderate agreement between the annotators, although the definition of persuasiveness annotation might seem a bit fuzzy. We found different amounts of persuasion in the specific topics. For instance, prayer in schools or private vs. public schools attract persuasive discourse, while other discussed controversies often contain non-persuasive discussions, represented by redshirting and mainstreaming. Although these two topics are also highly controversial, the participants of on-line discussions seem to not attempt to persuade but they rather exchange information, support others in their decisions, etc. This was also confirmed by socio-psychological researchers. Ammari.et.al.2014 show that parents of children with special needs rely on discussion sites for accessing information and social support and that, in particular, posts containing humor, achievement, or treatment suggestions are perceived to be more socially appropriate than posts containing judgment, violence, or social comparisons. According to Nicholson.Leask.2012, in the online forum, parents of autistic children were seen to understand the issue because they had lived it. Assuming that participants in discussions related to young kids (e.g., redshirting, or mainstreaming) are usually females (mothers), the gender can also play a role. In a study of online persuasion, Guadagno.Cialdini.2002 conclude that women chose to bond rather than compete (women feel more comfortable cooperating, even in a competitive environment), whereas men are motivated to compete if necessary to achieve independence.
Annotation study 2: Annotating micro-structure of arguments
The goal of this study was to annotate documents on a detailed level with respect to an argumentation model. First, we will present the annotation scheme. Second, we will describe the annotation process. Finally, we will evaluate the agreement and draw some conclusions.
Given the theoretical background briefly introduced in section SECREF2 , we motivate our selection of the argumentation model by the following requirements. First, the scope of this work is to capture argumentation within a single document, thus focusing on micro-level models. Second, there should exist empirical evidence that such a model has been used for analyzing argumentation in previous works, so it is likely to be suitable for our purposes of argumentative discourse analysis in user-generated content. Regarding the first requirement, two typical examples of micro-level models are the Toulmin's model BIBREF36 and Walton's schemes BIBREF55 . Let us now elaborate on the second requirement.
Walton's argumentation schemes are claimed to be general and domain independent. Nevertheless, evidence from the computational linguistics field shows that the schemes lack coverage for analyzing real argumentation in natural language texts. In examining real-world political argumentation from BIBREF56 , Walton.2012 found out that 37.1% of the arguments collected did not fit any of the fourteen schemes they chose so they created new schemes ad-hoc. Cabrio2013b selected five argumentation schemes from Walton and map these patterns to discourse relation categories in the Penn Discourse TreeBank (PDTB) BIBREF48 , but later they had to define two new argumentation schemes that they discovered in PDTB. Similarly, Song.et.al.2014 admitted that the schemes are ambiguous and hard to directly apply for annotation, therefore they modified the schemes and created new ones that matched the data.
Although Macagno.Konstantinidou.2012 show several examples of two argumentation schemes applied to few selected arguments in classroom experiments, empirical evidence presented by Anthony.Kim.2014 reveals many practical and theoretical difficulties of annotating dialogues with schemes in classroom deliberation, providing many details on the arbitrary selection of the sub-set of the schemes, the ambiguity of the scheme definitions, concluding that the presence of the authors during the experiment was essential for inferring and identifying the argument schemes BIBREF57 .
Although this model (refer to section SECREF21 ) was designed to be applicable to real-life argumentation, there are numerous studies criticizing both the clarity of the model definition and the differentiation between elements of the model. Ball1994 claims that the model can be used only for the most simple arguments and fails on the complex ones. Also Freeman1991 and other argumentation theorists criticize the usefulness of Toulmin's framework for the description of real-life argumentative texts. However, others have advocated the model and claimed that it can be applied to the people's ordinary argumenation BIBREF58 , BIBREF59 .
A number of studies (outside the field of computational linguistics) used Toulmin's model as their backbone argumentation framework. Chambliss1995 experimented with analyzing 20 written documents in a classroom setting in order to find the argument patterns and parts. Simosi2003 examined employees' argumentation to resolve conflicts. Voss2006 analyzed experts' protocols dealing with problem-solving.
The model has also been used in research on computer-supported collaborative learning. Erduran2004 adapt Toulmin's model for coding classroom argumentative discourse among teachers and students. Stegmann2011 builds on a simplified Toulmin's model for scripted construction of argument in computer-supported collaborative learning. Garcia-Mila2013 coded utterances into categories from Toulmin's model in persuasion and consensus-reaching among students. Weinberger.Fischer.2006 analyze asynchronous discussion boards in which learners engage in an argumentative discourse with the goal to acquire knowledge. For coding the argument dimension, they created a set of argumentative moves based on Toulmin's model. Given this empirical evidence, we decided to build upon the Toulmin's model.
In this annotation task, a sequence of tokens (e.g. a phrase, a sentence, or any arbitrary text span) is labeled with a corresponding argument component (such as the claim, the grounds, and others). There are no explicit relations between these annotation spans as the relations are implicitly encoded in the pragmatic function of the components in the Toulmin's model.
In order to prove the suitability of the Toulmin's model, we analyzed 40 random documents from the gold data persuasive dataset using the original Toulmin's model as presented in section SECREF21 . We took into account sever criteria for assessment, such as frequency of occurrence of the components or their importance for the task. We proposed some modifications of the model based on the following observations.
Authors do not state the degree of cogency (the probability of their claim, as proposed by Toulmin). Thus we omitted qualifier from the model due to its absence in the data.
The warrant as a logical explanation why one should accept the claim given the evidence is almost never stated. As pointed out by BIBREF37 , “data are appealed to explicitly, warrants implicitly.” This observation has also been made by Voss2006. Also, according to [p. 205]Eemeren.et.al.1987, the distinction of warrant is perfectly clear only in Toulmin’s examples, but the definitions fail in practice. We omitted warrant from the model.
Rebuttal is a statement that attacks the claim, thus playing a role of an opposing view. In reality, the authors often attack the presented rebuttals by another counter-rebuttal in order to keep the whole argument's position consistent. Thus we introduced a new component – refutation – which is used for attacking the rebuttal. Annotation of refutation was conditioned of explicit presence of rebuttal and enforced by the annotation guidelines. The chain rebuttal–refutation is also known as the procatalepsis figure in rhetoric, in which the speaker raises an objection to his own argument and then immediately answers it. By doing so, the speaker hopes to strengthen the argument by dealing with possible counter-arguments before the audience can raise them BIBREF43 .
The claim of the argument should always reflect the main standpoint with respect to the discussed controversy. We observed that this standpoint is not always explicitly expressed, but remains implicit and must be inferred by the reader. Therefore, we allow the claim to be implicit. In such a case, the annotators must explicitly write down the (inferred) stance of the author.
By definition, the Toulmin's model is intended to model single argument, with the claim in its center. However, we observed in our data, that some authors elaborate on both sides of the controversy equally and put forward an argument for each side (by argument here we mean the claim and its premises, backings, etc.). Therefore we allow multiple arguments to be annotated in one document. At the same time, we restrained the annotators from creating complex argument hierarchies.
Toulmin's grounds have an equivalent role to a premise in the classical view on an argument BIBREF15 , BIBREF60 in terms that they offer the reasons why one should accept the standpoint expressed by the claim. As this terminology has been used in several related works in the argumentation mining field BIBREF7 , BIBREF61 , BIBREF62 , BIBREF11 , we will keep this convention and denote the grounds as premises.
One of the main critiques of the original Toulmin's model was the vague distinction between grounds, warrant, and backing BIBREF63 , BIBREF64 , BIBREF65 . The role of backing is to give additional support to the warrant, but there is no warrant in our model anymore. However, what we observed during the analysis, was a presence of some additional evidence. Such evidence does not play the role of the grounds (premises) as it is not meant as a reason supporting the claim, but it also does not explain the reasoning, thus is not a warrant either. It usually supports the whole argument and is stated by the author as a certain fact. Therefore, we extended the scope of backing as an additional support to the whole argument.
The annotators were instructed to distinguish between premises and backing, so that premises should cover generally applicable reasons for the claim, whereas backing is a single personal experience or statements that give credibility or attribute certain expertise to the author. As a sanity check, the argument should still make sense after removing backing (would be only considered “weaker”).
We call the model as a modified Toulmin's model. It contains five argument components, namely claim, premise, backing, rebuttal, and refutation. When annotating a document, any arbitrary token span can be labeled with an argument component; the components do not overlap. The spans are not known in advance and the annotator thus chooses the span and the component type at the same time. All components are optional (they do not have to be present in the argument) except the claim, which is either explicit or implicit (see above). If a token span is not labeled by any argument component, it is not considered as a part of the argument and is later denoted as none (this category is not assigned by the annotators).
An example analysis of a forum post is shown in Figure FIGREF65 . Figure FIGREF66 then shows a diagram of the analysis from that example (the content of the argument components was shortened or rephrased).
The annotation experiment was split into three phases. All documents were annotated by three independent annotators, who participated in two training sessions. During the first phase, 50 random comments and forum posts were annotated. Problematic cases were resolved after discussion and the guidelines were refined. In the second phase, we wanted to extend the range of annotated registers, so we selected 148 comments and forum posts as well as 41 blog posts. After the second phase, the annotation guidelines were final.
In the final phase, we extended the range of annotated registers and added newswire articles from the raw corpus in order to test whether the annotation guidelines (and inherently the model) is general enough. Therefore we selected 96 comments/forum posts, 8 blog posts, and 8 articles for this phase. A detailed inter-annotator agreement study on documents from this final phase will be reported in section UID75 .
The annotations were very time-consuming. In total, each annotator spent 35 hours by annotating in the course of five weeks. Discussions and consolidation of the gold data took another 6 hours. Comments and forum posts required on average of 4 minutes per document to annotate, while blog posts and articles on average of 14 minutes per document. Examples of annotated documents from the gold data are listed in Appendix UID158 .
We discarded 11 documents out of the total 351 annotated documents. Five forum posts, although annotated as persuasive in the first annotation study, were at a deeper look a mixture of two or more posts with missing quotations, therefore unsuitable for analyzing argumentation. Three blog posts and two articles were found not to be argumentative (the authors took no stance to the discussed controversy) and one article was an interview, which the current model cannot capture (a dialogical argumentation model would be required).
For each of the 340 documents, the gold standard annotations were obtained using the majority vote. If simple majority voting was not possible (different boundaries of the argument component together with a different component label), the gold standard was set after discussion among the annotators. We will refer to this corpus as the gold standard Toulmin corpus. The distribution of topics and registers in this corpus in shown in Table TABREF71 , and Table TABREF72 presents some lexical statistics.
Based on pre-studies, we set the minimal unit for annotation as token. The documents were pre-segmented using the Stanford Core NLP sentence splitter BIBREF69 embedded in the DKPro Core framework BIBREF70 . Annotators were asked to stick to the sentence level by default and label entire pre-segmented sentences. They should switch to annotations on the token level only if (a) a particular sentence contained more than one argument component, or (b) if the automatic sentence segmentation was wrong. Given the “noise” in user-generated Web data (wrong or missing punctuation, casing, etc.), this was often the case.
Annotators were also asked to rephrase (summarize) each annotated argument component into a simple statement when applicable, as shown in Figure FIGREF66 . This was used as a first sanity checking step, as each argument component is expected to be a coherent discourse unit. For example, if a particular occurrence of a premise cannot be summarized/rephrased into one statement, this may require further splitting into two or more premises.
For the actual annotations, we developed a custom-made web-based application that allowed users to switch between different granularity of argument components (tokens or sentences), to annotate the same document in different argument “dimensions” (logos and pathos), and to write summary for each annotated argument component.
As a measure of annotation reliability, we rely on Krippendorff's unitized alpha ( INLINEFORM0 ) BIBREF71 . To the best of our knowledge, this is the only agreement measure that is applicable when both labels and boundaries of segments are to be annotated.
Although the measure has been used in related annotation works BIBREF61 , BIBREF7 , BIBREF72 , there is one important detail that has not been properly communicated. The INLINEFORM0 is computed over a continuum of the smallest units, such as tokens. This continuum corresponds to a single document in the original Krippendorff's work. However, there are two possible extensions to multiple documents (a corpus), namely (a) to compute INLINEFORM1 for each document first and then report an average value, or (b) to concatenate all documents into one large continuum and compute INLINEFORM2 over it. The first approach with averaging yielded extremely high the standard deviation of INLINEFORM3 (i.e., avg. = 0.253; std. dev. = 0.886; median = 0.476 for the claim). This says that some documents are easy to annotate while others are harder, but interpretation of such averaged value has no evidence either in BIBREF71 or other papers based upon it. Thus we use the other methodology and treat the whole corpus as a single long continuum (which yields in the example of claim 0.541 INLINEFORM4 ).
Table TABREF77 shows the inter-annotator agreement as measured on documents from the last annotation phase (see section UID67 ). The overall INLINEFORM0 for all register types, topics, and argument components is 0.48 in the logos dimension (annotated with the modified Toulmin's model). Such agreement can be considered as moderate by the measures proposed by Landis.Koch.1977, however, direct interpretation of the agreement value lacks consensus BIBREF54 . Similar inter-annotator agreement numbers were achieved in the relevant works in argumentation mining (refer to Table TABREF33 in section SECREF31 ; although most of the numbers are not directly comparable, as different inter-annotator metrics were used on different tasks).
There is a huge difference in INLINEFORM0 regarding the registers between comments + forums posts ( INLINEFORM1 0.60, Table TABREF77 a) and articles + blog posts ( INLINEFORM2 0.09, Table TABREF77 b) in the logos dimension. If we break down the value with respect to the individual argument components, the agreement on claim and premise is substantial in the case of comments and forum posts (0.59 and 0.69, respectively). By contrast, these argument components were annotated only with a fair agreement in articles and blog posts (0.22 and 0.24, respectively).
As can be also observed from Table TABREF77 , the annotation agreement in the logos dimension varies regarding the document topic. While it is substantial/moderate for prayer in schools (0.68) or private vs. public schools (0.44), for some topics it remains rather slight, such as in the case of redshirting (0.14) or mainstreaming (0.08).
First, we examine the disagreement in annotations by posing the following research question: are there any measurable properties of the annotated documents that might systematically cause low inter-annotator agreement? We use Pearson's correlation coefficient between INLINEFORM0 on each document and the particular property under investigation. We investigated the following set of measures.
Full sentence coverage ratio represents a ratio of argument component boundaries that are aligned to sentence boundaries. The value is 1.0 if all annotations in the particular document are aligned to sentences and 0.0 if no annotations match the sentence boundaries. Our hypothesis was that automatic segmentation to sentences was often incorrect, therefore annotators had to switch to the token level annotations and this might have increased disagreement on boundaries of the argument components.
Document length, paragraph length and average sentence length. Our hypotheses was that the length of documents, paragraphs, or sentences negatively affects the agreement.
Readability measures. We tested four standard readability measures, namely Ari BIBREF73 , Coleman-Liau BIBREF74 , Flesch BIBREF75 , and Lix BIBREF76 to find out whether readability of the documents plays any role in annotation agreement.
Correlation results are listed in Table TABREF82 . We observed the following statistically significant ( INLINEFORM0 ) correlations. First, document length negatively correlates with agreement in comments. The longer the comment was the lower the agreement was. Second, average paragraph length negatively correlates with agreement in blog posts. The longer the paragraphs in blogs were, the lower agreement was reached. Third, all readability scores negatively correlate with agreement in the public vs. private school domain, meaning that the more complicated the text in terms of readability is, the lower agreement was reached. We observed no significant correlation in sentence coverage and average sentence length measures. We cannot draw any general conclusion from these results, but we can state that some registers and topics, given their properties, are more challenging to annotate than others.
Another qualitative analysis of disagreements between annotators was performed by constructing a probabilistic confusion matrix BIBREF77 on the token level. The biggest disagreements, as can be seen in Table TABREF85 , is caused by rebuttal and refutation confused with none (0.27 and 0.40, respectively). This is another sign that these two argument components were very hard to annotate. As shown in Table TABREF77 , the INLINEFORM5 was also low – 0.08 for rebuttal and 0.17 for refutation.
We analyzed the annotations and found the following phenomena that usually caused disagreements between annotators.
Each argument component (e.g., premise or backing) should express one consistent and coherent piece of information, for example a single reason in case of the premise (see Section UID73 ). However, the decision whether a longer text should be kept as a single argument component or segmented into multiple components is subjective and highly text-specific.
While rhetorical questions have been researched extensively in linguistics BIBREF78 , BIBREF79 , BIBREF80 , BIBREF81 , their role in argumentation represents a substantial research question BIBREF82 , BIBREF83 , BIBREF84 , BIBREF85 , BIBREF86 . Teninbaum.2011 provides a brief history of rhetorical questions in persuasion. In short, rhetorical questions should provoke the reader. From the perspective of our argumentation model, rhetorical questions might fall both into the logos dimension (and thus be labeled as, e.g., claim, premise, etc.) or into the pathos dimension (refer to Section SECREF20 ). Again, the decision is usually not clear-cut.
As introduced in section UID55 , rebuttal attacks the claim by presenting an opponent's view. In most cases, the rebuttal is again attacked by the author using refutation. From the pragmatic perspective, refutation thus supports the author's stance expressed by the claim. Therefore, it can be easily confused with premises, as the function of both is to provide support for the claim. Refutation thus only takes place if it is meant as a reaction to the rebuttal. It follows the discussed matter and contradicts it. Such a discourse is usually expressed as:
[claim: My claim.] [rebuttal: On the other hand, some people claim XXX which makes my claim wrong.] [refutation: But this is not true, because of YYY.]
However, the author might also take the following defensible approach to formulate the argument:
[rebuttal: Some people claim XXX-1 which makes my claim wrong.] [refutation: But this is not true, because of YYY-1.] [rebuttal: Some people claim XXX-2 which makes my claim wrong.] [refutation: But this is not true, because of YYY-2.] [claim: Therefore my claim.]
If this argument is formulated without stating the rebuttals, it would be equivalent to the following:
[premise: YYY-1.] [premise: YYY-2.] [claim: Therefore my claim.]
This example shows that rebuttal and refutation represent a rhetorical device to produce arguments, but the distinction between refutation and premise is context-dependent and on the functional level both premise and refutation have very similar role – to support the author's standpoint. Although introducing dialogical moves into monological model and its practical consequences, as described above, can be seen as a shortcoming of our model, this rhetoric figure has been identified by argumentation researchers as procatalepsis BIBREF43 . A broader view on incorporating opposing views (or lack thereof) is discussed under the term confirmation bias by BIBREF21 who claim that “[...] people are trying to convince others. They are typically looking for arguments and evidence to confirm their own claim, and ignoring negative arguments and evidence unless they anticipate having to rebut them.” The dialectical attack of possible counter-arguments may thus strengthen one's own argument.
One possible solution would be to refrain from capturing this phenomena completely and to simplify the model to claims and premises, for instance. However, the following example would then miss an important piece of information, as the last two clauses would be left un-annotated. At the same time, annotating the last clause as premise would be misleading, because it does not support the claim (in fact, it supports it only indirectly by attacking the rebuttal; this can be seen as a support is considered as an admissible extension of abstract argument graph by BIBREF87 ).
Doc#422 (forumpost, homeschooling) [claim: I try not to be anti-homeschooling, but... it's just hard for me.] [premise: I really haven't met any homeschoolers who turned out quite right, including myself.] I apologize if what I'm saying offends any of you - that's not my intention, [rebuttal: I know that there are many homeschooled children who do just fine,] but [refutation: that hasn't been my experience.]
To the best of our knowledge, these context-dependent dialogical properties of argument components using Toulmin's model have not been solved in the literature on argumentation theory and we suggest that these observations should be taken into account in the future research in monological argumentation.
Appeal to emotion, sarcasm, irony, or jokes are common in argumentation in user-generated Web content. We also observed documents in our data that were purely sarcastic (the pathos dimension), therefore logical analysis of the argument (the logos dimension) would make no sense. However, given the structure of such documents, some claims or premises might be also identified. Such an argument is a typical example of fallacious argumentation, which intentionally pretends to present a valid argument, but its persuasion is conveyed purely for example by appealing to emotions of the reader BIBREF88 .
We present some statistics of the annotated data that are important from the argumentation research perspective. Regardless of the register, 48% of claims are implicit. This means that the authors assume that their standpoint towards the discussed controversy can be inferred by the reader and give only reasons for that standpoint. Also, explicit claims are mainly written just once, only in 3% of the documents the claim was rephrased and occurred multiple times.
In 6% of the documents, the reasons for an implicit claim are given only in the pathos dimension, making the argument purely persuasive without logical argumentation.
The “myside bias”, defined as a bias against information supporting another side of an argument BIBREF89 , BIBREF90 , can be observed by the presence of rebuttals to the author's claim or by formulating arguments for both sides when the overall stance is neutral. While 85% of the documents do not consider any opposing side, only 8% documents present a rebuttal, which is then attacked by refutation in 4% of the documents. Multiple rebuttals and refutations were found in 3% of the documents. Only 4% of the documents were overall neutral and presented arguments for both sides, mainly in blog posts.
We were also interested whether mitigating linguistic devices are employed in the annotated arguments, namely in their main stance-taking components, the claims. Such devices typically include parenthetical verbs, syntactic constructions, token agreements, hedges, challenge questions, discourse markers, and tag questions, among others BIBREF91 . In particular, [p. 1]Kaltenbock.et.al.2010 define hedging as a discourse strategy that reduces the force or truth of an utterance and thus reduces the risk a speaker runs when uttering a strong or firm assertion or other speech act. We manually examined the use of hedging in the annotated claims.
Our main observation is that hedging is used differently across topics. For instance, about 30-35% of claims in homeschooling and mainstreaming signal the lack of a full commitment to the expressed stance, in contrast to prayer in schools (15%) or public vs. private schools (about 10%). Typical hedging cues include speculations and modality (“If I have kids, I will probably homeschool them.”), statements as neutral observations (“It's not wrong to hold the opinion that in general it's better for kids to go to school than to be homeschooled.”), or weasel phrases BIBREF92 (“In some cases, inclusion can work fantastically well.”, “For the majority of the children in the school, mainstream would not have been a suitable placement.”).
On the other hand, most claims that are used for instance in the prayer in schools arguments are very direct, without trying to diminish its commitment to the conveyed belief (for example, “NO PRAYER IN SCHOOLS!... period.”, “Get it out of public schools”, “Pray at home.”, or “No organized prayers or services anywhere on public school board property - FOR ANYONE.”). Moreover, some claims are clearly offensive, persuading by direct imperative clauses towards the opponents/audience (“TAKE YOUR KIDS PRIVATE IF YOU CARE AS I DID”, “Run, don't walk, to the nearest private school.”) or even accuse the opponents for taking a certain stance (“You are a bad person if you send your children to private school.”).
These observations are consistent with the findings from the first annotation study on persuasion (see section UID48 ), namely that some topics attract heated argumentation where participant take very clear and reserved standpoints (such as prayer in schools or private vs. public schools), while discussions about other topics are rather milder. It has been shown that the choices a speaker makes to express a position are informed by their social and cultural background, as well as their ability to speak the language BIBREF93 , BIBREF94 , BIBREF91 . However, given the uncontrolled settings of the user-generated Web content, we cannot infer any similar conclusions in this respect.
We investigated premises across all topics in order to find the type of support used in the argument. We followed the approach of Park.Cardie.2014, who distinguished three types of propositions in their study, namely unverifiable, verifiable non-experiential, and verifiable experiential.
Verifiable non-experiential and verifiable experiential propositions, unlike unverifiable propositions, contain an objective assertion, where objective means “expressing or dealing with facts or conditions as perceived without distortion by personal feelings, prejudices, or interpretations.” Such assertions have truth values that can be proved or disproved with objective evidence; the correctness of the assertion or the availability of the objective evidence does not matter BIBREF8 . A verifiable proposition can further be distinguished as experiential or not, depending on whether the proposition is about the writer's personal state or experience or something non-experiential. Verifiable experiential propositions are sometimes referred to as anectotal evidence, provide the novel knowledge that readers are seeking BIBREF8 .
Table TABREF97 shows the distribution of the premise types with examples for each topic from the annotated corpus. As can be seen in the first row, arguments in prayer in schools contain majority (73%) of unverifiable premises. Closer examination reveals that their content vary from general vague propositions to obvious fallacies, such as a hasty generalization, straw men, or slippery slope. As Nieminen.Mustonen.2014 found out, fallacies are very common in argumentation about religion-related issues. On the other side of the spectrum, arguments about redshirting rely mostly on anecdotal evidence (61% of verifiable experiential propositions). We will discuss the phenomena of narratives in argumentation in more detail later in section UID98 . All the topics except private vs. public schools exhibit similar amount of verifiable non-experiential premises (9%–22%), usually referring to expert studies or facts. However, this type of premises has usually the lowest frequency.
Manually analyzing argumentative discourse and reconstructing (annotating) the underlying argument structure and its components is difficult. As [p. 267]Reed2006 point out, “the analysis of arguments is often hard, not only for students, but for experts too.” According to [p. 81]Harrell.2011b, argumentation is a skill and “even for simple arguments, untrained college students can identify the conclusion but without prompting are poor at both identifying the premises and how the premises support the conclusion.” [p. 81]Harrell.2011 further claims that “a wide literature supports the contention that the particular skills of understanding, evaluating, and producing arguments are generally poor in the population of people who have not had specific training and that specific training is what improves these skills.” Some studies, for example, show that students perform significantly better on reasoning tasks when they have learned to identify premises and conclusions BIBREF95 or have learned some standard argumentation norms BIBREF96 .
One particular extra challenge in analyzing argumentation in Web user-generated discourse is that the authors produce their texts probably without any existing argumentation theory or model in mind. We assume that argumentation or persuasion is inherent when users discuss controversial topics, but the true reasons why people participate in on-line communities and what drives their behavior is another research question BIBREF97 , BIBREF98 , BIBREF99 , BIBREF100 . When the analyzed texts have a clear intention to produce argumentative discourse, such as in argumentative essays BIBREF7 , the argumentation is much more explicit and a substantially higher inter-annotator agreement can be achieved.
The model seems to be suitable for short persuasive documents, such as comments and forum posts. Its applicability to longer documents, such as articles or blog posts, is problematic for several reasons.
The argument components of the (modified) Toulmin's model and their roles are not expressive enough to capture argumentation that not only conveys the logical structure (in terms of reasons put forward to support the claim), but also relies heavily on the rhetorical power. This involves various stylistic devices, pervading narratives, direct and indirect speech, or interviews. While in some cases the argument components are easily recognizable, the vast majority of the discourse in articles and blog posts does not correspond to any distinguishable argumentative function in the logos dimension. As the purpose of such discourse relates more to rhetoric than to argumentation, unambiguous analysis of such phenomena goes beyond capabilities of the current argumentation model. For a discussion about metaphors in Toulmin's model of argumentation see, e.g., BIBREF102 , BIBREF103 .
Articles without a clear standpoint towards the discussed controversy cannot be easily annotated with the model either. Although the matter is viewed from both sides and there might be reasons presented for either of them, the overall persuasive intention is missing and fitting such data to the argumentation framework causes disagreements. One solution might be to break the document down to paragraphs and annotate each paragraph separately, examining argumentation on a different level of granularity.
As introduced in section SECREF20 , there are several dimensions of an argument. The Toulmin's model focuses solely on the logos dimension. We decided to ignore the ethos dimension, because dealing with the author's credibility remains unclear, given the variety of the source web data. However, exploiting the pathos dimension of an argument is prevalent in the web data, for example as an appeal to emotions. Therefore we experimented with annotating appeal to emotions as a separate category independent of components in the logos dimension. We defined some features for the annotators how to distinguish appeal to emotions. Figurative language such as hyperbole, sarcasm, or obvious exaggerating to “spice up” the argument are the typical signs of pathos. In an extreme case, the whole argument might be purely emotional, as in the following example.
Doc#1698 (comment, prayer in schools) [app-to-emot: Prayer being removed from school is just the leading indicator of a nation that is ‘Falling Away’ from Jehovah. [...] And the disasters we see today are simply God’s finger writing on the wall: Mene, mene, Tekel, Upharsin; that is, God has weighed America in the balances, and we’ve been found wanting. No wonder 50 million babies have been aborted since 1973. [...]]
We kept annotations on the pathos dimension as simple as possible (with only one appeal to emotions label), but the resulting agreement was unsatisfying ( INLINEFORM0 0.30) even after several annotation iterations. Appeal to emotions is considered as a type of fallacy BIBREF104 , BIBREF18 . Given the results, we assume that more carefully designed approach to fallacy annotation should be applied. To the best of our knowledge, there have been very few research works on modeling fallacies similarly to arguments on the discourse level BIBREF105 . Therefore the question, in which detail and structure fallacies should be annotated, remains open. For the rest of the paper, we thus focus on the logos dimension solely.
Some of the educational topics under examination relate to young children (e.g., redshirting or mainstreaming); therefore we assume that the majority of participants in discussions are their parents. We observed that many documents related to these topics contain narratives. Sometimes the story telling is meant as a support for the argument, but there are documents where the narrative has no intention to persuade and is simply a story sharing.
There is no widely accepted theory of the role of narratives among argumentation scholars. According to Fisher.1987, humans are storytellers by nature, and the “reason” in argumentation is therefore better understood in and through the narratives. He found that good reasons often take the form of narratives. Hoeken.Fikkers.2014 investigated how integration of explicit argumentative content into narratives influences issue-relevant thinking and concluded that identifying with the character being in favor of the issue yielded a more positive attitude toward the issue. In a recent research, Bex.2011 proposes an argumentative-narrative model of reasoning with evidence, further elaborated in BIBREF106 ; also Niehaus.et.al.2012 proposes a computational model of narrative persuasion.
Stemming from another research field, LeytonEscobar2014 found that online community members who use and share narratives have higher participation levels and that narratives are useful tools to build cohesive cultures and increase participation. Betsch.et.al.2010 examined influencing vaccine intentions among parents and found that narratives carry more weight than statistics.
Summary of annotation studies
This section described two annotation studies that deal with argumentation in user-generated Web content on different levels of detail. In section SECREF44 , we argued for a need of document-level distinction of persuasiveness. We annotated 990 comments and forum posts, reaching moderate inter-annotator agreement (Fleiss' INLINEFORM0 0.59). Section SECREF51 motivated the selection of a model for micro-level argument annotation, proposed its extension based on pre-study observations, and outlined the annotation set-up. This annotation study resulted into 340 documents annotated with the modified Toulmin's model and reached moderate inter-annotator agreement in the logos dimension (Krippendorff's INLINEFORM1 0.48). These results make the annotated corpora suitable for training and evaluation computational models and each of these two annotation studies will have their experimental counterparts in the following section.
Experiments
This section presents experiments conducted on the annotated corpora introduced in section SECREF4 . We put the main focus on identifying argument components in the discourse. To comply with the machine learning terminology, in this section we will use the term domain as an equivalent to a topic (remember that our dataset includes six different topics; see section SECREF38 ).
We evaluate three different scenarios. First, we report ten-fold cross validation over a random ordering of the entire data set. Second, we deal with in-domain ten-fold cross validation for each of the six domains. Third, in order to evaluate the domain portability of our approach, we train the system on five domains and test on the remaining one for all six domains (which we report as cross-domain validation).
Identification of argument components
In the following experiment, we focus on automatic identification of arguments in the discourse. Our approach is based on supervised and semi-supervised machine learning methods on the gold data Toulmin dataset introduced in section SECREF51 .
An argument consists of different components (such as premises, backing, etc.) which are implicitly linked to the claim. In principle one document can contain multiple independent arguments. However, only 4% of the documents in our dataset contain arguments for both sides of the issue. Thus we simplify the task and assume there is only one argument per document.
Given the low inter-annotator agreement on the pathos dimension (Table TABREF77 ), we focus solely on recognizing the logical dimension of argument. The pathos dimension of argument remains an open problem for a proper modeling as well as its later recognition.
Since the smallest annotation unit is a token and the argument components do not overlap, we approach identification of argument components as a sequence labeling problem. We use the BIO encoding, so each token belongs to one of the following 11 classes: O (not a part of any argument component), Backing-B, Backing-I, Claim-B, Claim-I, Premise-B, Premise-I, Rebuttal-B, Rebuttal-I, Refutation-B, Refutation-I. This is the minimal encoding that is able to distinguish two adjacent argument components of the same type. In our data, 48% of all adjacent argument components of the same type are direct neighbors (there are no "O" tokens in between).
We report Macro- INLINEFORM0 score and INLINEFORM1 scores for each of the 11 classes as the main evaluation metric. This evaluation is performed on the token level, and for each token the predicted label must exactly match the gold data label (classification of tokens into 11 classes).
As instances for the sequence labeling model, we chose sentences rather than tokens. During our initial experiments, we observed that building a sequence labeling model for recognizing argument components as sequences of tokens is too fine-grained, as a single token does not convey enough information that could be encoded as features for a machine learner. However, as discussed in section UID73 , the annotations were performed on data pre-segmented to sentences and annotating tokens was necessary only when the sentence segmentation was wrong or one sentence contained multiple argument components. Our corpus consists of 3899 sentences, from which 2214 sentences (57%) contain no argument component. From the remaining ones, only 50 sentences (1%) have more than one argument component. Although in 19 cases (0.5%) the sentence contains a Claim-Premise pair which is an important distinction from the argumentation perspective, given the overall small number of such occurrences, we simplify the task by treating each sentence as if it has either one argument component or none.
The approximation with sentence-level units is explained in the example in Figure FIGREF112 . In order to evaluate the expected performance loss using this approximation, we used an oracle that always predicts the correct label for the unit (sentence) and evaluated it against the true labels (recall that the evaluation against the true gold labels is done always on token level). We lose only about 10% of macro INLINEFORM0 score (0.906) and only about 2% of accuracy (0.984). This performance is still acceptable, while allowing to model sequences where the minimal unit is a sentence.
Table TABREF114 shows the distribution of the classes in the gold data Toulmin, where the labeling was already mapped to the sentences. The little presence of rebuttal and refutation (4 classes account only for 3.4% of the data) makes this dataset very unbalanced.
We chose SVMhmm BIBREF111 implementation of Structural Support Vector Machines for sequence labeling. Each sentence ( INLINEFORM0 ) is represented as a vector of real-valued features.
We defined the following feature sets:
FS0: Baseline lexical features
word uni-, bi-, and tri-grams (binary)
FS1: Structural, morphological, and syntactic features
First and last 3 tokens. Motivation: these tokens may contain discourse markers or other indicators for argument components, such as “therefore” and “since” for premises or “think” and “believe” for claims.
Relative position in paragraph and relative position in document. Motivation: We expect that claims are more likely to appear at the beginning or at the end of the document.
Number of POS 1-3 grams, dependency tree depth, constituency tree production rules, and number of sub-clauses. Based on BIBREF113 .
FS2: Topic and sentiment features
30 features taken from a vector representation of the sentence obtained by using Gibbs sampling on LDA model BIBREF114 , BIBREF115 with 30 topics trained on unlabeled data from the raw corpus. Motivation: Topic representation of a sentence might be valuable for detecting off-topic sentences, namely non-argument components.
Scores for five sentiment categories (from very negative to very positive) obtained from Stanford sentiment analyzer BIBREF116 . Motivation: Claims usually express opinions and carry sentiment.
FS3: Semantic, coreference, and discourse features
Binary features from Clear NLP Semantic Role Labeler BIBREF117 . Namely, we extract agent, predicate + agent, predicate + agent + patient + (optional) negation, argument type + argument value, and discourse marker which are based on PropBank semantic role labels. Motivation: Exploit the semantics of Capturing the semantics of the sentences.
Binary features from Stanford Coreference Chain Resolver BIBREF118 , e.g., presence of the sentence in a chain, transition type (i.e., nominal–pronominal), distance to previous/next sentences in the chain, or number of inter-sentence coreference links. Motivation: Presence of coreference chains indicates links outside the sentence and thus may be informative, for example, for classifying whether the sentence is a part of a larger argument component.
Results of a PTDB-style discourse parser BIBREF119 , namely the type of discourse relation (explicit, implicit), presence of discourse connectives, and attributions. Motivation: It has been claimed that discourse relations play a role in argumentation mining BIBREF120 .
FS4: Embedding features
300 features from word embedding vectors using word embeddings trained on part of the Google News dataset BIBREF121 . In particular, we sum up embedding vectors (dimensionality 300) of each word, resulting into a single vector for the entire sentence. This vector is then directly used as a feature vector. Motivation: Embeddings helped to achieve state-of-the-art results in various NLP tasks BIBREF116 , BIBREF122 .
Except the baseline lexical features, all feature types are extracted not only for the current sentence INLINEFORM0 , but also for INLINEFORM1 preceding and subsequent sentences, namely INLINEFORM2 , INLINEFORM3 , INLINEFORM4 INLINEFORM5 , INLINEFORM6 , where INLINEFORM7 was empirically set to 4. Each feature is then represented with a prefix to determine its relative position to the current sequence unit.
Let us first discuss the upper bounds of the system. Performance of the three human annotators is shown in the first column of Table TABREF139 (results are obtained from a cumulative confusion matrix). The overall Macro- INLINEFORM0 score is 0.602 (accuracy 0.754). If we look closer at the different argument components, we observe that humans are good at predicting claims, premises, backing and non-argumentative text (about 0.60-0.80 INLINEFORM1 ), but on rebuttal and refutation they achieve rather low scores. Without these two components, the overall human Macro- INLINEFORM2 would be 0.707. This trend follows the inter-annotator agreement scores, as discussed in section UID75 .
In our experiments, the feature sets were combined in the bottom-up manner, starting with the simple lexical features (FS0), adding structural and syntactic features (FS1), then adding topic and sentiment features (FS2), then features reflecting the discourse structure (FS3), and finally enriched with completely unsupervised latent vector space representation (FS4). In addition, we were gradually removing the simple features (e.g., without lexical features, without syntactic features, etc.) to test the system with more “abstract” feature sets (feature ablation). The results are shown in Table TABREF139 .
The overall best performance (Macro- INLINEFORM0 0.251) was achieved using the rich feature sets (01234 and 234) and significantly outperformed the baseline as well as other feature sets. Classification of non-argumentative text (the "O" class) yields about 0.7 INLINEFORM1 score even in the baseline setting. The boundaries of claims (Cla-B), premises (Pre-B), and backing (Bac-B) reach in average lower scores then their respective inside tags (Cla-I, Pre-I, Bac-I). It can be interpreted such that the system is able to classify that a certain sentence belongs to a certain argument component, but the distinction whether it is a beginning of the argument component is harder. The very low numbers for rebuttal and refutation have two reasons. First, these two argument components caused many disagreements in the annotations, as discussed in section UID86 , and were hard to recognize for the humans too. Second, these four classes have very few instances in the corpus (about 3.4%, see Table TABREF114 ), so the classifier suffers from the lack of training data.
The results for the in-domain cross validation scenario are shown in Table TABREF140 . Similarly to the cross-validation scenario, the overall best results were achieved using the largest feature set (01234). For mainstreaming and red-shirting, the best results were achieved using only the feature set 4 (embeddings). These two domains contain also fewer documents, compared to other domains (refer to Table TABREF71 ). We suspect that embeddings-based features convey important information when not enough in-domain data are available. This observation will become apparent in the next experiment.
The cross-domain experiments yield rather poor results for most of the feature combinations (Table TABREF141 ). However, using only feature set 4 (embeddings), the system performance increases rapidly, so it is even comparable to numbers achieved in the in-domain scenario. These results indicate that embedding features generalize well across domains in our task of argument component identification. We leave investigating better performing vector representations, such as paragraph vectors BIBREF123 , for future work.
Error analysis based on the probabilistic confusion matrix BIBREF124 shown in Table TABREF142 reveals further details. About a half of the instances for each class are misclassified as non-argumentative (the "O" prediction).
Backing-B is often confused with Premise-B (12%) and Backing-I with Premise-I (23%). Similarly, Premise-I is misclassified as Backing-I in 9%. This shows that distinguishing between backing and premises is not easy because these two components are similar such that they support the claim, as discussed in section UID86 . We can also see that the misclassification is consistent among *-B and *-I tags.
Rebuttal is often misclassified as Premise (28% for Rebuttal-I and 18% for Rebuttal-B; notice again the consistency in *-B and *-I tags). This is rather surprising, as one would expect that rebuttal would be confused with a claim, because its role is to provide an opposing view.
Refutation-B and Refutation-I is misclassified as Premise-I in 19% and 27%, respectively. This finding confirms the discussion in section UID86 , because the role of refutation is highly context-dependent. In a pragmatic perspective, it is put forward to indirectly support the claim by attacking the rebuttal, thus having a similar function to the premise.
We manually examined miss-classified examples produced the best-performing system to find out which phenomena pose biggest challenges. Properly detecting boundaries of argument components caused problems, as shown in Figure FIGREF146 (a). This goes in line with the granularity annotation difficulties discussed in section UID86 . The next example in Figure FIGREF146 (b) shows that even if boundaries of components were detected precisely, the distinction between premise and backing fails. The example also shows that in some cases, labeling on clause level is required (left-hand side claim and premise) but the approximation in the system cannot cope with this level of detail (as explained in section UID111 ). Confusing non-argumentative text and argument components by the system is sometimes plausible, as is the case of the last rhetorical question in Figure FIGREF146 (c). On the other hand, the last example in Figure FIGREF146 (d) shows that some claims using figurative language were hard to be identified. The complete predictions along with the gold data are publicly available.
SVMhmm offers many hyper-parameters with suggested default values, from which three are of importance. Parameter INLINEFORM0 sets the order of dependencies of transitions in HMM, parameter INLINEFORM1 sets the order of dependencies of emissions in HMM, and parameter INLINEFORM2 represents a trading-off slack versus magnitude of the weight-vector. For all experiments, we set all the hyper-parameters to their default values ( INLINEFORM3 , INLINEFORM4 , INLINEFORM5 ). Using the best performing feature set from Table TABREF139 , we experimented with a grid search over different values ( INLINEFORM6 , INLINEFORM7 , INLINEFORM8 ) but the results did not outperform the system trained with default parameter values.
The INLINEFORM0 scores might seem very low at the first glance. One obvious reason is the actual performance of the system, which gives a plenty of room for improvement in the future. But the main cause of low INLINEFORM2 numbers is the evaluation measure — using 11 classes on the token level is very strict, as it penalizes a mismatch in argument component boundaries the same way as a wrongly predicted argument component type. Therefore we also report two another evaluation metrics that help to put our results into a context.
Krippendorff's INLINEFORM0 — It was also used for evaluating inter-annotator agreement (see section UID75 ).
Boundary similarity BIBREF125 — Using this metric, the problem is treated solely as a segmentation task without recognizing the argument component types.
As shown in Table TABREF157 (the Macro- INLINEFORM0 scores are repeated from Table TABREF139 ), the best-performing system achieves 0.30 score using Krippendorf's INLINEFORM1 , which is in the middle between the baseline and the human performance (0.48) but is considered as poor from the inter-annotator agreement point of view BIBREF54 . The boundary similarity metrics is not directly suitable for evaluating argument component classification, but reveals a sub-task of finding the component boundaries. The best system achieved 0.32 on this measure. Vovk2013MT used this measure to annotate argument spans and his annotators achieved 0.36 boundary similarity score. Human annotators in BIBREF125 reached 0.53 boundary similarity score.
The overall performance of the system is also affected by the accuracy of individual NPL tools used for extracting features. One particular problem is that the preprocessing models we rely on (POS, syntax, semantic roles, coreference, discourse; see section UID115 ) were trained on newswire corpora, so one has to expect performance drop when applied on user-generated content. This is however a well-known issue in NLP BIBREF126 , BIBREF127 , BIBREF128 .
To get an impression of the actual performance of the system on the data, we also provide the complete output of our best performing system in one PDF document together with the gold annotations in the logos dimension side by side in the accompanying software package. We believe this will help the community to see the strengths of our model as well as possible limitations of our current approaches.
Conclusions
Let us begin with summarizing answers to the research questions stated in the introduction. First, as we showed in section UID55 , existing argumentation theories do offer models for capturing argumentation in user-generated content on the Web. We built upon the Toulmin's model and proposed some extensions.
Second, as compared to the negative experiences with annotating using Walton's schemes (see sections UID52 and SECREF31 ), our modified Toulmin's model offers a trade-off between its expressiveness and annotation reliability. However, we found that the capabilities of the model to capture argumentation depend on the register and topic, the length of the document, and inherently on the literary devices and structures used for expressing argumentation as these properties influenced the agreement among annotators.
Third, there are aspects of online argumentation that lack their established theoretical counterparts, such as rhetorical questions, figurative language, narratives, and fallacies in general. We tried to model some of them in the pathos dimension of argument (section UID103 ), but no satisfying agreement was reached. Furthermore, we dealt with a step that precedes argument analysis by filtering documents given their persuasiveness with respect to the controversy. Finally, we proposed a computational model based on machine learning for identifying argument components (section SECREF108 ). In this identification task, we experimented with a wide range of linguistically motivated features and found that (1) the largest feature set (including n-grams, structural features, syntactic features, topic distribution, sentiment distribution, semantic features, coreference feaures, discourse features, and features based on word embeddings) performs best in both in-domain and all-data cross validation, while (2) features based only on word embeddings yield best results in cross-domain evaluation.
Since there is no one-size-fits-all argumentation theory to be applied to actual data on the Web, the argumentation model and an annotation scheme for argumentation mining is a function of the task requirements and the corpus properties. Its selection should be based on the data at hand and the desired application. Given the proposed use-case scenarios (section SECREF1 ) and the results of our annotation study (section SECREF51 ), we recommend a scheme based on Toulmin's model for short documents, such as comments or forum posts.
| [
"claim, premise, backing, rebuttal, and refutation",
"claim, premise, backing, rebuttal, refutation"
] | 14,472 | qasper | en | null | 90b2dc5aba0e95b2f541da4efa36c414b93bb33d56f2c31e | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What argument components do the ML methods aim to identify?
Answer: | [
"claim, premise, backing, rebuttal, and refutation",
"claim, premise, backing, rebuttal, refutation"
] | qasper | 128 | 54 | |
Ngrams of which length are aligned using PARENT? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The task of generating natural language descriptions of structured data (such as tables) BIBREF2 , BIBREF3 , BIBREF4 has seen a growth in interest with the rise of sequence to sequence models that provide an easy way of encoding tables and generating text from them BIBREF0 , BIBREF1 , BIBREF5 , BIBREF6 .
For text generation tasks, the only gold standard metric is to show the output to humans for judging its quality, but this is too expensive to apply repeatedly anytime small modifications are made to a system. Hence, automatic metrics that compare the generated text to one or more reference texts are routinely used to compare models BIBREF7 . For table-to-text generation, automatic evaluation has largely relied on BLEU BIBREF8 and ROUGE BIBREF9 . The underlying assumption behind these metrics is that the reference text is gold-standard, i.e., it is the ideal target text that a system should generate. In practice, however, when datasets are collected automatically and heuristically, the reference texts are often not ideal. Figure FIGREF2 shows an example from the WikiBio dataset BIBREF0 . Here the reference contains extra information which no system can be expected to produce given only the associated table. We call such reference texts divergent from the table.
We show that existing automatic metrics, including BLEU, correlate poorly with human judgments when the evaluation sets contain divergent references (§ SECREF36 ). For many table-to-text generation tasks, the tables themselves are in a pseudo-natural language format (e.g., WikiBio, WebNLG BIBREF6 , and E2E-NLG BIBREF10 ). In such cases we propose to compare the generated text to the underlying table as well to improve evaluation. We develop a new metric, PARENT (Precision And Recall of Entailed N-grams from the Table) (§ SECREF3 ). When computing precision, PARENT effectively uses a union of the reference and the table, to reward correct information missing from the reference. When computing recall, it uses an intersection of the reference and the table, to ignore extra incorrect information in the reference. The union and intersection are computed with the help of an entailment model to decide if a text n-gram is entailed by the table. We show that this method is more effective than using the table as an additional reference. Our main contributions are:
Table-to-Text Generation
We briefly review the task of generating natural language descriptions of semi-structured data, which we refer to as tables henceforth BIBREF11 , BIBREF12 . Tables can be expressed as set of records INLINEFORM0 , where each record is a tuple (entity, attribute, value). When all the records are about the same entity, we can truncate the records to (attribute, value) pairs. For example, for the table in Figure FIGREF2 , the records are {(Birth Name, Michael Dahlquist), (Born, December 22 1965), ...}. The task is to generate a text INLINEFORM1 which summarizes the records in a fluent and grammatical manner. For training and evaluation we further assume that we have a reference description INLINEFORM2 available for each table. We let INLINEFORM3 denote an evaluation set of tables, references and texts generated from a model INLINEFORM4 , and INLINEFORM5 , INLINEFORM6 denote the collection of n-grams of order INLINEFORM7 in INLINEFORM8 and INLINEFORM9 , respectively. We use INLINEFORM10 to denote the count of n-gram INLINEFORM11 in INLINEFORM12 , and INLINEFORM13 to denote the minimum of its counts in INLINEFORM14 and INLINEFORM15 . Our goal is to assign a score to the model, which correlates highly with human judgments of the quality of that model.
PARENT
PARENT evaluates each instance INLINEFORM0 separately, by computing the precision and recall of INLINEFORM1 against both INLINEFORM2 and INLINEFORM3 .
Evaluation via Information Extraction
BIBREF1 proposed to use an auxiliary model, trained to extract structured records from text, for evaluation. However, the extraction model presented in that work is limited to the closed-domain setting of basketball game tables and summaries. In particular, they assume that each table has exactly the same set of attributes for each entity, and that the entities can be identified in the text via string matching. These assumptions are not valid for the open-domain WikiBio dataset, and hence we train our own extraction model to replicate their evaluation scheme.
Our extraction system is a pointer-generator network BIBREF19 , which learns to produce a linearized version of the table from the text. The network learns which attributes need to be populated in the output table, along with their values. It is trained on the training set of WikiBio. At test time we parsed the output strings into a set of (attribute, value) tuples and compare it to the ground truth table. The F-score of this text-to-table system was INLINEFORM0 , which is comparable to other challenging open-domain settings BIBREF20 . More details are included in the Appendix SECREF52 .
Given this information extraction system, we consider the following metrics for evaluation, along the lines of BIBREF1 . Content Selection (CS): F-score for the (attribute, value) pairs extracted from the generated text compared to those extracted from the reference. Relation Generation (RG): Precision for the (attribute, value) pairs extracted from the generated text compared to those in the ground truth table. RG-F: Since our task emphasizes the recall of information from the table as well, we consider another variant which computes the F-score of the extracted pairs to those in the table. We omit the content ordering metric, since our extraction system does not align records to the input text.
Experiments & Results
In this section we compare several automatic evaluation metrics by checking their correlation with the scores assigned by humans to table-to-text models. Specifically, given INLINEFORM0 models INLINEFORM1 , and their outputs on an evaluation set, we show these generated texts to humans to judge their quality, and obtain aggregated human evaluation scores for all the models, INLINEFORM2 (§ SECREF33 ). Next, to evaluate an automatic metric, we compute the scores it assigns to each model, INLINEFORM3 , and check the Pearson correlation between INLINEFORM4 and INLINEFORM5 BIBREF21 .
Data & Models
Our main experiments are on the WikiBio dataset BIBREF0 , which is automatically constructed and contains many divergent references. In § SECREF47 we also present results on the data released as part of the WebNLG challenge.
We developed several models of varying quality for generating text from the tables in WikiBio. This gives us a diverse set of outputs to evaluate the automatic metrics on. Table TABREF32 lists the models along with their hyperparameter settings and their scores from the human evaluation (§ SECREF33 ). Our focus is primarily on neural sequence-to-sequence methods since these are most widely used, but we also include a template-based baseline. All neural models were trained on the WikiBio training set. Training details and sample outputs are included in Appendices SECREF56 & SECREF57 .
We divide these models into two categories and measure correlation separately for both the categories. The first category, WikiBio-Systems, includes one model each from the four families listed in Table TABREF32 . This category tests whether a metric can be used to compare different model families with a large variation in the quality of their outputs. The second category, WikiBio-Hyperparams, includes 13 different hyperparameter settings of PG-Net BIBREF19 , which was the best performing system overall. 9 of these were obtained by varying the beam size and length normalization penalty of the decoder network BIBREF23 , and the remaining 4 were obtained by re-scoring beams of size 8 with the information extraction model described in § SECREF4 . All the models in this category produce high quality fluent texts, and differ primarily on the quantity and accuracy of the information they express. Here we are testing whether a metric can be used to compare similar systems with a small variation in performance. This is an important use-case as metrics are often used to tune hyperparameters of a model.
Human Evaluation
We collected human judgments on the quality of the 16 models trained for WikiBio, plus the reference texts. Workers on a crowd-sourcing platform, proficient in English, were shown a table with pairs of generated texts, or a generated text and the reference, and asked to select the one they prefer. Figure FIGREF34 shows the instructions they were given. Paired comparisons have been shown to be superior to rating scales for comparing generated texts BIBREF24 . However, for measuring correlation the comparisons need to be aggregated into real-valued scores, INLINEFORM0 , for each of the INLINEFORM1 models. For this, we use Thurstone's method BIBREF22 , which assigns a score to each model based on how many times it was preferred over an alternative.
The data collection was performed separately for models in the WikiBio-Systems and WikiBio-Hyperparams categories. 1100 tables were sampled from the development set, and for each table we got 8 different sentence pairs annotated across the two categories, resulting in a total of 8800 pairwise comparisons. Each pair was judged by one worker only which means there may be noise at the instance-level, but the aggregated system-level scores had low variance (cf. Table TABREF32 ). In total around 500 different workers were involved in the annotation. References were also included in the evaluation, and they received a lower score than PG-Net, highlighting the divergence in WikiBio.
Compared Metrics
Text only: We compare BLEU BIBREF8 , ROUGE BIBREF9 , METEOR BIBREF18 , CIDEr and CIDEr-D BIBREF25 using their publicly available implementations.
Information Extraction based: We compare the CS, RG and RG-F metrics discussed in § SECREF4 .
Text & Table: We compare a variant of BLEU, denoted as BLEU-T, where the values from the table are used as additional references. BLEU-T draws inspiration from iBLEU BIBREF26 but instead rewards n-grams which match the table rather than penalizing them. For PARENT, we compare both the word-overlap model (PARENT-W) and the co-occurrence model (PARENT-C) for determining entailment. We also compare versions where a single INLINEFORM0 is tuned on the entire dataset to maximize correlation with human judgments, denoted as PARENT*-W/C.
Correlation Comparison
We use bootstrap sampling (500 iterations) over the 1100 tables for which we collected human annotations to get an idea of how the correlation of each metric varies with the underlying data. In each iteration, we sample with replacement, tables along with their references and all the generated texts for that table. Then we compute aggregated human evaluation and metric scores for each of the models and compute the correlation between the two. We report the average correlation across all bootstrap samples for each metric in Table TABREF37 . The distribution of correlations for the best performing metrics are shown in Figure FIGREF38 .
Table TABREF37 also indicates whether PARENT is significantly better than a baseline metric. BIBREF21 suggest using the William's test for this purpose, but since we are computing correlations between only 4/13 systems at a time, this test has very weak power in our case. Hence, we use the bootstrap samples to obtain a INLINEFORM0 confidence interval of the difference in correlation between PARENT and any other metric and check whether this is above 0 BIBREF27 .
Correlations are higher for the systems category than the hyperparams category. The latter is a more difficult setting since very similar models are compared, and hence the variance of the correlations is also high. Commonly used metrics which only rely on the reference (BLEU, ROUGE, METEOR, CIDEr) have only weak correlations with human judgments. In the hyperparams category, these are often negative, implying that tuning models based on these may lead to selecting worse models. BLEU performs the best among these, and adding n-grams from the table as references improves this further (BLEU-T).
Among the extractive evaluation metrics, CS, which also only relies on the reference, has poor correlation in the hyperparams category. RG-F, and both variants of the PARENT metric achieve the highest correlation for both settings. There is no significant difference among these for the hyperparams category, but for systems, PARENT-W is significantly better than the other two. While RG-F needs a full information extraction pipeline in its implementation, PARENT-C only relies on co-occurrence counts, and PARENT-W can be used out-of-the-box for any dataset. To our knowledge, this is the first rigorous evaluation of using information extraction for generation evaluation.
On this dataset, the word-overlap model showed higher correlation than the co-occurrence model for entailment. In § SECREF47 we will show that for the WebNLG dataset, where more paraphrasing is involved between the table and text, the opposite is true. Lastly, we note that the heuristic for selecting INLINEFORM0 is sufficient to produce high correlations for PARENT, however, if human annotations are available, this can be tuned to produce significantly higher correlations (PARENT*-W/C).
Analysis
In this section we further analyze the performance of PARENT-W under different conditions, and compare to the other best metrics from Table TABREF37 .
To study the correlation as we vary the number of divergent references, we also collected binary labels from workers for whether a reference is entailed by the corresponding table. We define a reference as entailed when it mentions only information which can be inferred from the table. Each table and reference pair was judged by 3 independent workers, and we used the majority vote as the label for that pair. Overall, only INLINEFORM0 of the references were labeled as entailed by the table. Fleiss' INLINEFORM1 was INLINEFORM2 , which indicates a fair agreement. We found the workers sometimes disagreed on what information can be reasonably entailed by the table.
Figure FIGREF40 shows the correlations as we vary the percent of entailed examples in the evaluation set of WikiBio. Each point is obtained by fixing the desired proportion of entailed examples, and sampling subsets from the full set which satisfy this proportion. PARENT and RG-F remain stable and show a high correlation across the entire range, whereas BLEU and BLEU-T vary a lot. In the hyperparams category, the latter two have the worst correlation when the evaluation set contains only entailed examples, which may seem surprising. However, on closer examination we found that this subset tends to omit a lot of information from the tables. Systems which produce more information than these references are penalized by BLEU, but not in the human evaluation. PARENT overcomes this issue by measuring recall against the table in addition to the reference.
We check how different components in the computation of PARENT contribute to its correlation to human judgments. Specifically, we remove the probability INLINEFORM0 of an n-gram INLINEFORM1 being entailed by the table from Eqs. EQREF19 and EQREF23 . The average correlation for PARENT-W drops to INLINEFORM5 in this case. We also try a variant of PARENT with INLINEFORM6 , which removes the contribution of Table Recall (Eq. EQREF22 ). The average correlation is INLINEFORM7 in this case. With these components, the correlation is INLINEFORM8 , showing that they are crucial to the performance of PARENT.
BIBREF28 point out that hill-climbing on an automatic metric is meaningless if that metric has a low instance-level correlation to human judgments. In Table TABREF46 we show the average accuracy of the metrics in making the same judgments as humans between pairs of generated texts. Both variants of PARENT are significantly better than the other metrics, however the best accuracy is only INLINEFORM0 for the binary task. This is a challenging task, since there are typically only subtle differences between the texts. Achieving higher instance-level accuracies will require more sophisticated language understanding models for evaluation.
WebNLG Dataset
To check how PARENT correlates with human judgments when the references are elicited from humans (and less likely to be divergent), we check its correlation with the human ratings provided for the systems competing in the WebNLG challenge BIBREF6 . The task is to generate text describing 1-5 RDF triples (e.g. John E Blaha, birthPlace, San Antonio), and human ratings were collected for the outputs of 9 participating systems on 223 instances. These systems include a mix of pipelined, statistical and neural methods. Each instance has upto 3 reference texts associated with the RDF triples, which we use for evaluation.
The human ratings were collected on 3 distinct aspects – grammaticality, fluency and semantics, where semantics corresponds to the degree to which a generated text agrees with the meaning of the underlying RDF triples. We report the correlation of several metrics with these ratings in Table TABREF48 . Both variants of PARENT are either competitive or better than the other metrics in terms of the average correlation to all three aspects. This shows that PARENT is applicable for high quality references as well.
While BLEU has the highest correlation for the grammar and fluency aspects, PARENT does best for semantics. This suggests that the inclusion of source tables into the evaluation orients the metric more towards measuring the fidelity of the content of the generation. A similar trend is seen comparing BLEU and BLEU-T. As modern neural text generation systems are typically very fluent, measuring their fidelity is of increasing importance. Between the two entailment models, PARENT-C is better due to its higher correlation with the grammaticality and fluency aspects.
The INLINEFORM0 parameter in the calculation of PARENT decides whether to compute recall against the table or the reference (Eq. EQREF22 ). Figure FIGREF50 shows the distribution of the values taken by INLINEFORM1 using the heuristic described in § SECREF3 for instances in both WikiBio and WebNLG. For WikiBio, the recall of the references against the table is generally low, and hence the recall of the generated text relies more on the table. For WebNLG, where the references are elicited from humans, this recall is much higher (often INLINEFORM2 ), and hence the recall of the generated text relies more on the reference.
Related Work
Over the years several studies have evaluated automatic metrics for measuring text generation performance BIBREF29 , BIBREF30 , BIBREF31 , BIBREF32 , BIBREF33 , BIBREF34 , BIBREF35 . The only consensus from these studies seems to be that no single metric is suitable across all tasks. A recurring theme is that metrics like BLEU and NIST BIBREF36 are not suitable for judging content quality in NLG. Recently, BIBREF37 did a comprehensive study of several metrics on the outputs of state-of-the-art NLG systems, and found that while they showed acceptable correlation with human judgments at the system level, they failed to show any correlation at the sentence level. Ours is the first study which checks the quality of metrics when table-to-text references are divergent. We show that in this case even system level correlations can be unreliable.
Hallucination BIBREF38 , BIBREF39 refers to when an NLG system generates text which mentions extra information than what is present in the source from which it is generated. Divergence can be viewed as hallucination in the reference text itself. PARENT deals with hallucination by discounting n-grams which do not overlap with either the reference or the table.
PARENT draws inspiration from iBLEU BIBREF26 , a metric for evaluating paraphrase generation, which compares the generated text to both the source text and the reference. While iBLEU penalizes texts which match the source, here we reward such texts since our task values accuracy of generated text more than the need for paraphrasing the tabular content BIBREF40 . Similar to SARI for text simplification BIBREF41 and Q-BLEU for question generation BIBREF42 , PARENT falls under the category of task-specific metrics.
Conclusions
We study the automatic evaluation of table-to-text systems when the references diverge from the table. We propose a new metric, PARENT, which shows the highest correlation with humans across a range of settings with divergent references in WikiBio. We also perform the first empirical evaluation of information extraction based metrics BIBREF1 , and find RG-F to be effective. Lastly, we show that PARENT is comparable to the best existing metrics when references are elicited by humans on the WebNLG data.
Acknowledgements
Bhuwan Dhingra is supported by a fellowship from Siemens, and by grants from Google. We thank Maruan Al-Shedivat, Ian Tenney, Tom Kwiatkowski, Michael Collins, Slav Petrov, Jason Baldridge, David Reitter and other members of the Google AI Language team for helpful discussions and suggestions. We thank Sam Wiseman for sharing data for an earlier version of this paper. We also thank the anonymous reviewers for their feedback.
Information Extraction System
For evaluation via information extraction BIBREF1 we train a model for WikiBio which accepts text as input and generates a table as the output. Tables in WikiBio are open-domain, without any fixed schema for which attributes may be present or absent in an instance. Hence we employ the Pointer-Generator Network (PG-Net) BIBREF19 for this purpose. Specifically, we use a sequence-to-sequence model, whose encoder and decoder are both single-layer bi-directional LSTMs. The decoder is augmented with an attention mechanism over the states of the encoder. Further, it also uses a copy mechanism to optionally copy tokens directly from the source text. We do not use the coverage mechanism of BIBREF19 since that is specific to the task of summarization they study. The decoder is trained to produce a linearized version of the table where the rows and columns are flattened into a sequence, and separate by special tokens. Figure FIGREF53 shows an example.
Clearly, since the references are divergent, the model cannot be expected to produce the entire table, and we see some false information being hallucinated after training. Nevertheless, as we show in § SECREF36 , this system can be used for evaluating generated texts. After training, we can parse the output sequence along the special tokens INLINEFORM0 R INLINEFORM1 and INLINEFORM2 C INLINEFORM3 to get a set of (attribute, value) pairs. Table TABREF54 shows the precision, recall and F-score of these extracted pairs against the ground truth tables, where the attributes and values are compared using an exact string match.
Hyperparameters
After tuning we found the same set of hyperparameters to work well for both the table-to-text PG-Net, and the inverse information extraction PG-Net. The hidden state size of the biLSTMs was set to 200. The input and output vocabularies were set to 50000 most common words in the corpus, with additional special symbols for table attribute names (such as “birth-date”). The embeddings of the tokens in the vocabulary were initialized with Glove BIBREF43 . Learning rate of INLINEFORM0 was used during training, with the Adam optimizer, and a dropout of INLINEFORM1 was also applied to the outputs of the biLSTM. Models were trained till the loss on the dev set stopped dropping. Maximum length of a decoded text was set to 40 tokens, and that of the tables was set to 120 tokens. Various beam sizes and length normalization penalties were applied for the table-to-text system, which are listed in the main paper. For the information extraction system, we found a beam size of 8 and no length penalty to produce the highest F-score on the dev set.
Sample Outputs
Table TABREF55 shows some sample references and the corresponding predictions from the best performing model, PG-Net for WikiBio.
| [
"Unanswerable",
"Answer with content missing: (Parent subsections) combine precisions for n-gram orders 1-4"
] | 3,827 | qasper | en | null | 7e339da53449dbd58a1e815d55e440e712af6a1950f342b7 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Ngrams of which length are aligned using PARENT?
Answer: | [
"Unanswerable",
"Answer with content missing: (Parent subsections) combine precisions for n-gram orders 1-4"
] | qasper | 128 | 55 | |
How large is the Twitter dataset? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Recently, people have started looking at online forums either as a primary or secondary source of counseling services BIBREF0. BIBREF1 reported that over the first five years of operation (2011-2016), ReachOut.com – Ireland's online youth mental health service – 62% of young people would visit a website for support when going through a tough time. With the expansion of the Internet, there has been a substantial growth in the number of users looking for psychological support online.
The importance of the on-line life of patients has been recognized in research as well. BIBREF2 stated that the online life of patients constitutes a major influence on their self-definition. Furthermore, according to BIBREF3, the social networking activities of an individual, offer an important reflection of their personality. While dealing with patients suffering from psychological problems, it is important that therapists do not ignore this pivotal source of information which can provide deep insights into their patients' mental conditions.
Acceptance of on-line support groups (OSG) by Mental Health Professionals is still not established BIBREF4. Since OSG can have double-edged effects on patients and the presence of professionals is often limited, we argue that their properties should be further studied. According to BIBREF5 OSG effectiveness is hard to assess, while some studies showed OSG's potential to change participants' attitudes, no such effect was observed in other studies (see Related Work Section for more details). Furthermore the scope of previous work on analysis of users' behaviour in OSG has been limited by the fact that they relied on expert annotation of posts and comments BIBREF6.
We present a novel approach for automatically analysing online conversations for the presence of therapeutic factors of group therapy defined by BIBREF7 as “the actual mechanisms of effecting change in the patient”. The authors have identified 11 therapeutic factors in group therapy: Universality, Altruism, Instillation of Hope, Guidance, Imparting information, Developing social skills, Interpersonal learning, Cohesion, Catharsis, Existential factors, Imitative behavior and Corrective recapitulation of family of origin issues. In this paper, we focus on 3 therapeutic factors: Universality, Altruism and Instillation of Hope (listed below), as we believe that these can be approximated by using established NLP techniques (e.g. Sentiment Analysis, Dialogue Act tagging etc.).
Universality: the disconfirmation of a user's feelings of uniqueness of their mental health condition.
Altruism: others offer support, reassurance, suggestions and insight.
Instillation of Hope: inspiration provided to participants by their peers.
The selected therapeutic factors are analysed in terms of illocutionary force and attitude. Due to the multi-party and asynchronous nature of on-line social media conversations, prior to the analysis, we extract conversation threads among users – an essential prerequisite for any kind of higher-level dialogue analysis BIBREF10. Afterwards, the illocutionary force is identified using Dialogue Act tagging, whereas the attitude by using Sentiment Analysis. The quantitative analysis is then performed on these processed conversations.
Ideally, the analysis would require experts to annotate each post and comment on the presence of therapeutic factors. However, due to time and cost demands of this task, it is feasible to analyse only a small fraction of the available data. Compared to previous studies (e.g. BIBREF6) that analysed few tens of conversations and several thousand lines of chat; using the proposed approach – application of Dialogue Acts and Sentiment Analysis – we were able to automatically analyse approximately 300 thousands conversations (roughly 1.5 million comments). The rest of the paper is structured as follows. In Section 2 we introduce related work. Next, in Section 3 we describe the pre-processing pipeline and the methodology to perform thread extraction on asynchronous multi-party conversations. In Section 4 we provide the describe the final dataset used for the analysis, and in Section 5 we present the results of our analysis. Finally, in Section 6 we provide concluding remarks and future research directions.
Related Work
On-line support groups have been analyzed for various factors before. For instance, BIBREF11 analysed stress reduction in on-line support group chat-rooms, and the effects of on-line social interactions. Such studies mostly relied on questionnaires and were based on a small number of users. Nevertheless, in BIBREF11, the author showed that social support facilitates coping with distress, improves mood and expedites recovery from it. These findings highlight that, overall, on-line discussion boards appear to be therapeutic and constructive for individuals suffering alcohol-abuse.
Application of NLP to the analysis of mental health-related conversation has been studied as well (e.g. BIBREF12, BIBREF13). BIBREF6 applied sentiment-analysis combined with extensive turn-level annotation to investigate stress reduction in on-line support group chat-rooms, showing that sentiment-analysis is a good predictor of entrance stress level. Furthermore, similar to our setting, they applied automatic thread-extraction to determine conversation threads.
BIBREF14 have shown that on-line support group therapy increased the quality of life of patients with metastatic breast cancer. Since many original posters reported the benefits of group therapy on patients BIBREF15, BIBREF2, BIBREF16, BIBREF17, BIBREF18, BIBREF7, we evaluate the effect of the user interaction using sentiment scores of comments in on-line support groups.
According to BIBREF6, users with high incoming stress tend to request less information from others, as a percentage of their time, and share much more information, in absolute terms. In addition, high information sharing has been shown to be a good predictor of stress reduction at the end of the chat BIBREF6. Regarding information sharing, we rely on Dialogue Acts BIBREF19 to model the speaker's intention in producing an utterance. In particular, we are interested in Dialogue Act label that is defined to represent descriptive, narrative, or personal information – the statement.
Dialogue Acts have been applied to the analysis of spoken BIBREF20, BIBREF21 as well as on-line written synchronous conversations BIBREF22. We apply Dialogue Act tag set defined in BIBREF22 to the analysis of our on-line asynchronous conversations. We argue that Dialogue Acts can be used to analyse user behaviour in social media and verify the presence of therapeutic factors.
Methodology
We select the three therapeutic factors – Universality, Altruism and Instillation of Hope – that can be best approximated using NLP techniques: Sentiment Analysis and Dialogue Act tagging. We discuss each one of the selected therapeutic factors and the identified necessary conditions. The listed conditions, however, are not sufficient to attribute the presence of a therapeutic factor with high confidence, which only can be obtained using expert annotation. Our analysis focuses on the structure of conversations; though content plays an important role as well.
Universality consists in the disconfirmation of patients' belief of uniqueness of their disease. This therapeutic factor is shown to be a powerful source of relief for the patient, according to BIBREF7. From this definition, we can draw the following conditions that are applicable to our environment:
improvement of original poster's sentiment: we hypothesize that the discovery that other people passed through similar issues leads to a higher sentiment score;
posts containing negative personal experiences: to disconfirm the belief of uniqueness users have to share their story;
comments containing negative statements: to disconfirm the patient's feelings of uniqueness, the commenting user must tell a similar negative personal experience. This condition requires two sub-conditions: high presence of statements in comments and the presence of negative comments replying to negative posts.
Instillation of Hope is based on inspiration provided to participants by their peers. Through the inspiration provided by their peers, patients can increase their expectation on the therapy outcome. BIBREF7 in several studies have demonstrated that a high expectation of help before the start of a therapy is significantly correlated with a positive therapy outcome. The author states that many patients pointed out the importance of having observed the improvement of others. Therefore, the three main conditions are the following:
improvement of original poster's sentiment: we hypothesize that instillation of hope leads to a higher sentiment score;
posts containing negative personal experiences: hope can be instilled in someone who shares a negative personal experience;
comments containing positive personal experiences: in order to instill hope, commenting posters must show to original posters an overall positive personal experience. To detect positive personal experience, we require the presence of statements in comments and a positive sentiment of comments replying to negative posts.
Altruism consists of peers offering support, reassurance, suggestions and insight, since they share similar problems with one another BIBREF7. The experience of finding that a patient can be of value to others is refreshing and boosts self-esteem BIBREF7. However, in the current study we focus on testing whether commenting posters are altruists or not. We do not test whether the altruistic behavior leads to an improvement on the altruist itself. For these reasons, we define three main conditions:
improvement of original poster's sentiment: we hypothesize that supportive and reassuring statements improve the sentiment score of the original poster;
posts contains negative personal experiences: users offer support, reassurance and suggestion when facing a negative personal experience of the original poster;
comments containing positive statements: either supportive or reassuring statements show by definition a positive intended emotional communication. Thus comments to the post should consist of positive sentiment statements.
Consequently, a conversation containing the aforementioned therapeutic factors should satisfy the following conditions in terms of NLP: Sentiment Analysis and Dialogue Acts.
original posters have a higher sentiment score at the end of the thread than at the beginning;
the original post consists mostly of polarised statements;
the presence of a significant amount of statements in comments, since both support and sharing similar negative experiences can be represented as statements;
both negative and positive statements in comments lead to higher final sentiment score of the original poster.
Datasets
We verify the presence of therapeutic factors in two social media datasets: OSG and Twitter. The first dataset is crawled from an on-line support groups website, and the second dataset consists of a small sample of Twitter conversation threads. Since the former consists of multi-threaded conversations, we apply a pre-processing to extract conversation threads to provide a fair comparison with the Twitter dataset. An example conversation from each data source is presented in Figure FIGREF19.
Datasets ::: Twitter
We have downloaded 1,873 Twitter conversation threads, roughly 14k tweets, from a publicly available resource that were previously pre-processed and have conversation threads extracted. A conversation in the dataset consists of at least 4 tweets. Even though, according to BIBREF23, Twitter is broadly applicable to public health research, our expectation is that it contains less therapeutic conversations in comparison to specialized on-line support forums.
Datasets ::: OSG
Our data has been developed by crawling and pre-processing an OSG web forum. The forum has a great variety of different groups such as depression, anxiety, stress, relationship, cancer, sexually transmitted diseases, etc. Each conversation starts with one post and can contain multiple comments. Each post or comment is represented by a poster, a timestamp, a list of users it is referencing to, thread id, a comment id and a conversation id. The thread id is the same for comments replying to each other, otherwise it is different. The thread id is increasing with time. Thus, it provides ordering among threads; whereas the timestamp provides ordering in the thread.
Each conversation can belong to multiple groups. Consequently, the dataset needs to be processed to remove duplicates. The dataset resulting after de-duplication contains 295 thousand conversations, each conversation contains on average 6 comments. In total, there are 1.5 million comments. Since the created dataset is multi-threaded, we need to extract conversation threads, to eliminate paths not relevant to the original post.
Datasets ::: OSG ::: Conversation Thread Extraction
The thread extraction algorithm is heuristic-based and consists of two steps: (1) creation of a tree, based on a post written by a user and the related comments and (2) transformation of the tree into a list of threads.
The tree creation is an extension of the approach of BIBREF24, where first a graph of conversation is constructed. In the approach, direct replies to a post are attached to the first nesting level and subsequent comments to increasing nesting levels. In our approach, we also exploit comments' features.
The tree creation is performed without processing the content of comments, which allows us to process posts and comments of any length efficiently. The heuristic used in the process is based on three simplifying assumptions:
Unless there is a specific reference to another comment or a user, comments are attached to the original post.
When replying, the commenting poster is always replying to the original post or some other comment. Unless specified otherwise, it is assumed that it is a response to the previous (in time) post/comment.
Subsequent comments by the same poster are part of the same thread.
To evaluate the performance of the thread extraction algorithm, 2 annotators have manually constructed the trees for 100 conversations. The performance of the algorithm on this set of 100 conversations is evaluated using accuracy and standard Information Retrieval evaluation metrics of precision, recall, and F$_1$ measure. The results are reported in Table TABREF28 together with random and majority baselines. The turn-level percent agreement between the 2 annotators is 97.99% and Cohen's Kappa Coefficient is 83.80%.
Datasets ::: Data Representation
For both data sources, Twitter and OSG with extracted threads, posts and comments are tokenized and sentence split. Each sentence is passed through Sentiment Analysis and Dialogue Act tagging. Since a post or a comment can contain multiple sentences, therefore multiple Dialogue Acts, it is represented as as a one-hot encoding, where each position represents a Dialogue Act.
For Sentiment Analysis we use a lexicon-based sentiment analyser introduced by BIBREF25. For Dialogue Act tagging, on the other hand, we make use of a model trained on NPSChat corpus BIBREF22 following the approach of BIBREF26.
Analysis
As we mentioned in Section 3, the presence of each of the therapeutic conditions under analysis is a necessary for a conversation to be considered to have therapeutic factors. In this section we present the results of our analysis with respect to these conditions.
Analysis ::: Change in Sentiment score of Original Posters
The first condition which we test is the sentiment change in conversation threads, comparing the initial and final sentiment scores (i.e. posts' scores) of the original poster. The results of the analysis are presented in Figure FIGREF33. In the figure we can observe that the distribution of the sentiment change in the two datasets is different. While in Twitter the amount of conversations that lead to the increase of sentiment score is roughly equal to the amount of conversations that lead to the decrease of sentiment score; the situation is different for OSG. In OSG, the amount of conversations that lead to the increase of sentiment score is considerably higher.
Figure FIGREF34 provides a more fine grained analysis, where we additionally analyse the sentiment change in nominal polarity terms – negative and positive. In OSG, the number of users that changed polarity from negative to positive is more than the double of the users that have changed the polarity from positive to negative. In Twitter, on the other hand, the users mostly changed polarity from positive to negative. Results of the analysis suggest that in OSG , sentiment increases and users tend to change polarity from negative to positive, whereas in Twitter sentiment tends to decrease. Verification of this condition alone indicates that the ratio of potentially therapeutic conversations in Twitter is lower.
Analysis ::: Structure of Posts and Comments
Table TABREF36 presents the distribution of automatically predicted per-sentence Dialogue Acts in the datasets. The most frequent tag is statement in both. In Table TABREF37, on the other hand, we present the distribution of post and comment structures in terms of automatically predicted Dialogue Act tags. The structure is an unordered set of tags in the post or comment. From the table we can observe that the distribution of tag sets is similar between posts and comments. In both cases the most common set is statement only. However, conversations containing only statement, emphasis or question posts and comments predominantly appear in Twitter. Which is expected due to the shorter length of Twitter posts and comments.
We can also observe that the original posters tend to ask more questions than the commenting posters – 19.83% for posts vs. 11.21% for comments (summed). This suggests that the original posters frequently ask either for suggestion or confirmation of their points of view or their disconfirmation. However, the high presence of personal experiences is supported by the high number of posts containing only statements.
High number of statement tags in comments suggests that users reply either with supporting or empathic statements or personal experience. However, 6.39% of comments contain accept and reject tags, which mark the degree to which a speaker accepts some previous proposal, plan, opinion, or statement BIBREF20. The described Dialogue Act tags are often used when commenting posters discuss original poster's point of view. For instance, “It's true. I felt the same.” – {Accept, Statement} or “Well no. You're not alone” – {Reject, Statement}. The datasets differ with respect to the distribution of these Dialogue Acts tags, they appear more frequently in OSG.
Analysis ::: Sentiment of Posts and Comments
Table TABREF39 presents the distribution of sentiment polarity in post and comment statements (i.e. sentences tagged as statement). For OSG, the predominant sentiment label of statements is positive and it is the highest for both posts and comments. However, the difference between the amounts of positive and negative statements is higher for the replying comments (34.5% vs. 42.5%). For Twitter, on the other hand, the predominant sentiment label of statements is neutral and the polarity distribution between posts and comments is very close. One particular observation is that the ratio of negative statements is higher in OSG for both posts and comments than in Twitter, which supports the idea of sharing negative experiences.
Further we analyze whether the sentiment of a comment (i.e. the replying user) is affected by the sentiment of the original post (i.e. the user being replied to), which will imply that the users adapt their behaviour with respect to the post's sentiment. For the analysis, we split the datasets into three buckets according to the posts' sentiment score – negative, neutral, or positive, and represent each conversation in terms of percentages of comments (replies) with each sentiment label. The buckets are then compared using t-test for statistically significant differences.
Table TABREF40 presents the distribution of sentiment labels with respect to the post's sentiment score. The patterns of distribution are similar across the datasets. We can observe that overall, replies tend to have a positive sentiment, which suggests that replying posters tend to have a positive attitude. However, the ratio of positive comments is higher for OSG than for Twitter.
The results of the Welch's t-test on OSG data reveal that there are statistically significant differences in the distribution of replying comments' sentiment between conversations with positive and negative starting posts. A positive post tends to get significantly more positive replies. Similarly, a negative post tends to get significantly more negative replies (both with $p < 0.01$).
Table TABREF41 presents the distribution of the sentiment labels of the final text provided by the original poster with respect to the sentiment polarity of the comments. The results indicate that OSG participants are more supportive, as the majority of conversations end in a positive final sentiment regardless of the sentiment of comments. We can also observe that negative comments in OSG lead to positive sentiment, which supports the idea of sharing the negative experiences, thus presence of therapeutic factors. For Twitter, on the other hand, only positive comments lead to the positive final sentiments, whereas other comments lead predominantly to neutral final sentiments.
Our analysis in terms of sentiment and Dialogue Acts supports the presence of the three selected therapeutic factors – Universality, Altruism and Instillation of Hope – in OSG more than in Twitter. The main contributors to this conclusion are the facts that there is more positive change in the sentiment of the original posters in OSG (people seeking support) and that in OSG even negative and neutral comments are likely to lead to positive changes.
Conclusion
In this work, we propose a methodology to automatically analyse online social platforms for the presence of therapeutic factors (i.e. Universality, Altruism and Instillation of Hope). We evaluate our approach on two on-line platforms, Twitter and an OSG web forum. We apply NLP techniques of Sentiment Analysis and Dialogue Act tagging to automatically verify the presence of therapeutic factors, which allows us to analyse larger amounts of conversational data (as compared to previous studies).
Our analysis indicates that OSG conversations satisfy higher number of conditions approximating therapeutic factors than Twitter conversations. Given this outcome, we postulate that users who join support group websites spontaneously seem to benefit from it. Indeed, as shown in Section SECREF5, the original posters who interact with others by replying to comments, have benefited from an improvement of their emotional state.
We would like to reemphasise that the conditions for the therapeutic factors are necessary but not sufficient; since our analysis focuses on the structure of conversations, being agnostic to the content. NLP, however, allows us to strengthen our approximations even further. Thus, the further extension of our work is also augmentation of our study with other language analysis metrics and their correlation with human annotation.
It should be noted that the proposed approach is an approximation of the tedious tasks of annotation of conversations by experts versed in the therapeutic factors and their associated theories. Even though we can use Sentiment Analysis to detect the existence of therapeutic factors, we cannot differentiate between Altruism and Instillation of Hope, as this requires differentiation between emotional state of the user and the intended emotional communication. Thus, the natural extensions of this work are differentiation between different therapeutic factors and comparison of the proposed analysis to the human evaluation.
Although we acknowledge that the proposed methodology does not serve as a replacement of manual analysis of OSG for the presence of therapeutic factors, we believe that it could facilitate and supplement this process. The method can serve as a tool for general practitioners and psychologists who can use it as an additional source of information regarding their patients’ condition and, in turn, offer a more personalised support that is better tailored to individual therapeutic needs.
| [
"1,873 Twitter conversation threads, roughly 14k tweets",
"1,873 Twitter conversation threads, roughly 14k tweets"
] | 3,721 | qasper | en | null | a3aadf47eb153a70559416ba375c57883aec1538a2b17fb5 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How large is the Twitter dataset?
Answer: | [
"1,873 Twitter conversation threads, roughly 14k tweets",
"1,873 Twitter conversation threads, roughly 14k tweets"
] | qasper | 128 | 56 | |
What are the 12 languages covered? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The lack of annotated training and evaluation data for many tasks and domains hinders the development of computational models for the majority of the world's languages BIBREF0, BIBREF1, BIBREF2. The necessity to guide and advance multilingual and cross-lingual NLP through annotation efforts that follow cross-lingually consistent guidelines has been recently recognized by collaborative initiatives such as the Universal Dependency (UD) project BIBREF3. The latest version of UD (as of March 2020) covers more than 70 languages. Crucially, this resource continues to steadily grow and evolve through the contributions of annotators from across the world, extending the UD's reach to a wide array of typologically diverse languages. Besides steering research in multilingual parsing BIBREF4, BIBREF5, BIBREF6 and cross-lingual parser transfer BIBREF7, BIBREF8, BIBREF9, the consistent annotations and guidelines have also enabled a range of insightful comparative studies focused on the languages' syntactic (dis)similarities BIBREF10, BIBREF11, BIBREF12.
Inspired by the UD work and its substantial impact on research in (multilingual) syntax, in this article we introduce Multi-SimLex, a suite of manually and consistently annotated semantic datasets for 12 different languages, focused on the fundamental lexical relation of semantic similarity BIBREF13, BIBREF14. For any pair of words, this relation measures whether their referents share the same (functional) features, as opposed to general cognitive association captured by co-occurrence patterns in texts (i.e., the distributional information). Datasets that quantify the strength of true semantic similarity between concept pairs such as SimLex-999 BIBREF14 or SimVerb-3500 BIBREF15 have been instrumental in improving models for distributional semantics and representation learning. Discerning between semantic similarity and relatedness/association is not only crucial for theoretical studies on lexical semantics (see §SECREF2), but has also been shown to benefit a range of language understanding tasks in NLP. Examples include dialog state tracking BIBREF16, BIBREF17, spoken language understanding BIBREF18, BIBREF19, text simplification BIBREF20, BIBREF21, BIBREF22, dictionary and thesaurus construction BIBREF23, BIBREF24.
Despite the proven usefulness of semantic similarity datasets, they are available only for a small and typologically narrow sample of resource-rich languages such as German, Italian, and Russian BIBREF25, whereas some language types and low-resource languages typically lack similar evaluation data. Even if some resources do exist, they are limited in their size (e.g., 500 pairs in Turkish BIBREF26, 500 in Farsi BIBREF27, or 300 in Finnish BIBREF28) and coverage (e.g., all datasets which originated from the original English SimLex-999 contain only high-frequent concepts, and are dominated by nouns). This is why, as our departure point, we introduce a larger and more comprehensive English word similarity dataset spanning 1,888 concept pairs (see §SECREF4).
Most importantly, semantic similarity datasets in different languages have been created using heterogeneous construction procedures with different guidelines for translation and annotation, as well as different rating scales. For instance, some datasets were obtained by directly translating the English SimLex-999 in its entirety BIBREF25, BIBREF16 or in part BIBREF28. Other datasets were created from scratch BIBREF26 and yet others sampled English concept pairs differently from SimLex-999 and then translated and reannotated them in target languages BIBREF27. This heterogeneity makes these datasets incomparable and precludes systematic cross-linguistic analyses. In this article, consolidating the lessons learned from previous dataset construction paradigms, we propose a carefully designed translation and annotation protocol for developing monolingual Multi-SimLex datasets with aligned concept pairs for typologically diverse languages. We apply this protocol to a set of 12 languages, including a mixture of major languages (e.g., Mandarin, Russian, and French) as well as several low-resource ones (e.g., Kiswahili, Welsh, and Yue Chinese). We demonstrate that our proposed dataset creation procedure yields data with high inter-annotator agreement rates (e.g., the average mean inter-annotator agreement for Welsh is 0.742).
The unified construction protocol and alignment between concept pairs enables a series of quantitative analyses. Preliminary studies on the influence that polysemy and cross-lingual variation in lexical categories (see §SECREF6) have on similarity judgments are provided in §SECREF5. Data created according to Multi-SimLex protocol also allow for probing into whether similarity judgments are universal across languages, or rather depend on linguistic affinity (in terms of linguistic features, phylogeny, and geographical location). We investigate this question in §SECREF25. Naturally, Multi-SimLex datasets can be used as an intrinsic evaluation benchmark to assess the quality of lexical representations based on monolingual, joint multilingual, and transfer learning paradigms. We conduct a systematic evaluation of several state-of-the-art representation models in §SECREF7, showing that there are large gaps between human and system performance in all languages. The proposed construction paradigm also supports the automatic creation of 66 cross-lingual Multi-SimLex datasets by interleaving the monolingual ones. We outline the construction of the cross-lingual datasets in §SECREF6, and then present a quantitative evaluation of a series of cutting-edge cross-lingual representation models on this benchmark in §SECREF8.
Contributions. We now summarize the main contributions of this work:
1) Building on lessons learned from prior work, we create a more comprehensive lexical semantic similarity dataset for the English language spanning a total of 1,888 concept pairs balanced with respect to similarity, frequency, and concreteness, and covering four word classes: nouns, verbs, adjectives and, for the first time, adverbs. This dataset serves as the main source for the creation of equivalent datasets in several other languages.
2) We present a carefully designed and rigorous language-agnostic translation and annotation protocol. These well-defined guidelines will facilitate the development of future Multi-SimLex datasets for other languages. The proposed protocol eliminates some crucial issues with prior efforts focused on the creation of multi-lingual semantic resources, namely: i) limited coverage; ii) heterogeneous annotation guidelines; and iii) concept pairs which are semantically incomparable across different languages.
3) We offer to the community manually annotated evaluation sets of 1,888 concept pairs across 12 typologically diverse languages, and 66 large cross-lingual evaluation sets. To the best of our knowledge, Multi-SimLex is the most comprehensive evaluation resource to date focused on the relation of semantic similarity.
4) We benchmark a wide array of recent state-of-the-art monolingual and cross-lingual word representation models across our sample of languages. The results can serve as strong baselines that lay the foundation for future improvements.
5) We present a first large-scale evaluation study on the ability of encoders pretrained on language modeling (such as bert BIBREF29 and xlm BIBREF30) to reason over word-level semantic similarity in different languages. To our own surprise, the results show that monolingual pretrained encoders, even when presented with word types out of context, are sometimes competitive with static word embedding models such as fastText BIBREF31 or word2vec BIBREF32. The results also reveal a huge gap in performance between massively multilingual pretrained encoders and language-specific encoders in favor of the latter: our findings support other recent empirical evidence related to the “curse of multilinguality” BIBREF33, BIBREF34 in representation learning.
6) We make all of these resources available on a website which facilitates easy creation, submission and sharing of Multi-Simlex-style datasets for a larger number of languages. We hope that this will yield an even larger repository of semantic resources that inspire future advances in NLP within and across languages.
In light of the success of Universal Dependencies BIBREF3, we hope that our initiative will instigate a collaborative public effort with established and clear-cut guidelines that will result in additional Multi-SimLex datasets in a large number of languages in the near future. Moreover, we hope that it will provide means to advance our understanding of distributional and lexical semantics across a large number of languages. All monolingual and cross-lingual Multi-SimLex datasets–along with detailed translation and annotation guidelines–are available online at: https://multisimlex.com/.
Lexical Semantic Similarity ::: Similarity and Association
The focus of the Multi-SimLex initiative is on the lexical relation of pure semantic similarity. For any pair of words, this relation measures whether their referents share the same features. For instance, graffiti and frescos are similar to the extent that they are both forms of painting and appear on walls. This relation can be contrasted with the cognitive association between two words, which often depends on how much their referents interact in the real world, or are found in the same situations. For instance, a painter is easily associated with frescos, although they lack any physical commonalities. Association is also known in the literature under other names: relatedness BIBREF13, topical similarity BIBREF35, and domain similarity BIBREF36.
Semantic similarity and association overlap to some degree, but do not coincide BIBREF37, BIBREF38. In fact, there exist plenty of pairs that are intuitively associated but not similar. Pairs where the converse is true can also be encountered, although more rarely. An example are synonyms where a word is common and the other infrequent, such as to seize and to commandeer. BIBREF14 revealed that while similarity measures based on the WordNet graph BIBREF39 and human judgments of association in the University of South Florida Free Association Database BIBREF40 do correlate, a number of pairs follow opposite trends. Several studies on human cognition also point in the same direction. For instance, semantic priming can be triggered by similar words without association BIBREF41. On the other hand, a connection with cue words is established more quickly for topically related words rather than for similar words in free association tasks BIBREF42.
A key property of semantic similarity is its gradience: pairs of words can be similar to a different degree. On the other hand, the relation of synonymy is binary: pairs of words are synonyms if they can be substituted in all contexts (or most contexts, in a looser sense), otherwise they are not. While synonyms can be conceived as lying on one extreme of the semantic similarity continuum, it is crucial to note that their definition is stated in purely relational terms, rather than invoking their referential properties BIBREF43, BIBREF44, BIBREF45. This makes behavioral studies on semantic similarity fundamentally different from lexical resources like WordNet BIBREF46, which include paradigmatic relations (such as synonymy).
Lexical Semantic Similarity ::: Similarity for NLP: Intrinsic Evaluation and Semantic Specialization
The ramifications of the distinction between similarity and association are profound for distributional semantics. This paradigm of lexical semantics is grounded in the distributional hypothesis, formulated by BIBREF47 and BIBREF48. According to this hypothesis, the meaning of a word can be recovered empirically from the contexts in which it occurs within a collection of texts. Since both pairs of topically related words and pairs of purely similar words tend to appear in the same contexts, their associated meaning confounds the two distinct relations BIBREF14, BIBREF49, BIBREF50. As a result, distributional methods obscure a crucial facet of lexical meaning. This limitation also reflects onto word embeddings (WEs), representations of words as low-dimensional vectors that have become indispensable for a wide range of NLP applications BIBREF51, BIBREF52, BIBREF53. In particular, it involves both static WEs learned from co-occurrence patterns BIBREF32, BIBREF54, BIBREF31 and contextualized WEs learned from modeling word sequences BIBREF55, BIBREF29. As a result, in the induced representations, geometrical closeness (measured e.g. through cosine distance) conflates genuine similarity with broad relatedness. For instance, the vectors for antonyms such as sober and drunk, by definition dissimilar, might be neighbors in the semantic space under the distributional hypothesis. BIBREF36, BIBREF56, and BIBREF53 demonstrated that different choices of hyper-parameters in WE algorithms (such as context window) emphasize different relations in the resulting representations. Likewise, BIBREF57 and BIBREF54 discovered that WEs learned from texts annotated with syntactic information mirror similarity better than simple local bag-of-words neighborhoods.
The failure of WEs to capture semantic similarity, in turn, affects model performance in several NLP applications where such knowledge is crucial. In particular, Natural Language Understanding tasks such as statistical dialog modeling, text simplification, or semantic text similarity BIBREF58, BIBREF18, BIBREF59, among others, suffer the most. As a consequence, resources providing clean information on semantic similarity are key in mitigating the side effects of the distributional signal. In particular, such databases can be employed for the intrinsic evaluations of specific WE models as a proxy of their reliability for downstream applications BIBREF60, BIBREF61, BIBREF14; intuitively, the more WEs are misaligned with human judgments of similarity, the more their performance on actual tasks is expected to be degraded. Moreover, word representations can be specialized (a.k.a. retrofitted) by disentangling word relations of similarity and association. In particular, linguistic constraints sourced from external databases (such as synonyms from WordNet) can be injected into WEs BIBREF62, BIBREF63, BIBREF16, BIBREF22, BIBREF64 in order to enforce a particular relation in a distributional semantic space while preserving the original adjacency properties.
Lexical Semantic Similarity ::: Similarity and Language Variation: Semantic Typology
In this work, we tackle the concept of (true) semantic similarity from a multilingual perspective. While the same meaning representations may be shared by all human speakers at a deep cognitive level, there is no one-to-one mapping between the words in the lexicons of different languages. This makes the comparison of similarity judgments across languages difficult, since the meaning overlap of translationally equivalent words is sometimes far less than exact. This results from the fact that the way languages `partition' semantic fields is partially arbitrary BIBREF65, although constrained cross-lingually by common cognitive biases BIBREF66. For instance, consider the field of colors: English distinguishes between green and blue, whereas Murle (South Sudan) has a single word for both BIBREF67.
In general, semantic typology studies the variation in lexical semantics across the world's languages. According to BIBREF68, the ways languages categorize concepts into the lexicon follow three main axes: 1) granularity: what is the number of categories in a specific domain?; 2) boundary location: where do the lines marking different categories lie?; 3) grouping and dissection: what are the membership criteria of a category; which instances are considered to be more prototypical? Different choices with respect to these axes lead to different lexicalization patterns. For instance, distinct senses in a polysemous word in English, such as skin (referring to both the body and fruit), may be assigned separate words in other languages such as Italian pelle and buccia, respectively BIBREF70. We later analyze whether similarity scores obtained from native speakers also loosely follow the patterns described by semantic typology.
Previous Work and Evaluation Data
Word Pair Datasets. Rich expert-created resources such as WordNet BIBREF46, BIBREF71, VerbNet BIBREF72, BIBREF73, or FrameNet BIBREF74 encode a wealth of semantic and syntactic information, but are expensive and time-consuming to create. The scale of this problem gets multiplied by the number of languages in consideration. Therefore, crowd-sourcing with non-expert annotators has been adopted as a quicker alternative to produce smaller and more focused semantic resources and evaluation benchmarks. This alternative practice has had a profound impact on distributional semantics and representation learning BIBREF14. While some prominent English word pair datasets such as WordSim-353 BIBREF75, MEN BIBREF76, or Stanford Rare Words BIBREF77 did not discriminate between similarity and relatedness, the importance of this distinction was established by BIBREF14 through the creation of SimLex-999. This inspired other similar datasets which focused on different lexical properties. For instance, SimVerb-3500 BIBREF15 provided similarity ratings for 3,500 English verbs, whereas CARD-660 BIBREF78 aimed at measuring the semantic similarity of infrequent concepts.
Semantic Similarity Datasets in Other Languages. Motivated by the impact of datasets such as SimLex-999 and SimVerb-3500 on representation learning in English, a line of related work focused on creating similar resources in other languages. The dominant approach is translating and reannotating the entire original English SimLex-999 dataset, as done previously for German, Italian, and Russian BIBREF25, Hebrew and Croatian BIBREF16, and Polish BIBREF79. Venekoski:2017nodalida apply this process only to a subset of 300 concept pairs from the English SimLex-999. On the other hand, BIBREF27 sampled a new set of 500 English concept pairs to ensure wider topical coverage and balance across similarity spectra, and then translated those pairs to German, Italian, Spanish, and Farsi (SEMEVAL-500). A similar approach was followed by BIBREF26 for Turkish, by BIBREF80 for Mandarin Chinese, and by BIBREF81 for Japanese. BIBREF82 translated the concatenation of SimLex-999, WordSim-353, and the English SEMEVAL-500 into Thai and then reannotated it. Finally, BIBREF83 translated English SimLex-999 and WordSim-353 to 11 resource-rich target languages (German, French, Russian, Italian, Dutch, Chinese, Portuguese, Swedish, Spanish, Arabic, Farsi), but they did not provide details concerning the translation process and the resolution of translation disagreements. More importantly, they also did not reannotate the translated pairs in the target languages. As we discussed in § SECREF6 and reiterate later in §SECREF5, semantic differences among languages can have a profound impact on the annotation scores; particulary, we show in §SECREF25 that these differences even roughly define language clusters based on language affinity.
A core issue with the current datasets concerns a lack of one unified procedure that ensures the comparability of resources in different languages. Further, concept pairs for different languages are sourced from different corpora (e.g., direct translation of the English data versus sampling from scratch in the target language). Moreover, the previous SimLex-based multilingual datasets inherit the main deficiencies of the English original version, such as the focus on nouns and highly frequent concepts. Finally, prior work mostly focused on languages that are widely spoken and do not account for the variety of the world's languages. Our long-term goal is devising a standardized methodology to extend the coverage also to languages that are resource-lean and/or typologically diverse (e.g., Welsh, Kiswahili as in this work).
Multilingual Datasets for Natural Language Understanding. The Multi-SimLex initiative and corresponding datasets are also aligned with the recent efforts on procuring multilingual benchmarks that can help advance computational modeling of natural language understanding across different languages. For instance, pretrained multilingual language models such as multilingual bert BIBREF29 or xlm BIBREF30 are typically probed on XNLI test data BIBREF84 for cross-lingual natural language inference. XNLI was created by translating examples from the English MultiNLI dataset, and projecting its sentence labels BIBREF85. Other recent multilingual datasets target the task of question answering based on reading comprehension: i) MLQA BIBREF86 includes 7 languages ii) XQuAD BIBREF87 10 languages; iii) TyDiQA BIBREF88 9 widely spoken typologically diverse languages. While MLQA and XQuAD result from the translation from an English dataset, TyDiQA was built independently in each language. Another multilingual dataset, PAWS-X BIBREF89, focused on the paraphrase identification task and was created translating the original English PAWS BIBREF90 into 6 languages. We believe that Multi-SimLex can substantially contribute to this endeavor by offering a comprehensive multilingual benchmark for the fundamental lexical level relation of semantic similarity. In future work, Multi-SimLex also offers an opportunity to investigate the correlations between word-level semantic similarity and performance in downstream tasks such as QA and NLI across different languages.
The Base for Multi-SimLex: Extending English SimLex-999
In this section, we discuss the design principles behind the English (eng) Multi-SimLex dataset, which is the basis for all the Multi-SimLex datasets in other languages, as detailed in §SECREF5. We first argue that a new, more balanced, and more comprehensive evaluation resource for lexical semantic similarity in English is necessary. We then describe how the 1,888 word pairs contained in the eng Multi-SimLex were selected in such a way as to represent various linguistic phenomena within a single integrated resource.
Construction Criteria. The following criteria have to be satisfied by any high-quality semantic evaluation resource, as argued by previous studies focused on the creation of such resources BIBREF14, BIBREF15, BIBREF91, BIBREF27:
(C1) Representative and diverse. The resource must cover the full range of diverse concepts occurring in natural language, including different word classes (e.g., nouns, verbs, adjectives, adverbs), concrete and abstract concepts, a variety of lexical fields, and different frequency ranges.
(C2) Clearly defined. The resource must provide a clear understanding of which semantic relation exactly is annotated and measured, possibly contrasting it with other relations. For instance, the original SimLex-999 and SimVerb-3500 explicitly focus on true semantic similarity and distinguish it from broader relatedness captured by datasets such as MEN BIBREF76 or WordSim-353 BIBREF75.
(C3) Consistent and reliable. The resource must ensure consistent annotations obtained from non-expert native speakers following simple and precise annotation guidelines.
In choosing the word pairs and constructing eng Multi-SimLex, we adhere to these requirements. Moreover, we follow good practices established by the research on related resources. In particular, since the introduction of the original SimLex-999 dataset BIBREF14, follow-up works have improved its construction protocol across several aspects, including: 1) coverage of more lexical fields, e.g., by relying on a diverse set of Wikipedia categories BIBREF27, 2) infrequent/rare words BIBREF78, 3) focus on particular word classes, e.g., verbs BIBREF15, 4) annotation quality control BIBREF78. Our goal is to make use of these improvements towards a larger, more representative, and more reliable lexical similarity dataset in English and, consequently, in all other languages.
The Final Output: English Multi-SimLex. In order to ensure that the criterion C1 is satisfied, we consolidate and integrate the data already carefully sampled in prior work into a single, comprehensive, and representative dataset. This way, we can control for diversity, frequency, and other properties while avoiding to perform this time-consuming selection process from scratch. Note that, on the other hand, the word pairs chosen for English are scored from scratch as part of the entire Multi-SimLex annotation process, introduced later in §SECREF5. We now describe the external data sources for the final set of word pairs:
1) Source: SimLex-999. BIBREF14. The English Multi-SimLex has been initially conceived as an extension of the original SimLex-999 dataset. Therefore, we include all 999 word pairs from SimLex, which span 666 noun pairs, 222 verb pairs, and 111 adjective pairs. While SimLex-999 already provides examples representing different POS classes, it does not have a sufficient coverage of different linguistic phenomena: for instance, it contains only very frequent concepts, and it does not provide a representative set of verbs BIBREF15.
2) Source: SemEval-17: Task 2 BIBREF92. We start from the full dataset of 500 concept pairs to extract a total of 334 concept pairs for English Multi-SimLex a) which contain only single-word concepts, b) which are not named entities, c) where POS tags of the two concepts are the same, d) where both concepts occur in the top 250K most frequent word types in the English Wikipedia, and e) do not already occur in SimLex-999. The original concepts were sampled as to span all the 34 domains available as part of BabelDomains BIBREF93, which roughly correspond to the main high-level Wikipedia categories. This ensures topical diversity in our sub-sample.
3) Source: CARD-660 BIBREF78. 67 word pairs are taken from this dataset focused on rare word similarity, applying the same selection criteria a) to e) employed for SEMEVAL-500. Words are controlled for frequency based on their occurrence counts from the Google News data and the ukWaC corpus BIBREF94. CARD-660 contains some words that are very rare (logboat), domain-specific (erythroleukemia) and slang (2mrw), which might be difficult to translate and annotate across a wide array of languages. Hence, we opt for retaining only the concept pairs above the threshold of top 250K most frequent Wikipedia concepts, as above.
4) Source: SimVerb-3500 BIBREF15 Since both CARD-660 and SEMEVAL-500 are heavily skewed towards noun pairs, and nouns also dominate the original SimLex-999, we also extract additional verb pairs from the verb-specific similarity dataset SimVerb-3500. We randomly sample 244 verb pairs from SimVerb-3500 that represent all similarity spectra. In particular, we add 61 verb pairs for each of the similarity intervals: $[0,1.5), [1.5,3), [3,4.5), [4.5, 6]$. Since verbs in SimVerb-3500 were originally chosen from VerbNet BIBREF95, BIBREF73, they cover a wide range of verb classes and their related linguistic phenomena.
5) Source: University of South Florida BIBREF96 norms, the largest database of free association for English. In order to improve the representation of different POS classes, we sample additional adjectives and adverbs from the USF norms following the procedure established by BIBREF14, BIBREF15. This yields additional 122 adjective pairs, but only a limited number of adverb pairs (e.g., later – never, now – here, once – twice). Therefore, we also create a set of adverb pairs semi-automatically by sampling adjectives that can be derivationally transformed into adverbs (e.g. adding the suffix -ly) from the USF, and assessing the correctness of such derivation in WordNet. The resulting pairs include, for instance, primarily – mainly, softly – firmly, roughly – reliably, etc. We include a total of 123 adverb pairs into the final English Multi-SimLex. Note that this is the first time adverbs are included into any semantic similarity dataset.
Fulfillment of Construction Criteria. The final eng Multi-SimLex dataset spans 1,051 noun pairs, 469 verb pairs, 245 adjective pairs, and 123 adverb pairs. As mentioned above, the criterion C1 has been fulfilled by relying only on word pairs that already underwent meticulous sampling processes in prior work, integrating them into a single resource. As a consequence, Multi-SimLex allows for fine-grained analyses over different POS classes, concreteness levels, similarity spectra, frequency intervals, relation types, morphology, lexical fields, and it also includes some challenging orthographically similar examples (e.g., infection – inflection). We ensure that the criteria C2 and C3 are satisfied by using similar annotation guidelines as Simlex-999, SimVerb-3500, and SEMEVAL-500 that explicitly target semantic similarity. In what follows, we outline the carefully tailored process of translating and annotating Multi-SimLex datasets in all target languages.
Multi-SimLex: Translation and Annotation
We now detail the development of the final Multi-SimLex resource, describing our language selection process, as well as translation and annotation of the resource, including the steps taken to ensure and measure the quality of this resource. We also provide key data statistics and preliminary cross-lingual comparative analyses.
Language Selection. Multi-SimLex comprises eleven languages in addition to English. The main objective for our inclusion criteria has been to balance language prominence (by number of speakers of the language) for maximum impact of the resource, while simultaneously having a diverse suite of languages based on their typological features (such as morphological type and language family). Table TABREF10 summarizes key information about the languages currently included in Multi-SimLex. We have included a mixture of fusional, agglutinative, isolating, and introflexive languages that come from eight different language families. This includes languages that are very widely used such as Chinese Mandarin and Spanish, and low-resource languages such as Welsh and Kiswahili. We hope to further include additional languages and inspire other researchers to contribute to the effort over the lifetime of this project.
The work on data collection can be divided into two crucial phases: 1) a translation phase where the extended English language dataset with 1,888 pairs (described in §SECREF4) is translated into eleven target languages, and 2) an annotation phase where human raters scored each pair in the translated set as well as the English set. Detailed guidelines for both phases are available online at: https://multisimlex.com.
Multi-SimLex: Translation and Annotation ::: Word Pair Translation
Translators for each target language were instructed to find direct or approximate translations for the 1,888 word pairs that satisfy the following rules. (1) All pairs in the translated set must be unique (i.e., no duplicate pairs); (2) Translating two words from the same English pair into the same word in the target language is not allowed (e.g., it is not allowed to translate car and automobile to the same Spanish word coche). (3) The translated pairs must preserve the semantic relations between the two words when possible. This means that, when multiple translations are possible, the translation that best conveys the semantic relation between the two words found in the original English pair is selected. (4) If it is not possible to use a single-word translation in the target language, then a multi-word expression (MWE) can be used to convey the nearest possible semantics given the above points (e.g., the English word homework is translated into the Polish MWE praca domowa).
Satisfying the above rules when finding appropriate translations for each pair–while keeping to the spirit of the intended semantic relation in the English version–is not always straightforward. For instance, kinship terminology in Sinitic languages (Mandarin and Yue) uses different terms depending on whether the family member is older or younger, and whether the family member comes from the mother’s side or the father’s side. UTF8gbsn In Mandarin, brother has no direct translation and can be translated as either: 哥哥 (older brother) or 弟弟 (younger brother). Therefore, in such cases, the translators are asked to choose the best option given the semantic context (relation) expressed by the pair in English, otherwise select one of the translations arbitrarily. This is also used to remove duplicate pairs in the translated set, by differentiating the duplicates using a variant at each instance. Further, many translation instances were resolved using near-synonymous terms in the translation. For example, the words in the pair: wood – timber can only be directly translated in Estonian to puit, and are not distinguishable. Therefore, the translators approximated the translation for timber to the compound noun puitmaterjal (literally: wood material) in order to produce a valid pair in the target language. In some cases, a direct transliteration from English is used. For example, the pair: physician and doctor both translate to the same word in Estonian (arst); the less formal word doktor is used as a translation of doctor to generate a valid pair.
We measure the quality of the translated pairs by using a random sample set of 100 pairs (from the 1,888 pairs) to be translated by an independent translator for each target language. The sample is proportionally stratified according to the part-of-speech categories. The independent translator is given identical instructions to the main translator; we then measure the percentage of matched translated words between the two translations of the sample set. Table TABREF12 summarizes the inter-translator agreement results for all languages and by part-of-speech subsets. Overall across all languages, the agreement is 84.8%, which is similar to prior work BIBREF27, BIBREF97.
Multi-SimLex: Translation and Annotation ::: Guidelines and Word Pair Scoring
Across all languages, 145 human annotators were asked to score all 1,888 pairs (in their given language). We finally collect at least ten valid annotations for each word pair in each language. All annotators were required to abide by the following instructions:
1. Each annotator must assign an integer score between 0 and 6 (inclusive) indicating how semantically similar the two words in a given pair are. A score of 6 indicates very high similarity (i.e., perfect synonymy), while zero indicates no similarity.
2. Each annotator must score the entire set of 1,888 pairs in the dataset. The pairs must not be shared between different annotators.
3. Annotators are able to break the workload over a period of approximately 2-3 weeks, and are able to use external sources (e.g. dictionaries, thesauri, WordNet) if required.
4. Annotators are kept anonymous, and are not able to communicate with each other during the annotation process.
The selection criteria for the annotators required that all annotators must be native speakers of the target language. Preference to annotators with university education was given, but not required. Annotators were asked to complete a spreadsheet containing the translated pairs of words, as well as the part-of-speech, and a column to enter the score. The annotators did not have access to the original pairs in English.
To ensure the quality of the collected ratings, we have employed an adjudication protocol similar to the one proposed and validated by Pilehvar:2018emnlp. It consists of the following three rounds:
Round 1: All annotators are asked to follow the instructions outlined above, and to rate all 1,888 pairs with integer scores between 0 and 6.
Round 2: We compare the scores of all annotators and identify the pairs for each annotator that have shown the most disagreement. We ask the annotators to reconsider the assigned scores for those pairs only. The annotators may chose to either change or keep the scores. As in the case with Round 1, the annotators have no access to the scores of the other annotators, and the process is anonymous. This process gives a chance for annotators to correct errors or reconsider their judgments, and has been shown to be very effective in reaching consensus, as reported by BIBREF78. We used a very similar procedure as BIBREF78 to identify the pairs with the most disagreement; for each annotator, we marked the $i$th pair if the rated score $s_i$ falls within: $s_i \ge \mu _i + 1.5$ or $s_i \le \mu _i - 1.5$, where $\mu _i$ is the mean of the other annotators' scores.
Round 3: We compute the average agreement for each annotator (with the other annotators), by measuring the average Spearman's correlation against all other annotators. We discard the scores of annotators that have shown the least average agreement with all other annotators, while we maintain at least ten annotators per language by the end of this round. The actual process is done in multiple iterations: (S1) we measure the average agreement for each annotator with every other annotator (this corresponds to the APIAA measure, see later); (S2) if we still have more than 10 valid annotators and the lowest average score is higher than in the previous iteration, we remove the lowest one, and rerun S1. Table TABREF14 shows the number of annotators at both the start (Round 1) and end (Round 3) of our process for each language.
We measure the agreement between annotators using two metrics, average pairwise inter-annotator agreement (APIAA), and average mean inter-annotator agreement (AMIAA). Both of these use Spearman's correlation ($\rho $) between annotators scores, the only difference is how they are averaged. They are computed as follows:
where $\rho (s_i,s_j)$ is the Spearman's correlation between annotators $i$ and $j$'s scores ($s_i$,$s_j$) for all pairs in the dataset, and $N$ is the number of annotators. APIAA has been used widely as the standard measure for inter-annotator agreement, including in the original SimLex paper BIBREF14. It simply averages the pairwise Spearman's correlation between all annotators. On the other hand, AMIAA compares the average Spearman's correlation of one held-out annotator with the average of all the other $N-1$ annotators, and then averages across all $N$ `held-out' annotators. It smooths individual annotator effects and arguably serves as a better upper bound than APIAA BIBREF15, BIBREF91, BIBREF78.
We present the respective APIAA and AMIAA scores in Table TABREF16 and Table TABREF17 for all part-of-speech subsets, as well as the agreement for the full datasets. As reported in prior work BIBREF15, BIBREF91, AMIAA scores are typically higher than APIAA scores. Crucially, the results indicate `strong agreement' (across all languages) using both measurements. The languages with the highest annotator agreement were French (fra) and Yue Chinese (yue), while Russian (rus) had the lowest overall IAA scores. These scores, however, are still considered to be `moderately strong agreement'.
Multi-SimLex: Translation and Annotation ::: Data Analysis
Similarity Score Distributions. Across all languages, the average score (mean $=1.61$, median$=1.1$) is on the lower side of the similarity scale. However, looking closer at the scores of each language in Table TABREF19, we indicate notable differences in both the averages and the spread of scores. Notably, French has the highest average of similarity scores (mean$=2.61$, median$=2.5$), while Kiswahili has the lowest average (mean$=1.28$, median$=0.5$). Russian has the lowest spread ($\sigma =1.37$), while Polish has the largest ($\sigma =1.62$). All of the languages are strongly correlated with each other, as shown in Figure FIGREF20, where all of the Spearman's correlation coefficients are greater than 0.6 for all language pairs. Languages that share the same language family are highly correlated (e.g, cmn-yue, rus-pol, est-fin). In addition, we observe high correlations between English and most other languages, as expected. This is due to the effect of using English as the base/anchor language to create the dataset. In simple words, if one translates to two languages $L_1$ and $L_2$ starting from the same set of pairs in English, it is higly likely that $L_1$ and $L_2$ will diverge from English in different ways. Therefore, the similarity between $L_1$-eng and $L_2$-eng is expected to be higher than between $L_1$-$L_2$, especially if $L_1$ and $L_2$ are typologically dissimilar languages (e.g., heb-cmn, see Figure FIGREF20). This phenomenon is well documented in related prior work BIBREF25, BIBREF27, BIBREF16, BIBREF97. While we acknowledge this as a slight artifact of the dataset design, it would otherwise be impossible to construct a semantically aligned and comprehensive dataset across a large number of languages.
We also report differences in the distribution of the frequency of words among the languages in Multi-SimLex. Figure FIGREF22 shows six example languages, where each bar segment shows the proportion of words in each language that occur in the given frequency range. For example, the 10K-20K segment of the bars represents the proportion of words in the dataset that occur in the list of most frequent words between the frequency rank of 10,000 and 20,000 in that language; likewise with other intervals. Frequency lists for the presented languages are derived from Wikipedia and Common Crawl corpora. While many concept pairs are direct or approximate translations of English pairs, we can see that the frequency distribution does vary across different languages, and is also related to inherent language properties. For instance, in Finnish and Russian, while we use infinitive forms of all verbs, conjugated verb inflections are often more frequent in raw corpora than the corresponding infinitive forms. The variance can also be partially explained by the difference in monolingual corpora size used to derive the frequency rankings in the first place: absolute vocabulary sizes are expected to fluctuate across different languages. However, it is also important to note that the datasets also contain subsets of lower-frequency and rare words, which can be used for rare word evaluations in multiple languages, in the spirit of Pilehvar:2018emnlp's English rare word dataset.
Cross-Linguistic Differences. Table TABREF23 shows some examples of average similarity scores of English, Spanish, Kiswahili and Welsh concept pairs. Remember that the scores range from 0 to 6: the higher the score, the more similar the participants found the concepts in the pair. The examples from Table TABREF23 show evidence of both the stability of average similarity scores across languages (unlikely – friendly, book – literature, and vanish – disappear), as well as language-specific differences (care – caution). Some differences in similarity scores seem to group languages into clusters. For example, the word pair regular – average has an average similarity score of 4.0 and 4.1 in English and Spanish, respectively, whereas in Kiswahili and Welsh the average similarity score of this pair is 0.5 and 0.8. We analyze this phenomenon in more detail in §SECREF25.
There are also examples for each of the four languages having a notably higher or lower similarity score for the same concept pair than the three other languages. For example, large – big in English has an average similarity score of 5.9, whereas Spanish, Kiswahili and Welsh speakers rate the closest concept pair in their native language to have a similarity score between 2.7 and 3.8. What is more, woman – wife receives an average similarity of 0.9 in English, 2.9 in Spanish, and greater than 4.0 in Kiswahili and Welsh. The examples from Spanish include banco – asiento (bank – seat) which receives an average similarity score 5.1, while in the other three languages the similarity score for this word pair does not exceed 0.1. At the same time, the average similarity score of espantosamente – fantásticamente (amazingly – fantastically) is much lower in Spanish (0.4) than in other languages (4.1 – 5.1). In Kiswahili, an example of a word pair with a higher similarity score than the rest would be machweo – jioni (sunset – evening), having an average score of 5.5, while the other languages receive 2.8 or less, and a notably lower similarity score is given to wa ajabu - mkubwa sana (wonderful – terrific), getting 0.9, while the other languages receive 5.3 or more. Welsh examples include yn llwyr - yn gyfan gwbl (purely – completely), which scores 5.4 among Welsh speakers but 2.3 or less in other languages, while addo – tyngu (promise – swear) is rated as 0 by all Welsh annotators, but in the other three languages 4.3 or more on average.
There can be several explanations for the differences in similarity scores across languages, including but not limited to cultural context, polysemy, metonymy, translation, regional and generational differences, and most commonly, the fact that words and meanings do not exactly map onto each other across languages. For example, it is likely that the other three languages do not have two separate words for describing the concepts in the concept pair: big – large, and the translators had to opt for similar lexical items that were more distant in meaning, explaining why in English the concept pair received a much higher average similarity score than in other languages. A similar issue related to the mapping problem across languages arose in the Welsh concept pair yn llwye – yn gyfan gwbl, where Welsh speakers agreed that the two concepts are very similar. When asked, bilingual speakers considered the two Welsh concepts more similar than English equivalents purely – completely, potentially explaining why a higher average similarity score was reached in Welsh. The example of woman – wife can illustrate cultural differences or another translation-related issue where the word `wife' did not exist in some languages (for example, Estonian), and therefore had to be described using other words, affecting the comparability of the similarity scores. This was also the case with the football – soccer concept pair. The pair bank – seat demonstrates the effect of the polysemy mismatch across languages: while `bank' has two different meanings in English, neither of them is similar to the word `seat', but in Spanish, `banco' can mean `bank', but it can also mean `bench'. Quite naturally, Spanish speakers gave the pair banco – asiento a higher similarity score than the speakers of languages where this polysemy did not occur.
An example of metonymy affecting the average similarity score can be seen in the Kiswahili version of the word pair: sunset – evening (machweo – jioni). The average similarity score for this pair is much higher in Kiswahili, likely because the word `sunset' can act as a metonym of `evening'. The low similarity score of wonderful – terrific in Kiswahili (wa ajabu - mkubwa sana) can be explained by the fact that while `mkubwa sana' can be used as `terrific' in Kiswahili, it technically means `very big', adding to the examples of translation- and mapping-related effects. The word pair amazingly – fantastically (espantosamente – fantásticamente) brings out another translation-related problem: the accuracy of the translation. While `espantosamente' could arguably be translated to `amazingly', more common meanings include: `frightfully', `terrifyingly', and `shockingly', explaining why the average similarity score differs from the rest of the languages. Another problem was brought out by addo – tyngu (promise – swear) in Welsh, where the `tyngu' may not have been a commonly used or even a known word choice for annotators, pointing out potential regional or generational differences in language use.
Table TABREF24 presents examples of concept pairs from English, Spanish, Kiswahili, and Welsh on which the participants agreed the most. For example, in English all participants rated the similarity of trial – test to be 4 or 5. In Spanish and Welsh, all participants rated start – begin to correspond to a score of 5 or 6. In Kiswahili, money – cash received a similarity rating of 6 from every participant. While there are numerous examples of concept pairs in these languages where the participants agreed on a similarity score of 4 or higher, it is worth noting that none of these languages had a single pair where all participants agreed on either 1-2, 2-3, or 3-4 similarity rating. Interestingly, in English all pairs where all the participants agreed on a 5-6 similarity score were adjectives.
Multi-SimLex: Translation and Annotation ::: Effect of Language Affinity on Similarity Scores
Based on the analysis in Figure FIGREF20 and inspecting the anecdotal examples in the previous section, it is evident that the correlation between similarity scores across languages is not random. To corroborate this intuition, we visualize the vectors of similarity scores for each single language by reducing their dimensionality to 2 via Principal Component Analysis BIBREF98. The resulting scatter plot in Figure FIGREF26 reveals that languages from the same family or branch have similar patterns in the scores. In particular, Russian and Polish (both Slavic), Finnish and Estonian (both Uralic), Cantonese and Mandarin Chinese (both Sinitic), and Spanish and French (both Romance) are all neighbors.
In order to quantify exactly the effect of language affinity on the similarity scores, we run correlation analyses between these and language features. In particular, we extract feature vectors from URIEL BIBREF99, a massively multilingual typological database that collects and normalizes information compiled by grammarians and field linguists about the world's languages. In particular, we focus on information about geography (the areas where the language speakers are concentrated), family (the phylogenetic tree each language belongs to), and typology (including syntax, phonological inventory, and phonology). Moreover, we consider typological representations of languages that are not manually crafted by experts, but rather learned from texts. BIBREF100 proposed to construct such representations by training language-identifying vectors end-to-end as part of neural machine translation models.
The vector for similarity judgments and the vector of linguistic features for a given language have different dimensionality. Hence, we first construct a distance matrix for each vector space, such that both columns and rows are language indices, and each cell value is the cosine distance between the vectors of the corresponding language pair. Given a set of L languages, each resulting matrix $S$ has dimensionality of $\mathbb {R}^{|L| \times |L|}$ and is symmetrical. To estimate the correlation between the matrix for similarity judgments and each of the matrices for linguistic features, we run a Mantel test BIBREF101, a non-parametric statistical test based on matrix permutations that takes into account inter-dependencies among pairwise distances.
The results of the Mantel test reported in Table FIGREF26 show that there exist statistically significant correlations between similarity judgments and geography, family, and syntax, given that $p < 0.05$ and $z > 1.96$. The correlation coefficient is particularly strong for geography ($r = 0.647$) and syntax ($r = 0.649$). The former result is intuitive, because languages in contact easily borrow and loan lexical units, and cultural interactions may result in similar cognitive categorizations. The result for syntax, instead, cannot be explained so easily, as formal properties of language do not affect lexical semantics. Instead, we conjecture that, while no causal relation is present, both syntactic features and similarity judgments might be linked to a common explanatory variable (such as geography). In fact, several syntactic properties are not uniformly spread across the globe. For instance, verbs with Verb–Object–Subject word order are mostly concentrated in Oceania BIBREF102. In turn, geographical proximity leads to similar judgment patterns, as mentioned above. On the other hand, we find no correlation with phonology and inventory, as expected, nor with the bottom-up typological features from BIBREF100.
Cross-Lingual Multi-SimLex Datasets
A crucial advantage of having semantically aligned monolingual datasets across different languages is the potential to create cross-lingual semantic similarity datasets. Such datasets allow for probing the quality of cross-lingual representation learning algorithms BIBREF27, BIBREF103, BIBREF104, BIBREF105, BIBREF106, BIBREF30, BIBREF107 as an intrinsic evaluation task. However, the cross-lingual datasets previous work relied upon BIBREF27 were limited to a homogeneous set of high-resource languages (e.g., English, German, Italian, Spanish) and a small number of concept pairs (all less than 1K pairs). We address both problems by 1) using a typologically more diverse language sample, and 2) relying on a substantially larger English dataset as a source for the cross-lingual datasets: 1,888 pairs in this work versus 500 pairs in the work of Camacho:2017semeval. As a result, each of our cross-lingual datasets contains a substantially larger number of concept pairs, as shown in Table TABREF30. The cross-lingual Multi-Simlex datasets are constructed automatically, leveraging word pair translations and annotations collected in all 12 languages. This yields a total of 66 cross-lingual datasets, one for each possible combination of languages. Table TABREF30 provides the final number of concept pairs, which lie between 2,031 and 3,480 pairs for each cross-lingual dataset, whereas Table TABREF29 shows some sample pairs with their corresponding similarity scores.
The automatic creation and verification of cross-lingual datasets closely follows the procedure first outlined by Camacho:2015acl and later adopted by Camacho:2017semeval (for semantic similarity) and Vulic:2019acl (for graded lexical entailment). First, given two languages, we intersect their aligned concept pairs obtained through translation. For instance, starting from the aligned pairs attroupement – foule in French and rahvasumm – rahvahulk in Estonian, we construct two cross-lingual pairs attroupement – rahvaluk and rahvasumm – foule. The scores of cross-lingual pairs are then computed as averages of the two corresponding monolingual scores. Finally, in order to filter out concept pairs whose semantic meaning was not preserved during this operation, we retain only cross-lingual pairs for which the corresponding monolingual scores $(s_s, s_t)$ differ at most by one fifth of the full scale (i.e., $\mid s_s - s_t \mid \le 1.2$). This heuristic mitigates the noise due to cross-lingual semantic shifts BIBREF27, BIBREF97. We refer the reader to the work of Camacho:2015acl for a detailed technical description of the procedure. UTF8gbsn
To assess the quality of the resulting cross-lingual datasets, we have conducted a verification experiment similar to Vulic:2019acl. We randomly sampled 300 concept pairs in the English-Spanish, English-French, and English-Mandarin cross-lingual datasets. Subsequently, we asked bilingual native speakers to provide similarity judgments of each pair. The Spearman's correlation score $\rho $ between automatically induced and manually collected ratings achieves $\rho \ge 0.90$ on all samples, which confirms the viability of the automatic construction procedure.
Score and Class Distributions. The summary of score and class distributions across all 66 cross-lingual datasets are provided in Figure FIGREF31 and Figure FIGREF31, respectively. First, it is obvious that the distribution over the four POS classes largely adheres to that of the original monolingual Multi-SimLex datasets, and that the variance is quite low: e.g., the eng-fra dataset contains the lowest proportion of nouns (49.21%) and the highest proportion of verbs (27.1%), adjectives (15.28%), and adverbs (8.41%). On the other hand, the distribution over similarity intervals in Figure FIGREF31 shows a much greater variance. This is again expected as this pattern resurfaces in monolingual datasets (see Table TABREF19). It is also evident that the data are skewed towards lower-similarity concept pairs. However, due to the joint size of all cross-lingual datasets (see Table TABREF30), even the least represented intervals contain a substantial number of concept pairs. For instance, the rus-yue dataset contains the least highly similar concept pairs (in the interval $[4,6]$) of all 66 cross-lingual datasets. Nonetheless, the absolute number of pairs (138) in that interval for rus-yue is still substantial. If needed, this makes it possible to create smaller datasets which are balanced across the similarity spectra through sub-sampling.
Monolingual Evaluation of Representation Learning Models
After the numerical and qualitative analyses of the Multi-SimLex datasets provided in §§ SECREF18–SECREF25, we now benchmark a series of representation learning models on the new evaluation data. We evaluate standard static word embedding algorithms such as fastText BIBREF31, as well as a range of more recent text encoders pretrained on language modeling such as multilingual BERT BIBREF29. These experiments provide strong baseline scores on the new Multi-SimLex datasets and offer a first large-scale analysis of pretrained encoders on word-level semantic similarity across diverse languages. In addition, the experiments now enabled by Multi-SimLex aim to answer several important questions. (Q1) Is it viable to extract high-quality word-level representations from pretrained encoders receiving subword-level tokens as input? Are such representations competitive with standard static word-level embeddings? (Q2) What are the implications of monolingual pretraining versus (massively) multilingual pretraining for performance? (Q3) Do lightweight unsupervised post-processing techniques improve word representations consistently across different languages? (Q4) Can we effectively transfer available external lexical knowledge from resource-rich languages to resource-lean languages in order to learn word representations that distinguish between true similarity and conceptual relatedness (see the discussion in §SECREF6)?
Monolingual Evaluation of Representation Learning Models ::: Models in Comparison
Static Word Embeddings in Different Languages. First, we evaluate a standard method for inducing non-contextualized (i.e., static) word embeddings across a plethora of different languages: fastText (ft) vectors BIBREF31 are currently the most popular and robust choice given 1) the availability of pretrained vectors in a large number of languages BIBREF108 trained on large Common Crawl (CC) plus Wikipedia (Wiki) data, and 2) their superior performance across a range of NLP tasks BIBREF109. In fact, fastText is an extension of the standard word-level CBOW and skip-gram word2vec models BIBREF32 that takes into account subword-level information, i.e. the contituent character n-grams of each word BIBREF110. For this reason, fastText is also more suited for modeling rare words and morphologically rich languages.
We rely on 300-dimensional ft word vectors trained on CC+Wiki and available online for 157 languages. The word vectors for all languages are obtained by CBOW with position-weights, with character n-grams of length 5, a window of size 5, 10 negative examples, and 10 training epochs. We also probe another (older) collection of ft vectors, pretrained on full Wikipedia dumps of each language.. The vectors are 300-dimensional, trained with the skip-gram objective for 5 epochs, with 5 negative examples, a window size set to 5, and relying on all character n-grams from length 3 to 6. Following prior work, we trim the vocabularies for all languages to the 200K most frequent words and compute representations for multi-word expressions by averaging the vectors of their constituent words.
Unsupervised Post-Processing. Further, we consider a variety of unsupervised post-processing steps that can be applied post-training on top of any pretrained input word embedding space without any external lexical semantic resource. So far, the usefulness of such methods has been verified only on the English language through benchmarks for lexical semantics and sentence-level tasks BIBREF113. In this paper, we assess if unsupervised post-processing is beneficial also in other languages. To this end, we apply the following post-hoc transformations on the initial word embeddings:
1) Mean centering (mc) is applied after unit length normalization to ensure that all vectors have a zero mean, and is commonly applied in data mining and analysis BIBREF114, BIBREF115.
2) All-but-the-top (abtt) BIBREF113, BIBREF116 eliminates the common mean vector and a few top dominating directions (according to PCA) from the input distributional word vectors, since they do not contribute towards distinguishing the actual semantic meaning of different words. The method contains a single (tunable) hyper-parameter $dd_{A}$ which denotes the number of the dominating directions to remove from the initial representations. Previous work has verified the usefulness of abtt in several English lexical semantic tasks such as semantic similarity, word analogies, and concept categorization, as well as in sentence-level text classification tasks BIBREF113.
3) uncovec BIBREF117 adjusts the similarity order of an arbitrary input word embedding space, and can emphasize either syntactic or semantic information in the transformed vectors. In short, it transforms the input space $\mathbf {X}$ into an adjusted space $\mathbf {X}\mathbf {W}_{\alpha }$ through a linear map $\mathbf {W}_{\alpha }$ controlled by a single hyper-parameter $\alpha $. The $n^{\text{th}}$-order similarity transformation of the input word vector space $\mathbf {X}$ (for which $n=1$) can be obtained as $\mathbf {M}_{n}(\mathbf {X}) = \mathbf {M}_1(\mathbf {X}\mathbf {W}_{(n-1)/2})$, with $\mathbf {W}_{\alpha }=\mathbf {Q}\mathbf {\Gamma }^{\alpha }$, where $\mathbf {Q}$ and $\mathbf {\Gamma }$ are the matrices obtained via eigendecomposition of $\mathbf {X}^T\mathbf {X}=\mathbf {Q}\mathbf {\Gamma }\mathbf {Q}^T$. $\mathbf {\Gamma }$ is a diagonal matrix containing eigenvalues of $\mathbf {X}^T\mathbf {X}$; $\mathbf {Q}$ is an orthogonal matrix with eigenvectors of $\mathbf {X}^T\mathbf {X}$ as columns. While the motivation for the uncovec methods does originate from adjusting discrete similarity orders, note that $\alpha $ is in fact a continuous real-valued hyper-parameter which can be carefully tuned. For more technical details we refer the reader to the original work of BIBREF117.
As mentioned, all post-processing methods can be seen as unsupervised retrofitting methods that, given an arbitrary input vector space $\mathbf {X}$, produce a perturbed/transformed output vector space $\mathbf {X}^{\prime }$, but unlike common retrofitting methods BIBREF62, BIBREF16, the perturbation is completely unsupervised (i.e., self-contained) and does not inject any external (semantic similarity oriented) knowledge into the vector space. Note that different perturbations can also be stacked: e.g., we can apply uncovec and then use abtt on top the output uncovec vectors. When using uncovec and abtt we always length-normalize and mean-center the data first (i.e., we apply the simple mc normalization). Finally, we tune the two hyper-parameters $d_A$ (for abtt) and $\alpha $ (uncovec) on the English Multi-SimLex and use the same values on the datasets of all other languages; we report results with $dd_A = 3$ or $dd_A = 10$, and $\alpha =-0.3$.
Contextualized Word Embeddings. We also evaluate the capacity of unsupervised pretraining architectures based on language modeling objectives to reason over lexical semantic similarity. To the best of our knowledge, our article is the first study performing such analyses. State-of-the-art models such as bert BIBREF29, xlm BIBREF30, or roberta BIBREF118 are typically very deep neural networks based on the Transformer architecture BIBREF119. They receive subword-level tokens as inputs (such as WordPieces BIBREF120) to tackle data sparsity. In output, they return contextualized embeddings, dynamic representations for words in context.
To represent words or multi-word expressions through a pretrained model, we follow prior work BIBREF121 and compute an input item's representation by 1) feeding it to a pretrained model in isolation; then 2) averaging the $H$ last hidden representations for each of the item’s constituent subwords; and then finally 3) averaging the resulting subword representations to produce the final $d$-dimensional representation, where $d$ is the embedding and hidden-layer dimensionality (e.g., $d=768$ with bert). We opt for this approach due to its proven viability and simplicity BIBREF121, as it does not require any additional corpora to condition the induction of contextualized embeddings. Other ways to extract the representations from pretrained models BIBREF122, BIBREF123, BIBREF124 are beyond the scope of this work, and we will experiment with them in the future.
In other words, we treat each pretrained encoder enc as a black-box function to encode a single word or a multi-word expression $x$ in each language into a $d$-dimensional contextualized representation $\mathbf {x}_{\textsc {enc}} \in \mathbb {R}^d = \textsc {enc}(x)$ (e.g., $d=768$ with bert). As multilingual pretrained encoders, we experiment with the multilingual bert model (m-bert) BIBREF29 and xlm BIBREF30. m-bert is pretrained on monolingual Wikipedia corpora of 102 languages (comprising all Multi-SimLex languages) with a 12-layer Transformer network, and yields 768-dimensional representations. Since the concept pairs in Multi-SimLex are lowercased, we use the uncased version of m-bert. m-bert comprises all Multi-SimLex languages, and its evident ability to perform cross-lingual transfer BIBREF12, BIBREF125, BIBREF126 also makes it a convenient baseline model for cross-lingual experiments later in §SECREF8. The second multilingual model we consider, xlm-100, is pretrained on Wikipedia dumps of 100 languages, and encodes each concept into a $1,280$-dimensional representation. In contrast to m-bert, xlm-100 drops the next-sentence prediction objective and adds a cross-lingual masked language modeling objective. For both encoders, the representations of each concept are computed as averages over the last $H=4$ hidden layers in all experiments, as suggested by Wu:2019arxiv.
Besides m-bert and xlm, covering multiple languages, we also analyze the performance of “language-specific” bert and xlm models for the languages where they are available: Finnish, Spanish, English, Mandarin Chinese, and French. The main goal of this comparison is to study the differences in performance between multilingual “one-size-fits-all” encoders and language-specific encoders. For all experiments, we rely on the pretrained models released in the Transformers repository BIBREF127.
Unsupervised post-processing steps devised for static word embeddings (i.e., mean-centering, abtt, uncovec) can also be applied on top of contextualized embeddings if we predefine a vocabulary of word types $V$ that will be represented in a word vector space $\mathbf {X}$. We construct such $V$ for each language as the intersection of word types covered by the corresponding CC+Wiki fastText vectors and the (single-word or multi-word) expressions appearing in the corresponding Multi-SimLex dataset.
Finally, note that it is not feasible to evaluate a full range of available pretrained encoders within the scope of this work. Our main intention is to provide the first set of baseline results on Multi-SimLex by benchmarking a sample of most popular encoders, at the same time also investigating other important questions such as performance of static versus contextualized word embeddings, or multilingual versus language-specific pretraining. Another purpose of the experiments is to outline the wide potential and applicability of the Multi-SimLex datasets for multilingual and cross-lingual representation learning evaluation.
Monolingual Evaluation of Representation Learning Models ::: Results and Discussion
The results we report are Spearman's $\rho $ coefficients of the correlation between the ranks derived from the scores of the evaluated models and the human scores provided in each Multi-SimLex dataset. The main results with static and contextualized word vectors for all test languages are summarized in Table TABREF43. The scores reveal several interesting patterns, and also pinpoint the main challenges for future work.
State-of-the-Art Representation Models. The absolute scores of CC+Wiki ft, Wiki ft, and m-bert are not directly comparable, because these models have different coverage. In particular, Multi-SimLex contains some out-of-vocabulary (OOV) words whose static ft embeddings are not available. On the other hand, m-bert has perfect coverage. A general comparison between CC+Wiki and Wiki ft vectors, however, supports the intuition that larger corpora (such as CC+Wiki) yield higher correlations. Another finding is that a single massively multilingual model such as m-bert cannot produce semantically rich word-level representations. Whether this actually happens because the training objective is different—or because the need to represent 100+ languages reduces its language-specific capacity—is investigated further below.
The overall results also clearly indicate that (i) there are differences in performance across different monolingual Multi-SimLex datasets, and (ii) unsupervised post-processing is universally useful, and can lead to huge improvements in correlation scores for many languages. In what follows, we also delve deeper into these analyses.
Impact of Unsupervised Post-Processing. First, the results in Table TABREF43 suggest that applying dimension-wise mean centering to the initial vector spaces has positive impact on word similarity scores in all test languages and for all models, both static and contextualized (see the +mc rows in Table TABREF43). BIBREF128 show that distributional word vectors have a tendency towards narrow clusters in the vector space (i.e., they occupy a narrow cone in the vector space and are therefore anisotropic BIBREF113, BIBREF129), and are prone to the undesired effect of hubness BIBREF130, BIBREF131. Applying dimension-wise mean centering has the effect of spreading the vectors across the hyper-plane and mitigating the hubness issue, which consequently improves word-level similarity, as it emerges from the reported results. Previous work has already validated the importance of mean centering for clustering-based tasks BIBREF132, bilingual lexicon induction with cross-lingual word embeddings BIBREF133, BIBREF134, BIBREF135, and for modeling lexical semantic change BIBREF136. However, to the best of our knowledge, the results summarized in Table TABREF43 are the first evidence that also confirms its importance for semantic similarity in a wide array of languages. In sum, as a general rule of thumb, we suggest to always mean-center representations for semantic tasks.
The results further indicate that additional post-processing methods such as abtt and uncovec on top of mean-centered vector spaces can lead to further gains in most languages. The gains are even visible for languages which start from high correlation scores: for instance., cmn with CC+Wiki ft increases from 0.534 to 0.583, from 0.315 to 0.526 with Wiki ft, and from 0.408 to 0.487 with m-bert. Similarly, for rus with CC+Wiki ft we can improve from 0.422 to 0.500, and for fra the scores improve from 0.578 to 0.613. There are additional similar cases reported in Table TABREF43.
Overall, the unsupervised post-processing techniques seem universally useful across languages, but their efficacy and relative performance does vary across different languages. Note that we have not carefully fine-tuned the hyper-parameters of the evaluated post-processing methods, so additional small improvements can be expected for some languages. The main finding, however, is that these post-processing techniques are robust to semantic similarity computations beyond English, and are truly language independent. For instance, removing dominant latent (PCA-based) components from word vectors emphasizes semantic differences between different concepts, as only shared non-informative latent semantic knowledge is removed from the representations.
In summary, pretrained word embeddings do contain more information pertaining to semantic similarity than revealed in the initial vectors. This way, we have corroborated the hypotheses from prior work BIBREF113, BIBREF117 which were not previously empirically verified on other languages due to a shortage of evaluation data; this gap has now been filled with the introduction of the Multi-SimLex datasets. In all follow-up experiments, we always explicitly denote which post-processing configuration is used in evaluation.
POS-Specific Subsets. We present the results for subsets of word pairs grouped by POS class in Table TABREF46. Prior work based on English data showed that representations for nouns are typically of higher quality than those for the other POS classes BIBREF49, BIBREF137, BIBREF50. We observe a similar trend in other languages as well. This pattern is consistent across different representation models and can be attributed to several reasons. First, verb representations need to express a rich range of syntactic and semantic behaviors rather than purely referential features BIBREF138, BIBREF139, BIBREF73. Second, low correlation scores on the adjective and adverb subsets in some languages (e.g., pol, cym, swa) might be due to their low frequency in monolingual texts, which yields unreliable representations. In general, the variance in performance across different word classes warrants further research in class-specific representation learning BIBREF140, BIBREF50. The scores further attest the usefulness of unsupervised post-processing as almost all class-specific correlation scores are improved by applying mean-centering and abtt. Finally, the results for m-bert and xlm-100 in Table TABREF46 further confirm that massively multilingual pretraining cannot yield reasonable semantic representations for many languages: in fact, for some classes they display no correlation with human ratings at all.
Differences across Languages. Naturally, the results from Tables TABREF43 and TABREF46 also reveal that there is variation in performance of both static word embeddings and pretrained encoders across different languages. Among other causes, the lowest absolute scores with ft are reported for languages with least resources available to train monolingual word embeddings, such as Kiswahili, Welsh, and Estonian. The low performance on Welsh is especially indicative: Figure FIGREF20 shows that the ratings in the Welsh dataset match up very well with the English ratings, but we cannot achieve the same level of correlation in Welsh with Welsh ft word embeddings. Difference in performance between two closely related languages, est (low-resource) and fin (high-resource), provides additional evidence in this respect.
The highest reported scores with m-bert and xlm-100 are obtained for Mandarin Chinese and Yue Chinese: this effectively points to the weaknesses of massively multilingual training with a joint subword vocabulary spanning 102 and 100 languages. Due to the difference in scripts, “language-specific” subwords for yue and cmn do not need to be shared across a vast amount of languages and the quality of their representation remains unscathed. This effectively means that m-bert's subword vocabulary contains plenty of cmn-specific and yue-specific subwords which are exploited by the encoder when producing m-bert-based representations. Simultaneously, higher scores with m-bert (and xlm in Table TABREF46) are reported for resource-rich languages such as French, Spanish, and English, which are better represented in m-bert's training data. We also observe lower absolute scores (and a larger number of OOVs) for languages with very rich and productive morphological systems such as the two Slavic languages (Polish and Russian) and Finnish. Since Polish and Russian are known to have large Wikipedias and Common Crawl data BIBREF33 (e.g., their Wikipedias are in the top 10 largest Wikipedias worldwide), the problem with coverage can be attributed exactly to the proliferation of morphological forms in those languages.
Finally, while Table TABREF43 does reveal that unsupervised post-processing is useful for all languages, it also demonstrates that peak scores are achieved with different post-processing configurations. This finding suggests that a more careful language-specific fine-tuning is indeed needed to refine word embeddings towards semantic similarity. We plan to inspect the relationship between post-processing techniques and linguistic properties in more depth in future work.
Multilingual vs. Language-Specific Contextualized Embeddings. Recent work has shown that—despite the usefulness of massively multilingual models such as m-bert and xlm-100 for zero-shot cross-lingual transfer BIBREF12, BIBREF125—stronger results in downstream tasks for a particular language can be achieved by pretraining language-specific models on language-specific data. In this experiment, motivated by the low results of m-bert and xlm-100 (see again Table TABREF46), we assess if monolingual pretrained encoders can produce higher-quality word-level representations than multilingual models. Therefore, we evaluate language-specific bert and xlm models for a subset of the Multi-SimLex languages for which such models are currently available: Finnish BIBREF141 (bert-base architecture, uncased), French BIBREF142 (the FlauBERT model based on xlm), English (bert-base, uncased), Mandarin Chinese (bert-base) BIBREF29 and Spanish (bert-base, uncased). In addition, we also evaluate a series of pretrained encoders available for English: (i) bert-base, bert-large, and bert-large with whole word masking (wwm) from the original work on BERT BIBREF29, (ii) monolingual “English-specific” xlm BIBREF30, and (iii) two models which employ parameter reduction techniques to build more compact encoders: albert-b uses a configuration similar to bert-base, while albert-l is similar to bert-large, but with an $18\times $ reduction in the number of parameters BIBREF143.
From the results in Table FIGREF49, it is clear that monolingual pretrained encoders yield much more reliable word-level representations. The gains are visible even for languages such as cmn which showed reasonable performance with m-bert and are substantial on all test languages. This further confirms the validity of language-specific pretraining in lieu of multilingual training, if sufficient monolingual data are available. Moreover, a comparison of pretrained English encoders in Figure FIGREF49 largely follows the intuition: the larger bert-large model yields slight improvements over bert-base, and we can improve a bit more by relying on word-level (i.e., lexical-level) masking.Finally, light-weight albert model variants are quite competitive with the original bert models, with only modest drops reported, and albert-l again outperforms albert-b. Overall, it is interesting to note that the scores obtained with monolingual pretrained encoders are on a par with or even outperform static ft word embeddings: this is a very intriguing finding per se as it shows that such subword-level models trained on large corpora can implicitly capture rich lexical semantic knowledge.
Similarity-Specialized Word Embeddings. Conflating distinct lexico-semantic relations is a well-known property of distributional representations BIBREF144, BIBREF53. Semantic specialization fine-tunes distributional spaces to emphasize a particular lexico-semantic relation in the transformed space by injecting external lexical knowledge BIBREF145. Explicitly discerning between true semantic similarity (as captured in Multi-SimLex) and broad conceptual relatedness benefits a number of tasks, as discussed in §SECREF4. Since most languages lack dedicated lexical resources, however, one viable strategy to steer monolingual word vector spaces to emphasize semantic similarity is through cross-lingual transfer of lexical knowledge, usually through a shared cross-lingual word vector space BIBREF106. Therefore, we evaluate the effectiveness of specialization transfer methods using Multi-SimLex as our multilingual test bed.
We evaluate a current state-of-the-art cross-lingual specialization transfer method with minimal requirements, put forth recently by Ponti:2019emnlp. In a nutshell, their li-postspec method is a multi-step procedure that operates as follows. First, the knowledge about semantic similarity is extracted from WordNet in the form of triplets, that is, linguistic constraints $(w_1, w_2, r)$, where $w_1$ and $w_2$ are two concepts, and $r$ is a relation between them obtained from WordNet (e.g., synonymy or antonymy). The goal is to “attract” synonyms closer to each other in the transformed vector space as they reflect true semantic similarity, and “repel” antonyms further apart. In the second step, the linguistic constraints are translated from English to the target language via a shared cross-lingual word vector space. To this end, following Ponti:2019emnlp we rely on cross-lingual word embeddings (CLWEs) BIBREF146 available online, which are based on Wiki ft vectors. Following that, a constraint refinement step is applied in the target language which aims to eliminate the noise inserted during the translation process. This is done by training a relation classification tool: it is trained again on the English linguistic constraints and then used on the translated target language constraints, where the transfer is again enabled via a shared cross-lingual word vector space. Finally, a state-of-the-art monolingual specialization procedure from Ponti:2018emnlp injects the (now target language) linguistic constraints into the target language distributional space.
The scores are summarized in Table TABREF56. Semantic specialization with li-postspec leads to substantial improvements in correlation scores for the majority of the target languages, demonstrating the importance of external semantic similarity knowledge for semantic similarity reasoning. However, we also observe deteriorated performance for the three target languages which can be considered the lowest-resource ones in our set: cym, swa, yue. We hypothesize that this occurs due to the inferior quality of the underlying monolingual Wikipedia word embeddings, which generates a chain of error accumulations. In particular, poor distributional word estimates compromise the alignment of the embedding spaces, which in turn results in increased translation noise, and reduced refinement ability of the relation classifier. On a high level, this “poor get poorer” observation again points to the fact that one of the primary causes of low performance of resource-low languages in semantic tasks is the sheer lack of even unlabeled data for distributional training. On the other hand, as we see from Table TABREF46, typological dissimilarity between the source and the target does not deteriorate the effectiveness of semantic specialization. In fact, li-postspec does yield substantial gains also for the typologically distant targets such as heb, cmn, and est. The critical problem indeed seems to be insufficient raw data for monolingual distributional training.
Cross-Lingual Evaluation
Similar to monolingual evaluation in §SECREF7, we now evaluate several state-of-the-art cross-lingual representation models on the suite of 66 automatically constructed cross-lingual Multi-SimLex datasets. Again, note that evaluating a full range of cross-lingual models available in the rich prior work on cross-lingual representation learning is well beyond the scope of this article. We therefore focus our cross-lingual analyses on several well-established and indicative state-of-the-art cross-lingual models, again spanning both static and contextualized cross-lingual word embeddings.
Cross-Lingual Evaluation ::: Models in Comparison
Static Word Embeddings. We rely on a state-of-the-art mapping-based method for the induction of cross-lingual word embeddings (CLWEs): vecmap BIBREF148. The core idea behind such mapping-based or projection-based approaches is to learn a post-hoc alignment of independently trained monolingual word embeddings BIBREF106. Such methods have gained popularity due to their conceptual simplicity and competitive performance coupled with reduced bilingual supervision requirements: they support CLWE induction with only as much as a few thousand word translation pairs as the bilingual supervision BIBREF149, BIBREF150, BIBREF151, BIBREF107. More recent work has shown that CLWEs can be induced with even weaker supervision from small dictionaries spanning several hundred pairs BIBREF152, BIBREF135, identical strings BIBREF153, or even only shared numerals BIBREF154. In the extreme, fully unsupervised projection-based CLWEs extract such seed bilingual lexicons from scratch on the basis of monolingual data only BIBREF103, BIBREF148, BIBREF155, BIBREF156, BIBREF104, BIBREF157.
Recent empirical studies BIBREF158, BIBREF135, BIBREF159 have compared a variety of unsupervised and weakly supervised mapping-based CLWE methods, and vecmap emerged as the most robust and very competitive choice. Therefore, we focus on 1) its fully unsupervised variant (unsuper) in our comparisons. For several language pairs, we also report scores with two other vecmap model variants: 2) a supervised variant which learns a mapping based on an available seed lexicon (super), and 3) a supervised variant with self-learning (super+sl) which iteratively increases the seed lexicon and improves the mapping gradually. For a detailed description of these variants, we refer the reader to recent work BIBREF148, BIBREF135. We again use CC+Wiki ft vectors as initial monolingual word vectors, except for yue where Wiki ft is used. The seed dictionaries of two different sizes (1k and 5k translation pairs) are based on PanLex BIBREF160, and are taken directly from prior work BIBREF135, or extracted from PanLex following the same procedure as in the prior work.
Contextualized Cross-Lingual Word Embeddings. We again evaluate the capacity of (massively) multilingual pretrained language models, m-bert and xlm-100, to reason over cross-lingual lexical similarity. Implicitly, such an evaluation also evaluates “the intrinsic quality” of shared cross-lingual word-level vector spaces induced by these methods, and their ability to boost cross-lingual transfer between different language pairs. We rely on the same procedure of aggregating the models' subword-level parameters into word-level representations, already described in §SECREF34.
As in monolingual settings, we can apply unsupervised post-processing steps such as abtt to both static and contextualized cross-lingual word embeddings.
Cross-Lingual Evaluation ::: Results and Discussion
Main Results and Differences across Language Pairs. A summary of the results on the 66 cross-lingual Multi-SimLex datasets are provided in Table TABREF59 and Figure FIGREF60. The findings confirm several interesting findings from our previous monolingual experiments (§SECREF44), and also corroborate several hypotheses and findings from prior work, now on a large sample of language pairs and for the task of cross-lingual semantic similarity.
First, we observe that the fully unsupervised vecmap model, despite being the most robust fully unsupervised method at present, fails to produce a meaningful cross-lingual word vector space for a large number of language pairs (see the bottom triangle of Table TABREF59): many correlation scores are in fact no-correlation results, accentuating the problem of fully unsupervised cross-lingual learning for typologically diverse languages and with fewer amounts of monolingual data BIBREF135. The scores are particularly low across the board for lower-resource languages such as Welsh and Kiswahili. It also seems that the lack of monolingual data is a larger problem than typological dissimilarity between language pairs, as we do observe reasonably high correlation scores with vecmap for language pairs such as cmn-spa, heb-est, and rus-fin. However, typological differences (e.g., morphological richness) still play an important role as we observe very low scores when pairing cmn with morphologically rich languages such fin, est, pol, and rus. Similar to prior work of Vulic:2019we and doval2019onthe, given the fact that unsupervised vecmap is the most robust unsupervised CLWE method at present BIBREF158, our results again question the usefulness of fully unsupervised approaches for a large number of languages, and call for further developments in the area of unsupervised and weakly supervised cross-lingual representation learning.
The scores of m-bert and xlm-100 lead to similar conclusions as in the monolingual settings. Reasonable correlation scores are achieved only for a small subset of resource-rich language pairs (e.g., eng, fra, spa, cmn) which dominate the multilingual m-bert training. Interestingly, the scores indicate a much higher performance of language pairs where yue is one of the languages when we use m-bert instead of vecmap. This boils down again to the fact that yue, due to its specific language script, has a good representation of its words and subwords in the shared m-bert vocabulary. At the same time, a reliable vecmap mapping between yue and other languages cannot be found due to a small monolingual yue corpus. In cases when vecmap does not yield a degenerate cross-lingual vector space starting from two monolingual ones, the final correlation scores seem substantially higher than the ones obtained by the single massively multilingual m-bert model.
Finally, the results in Figure FIGREF60 again verify the usefulness of unsupervised post-processing also in cross-lingual settings. We observe improved performance with both m-bert and xlm-100 when mean centering (+mc) is applied, and further gains can be achieved by using abtt on the mean-centered vector spaces. A similar finding also holds for static cross-lingual word embeddings, where applying abbt (-10) yields higher scores on 61/66 language pairs.
Fully Unsupervised vs. Weakly Supervised Cross-Lingual Embeddings. The results in Table TABREF59 indicate that fully unsupervised cross-lingual learning fails for a large number of language pairs. However, recent work BIBREF135 has noted that these sub-optimal non-alignment solutions with the unsuper model can be avoided by relying on (weak) cross-lingual supervision spanning only several thousands or even hundreds of word translation pairs. Therefore, we examine 1) if we can further improve the results on cross-lingual Multi-SimLex resorting to (at least some) cross-lingual supervision for resource-rich language pairs; and 2) if such available word-level supervision can also be useful for a range of languages which displayed near-zero performance in Table TABREF59. In other words, we test if recent “tricks of the trade” used in the rich literature on CLWE learning reflect in gains on cross-lingual Multi-SimLex datasets.
First, we reassess the findings established on the bilingual lexicon induction task BIBREF161, BIBREF135: using at least some cross-lingual supervision is always beneficial compared to using no supervision at all. We report improvements over the unsuper model for all 10 language pairs in Table TABREF66, even though the unsuper method initially produced strong correlation scores. The importance of self-learning increases with decreasing available seed dictionary size, and the +sl model always outperforms unsuper with 1k seed pairs; we observe the same patterns also with even smaller dictionary sizes than reported in Table TABREF66 (250 and 500 seed pairs). Along the same line, the results in Table TABREF67 indicate that at least some supervision is crucial for the success of static CLWEs on resource-leaner language pairs. We note substantial improvements on all language pairs; in fact, the vecmap model is able to learn a more reliable mapping starting from clean supervision. We again note large gains with self-learning.
Multilingual vs. Bilingual Contextualized Embeddings. Similar to the monolingual settings, we also inspect if massively multilingual training in fact dilutes the knowledge necessary for cross-lingual reasoning on a particular language pair. Therefore, we compare the 100-language xlm-100 model with i) a variant of the same model trained on a smaller set of 17 languages (xlm-17); ii) a variant of the same model trained specifically for the particular language pair (xlm-2); and iii) a variant of the bilingual xlm-2 model that also leverages bilingual knowledge from parallel data during joint training (xlm-2++). We again use the pretrained models made available by Conneau:2019nips, and we refer to the original work for further technical details.
The results are summarized in Figure FIGREF60, and they confirm the intuition that massively multilingual pretraining can damage performance even on resource-rich languages and language pairs. We observe a steep rise in performance when the multilingual model is trained on a much smaller set of languages (17 versus 100), and further improvements can be achieved by training a dedicated bilingual model. Finally, leveraging bilingual parallel data seems to offer additional slight gains, but a tiny difference between xlm-2 and xlm-2++ also suggests that this rich bilingual information is not used in the optimal way within the xlm architecture for semantic similarity.
In summary, these results indicate that, in order to improve performance in cross-lingual transfer tasks, more work should be invested into 1) pretraining dedicated language pair-specific models, and 2) creative ways of leveraging available cross-lingual supervision (e.g., word translation pairs, parallel or comparable corpora) BIBREF121, BIBREF123, BIBREF124 with pretraining paradigms such as bert and xlm. Using such cross-lingual supervision could lead to similar benefits as indicated by the results obtained with static cross-lingual word embeddings (see Table TABREF66 and Table TABREF67). We believe that Multi-SimLex can serve as a valuable means to track and guide future progress in this research area.
Conclusion and Future Work
We have presented Multi-SimLex, a resource containing human judgments on the semantic similarity of word pairs for 12 monolingual and 66 cross-lingual datasets. The languages covered are typologically diverse and include also under-resourced ones, such as Welsh and Kiswahili. The resource covers an unprecedented amount of 1,888 word pairs, carefully balanced according to their similarity score, frequency, concreteness, part-of-speech class, and lexical field. In addition to Multi-Simlex, we release the detailed protocol we followed to create this resource. We hope that our consistent guidelines will encourage researchers to translate and annotate Multi-Simlex -style datasets for additional languages. This can help and create a hugely valuable, large-scale semantic resource for multilingual NLP research.
The core Multi-SimLex we release with this paper already enables researchers to carry out novel linguistic analysis as well as establishes a benchmark for evaluating representation learning models. Based on our preliminary analyses, we found that speakers of closely related languages tend to express equivalent similarity judgments. In particular, geographical proximity seems to play a greater role than family membership in determining the similarity of judgments across languages. Moreover, we tested several state-of-the-art word embedding models, both static and contextualized representations, as well as several (supervised and unsupervised) post-processing techniques, on the newly released Multi-SimLex. This enables future endeavors to improve multilingual representation learning with challenging baselines. In addition, our results provide several important insights for research on both monolingual and cross-lingual word representations:
1) Unsupervised post-processing techniques (mean centering, elimination of top principal components, adjusting similarity orders) are always beneficial independently of the language, although the combination leading to the best scores is language-specific and hence needs to be tuned.
2) Similarity rankings obtained from word embeddings for nouns are better aligned with human judgments than all the other part-of-speech classes considered here (verbs, adjectives, and, for the first time, adverbs). This confirms previous generalizations based on experiments on English.
3) The factor having the greatest impact on the quality of word representations is the availability of raw texts to train them in the first place, rather than language properties (such as family, geographical area, typological features).
4) Massively multilingual pretrained encoders such as m-bert BIBREF29 and xlm-100 BIBREF30 fare quite poorly on our benchmark, whereas pretrained encoders dedicated to a single language are more competitive with static word embeddings such as fastText BIBREF31. Moreover, for language-specific encoders, parameter reduction techniques reduce performance only marginally.
5) Techniques to inject clean lexical semantic knowledge from external resources into distributional word representations were proven to be effective in emphasizing the relation of semantic similarity. In particular, methods capable of transferring such knowledge from resource-rich to resource-lean languages BIBREF59 increased the correlation with human judgments for most languages, except for those with limited unlabelled data.
Future work can expand our preliminary, yet large-scale study on the ability of pretrained encoders to reason over word-level semantic similarity in different languages. For instance, we have highlighted how sharing the same encoder parameters across multiple languages may harm performance. However, it remains unclear if, and to what extent, the input language embeddings present in xlm-100 but absent in m-bert help mitigate this issue. In addition, pretrained language embeddings can be obtained both from typological databases BIBREF99 and from neural architectures BIBREF100. Plugging these embeddings into the encoders in lieu of embeddings trained end-to-end as suggested by prior work BIBREF162, BIBREF163, BIBREF164 might extend the coverage to more resource-lean languages.
Another important follow-up analysis might involve the comparison of the performance of representation learning models on multilingual datasets for both word-level semantic similarity and sentence-level Natural Language Understanding. In particular, Multi-SimLex fills a gap in available resources for multilingual NLP and might help understand how lexical and compositional semantics interact if put alongside existing resources such as XNLI BIBREF84 for natural language inference or PAWS-X BIBREF89 for cross-lingual paraphrase identification. Finally, the Multi-SimLex annotation could turn out to be a unique source of evidence to study the effects of polysemy in human judgments on semantic similarity: for equivalent word pairs in multiple languages, are the similarity scores affected by how many senses the two words (or multi-word expressions) incorporate?
In light of the success of initiatives like Universal Dependencies for multilingual treebanks, we hope that making Multi-SimLex and its guidelines available will encourage other researchers to expand our current sample of languages. We particularly encourage creation and submission of comparable Multi-SimLex datasets for under-resourced and typologically diverse languages in future work. In particular, we have made a Multi-Simlex community website available to facilitate easy creation, gathering, dissemination, and use of annotated datasets: https://multisimlex.com/.
This work is supported by the ERC Consolidator Grant LEXICAL: Lexical Acquisition Across Languages (no 648909). Thierry Poibeau is partly supported by a PRAIRIE 3IA Institute fellowship ("Investissements d'avenir" program, reference ANR-19-P3IA-0001).
| [
"Chinese Mandarin, Welsh, English, Estonian, Finnish, French, Hebrew, Polish, Russian, Spanish, Kiswahili, Yue Chinese",
"Chinese Mandarin, Welsh, English, Estonian, Finnish, French, Hebrew, Polish, Russian, Spanish, Kiswahili, Yue Chinese"
] | 14,660 | qasper | en | null | 388c857ed1b4175e114a0ce9f3489766797a32cd183f10ff | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What are the 12 languages covered?
Answer: | [
"Chinese Mandarin, Welsh, English, Estonian, Finnish, French, Hebrew, Polish, Russian, Spanish, Kiswahili, Yue Chinese",
"Chinese Mandarin, Welsh, English, Estonian, Finnish, French, Hebrew, Polish, Russian, Spanish, Kiswahili, Yue Chinese"
] | qasper | 128 | 57 | |
What are two datasets model is applied to? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
“Ché saetta previsa vien più lenta.”
– Dante Alighieri, Divina Commedia, Paradiso
Antisocial behavior is a persistent problem plaguing online conversation platforms; it is both widespread BIBREF0 and potentially damaging to mental and emotional health BIBREF1, BIBREF2. The strain this phenomenon puts on community maintainers has sparked recent interest in computational approaches for assisting human moderators.
Prior work in this direction has largely focused on post-hoc identification of various kinds of antisocial behavior, including hate speech BIBREF3, BIBREF4, harassment BIBREF5, personal attacks BIBREF6, and general toxicity BIBREF7. The fact that these approaches only identify antisocial content after the fact limits their practicality as tools for assisting pre-emptive moderation in conversational domains.
Addressing this limitation requires forecasting the future derailment of a conversation based on early warning signs, giving the moderators time to potentially intervene before any harm is done (BIBREF8 BIBREF8, BIBREF9 BIBREF9, see BIBREF10 BIBREF10 for a discussion). Such a goal recognizes derailment as emerging from the development of the conversation, and belongs to the broader area of conversational forecasting, which includes future-prediction tasks such as predicting the eventual length of a conversation BIBREF11, whether a persuasion attempt will eventually succeed BIBREF12, BIBREF13, BIBREF14, whether team discussions will eventually lead to an increase in performance BIBREF15, or whether ongoing counseling conversations will eventually be perceived as helpful BIBREF16.
Approaching such conversational forecasting problems, however, requires overcoming several inherent modeling challenges. First, conversations are dynamic and their outcome might depend on how subsequent comments interact with each other. Consider the example in Figure FIGREF2: while no individual comment is outright offensive, a human reader can sense a tension emerging from their succession (e.g., dismissive answers to repeated questioning). Thus a forecasting model needs to capture not only the content of each individual comment, but also the relations between comments. Previous work has largely relied on hand-crafted features to capture such relations—e.g., similarity between comments BIBREF16, BIBREF12 or conversation structure BIBREF17, BIBREF18—, though neural attention architectures have also recently shown promise BIBREF19.
The second modeling challenge stems from the fact that conversations have an unknown horizon: they can be of varying lengths, and the to-be-forecasted event can occur at any time. So when is it a good time to make a forecast? Prior work has largely proposed two solutions, both resulting in important practical limitations. One solution is to assume (unrealistic) prior knowledge of when the to-be-forecasted event takes place and extract features up to that point BIBREF20, BIBREF8. Another compromising solution is to extract features from a fixed-length window, often at the start of the conversation BIBREF21, BIBREF15, BIBREF16, BIBREF9. Choosing a catch-all window-size is however impractical: short windows will miss information in comments they do not encompass (e.g., a window of only two comments would miss the chain of repeated questioning in comments 3 through 6 of Figure FIGREF2), while longer windows risk missing the to-be-forecasted event altogether if it occurs before the end of the window, which would prevent early detection.
In this work we introduce a model for forecasting conversational events that overcomes both these inherent challenges by processing comments, and their relations, as they happen (i.e., in an online fashion). Our main insight is that models with these properties already exist, albeit geared toward generation rather than prediction: recent work in context-aware dialog generation (or “chatbots”) has proposed sequential neural models that make effective use of the intra-conversational dynamics BIBREF22, BIBREF23, BIBREF24, while concomitantly being able to process the conversation as it develops (see BIBREF25 for a survey).
In order for these systems to perform well in the generative domain they need to be trained on massive amounts of (unlabeled) conversational data. The main difficulty in directly adapting these models to the supervised domain of conversational forecasting is the relative scarcity of labeled data: for most forecasting tasks, at most a few thousands labeled examples are available, insufficient for the notoriously data-hungry sequential neural models.
To overcome this difficulty, we propose to decouple the objective of learning a neural representation of conversational dynamics from the objective of predicting future events. The former can be pre-trained on large amounts of unsupervised data, similarly to how chatbots are trained. The latter can piggy-back on the resulting representation after fine-tuning it for classification using relatively small labeled data. While similar pre-train-then-fine-tune approaches have recently achieved state-of-the-art performance in a number of NLP tasks—including natural language inference, question answering, and commonsense reasoning (discussed in Section SECREF2)—to the best of our knowledge this is the first attempt at applying this paradigm to conversational forecasting.
To test the effectiveness of this new architecture in forecasting derailment of online conversations, we develop and distribute two new datasets. The first triples in size the highly curated `Conversations Gone Awry' dataset BIBREF9, where civil-starting Wikipedia Talk Page conversations are crowd-labeled according to whether they eventually lead to personal attacks; the second relies on in-the-wild moderation of the popular subreddit ChangeMyView, where the aim is to forecast whether a discussion will later be subject to moderator action due to “rude or hostile” behavior. In both datasets, our model outperforms existing fixed-window approaches, as well as simpler sequential baselines that cannot account for inter-comment relations. Furthermore, by virtue of its online processing of the conversation, our system can provide substantial prior notice of upcoming derailment, triggering on average 3 comments (or 3 hours) before an overtly toxic comment is posted.
To summarize, in this work we:
introduce the first model for forecasting conversational events that can capture the dynamics of a conversation as it develops;
build two diverse datasets (one entirely new, one extending prior work) for the task of forecasting derailment of online conversations;
compare the performance of our model against the current state-of-the-art, and evaluate its ability to provide early warning signs.
Our work is motivated by the goal of assisting human moderators of online communities by preemptively signaling at-risk conversations that might deserve their attention. However, we caution that any automated systems might encode or even amplify the biases existing in the training data BIBREF26, BIBREF27, BIBREF28, so a public-facing implementation would need to be exhaustively scrutinized for such biases BIBREF29.
Further Related Work
Antisocial behavior. Antisocial behavior online comes in many forms, including harassment BIBREF30, cyberbullying BIBREF31, and general aggression BIBREF32. Prior work has sought to understand different aspects of such behavior, including its effect on the communities where it happens BIBREF33, BIBREF34, the actors involved BIBREF35, BIBREF36, BIBREF37, BIBREF38 and connections to the outside world BIBREF39.
Post-hoc classification of conversations. There is a rich body of prior work on classifying the outcome of a conversation after it has concluded, or classifying conversational events after they happened. Many examples exist, but some more closely related to our present work include identifying the winner of a debate BIBREF40, BIBREF41, BIBREF42, identifying successful negotiations BIBREF21, BIBREF43, as well as detecting whether deception BIBREF44, BIBREF45, BIBREF46 or disagreement BIBREF47, BIBREF48, BIBREF49, BIBREF50, BIBREF51 has occurred.
Our goal is different because we wish to forecast conversational events before they happen and while the conversation is still ongoing (potentially allowing for interventions). Note that some post-hoc tasks can also be re-framed as forecasting tasks (assuming the existence of necessary labels); for instance, predicting whether an ongoing conversation will eventually spark disagreement BIBREF18, rather than detecting already-existing disagreement.
Conversational forecasting. As described in Section SECREF1, prior work on forecasting conversational outcomes and events has largely relied on hand-crafted features to capture aspects of conversational dynamics. Example feature sets include statistical measures based on similarity between utterances BIBREF16, sentiment imbalance BIBREF20, flow of ideas BIBREF20, increase in hostility BIBREF8, reply rate BIBREF11 and graph representations of conversations BIBREF52, BIBREF17. By contrast, we aim to automatically learn neural representations of conversational dynamics through pre-training.
Such hand-crafted features are typically extracted from fixed-length windows of the conversation, leaving unaddressed the problem of unknown horizon. While some work has trained multiple models for different window-lengths BIBREF8, BIBREF18, they consider these models to be independent and, as such, do not address the issue of aggregating them into a single forecast (i.e., deciding at what point to make a prediction). We implement a simple sliding windows solution as a baseline (Section SECREF5).
Pre-training for NLP. The use of pre-training for natural language tasks has been growing in popularity after recent breakthroughs demonstrating improved performance on a wide array of benchmark tasks BIBREF53, BIBREF54. Existing work has generally used a language modeling objective as the pre-training objective; examples include next-word prediction BIBREF55, sentence autoencoding, BIBREF56, and machine translation BIBREF57. BERT BIBREF58 introduces a variation on this in which the goal is to predict the next sentence in a document given the current sentence. Our pre-training objective is similar in spirit, but operates at a conversation level, rather than a document level. We hence view our objective as conversational modeling rather than (only) language modeling. Furthermore, while BERT's sentence prediction objective is framed as a multiple-choice task, our objective is framed as a generative task.
Derailment Datasets
We consider two datasets, representing related but slightly different forecasting tasks. The first dataset is an expanded version of the annotated Wikipedia conversations dataset from BIBREF9. This dataset uses carefully-controlled crowdsourced labels, strictly filtered to ensure the conversations are civil up to the moment of a personal attack. This is a useful property for the purposes of model analysis, and hence we focus on this as our primary dataset. However, we are conscious of the possibility that these strict labels may not fully capture the kind of behavior that moderators care about in practice. We therefore introduce a secondary dataset, constructed from the subreddit ChangeMyView (CMV) that does not use post-hoc annotations. Instead, the prediction task is to forecast whether the conversation will be subject to moderator action in the future.
Wikipedia data. BIBREF9's `Conversations Gone Awry' dataset consists of 1,270 conversations that took place between Wikipedia editors on publicly accessible talk pages. The conversations are sourced from the WikiConv dataset BIBREF59 and labeled by crowdworkers as either containing a personal attack from within (i.e., hostile behavior by one user in the conversation directed towards another) or remaining civil throughout.
A series of controls are implemented to prevent models from picking up on trivial correlations. To prevent models from capturing topic-specific information (e.g., political conversations are more likely to derail), each attack-containing conversation is paired with a clean conversation from the same talk page, where the talk page serves as a proxy for topic. To force models to actually capture conversational dynamics rather than detecting already-existing toxicity, human annotations are used to ensure that all comments preceding a personal attack are civil.
To the ends of more effective model training, we elected to expand the `Conversations Gone Awry' dataset, using the original annotation procedure. Since we found that the original data skewed towards shorter conversations, we focused this crowdsourcing run on longer conversations: ones with 4 or more comments preceding the attack. Through this additional crowdsourcing, we expand the dataset to 4,188 conversations, which we are publicly releasing as part of the Cornell Conversational Analysis Toolkit (ConvoKit).
We perform an 80-20-20 train/dev/test split, ensuring that paired conversations end up in the same split in order to preserve the topic control. Finally, we randomly sample another 1 million conversations from WikiConv to use for the unsupervised pre-training of the generative component.
Reddit CMV data. The CMV dataset is constructed from conversations collected via the Reddit API. In contrast to the Wikipedia-based dataset, we explicitly avoid the use of post-hoc annotation. Instead, we use as our label whether a conversation eventually had a comment removed by a moderator for violation of Rule 2: “Don't be rude or hostile to other users”.
Though the lack of post-hoc annotation limits the degree to which we can impose controls on the data (e.g., some conversations may contain toxic comments not flagged by the moderators) we do reproduce as many of the Wikipedia data's controls as we can. Namely, we replicate the topic control pairing by choosing pairs of positive and negative examples that belong to the same top-level post, following BIBREF12; and enforce that the removed comment was made by a user who was previously involved in the conversation. This process results in 6,842 conversations, to which we again apply a pair-preserving 80-20-20 split. Finally, we gather over 600,000 conversations that do not include any removed comment, for unsupervised pre-training.
Conversational Forecasting Model
We now describe our general model for forecasting future conversational events. Our model integrates two components: (a) a generative dialog model that learns to represent conversational dynamics in an unsupervised fashion; and (b) a supervised component that fine-tunes this representation to forecast future events. Figure FIGREF13 provides an overview of the proposed architecture, henceforth CRAFT (Conversational Recurrent Architecture for ForecasTing).
Terminology. For modeling purposes, we treat a conversation as a sequence of $N$ comments $C = \lbrace c_1,\dots ,c_N\rbrace $. Each comment, in turn, is a sequence of tokens, where the number of tokens may vary from comment to comment. For the $n$-th comment ($1 \le n \le N)$, we let $M_n$ denote the number of tokens. Then, a comment $c_n$ can be represented as a sequence of $M_n$ tokens: $c_n = \lbrace w_1,\dots ,w_{M_n}\rbrace $.
Generative component. For the generative component of our model, we use a hierarchical recurrent encoder-decoder (HRED) architecture BIBREF60, a modified version of the popular sequence-to-sequence (seq2seq) architecture BIBREF61 designed to account for dependencies between consecutive inputs. BIBREF23 showed that HRED can successfully model conversational context by encoding the temporal structure of previously seen comments, making it an ideal fit for our use case. Here, we provide a high-level summary of the HRED architecture, deferring deeper technical discussion to BIBREF60 and BIBREF23.
An HRED dialog model consists of three components: an utterance encoder, a context encoder, and a decoder. The utterance encoder is responsible for generating semantic vector representations of comments. It consists of a recurrent neural network (RNN) that reads a comment token-by-token, and on each token $w_m$ updates a hidden state $h^{\text{enc}}$ based on the current token and the previous hidden state:
where $f^{\text{RNN}}$ is a nonlinear gating function (our implementation uses GRU BIBREF62). The final hidden state $h^{\text{enc}}_M$ can be viewed as a vector encoding of the entire comment.
Running the encoder on each comment $c_n$ results in a sequence of $N$ vector encodings. A second encoder, the context encoder, is then run over this sequence:
Each hidden state $h^{\text{con}}_n$ can then be viewed as an encoding of the full conversational context up to and including the $n$-th comment. To generate a response to comment $n$, the context encoding $h^{\text{con}}_n$ is used to initialize the hidden state $h^{\text{dec}}_{0}$ of a decoder RNN. The decoder produces a response token by token using the following recurrence:
where $f^{\text{out}}$ is some function that outputs a probability distribution over words; we implement this using a simple feedforward layer. In our implementation, we further augment the decoder with attention BIBREF63, BIBREF64 over context encoder states to help capture long-term inter-comment dependencies. This generative component can be pre-trained using unlabeled conversational data.
Prediction component. Given a pre-trained HRED dialog model, we aim to extend the model to predict from the conversational context whether the to-be-forecasted event will occur. Our predictor consists of a multilayer perceptron (MLP) with 3 fully-connected layers, leaky ReLU activations between layers, and sigmoid activation for output. For each comment $c_n$, the predictor takes as input the context encoding $h^{\text{con}}_n$ and forwards it through the MLP layers, resulting in an output score that is interpreted as a probability $p_{\text{event}}(c_{n+1})$ that the to-be-forecasted event will happen (e.g., that the conversation will derail).
Training the predictive component starts by initializing the weights of the encoders to the values learned in pre-training. The main training loop then works as follows: for each positive sample—i.e., a conversation containing an instance of the to-be-forecasted event (e.g., derailment) at comment $c_e$—we feed the context $c_1,\dots ,c_{e-1}$ through the encoder and classifier, and compute cross-entropy loss between the classifier output and expected output of 1. Similarly, for each negative sample—i.e., a conversation where none of the comments exhibit the to-be-forecasted event and that ends with $c_N$—we feed the context $c_1,\dots ,c_{N-1}$ through the model and compute loss against an expected output of 0.
Note that the parameters of the generative component are not held fixed during this process; instead, backpropagation is allowed to go all the way through the encoder layers. This process, known as fine-tuning, reshapes the representation learned during pre-training to be more directly useful to prediction BIBREF55.
We implement the model and training code using PyTorch, and we are publicly releasing our implementation and the trained models together with the data as part of ConvoKit.
Forecasting Derailment
We evaluate the performance of CRAFT in the task of forecasting conversational derailment in both the Wikipedia and CMV scenarios. To this end, for each of these datasets we pre-train the generative component on the unlabeled portion of the data and fine-tune it on the labeled training split (data size detailed in Section SECREF3).
In order to evaluate our sequential system against conversational-level ground truth, we need to aggregate comment level predictions. If any comment in the conversation triggers a positive prediction—i.e., $p_{\text{event}}(c_{n+1})$ is greater than a threshold learned on the development split—then the respective conversation is predicted to derail. If this forecast is triggered in a conversation that actually derails, but before the derailment actually happens, then the conversation is counted as a true positive; otherwise it is a false positive. If no positive predictions are triggered for a conversation, but it actually derails then it counts as a false negative; if it does not derail then it is a true negative.
Fixed-length window baselines. We first seek to compare CRAFT to existing, fixed-length window approaches to forecasting. To this end, we implement two such baselines: Awry, which is the state-of-the-art method proposed in BIBREF9 based on pragmatic features in the first comment-reply pair, and BoW, a simple bag-of-words baseline that makes a prediction using TF-IDF weighted bag-of-words features extracted from the first comment-reply pair.
Online forecasting baselines. Next, we consider simpler approaches for making forecasts as the conversations happen (i.e., in an online fashion). First, we propose Cumulative BoW, a model that recomputes bag-of-words features on all comments seen thus far every time a new comment arrives. While this approach does exhibit the desired behavior of producing updated predictions for each new comment, it fails to account for relationships between comments.
This simple cumulative approach cannot be directly extended to models whose features are strictly based on a fixed number of comments, like Awry. An alternative is to use a sliding window: for a feature set based on a window of $W$ comments, upon each new comment we can extract features from a window containing that comment and the $W-1$ comments preceding it. We apply this to the Awry method and call this model Sliding Awry. For both these baselines, we aggregate comment-level predictions in the same way as in our main model.
CRAFT ablations. Finally, we consider two modified versions of the CRAFT model in order to evaluate the impact of two of its key components: (1) the pre-training step, and (2) its ability to capture inter-comment dependencies through its hierarchical memory.
To evaluate the impact of pre-training, we train the prediction component of CRAFT on only the labeled training data, without first pre-training the encoder layers with the unlabeled data. We find that given the relatively small size of labeled data, this baseline fails to successfully learn, and ends up performing at the level of random guessing. This result underscores the need for the pre-training step that can make use of unlabeled data.
To evaluate the impact of the hierarchical memory, we implement a simplified version of CRAFT where the memory size of the context encoder is zero (CRAFT $-$ CE), thus effectively acting as if the pre-training component is a vanilla seq2seq model. In other words, this model cannot capture inter-comment dependencies, and instead at each step makes a prediction based only on the utterance encoding of the latest comment.
Results. Table TABREF17 compares CRAFT to the baselines on the test splits (random baseline is 50%) and illustrates several key findings. First, we find that unsurprisingly, accounting for full conversational context is indeed helpful, with even the simple online baselines outperforming the fixed-window baselines. On both datasets, CRAFT outperforms all baselines (including the other online models) in terms of accuracy and F1. Furthermore, although it loses on precision (to CRAFT $-$ CE) and recall (to Cumulative BoW) individually on the Wikipedia data, CRAFT has the superior balance between the two, having both a visibly higher precision-recall curve and larger area under the curve (AUPR) than the baselines (Figure FIGREF20). This latter property is particularly useful in a practical setting, as it allows moderators to tune model performance to some desired precision without having to sacrifice as much in the way of recall (or vice versa) compared to the baselines and pre-existing solutions.
Analysis
We now examine the behavior of CRAFT in greater detail, to better understand its benefits and limitations. We specifically address the following questions: (1) How much early warning does the the model provide? (2) Does the model actually learn an order-sensitive representation of conversational context?
Early warning, but how early? The recent interest in forecasting antisocial behavior has been driven by a desire to provide pre-emptive, actionable warning to moderators. But does our model trigger early enough for any such practical goals?
For each personal attack correctly forecasted by our model, we count the number of comments elapsed between the time the model is first triggered and the attack. Figure FIGREF22 shows the distribution of these counts: on average, the model warns of an attack 3 comments before it actually happens (4 comments for CMV). To further evaluate how much time this early warning would give to the moderator, we also consider the difference in timestamps between the comment where the model first triggers and the comment containing the actual attack. Over 50% of conversations get at least 3 hours of advance warning (2 hours for CMV). Moreover, 39% of conversations get at least 12 hours of early warning before they derail.
Does order matter? One motivation behind the design of our model was the intuition that comments in a conversation are not independent events; rather, the order in which they appear matters (e.g., a blunt comment followed by a polite one feels intuitively different from a polite comment followed by a blunt one). By design, CRAFT has the capacity to learn an order-sensitive representation of conversational context, but how can we know that this capacity is actually used? It is conceivable that the model is simply computing an order-insensitive “bag-of-features”. Neural network models are notorious for their lack of transparency, precluding an analysis of how exactly CRAFT models conversational context. Nevertheless, through two simple exploratory experiments, we seek to show that it does not completely ignore comment order.
The first experiment for testing whether the model accounts for comment order is a prefix-shuffling experiment, visualized in Figure FIGREF23. For each conversation that the model predicts will derail, let $t$ denote the index of the triggering comment, i.e., the index where the model first made a derailment forecast. We then construct synthetic conversations by taking the first $t-1$ comments (henceforth referred to as the prefix) and randomizing their order. Finally, we count how often the model no longer predicts derailment at index $t$ in the synthetic conversations. If the model were ignoring comment order, its prediction should remain unchanged (as it remains for the Cumulative BoW baseline), since the actual content of the first $t$ comments has not changed (and CRAFT inference is deterministic). We instead find that in roughly one fifth of cases (12% for CMV) the model changes its prediction on the synthetic conversations. This suggests that CRAFT learns an order-sensitive representation of context, not a mere “bag-of-features”.
To more concretely quantify how much this order-sensitive context modeling helps with prediction, we can actively prevent the model from learning and exploiting any order-related dynamics. We achieve this through another type of shuffling experiment, where we go back even further and shuffle the comment order in the conversations used for pre-training, fine-tuning and testing. This procedure preserves the model's ability to capture signals present within the individual comments processed so far, as the utterance encoder is unaffected, but inhibits it from capturing any meaningful order-sensitive dynamics. We find that this hurts the model's performance (65% accuracy for Wikipedia, 59.5% for CMV), lowering it to a level similar to that of the version where we completely disable the context encoder.
Taken together, these experiments provide evidence that CRAFT uses its capacity to model conversational context in an order-sensitive fashion, and that it makes effective use of the dynamics within. An important avenue for future work would be developing more transparent models that can shed light on exactly what kinds of order-related features are being extracted and how they are used in prediction.
Conclusions and Future Work
In this work, we introduced a model for forecasting conversational events that processes comments as they happen and takes the full conversational context into account to make an updated prediction at each step. This model fills a void in the existing literature on conversational forecasting, simultaneously addressing the dual challenges of capturing inter-comment dynamics and dealing with an unknown horizon. We find that our model achieves state-of-the-art performance on the task of forecasting derailment in two different datasets that we release publicly. We further show that the resulting system can provide substantial prior notice of derailment, opening up the potential for preemptive interventions by human moderators BIBREF65.
While we have focused specifically on the task of forecasting derailment, we view this work as a step towards a more general model for real-time forecasting of other types of emergent properties of conversations. Follow-up work could adapt the CRAFT architecture to address other forecasting tasks mentioned in Section SECREF2—including those for which the outcome is extraneous to the conversation. We expect different tasks to be informed by different types of inter-comment dynamics, and further architecture extensions could add additional supervised fine-tuning in order to direct it to focus on specific dynamics that might be relevant to the task (e.g., exchange of ideas between interlocutors or stonewalling).
With respect to forecasting derailment, there remain open questions regarding what human moderators actually desire from an early-warning system, which would affect the design of a practical system based on this work. For instance, how early does a warning need to be in order for moderators to find it useful? What is the optimal balance between precision, recall, and false positive rate at which such a system is truly improving moderator productivity rather than wasting their time through false positives? What are the ethical implications of such a system? Follow-up work could run a user study of a prototype system with actual moderators to address these questions.
A practical limitation of the current analysis is that it relies on balanced datasets, while derailment is a relatively rare event for which a more restrictive trigger threshold would be appropriate. While our analysis of the precision-recall curve suggests the system is robust across multiple thresholds ($AUPR=0.7$), additional work is needed to establish whether the recall tradeoff would be acceptable in practice.
Finally, one major limitation of the present work is that it assigns a single label to each conversation: does it derail or not? In reality, derailment need not spell the end of a conversation; it is possible that a conversation could get back on track, suffer a repeat occurrence of antisocial behavior, or any number of other trajectories. It would be exciting to consider finer-grained forecasting of conversational trajectories, accounting for the natural—and sometimes chaotic—ebb-and-flow of human interactions.
Acknowledgements. We thank Caleb Chiam, Liye Fu, Lillian Lee, Alexandru Niculescu-Mizil, Andrew Wang and Justine Zhang for insightful conversations (with unknown horizon), Aditya Jha for his great help with implementing and running the crowd-sourcing tasks, Thomas Davidson and Claire Liang for exploratory data annotation, as well as the anonymous reviewers for their helpful comments. This work is supported in part by the NSF CAREER award IIS-1750615 and by the NSF Grant SES-1741441.
| [
" `Conversations Gone Awry' dataset, subreddit ChangeMyView",
"An expanded version of the existing 'Conversations Gone Awry' dataset and the ChangeMyView dataset, a subreddit whose only annotation is whether the conversation required action by the Reddit moderators. "
] | 4,718 | qasper | en | null | 2589b46ee58a7600e17fa89a0d4fffd9a0faf1df49b3c035 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What are two datasets model is applied to?
Answer: | [
" `Conversations Gone Awry' dataset, subreddit ChangeMyView",
"An expanded version of the existing 'Conversations Gone Awry' dataset and the ChangeMyView dataset, a subreddit whose only annotation is whether the conversation required action by the Reddit moderators. "
] | qasper | 128 | 58 | |
Were any of the pipeline components based on deep learning models? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The automatic identification, extraction and representation of the information conveyed in texts is a key task nowadays. In fact, this research topic is increasing its relevance with the exponential growth of social networks and the need to have tools that are able to automatically process them BIBREF0.
Some of the domains where it is more important to be able to perform this kind of action are the juridical and legal ones. Effectively, it is crucial to have the capability to analyse open access text sources, like social nets (Twitter and Facebook, for instance), blogs, online newspapers, and to be able to extract the relevant information and represent it in a knowledge base, allowing posterior inferences and reasoning.
In the context of this work, we will present results of the R&D project Agatha, where we developed a pipeline of processes that analyses texts (in Portuguese, Spanish, or English) and is able to populate a specialized ontology BIBREF1 (related to criminal law) for the representation of events, depicted in such texts. Events are represented by objects having associated actions, agents, elements, places and time. After having populated the event ontology, we have an automatic process linking the identified entities to external referents, creating, this way, a linked data knowledge base.
It is important to point out that, having the text information represented in an ontology allows us to perform complex queries and inferences, which can detect patterns of typical criminal actions.
Another axe of innovation in this research is the development, for the Portuguese language, of a pipeline of Natural Language Processing (NLP) processes, that allows us to fully process sentences and represent their content in an ontology. Although there are several tools for the processing of the Portuguese language, the combination of all these steps in a integrated tool is a new contribution.
Moreover, we have already explored other related research path, namely author profiling BIBREF2, aggression identification BIBREF3 and hate-speech detection BIBREF4 over social media, plus statute law retrieval and entailment for Japanese BIBREF5.
The remainder of this paper is organized as follows: Section SECREF2 describes our proposed architecture together with the Portuguese modules for its computational processing. Section SECREF3 discusses different design options and Section SECREF4 provides our conclusions together with some pointers for future work.
Framework for Processing Portuguese Text
The framework for processing Portuguese texts is depicted in Fig. FIGREF2, which illustrates how relevant pieces of information are extracted from the text. Namely, input files (Portuguese texts) go through a series of modules: part-of-speech tagging, named entity recognition, dependency parsing, semantic role labeling, subject-verb-object triple extraction, and lexicon matching.
The main goal of all the modules except lexicon matching is to identify events given in the text. These events are then used to populate an ontology.
The lexicon matching, on the other hand, was created to link words that are found in the text source with the data available not only on Eurovoc BIBREF6 thesaurus but also on the EU's terminology database IATE BIBREF7 (see Section SECREF12 for details).
Most of these modules are deeply related and are detailed in the subsequent subsections.
Framework for Processing Portuguese Text ::: Part-Of-Speech Tagging
Part-of-speech tagging happens after language detection. It labels each word with a tag that indicates its syntactic role in the sentence. For instance, a word could be a noun, verb, adjective or adverb (or other syntactic tag). We used Freeling BIBREF8 library to provide the tags. This library resorts to a Hidden Markov Model as described by Brants BIBREF9. The end result is a tag for each word as described by the EAGLES tagset .
Framework for Processing Portuguese Text ::: Named Entity Recognition
We use the named entity recognition module after part-of-speech tagging. This module labels each part of the sentence into different categories such as "PERSON", "LOCATION", or "ORGANIZATION". We also used Freeling to label the named entities and the details of the algorithm are shown in the paper by Carreras et al BIBREF10. Aside from the three aforementioned categories, we also extracted "DATE/TIME" and "CURRENCY" values by looking at the part-of-speech tags: date/time words have a tag of "W", while currencies have "Zm".
Framework for Processing Portuguese Text ::: Dependency Parsing
Dependency parsing involves tagging a word based on different features to indicate if it is dependent on another word. The Freeling library also has dependency parsing models for Portuguese. Since we wanted to build a SRL (Semantic Role Labeling) module on top of the dependency parser and the current released version of Freeling does not have an SRL module for Portuguese, we trained a different Portuguese dependency parsing model that was compatible (in terms of used tags) with the available annotated.
We used the dataset from System-T BIBREF11, which has SRL tags, as well as, the other preceding tags. It was necessary to do some pre-processing and tag mapping in order to make it viable to train a Portuguese model.
We made 589 tag conversions over 14 different categories. The breakdown of tag conversions per category is given by table TABREF7. These rules can be further seen in the corresponding Github repository BIBREF12. The modified training and development datasets are also available on another Github repositorypage BIBREF13 for further research and comparison purposes.
Framework for Processing Portuguese Text ::: Semantic Role Labeling
We execute the SRL (Semantic Role Labeling) module after obtaining the word dependencies. This module aims at giving a semantic role to a syntactic constituent of a sentence. The semantic role is always in relation to a verb and these roles could either be an actor, object, time, or location, which are then tagged as A0, A1, AM-TMP, AM-LOC, respectively. We trained a model for this module on top of the dependency parser described in the previous subsection using the modified dataset from System-T. The module also needs co-reference resolution to work and, to achieve this, we adapted the Spanish co-reference modules for Portuguese, changing the words that are equivalent (in total, we changed 253 words).
Framework for Processing Portuguese Text ::: SVO Extraction
From the yield of the SRL (Semantic Role Labeling) module, our framework can distinguish actors, actions, places, time and objects from the sentences. Utilizing this extracted data, we can distinguish subject-verb-object (SVO) triples using the SVO extraction algorithm BIBREF14. The algorithm finds, for each sentence, the verb and the tuples related to that verb using Semantic Role Labeling (subsection SECREF8). After the extraction of SVOs from texts, they are inserted into a specific event ontology (see section SECREF12 for the creation of a knowledge base).
Framework for Processing Portuguese Text ::: Lexicon Matching
The sole purpose of this module is to find important terms and/or concepts from the extracted text. To do this, we use Euvovoc BIBREF6, a multilingual thesaurus that was developed for and by the European Union. The Euvovoc has 21 fields and each field is further divided into a variable number of micro-thesauri. Here, due to the application of this work in the Agatha project (mentioned in Section SECREF1), we use the terms of the criminal law BIBREF15 micro-thesaurus. Further, we classified each term of the criminal law micro-thesaurus into four categories namely, actor, event, place and object. The term classification can be seen in Table TABREF11.
After the classification of these terms, we implemented two different matching algorithms between the extracted words and the criminal law micro-thesaurus terms. The first is an exact string match wherein lowercase equivalents of the words of the input sentences are matched exactly with lower case equivalents of the predefined terms. The second matching algorithm uses Levenshtein distance BIBREF16, allowing some near-matches that are close enough to the target term.
Framework for Processing Portuguese Text ::: Linked Data: Ontology, Thesaurus and Terminology
In the computer science field, an ontology can be defined has:
a formal specification of a conceptualization;
shared vocabulary and taxonomy which models a domain with the definition of objects and/or concepts and their properties and relations;
the representation of entities, ideas, and events, along with their properties and relations, according to a system of categories.
A knowledge base is one kind of repository typically used to store answers to questions or solutions to problems enabling rapid search, retrieval, and reuse, either by an ontology or directly by those requesting support. For a more detailed description of ontologies and knowledge bases, see for instance BIBREF17.
For designing the ontology adequate for our goals, we referred to the Simple Event Model (SEM) BIBREF18 as a baseline model. A pictorial representation of this ontology is given in Figure FIGREF16
Considering the criminal law domain case study, we made a few changes to the original SEM ontology. The entities of the model are:
Actor: person involved with event
Place: location of the event
Time: time of the event
Object: that actor act upon
Organization: organization involved with event
Currency: money involved with event
The proposed ontology was designed in such a manner that it can incorporate information extracted from multiple documents. In this context, suppose that the source of documents is aare a legal police department, where each document isare under the hood of a particular case/crime; furthermoreFurther, a single case can have documents from multiple languages. Now, considering case 1 has 100 documents and case 2 has 100 documents then there is not only a connection among the documents of a single case but rather among all the cases with all the combined 200 documents. In this way, the proposed method is able to produce a detailed and well-connected knowledge base.
Figure FIGREF23 shows the proposed ontology, which, in our evaluation procedure, was populated with 3121 events entries from 51 documents.
Protege BIBREF19 tool was used for creating the ontology and GraphDB BIBREF20 for populating & querying the data. GraphDB is an enterprise-ready Semantic Graph Database, compliant with W3C Standards. Semantic Graph Databases (also called RDF triplestores) provide the core infrastructure for solutions where modeling agility, data integration, relationship exploration, and cross-enterprise data publishing and consumption are important. GraphDB has a SPARQL (SQL-like query language) interface for RDF graph databases with the following types:
SELECT: returns tabular results
CONSTRUCT: creates a new RDF graph based on query results
ASK: returns "YES", if the query has a solution, otherwise "NO"
DESCRIBE: returns RDF data about a resource. This is useful when the RDF data structure in the data source is not known
INSERT: inserts triples into a graph
DELETE: deletes triples from a graph
Furthermore, we have extended the ontology BIBREF21 to connect the extracted terms with Eurovoc criminal law (discussed in subsection SECREF10) and IATE BIBREF7 terms. IATE (Interactive Terminology for Europe) is the EU's general terminology database and its aim is to provide a web-based infrastructure for all EU terminology resources, enhancing the availability and standardization of the information. The extended ontology has a number of sub-classes for Actor, Event, Object and Place classes detailed in Table TABREF30.
Discussion
We have defined a major design principle for our architecture: it should be modular and not rely on human made rules allowing, as much as possible, its independence from a specific language. In this way, its potential application to another language would be easier, simply by changing the modules or the models of specific modules. In fact, we have explored the use of already existing modules and adopted and integrated several of these tools into our pipeline.
It is important to point out that, as far as we know, there is no integrated architecture supporting the full processing pipeline for the Portuguese language. We evaluated several systems like Rembrandt BIBREF22 or LinguaKit: the former only has the initial steps of our proposal (until NER) and the later performed worse than our system.
This framework, developed within the context of the Agatha project (described in Section SECREF1) has the full processing pipeline for Portuguese texts: it receives sentences as input and outputs ontological information: a) first performs all NLP typical tasks until semantic role labelling; b) then, it extracts subject-verb-object triples; c) and, then, it performs ontology matching procedures. As a final result, the obtained output is inserted into a specialized ontology.
We are aware that each of the architecture modules can, and should, be improved but our main goal was the creation of a full working text processing pipeline for the Portuguese language.
Conclusions and Future Work
Besides the end–to–end NLP pipeline for the Portuguese language, the other main contributions of this work can be summarize as follows:
Development of an ontology for the criminal law domain;
Alignment of the Eurovoc thesaurus and IATE terminology with the ontology created;
Representation of the extracted events from texts in the linked knowledge base defined.
The obtained results support our claim that the proposed system can be used as a base tool for information extraction for the Portuguese language. Being composed by several modules, each of them with a high level of complexity, it is certain that our approach can be improved and an overall better performance can be achieved.
As future work we intend, not only to continue improving the individual modules, but also plan to extend this work to the:
automatic creation of event timelines;
incorporation in the knowledge base of information obtained from videos or pictures describing scenes relevant to criminal investigations.
Acknowledgments
The authors would like to thank COMPETE 2020, PORTUGAL 2020 Program, the European Union, and ALENTEJO 2020 for supporting this research as part of Agatha Project SI & IDT number 18022 (Intelligent analysis system of open of sources information for surveillance/crime control). The authors would also like to thank LISP - Laboratory of Informatics, Systems and Parallelism.
| [
"No",
"No"
] | 2,276 | qasper | en | null | 4c9552eec5c238657f8ed5237bf66067d3fdda2409a903b6 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Were any of the pipeline components based on deep learning models?
Answer: | [
"No",
"No"
] | qasper | 128 | 59 | |
How is the quality of the data empirically evaluated? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
End-to-end speech-to-text translation (ST) has attracted much attention recently BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6 given its simplicity against cascading automatic speech recognition (ASR) and machine translation (MT) systems. The lack of labeled data, however, has become a major blocker for bridging the performance gaps between end-to-end models and cascading systems. Several corpora have been developed in recent years. post2013improved introduced a 38-hour Spanish-English ST corpus by augmenting the transcripts of the Fisher and Callhome corpora with English translations. di-gangi-etal-2019-must created the largest ST corpus to date from TED talks but the language pairs involved are out of English only. beilharz2019librivoxdeen created a 110-hour German-English ST corpus from LibriVox audiobooks. godard-etal-2018-low created a Moboshi-French ST corpus as part of a rare language documentation effort. woldeyohannis provided an Amharic-English ST corpus in the tourism domain. boito2019mass created a multilingual ST corpus involving 8 languages from a multilingual speech corpus based on Bible readings BIBREF7. Previous work either involves language pairs out of English, very specific domains, very low resource languages or a limited set of language pairs. This limits the scope of study, including the latest explorations on end-to-end multilingual ST BIBREF8, BIBREF9. Our work is mostly similar and concurrent to iranzosnchez2019europarlst who created a multilingual ST corpus from the European Parliament proceedings. The corpus we introduce has larger speech durations and more translation tokens. It is diversified with multiple speakers per transcript/translation. Finally, we provide additional out-of-domain test sets.
In this paper, we introduce CoVoST, a multilingual ST corpus based on Common Voice BIBREF10 for 11 languages into English, diversified with over 11,000 speakers and over 60 accents. It includes a total 708 hours of French (Fr), German (De), Dutch (Nl), Russian (Ru), Spanish (Es), Italian (It), Turkish (Tr), Persian (Fa), Swedish (Sv), Mongolian (Mn) and Chinese (Zh) speeches, with French and German ones having the largest durations among existing public corpora. We also collect an additional evaluation corpus from Tatoeba for French, German, Dutch, Russian and Spanish, resulting in a total of 9.3 hours of speech. Both corpora are created at the sentence level and do not require additional alignments or segmentation. Using the official Common Voice train-development-test split, we also provide baseline models, including, to our knowledge, the first end-to-end many-to-one multilingual ST models. CoVoST is released under CC0 license and free to use. The Tatoeba evaluation samples are also available under friendly CC licenses. All the data can be acquired at https://github.com/facebookresearch/covost.
Data Collection and Processing ::: Common Voice (CoVo)
Common Voice BIBREF10 is a crowdsourcing speech recognition corpus with an open CC0 license. Contributors record voice clips by reading from a bank of donated sentences. Each voice clip was validated by at least two other users. Most of the sentences are covered by multiple speakers, with potentially different genders, age groups or accents.
Raw CoVo data contains samples that passed validation as well as those that did not. To build CoVoST, we only use the former one and reuse the official train-development-test partition of the validated data. As of January 2020, the latest CoVo 2019-06-12 release includes 29 languages. CoVoST is currently built on that release and covers the following 11 languages: French, German, Dutch, Russian, Spanish, Italian, Turkish, Persian, Swedish, Mongolian and Chinese.
Validated transcripts were sent to professional translators. Note that the translators had access to the transcripts but not the corresponding voice clips since clips would not carry additional information. Since transcripts were duplicated due to multiple speakers, we deduplicated the transcripts before sending them to translators. As a result, different voice clips of the same content (transcript) will have identical translations in CoVoST for train, development and test splits.
In order to control the quality of the professional translations, we applied various sanity checks to the translations BIBREF11. 1) For German-English, French-English and Russian-English translations, we computed sentence-level BLEU BIBREF12 with the NLTK BIBREF13 implementation between the human translations and the automatic translations produced by a state-of-the-art system BIBREF14 (the French-English system was a Transformer big BIBREF15 separately trained on WMT14). We applied this method to these three language pairs only as we are confident about the quality of the corresponding systems. Translations with a score that was too low were manually inspected and sent back to the translators when needed. 2) We manually inspected examples where the source transcript was identical to the translation. 3) We measured the perplexity of the translations using a language model trained on a large amount of clean monolingual data BIBREF14. We manually inspected examples where the translation had a high perplexity and sent them back to translators accordingly. 4) We computed the ratio of English characters in the translations. We manually inspected examples with a low ratio and sent them back to translators accordingly. 5) Finally, we used VizSeq BIBREF16 to calculate similarity scores between transcripts and translations based on LASER cross-lingual sentence embeddings BIBREF17. Samples with low scores were manually inspected and sent back for translation when needed.
We also sanity check the overlaps of train, development and test sets in terms of transcripts and voice clips (via MD5 file hashing), and confirm they are totally disjoint.
Data Collection and Processing ::: Tatoeba (TT)
Tatoeba (TT) is a community built language learning corpus having sentences aligned across multiple languages with the corresponding speech partially available. Its sentences are on average shorter than those in CoVoST (see also Table TABREF2) given the original purpose of language learning. Sentences in TT are licensed under CC BY 2.0 FR and part of the speeches are available under various CC licenses.
We construct an evaluation set from TT (for French, German, Dutch, Russian and Spanish) as a complement to CoVoST development and test sets. We collect (speech, transcript, English translation) triplets for the 5 languages and do not include those whose speech has a broken URL or is not CC licensed. We further filter these samples by sentence lengths (minimum 4 words including punctuations) to reduce the portion of short sentences. This makes the resulting evaluation set closer to real-world scenarios and more challenging.
We run the same quality checks for TT as for CoVoST but we do not find poor quality translations according to our criteria. Finally, we report the overlap between CoVo transcripts and TT sentences in Table TABREF5. We found a minimal overlap, which makes the TT evaluation set a suitable additional test set when training on CoVoST.
Data Analysis ::: Basic Statistics
Basic statistics for CoVoST and TT are listed in Table TABREF2 including (unique) sentence counts, speech durations, speaker demographics (partially available) as well as vocabulary and token statistics (based on Moses-tokenized sentences by sacreMoses) on both transcripts and translations. We see that CoVoST has over 327 hours of German speeches and over 171 hours of French speeches, which, to our knowledge, corresponds to the largest corpus among existing public ST corpora (the second largest is 110 hours BIBREF18 for German and 38 hours BIBREF19 for French). Moreover, CoVoST has a total of 18 hours of Dutch speeches, to our knowledge, contributing the first public Dutch ST resource. CoVoST also has around 27-hour Russian speeches, 37-hour Italian speeches and 67-hour Persian speeches, which is 1.8 times, 2.5 times and 13.3 times of the previous largest public one BIBREF7. Most of the sentences (transcripts) in CoVoST are covered by multiple speakers with potentially different accents, resulting in a rich diversity in the speeches. For example, there are over 1,000 speakers and over 10 accents in the French and German development / test sets. This enables good coverage of speech variations in both model training and evaluation.
Data Analysis ::: Speaker Diversity
As we can see from Table TABREF2, CoVoST is diversified with a rich set of speakers and accents. We further inspect the speaker demographics in terms of sample distributions with respect to speaker counts, accent counts and age groups, which is shown in Figure FIGREF6, FIGREF7 and FIGREF8. We observe that for 8 of the 11 languages, at least 60% of the sentences (transcripts) are covered by multiple speakers. Over 80% of the French sentences have at least 3 speakers. And for German sentences, even over 90% of them have at least 5 speakers. Similarly, we see that a large portion of sentences are spoken in multiple accents for French, German, Dutch and Spanish. Speakers of each language also spread widely across different age groups (below 20, 20s, 30s, 40s, 50s, 60s and 70s).
Baseline Results
We provide baselines using the official train-development-test split on the following tasks: automatic speech recognition (ASR), machine translation (MT) and speech translation (ST).
Baseline Results ::: Experimental Settings ::: Data Preprocessing
We convert raw MP3 audio files from CoVo and TT into mono-channel waveforms, and downsample them to 16,000 Hz. For transcripts and translations, we normalize the punctuation, we tokenize the text with sacreMoses and lowercase it. For transcripts, we further remove all punctuation markers except for apostrophes. We use character vocabularies on all the tasks, with 100% coverage of all the characters. Preliminary experimentation showed that character vocabularies provided more stable training than BPE. For MT, the vocabulary is created jointly on both transcripts and translations. We extract 80-channel log-mel filterbank features, computed with a 25ms window size and 10ms window shift using torchaudio. The features are normalized to 0 mean and 1.0 standard deviation. We remove samples having more than 3,000 frames or more than 256 characters for GPU memory efficiency (less than 25 samples are removed for all languages).
Baseline Results ::: Experimental Settings ::: Model Training
Our ASR and ST models follow the architecture in berard2018end, but have 3 decoder layers like that in pino2019harnessing. For MT, we use a Transformer base architecture BIBREF15, but with 3 encoder layers, 3 decoder layers and 0.3 dropout. We use a batch size of 10,000 frames for ASR and ST, and a batch size of 4,000 tokens for MT. We train all models using Fairseq BIBREF20 for up to 200,000 updates. We use SpecAugment BIBREF21 for ASR and ST to alleviate overfitting.
Baseline Results ::: Experimental Settings ::: Inference and Evaluation
We use a beam size of 5 for all models. We use the best checkpoint by validation loss for MT, and average the last 5 checkpoints for ASR and ST. For MT and ST, we report case-insensitive tokenized BLEU BIBREF22 using sacreBLEU BIBREF23. For ASR, we report word error rate (WER) and character error rate (CER) using VizSeq.
Baseline Results ::: Automatic Speech Recognition (ASR)
For simplicity, we use the same model architecture for ASR and ST, although we do not leverage ASR models to pretrain ST model encoders later. Table TABREF18 shows the word error rate (WER) and character error rate (CER) for ASR models. We see that French and German perform the best given they are the two highest resource languages in CoVoST. The other languages are relatively low resource (especially Turkish and Swedish) and the ASR models are having difficulties to learn from this data.
Baseline Results ::: Machine Translation (MT)
MT models take transcripts (without punctuation) as inputs and outputs translations (with punctuation). For simplicity, we do not change the text preprocessing methods for MT to correct this mismatch. Moreover, this mismatch also exists in cascading ST systems, where MT model inputs are the outputs of an ASR model. Table TABREF20 shows the BLEU scores of MT models. We notice that the results are consistent with what we see from ASR models. For example thanks to abundant training data, French has a decent BLEU score of 29.8/25.4. German doesn't perform well, because of less richness of content (transcripts). The other languages are low resource in CoVoST and it is difficult to train decent models without additional data or pre-training techniques.
Baseline Results ::: Speech Translation (ST)
CoVoST is a many-to-one multilingual ST corpus. While end-to-end one-to-many and many-to-many multilingual ST models have been explored very recently BIBREF8, BIBREF9, many-to-one multilingual models, to our knowledge, have not. We hence use CoVoST to examine this setting. Table TABREF22 and TABREF23 show the BLEU scores for both bilingual and multilingual end-to-end ST models trained on CoVoST. We observe that combining speeches from multiple languages is consistently bringing gains to low-resource languages (all besides French and German). This includes combinations of distant languages, such as Ru+Fr, Tr+Fr and Zh+Fr. Moreover, some combinations do bring gains to high-resource language (French) as well: Es+Fr, Tr+Fr and Mn+Fr. We simply provide the most basic many-to-one multilingual baselines here, and leave the full exploration of the best configurations to future work. Finally, we note that for some language pairs, absolute BLEU numbers are relatively low as we restrict model training to the supervised data. We encourage the community to improve upon those baselines, for example by leveraging semi-supervised training.
Baseline Results ::: Multi-Speaker Evaluation
In CoVoST, large portion of transcripts are covered by multiple speakers with different genders, accents and age groups. Besides the standard corpus-level BLEU scores, we also want to evaluate model output variance on the same content (transcript) but different speakers. We hence propose to group samples (and their sentence BLEU scores) by transcript, and then calculate average per-group mean and average coefficient of variation defined as follows:
and
where $G$ is the set of sentence BLEU scores grouped by transcript and $G^{\prime } = \lbrace g | g\in G, |g|>1, \textrm {Mean}(g) > 0 \rbrace $.
$\textrm {BLEU}_{MS}$ provides a normalized quality score as oppose to corpus-level BLEU or unnormalized average of sentence BLEU. And $\textrm {CoefVar}_{MS}$ is a standardized measure of model stability against different speakers (the lower the better). Table TABREF24 shows the $\textrm {BLEU}_{MS}$ and $\textrm {CoefVar}_{MS}$ of our ST models on CoVoST test set. We see that German and Persian have the worst $\textrm {CoefVar}_{MS}$ (least stable) given their rich speaker diversity in the test set and relatively small train set (see also Figure FIGREF6 and Table TABREF2). Dutch also has poor $\textrm {CoefVar}_{MS}$ because of the lack of training data. Multilingual models are consistantly more stable on low-resource languages. Ru+Fr, Tr+Fr, Fa+Fr and Zh+Fr even have better $\textrm {CoefVar}_{MS}$ than all individual languages.
Conclusion
We introduce a multilingual speech-to-text translation corpus, CoVoST, for 11 languages into English, diversified with over 11,000 speakers and over 60 accents. We also provide baseline results, including, to our knowledge, the first end-to-end many-to-one multilingual model for spoken language translation. CoVoST is free to use with a CC0 license, and the additional Tatoeba evaluation samples are also CC-licensed.
| [
"Validated transcripts were sent to professional translators., various sanity checks to the translations, sanity check the overlaps of train, development and test sets",
"computed sentence-level BLEU, We manually inspected examples where the source transcript was identical to the translation, measured the perple... | 2,435 | qasper | en | null | bee3efb92996570d990ab04bfdf054835ef720b9dd1c6f5a | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How is the quality of the data empirically evaluated?
Answer: | [
"Validated transcripts were sent to professional translators., various sanity checks to the translations, sanity check the overlaps of train, development and test sets",
"computed sentence-level BLEU, We manually inspected examples where the source transcript was identical to the translation, measured the perple... | qasper | 128 | 60 | |
How do they combine audio and text sequences in their RNN? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Recently, deep learning algorithms have successfully addressed problems in various fields, such as image classification, machine translation, speech recognition, text-to-speech generation and other machine learning related areas BIBREF0 , BIBREF1 , BIBREF2 . Similarly, substantial improvements in performance have been obtained when deep learning algorithms have been applied to statistical speech processing BIBREF3 . These fundamental improvements have led researchers to investigate additional topics related to human nature, which have long been objects of study. One such topic involves understanding human emotions and reflecting it through machine intelligence, such as emotional dialogue models BIBREF4 , BIBREF5 .
In developing emotionally aware intelligence, the very first step is building robust emotion classifiers that display good performance regardless of the application; this outcome is considered to be one of the fundamental research goals in affective computing BIBREF6 . In particular, the speech emotion recognition task is one of the most important problems in the field of paralinguistics. This field has recently broadened its applications, as it is a crucial factor in optimal human-computer interactions, including dialog systems. The goal of speech emotion recognition is to predict the emotional content of speech and to classify speech according to one of several labels (i.e., happy, sad, neutral, and angry). Various types of deep learning methods have been applied to increase the performance of emotion classifiers; however, this task is still considered to be challenging for several reasons. First, insufficient data for training complex neural network-based models are available, due to the costs associated with human involvement. Second, the characteristics of emotions must be learned from low-level speech signals. Feature-based models display limited skills when applied to this problem.
To overcome these limitations, we propose a model that uses high-level text transcription, as well as low-level audio signals, to utilize the information contained within low-resource datasets to a greater degree. Given recent improvements in automatic speech recognition (ASR) technology BIBREF7 , BIBREF2 , BIBREF8 , BIBREF9 , speech transcription can be carried out using audio signals with considerable skill. The emotional content of speech is clearly indicated by the emotion words contained in a sentence BIBREF10 , such as “lovely” and “awesome,” which carry strong emotions compared to generic (non-emotion) words, such as “person” and “day.” Thus, we hypothesize that the speech emotion recognition model will be benefit from the incorporation of high-level textual input.
In this paper, we propose a novel deep dual recurrent encoder model that simultaneously utilizes audio and text data in recognizing emotions from speech. Extensive experiments are conducted to investigate the efficacy and properties of the proposed model. Our proposed model outperforms previous state-of-the-art methods by 68.8% to 71.8% when applied to the IEMOCAP dataset, which is one of the most well-studied datasets. Based on an error analysis of the models, we show that our proposed model accurately identifies emotion classes. Moreover, the neutral class misclassification bias frequently exhibited by previous models, which focus on audio features, is less pronounced in our model.
Related work
Classical machine learning algorithms, such as hidden Markov models (HMMs), support vector machines (SVMs), and decision tree-based methods, have been employed in speech emotion recognition problems BIBREF11 , BIBREF12 , BIBREF13 . Recently, researchers have proposed various neural network-based architectures to improve the performance of speech emotion recognition. An initial study utilized deep neural networks (DNNs) to extract high-level features from raw audio data and demonstrated its effectiveness in speech emotion recognition BIBREF14 . With the advancement of deep learning methods, more complex neural-based architectures have been proposed. Convolutional neural network (CNN)-based models have been trained on information derived from raw audio signals using spectrograms or audio features such as Mel-frequency cepstral coefficients (MFCCs) and low-level descriptors (LLDs) BIBREF15 , BIBREF16 , BIBREF17 . These neural network-based models are combined to produce higher-complexity models BIBREF18 , BIBREF19 , and these models achieved the best-recorded performance when applied to the IEMOCAP dataset.
Another line of research has focused on adopting variant machine learning techniques combined with neural network-based models. One researcher utilized the multiobject learning approach and used gender and naturalness as auxiliary tasks so that the neural network-based model learned more features from a given dataset BIBREF20 . Another researcher investigated transfer learning methods, leveraging external data from related domains BIBREF21 .
As emotional dialogue is composed of sound and spoken content, researchers have also investigated the combination of acoustic features and language information, built belief network-based methods of identifying emotional key phrases, and assessed the emotional salience of verbal cues from both phoneme sequences and words BIBREF22 , BIBREF23 . However, none of these studies have utilized information from speech signals and text sequences simultaneously in an end-to-end learning neural network-based model to classify emotions.
Model
This section describes the methodologies that are applied to the speech emotion recognition task. We start by introducing the recurrent encoder model for the audio and text modalities individually. We then propose a multimodal approach that encodes both audio and textual information simultaneously via a dual recurrent encoder.
Audio Recurrent Encoder (ARE)
Motivated by the architecture used in BIBREF24 , BIBREF25 , we build an audio recurrent encoder (ARE) to predict the class of a given audio signal. Once MFCC features have been extracted from an audio signal, a subset of the sequential features is fed into the RNN (i.e., gated recurrent units (GRUs)), which leads to the formation of the network's internal hidden state INLINEFORM0 to model the time series patterns. This internal hidden state is updated at each time step with the input data INLINEFORM1 and the hidden state of the previous time step INLINEFORM2 as follows: DISPLAYFORM0
where INLINEFORM0 is the RNN function with weight parameter INLINEFORM1 , INLINEFORM2 represents the hidden state at t- INLINEFORM3 time step, and INLINEFORM4 represents the t- INLINEFORM5 MFCC features in INLINEFORM6 . After encoding the audio signal INLINEFORM7 with the RNN, the last hidden state of the RNN, INLINEFORM8 , is considered to be the representative vector that contains all of the sequential audio data. This vector is then concatenated with another prosodic feature vector, INLINEFORM9 , to generate a more informative vector representation of the signal, INLINEFORM10 . The MFCC and the prosodic features are extracted from the audio signal using the openSMILE toolkit BIBREF26 , INLINEFORM11 , respectively. Finally, the emotion class is predicted by applying the softmax function to the vector INLINEFORM12 . For a given audio sample INLINEFORM13 , we assume that INLINEFORM14 is the true label vector, which contains all zeros but contains a one at the correct class, and INLINEFORM15 is the predicted probability distribution from the softmax layer. The training objective then takes the following form: DISPLAYFORM0
where INLINEFORM0 is the calculated representative vector of the audio signal with dimensionality INLINEFORM1 . The INLINEFORM2 and the bias INLINEFORM3 are learned model parameters. C is the total number of classes, and N is the total number of samples used in training. The upper part of Figure shows the architecture of the ARE model.
Text Recurrent Encoder (TRE)
We assume that speech transcripts can be extracted from audio signals with high accuracy, given the advancement of ASR technologies BIBREF7 . We attempt to use the processed textual information as another modality in predicting the emotion class of a given signal. To use textual information, a speech transcript is tokenized and indexed into a sequence of tokens using the Natural Language Toolkit (NLTK) BIBREF27 . Each token is then passed through a word-embedding layer that converts a word index to a corresponding 300-dimensional vector that contains additional contextual meaning between words. The sequence of embedded tokens is fed into a text recurrent encoder (TRE) in such a way that the audio MFCC features are encoded using the ARE represented by equation EQREF2 . In this case, INLINEFORM0 is the t- INLINEFORM1 embedded token from the text input. Finally, the emotion class is predicted from the last hidden state of the text-RNN using the softmax function.
We use the same training objective as the ARE model, and the predicted probability distribution for the target class is as follows: DISPLAYFORM0
where INLINEFORM0 is last hidden state of the text-RNN, INLINEFORM1 , and the INLINEFORM2 and bias INLINEFORM3 are learned model parameters. The lower part of Figure indicates the architecture of the TRE model.
Multimodal Dual Recurrent Encoder (MDRE)
We present a novel architecture called the multimodal dual recurrent encoder (MDRE) to overcome the limitations of existing approaches. In this study, we consider multiple modalities, such as MFCC features, prosodic features and transcripts, which contain sequential audio information, statistical audio information and textual information, respectively. These types of data are the same as those used in the ARE and TRE cases. The MDRE model employs two RNNs to encode data from the audio signal and textual inputs independently. The audio-RNN encodes MFCC features from the audio signal using equation EQREF2 . The last hidden state of the audio-RNN is concatenated with the prosodic features to form the final vector representation INLINEFORM0 , and this vector is then passed through a fully connected neural network layer to form the audio encoding vector A. On the other hand, the text-RNN encodes the word sequence of the transcript using equation EQREF2 . The final hidden states of the text-RNN are also passed through another fully connected neural network layer to form a textual encoding vector T. Finally, the emotion class is predicted by applying the softmax function to the concatenation of the vectors A and T. We use the same training objective as the ARE model, and the predicted probability distribution for the target class is as follows: DISPLAYFORM0
where INLINEFORM0 is the feed-forward neural network with weight parameter INLINEFORM1 , and INLINEFORM2 , INLINEFORM3 are final encoding vectors from the audio-RNN and text-RNN, respectively. INLINEFORM4 and the bias INLINEFORM5 are learned model parameters.
Multimodal Dual Recurrent Encoder with Attention (MDREA)
Inspired by the concept of the attention mechanism used in neural machine translation BIBREF28 , we propose a novel multimodal attention method to focus on the specific parts of a transcript that contain strong emotional information, conditioning on the audio information. Figure shows the architecture of the MDREA model. First, the audio data and text data are encoded with the audio-RNN and text-RNN using equation EQREF2 . We then consider the final audio encoding vector INLINEFORM0 as a context vector. As seen in equation EQREF9 , during each time step t, the dot product between the context vector e and the hidden state of the text-RNN at each t-th sequence INLINEFORM1 is evaluated to calculate a similarity score INLINEFORM2 . Using this score INLINEFORM3 as a weight parameter, the weighted sum of the sequences of the hidden state of the text-RNN, INLINEFORM4 , is calculated to generate an attention-application vector Z. This attention-application vector is concatenated with the final encoding vector of the audio-RNN INLINEFORM5 (equation EQREF7 ), which will be passed through the softmax function to predict the emotion class. We use the same training objective as the ARE model, and the predicted probability distribution for the target class is as follows: DISPLAYFORM0
where INLINEFORM0 and the bias INLINEFORM1 are learned model parameters.
Dataset
We evaluate our model using the Interactive Emotional Dyadic Motion Capture (IEMOCAP) BIBREF18 dataset. This dataset was collected following theatrical theory in order to simulate natural dyadic interactions between actors. We use categorical evaluations with majority agreement. We use only four emotional categories happy, sad, angry, and neutral to compare the performance of our model with other research using the same categories. The IEMOCAP dataset includes five sessions, and each session contains utterances from two speakers (one male and one female). This data collection process resulted in 10 unique speakers. For consistent comparison with previous work, we merge the excitement dataset with the happiness dataset. The final dataset contains a total of 5531 utterances (1636 happy, 1084 sad, 1103 angry, 1708 neutral).
Feature extraction
To extract speech information from audio signals, we use MFCC values, which are widely used in analyzing audio signals. The MFCC feature set contains a total of 39 features, which include 12 MFCC parameters (1-12) from the 26 Mel-frequency bands and log-energy parameters, 13 delta and 13 acceleration coefficients The frame size is set to 25 ms at a rate of 10 ms with the Hamming function. According to the length of each wave file, the sequential step of the MFCC features is varied. To extract additional information from the data, we also use prosodic features, which show effectiveness in affective computing. The prosodic features are composed of 35 features, which include the F0 frequency, the voicing probability, and the loudness contours. All of these MFCC and prosodic features are extracted from the data using the OpenSMILE toolkit BIBREF26 .
Implementation details
Among the variants of the RNN function, we use GRUs as they yield comparable performance to that of the LSTM and include a smaller number of weight parameters BIBREF29 . We use a max encoder step of 750 for the audio input, based on the implementation choices presented in BIBREF30 and 128 for the text input because it covers the maximum length of the transcripts. The vocabulary size of the dataset is 3,747, including the “_UNK_" token, which represents unknown words, and the “_PAD_" token, which is used to indicate padding information added while preparing mini-batch data. The number of hidden units and the number of layers in the RNN for each model (ARE, TRE, MDRE and MDREA) are selected based on extensive hyperparameter search experiments. The weights of the hidden units are initialized using orthogonal weights BIBREF31 ], and the text embedding layer is initialized from pretrained word-embedding vectors BIBREF32 .
In preparing the textual dataset, we first use the released transcripts of the IEMOCAP dataset for simplicity. To investigate the practical performance, we then process all of the IEMOCAP audio data using an ASR system (the Google Cloud Speech API) and retrieve the transcripts. The performance of the Google ASR system is reflected by its word error rate (WER) of 5.53%.
Performance evaluation
As the dataset is not explicitly split beforehand into training, development, and testing sets, we perform 5-fold cross validation to determine the overall performance of the model. The data in each fold are split into training, development, and testing datasets (8:0.5:1.5, respectively). After training the model, we measure the weighted average precision (WAP) over the 5-fold dataset. We train and evaluate the model 10 times per fold, and the model performance is assessed in terms of the mean score and standard deviation.
We examine the WAP values, which are shown in Table 1. First, our ARE model shows the baseline performance because we use minimal audio features, such as the MFCC and prosodic features with simple architectures. On the other hand, the TRE model shows higher performance gain compared to the ARE. From this result, we note that textual data are informative in emotion prediction tasks, and the recurrent encoder model is effective in understanding these types of sequential data. Second, the newly proposed model, MDRE, shows a substantial performance gain. It thus achieves the state-of-the-art performance with a WAP value of 0.718. This result shows that multimodal information is a key factor in affective computing. Lastly, the attention model, MDREA, also outperforms the best existing research results (WAP 0.690 to 0.688) BIBREF19 . However, the MDREA model does not match the performance of the MDRE model, even though it utilizes a more complex architecture. We believe that this result arises because insufficient data are available to properly determine the complex model parameters in the MDREA model. Moreover, we presume that this model will show better performance when the audio signals are aligned with the textual sequence while applying the attention mechanism. We leave the implementation of this point as a future research direction.
To investigate the practical performance of the proposed models, we conduct further experiments with the ASR-processed transcript data (see “-ASR” models in Table ). The label accuracy of the processed transcripts is 5.53% WER. The TRE-ASR, MDRE-ASR and MDREA-ASR models reflect degraded performance compared to that of the TRE, MDRE and MDREA models. However, the performance of these models is still competitive; in particular, the MDRE-ASR model outperforms the previous best-performing model, 3CNN-LSTM10H (WAP 0.691 to 0.688).
Error analysis
We analyze the predictions of the ARE, TRE, and MDRE models. Figure shows the confusion matrix of each model. The ARE model (Fig. ) incorrectly classifies most instances of happy as neutral (43.51%); thus, it shows reduced accuracy (35.15%) in predicting the the happy class. Overall, most of the emotion classes are frequently confused with the neutral class. This observation is in line with the findings of BIBREF30 , who noted that the neutral class is located in the center of the activation-valence space, complicating its discrimination from the other classes.
Interestingly, the TRE model (Fig. ) shows greater prediction gains in predicting the happy class when compared to the ARE model (35.15% to 75.73%). This result seems plausible because the model can benefit from the differences among the distributions of words in happy and neutral expressions, which gives more emotional information to the model than that of the audio signal data. On the other hand, it is striking that the TRE model incorrectly predicts instances of the sad class as the happy class 16.20% of the time, even though these emotional states are opposites of one another.
The MDRE model (Fig. ) compensates for the weaknesses of the previous two models (ARE and TRE) and benefits from their strengths to a surprising degree. The values arranged along the diagonal axis show that all of the accuracies of the correctly predicted class have increased. Furthermore, the occurrence of the incorrect “sad-to-happy" cases in the TRE model is reduced from 16.20% to 9.15%.
Conclusions
In this paper, we propose a novel multimodal dual recurrent encoder model that simultaneously utilizes text data, as well as audio signals, to permit the better understanding of speech data. Our model encodes the information from audio and text sequences using dual RNNs and then combines the information from these sources using a feed-forward neural model to predict the emotion class. Extensive experiments show that our proposed model outperforms other state-of-the-art methods in classifying the four emotion categories, and accuracies ranging from 68.8% to 71.8% are obtained when the model is applied to the IEMOCAP dataset. In particular, it resolves the issue in which predictions frequently incorrectly yield the neutral class, as occurs in previous models that focus on audio features.
In the future work, we aim to extend the modalities to audio, text and video inputs. Furthermore, we plan to investigate the application of the attention mechanism to data derived from multiple modalities. This approach seems likely to uncover enhanced learning schemes that will increase performance in both speech emotion recognition and other multimodal classification tasks.
Acknowledgments
K. Jung is with the Department of Electrical and Computer Engineering, ASRI, Seoul National University, Seoul, Korea. This work was supported by the Ministry of Trade, Industry & Energy (MOTIE, Korea) under Industrial Technology Innovation Program (No.10073144).
| [
"combines the information from these sources using a feed-forward neural model",
"encodes the information from audio and text sequences using dual RNNs and then combines the information from these sources using a feed-forward neural model"
] | 3,201 | qasper | en | null | 7e2663bb13493e18205f6aab469fee8b2a9df281a0ba0e2a | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How do they combine audio and text sequences in their RNN?
Answer: | [
"combines the information from these sources using a feed-forward neural model",
"encodes the information from audio and text sequences using dual RNNs and then combines the information from these sources using a feed-forward neural model"
] | qasper | 128 | 61 | |
by how much did their model improve? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Text simplification aims to reduce the lexical and structural complexity of a text, while still retaining the semantic meaning, which can help children, non-native speakers, and people with cognitive disabilities, to understand text better. One of the methods of automatic text simplification can be generally divided into three categories: lexical simplification (LS) BIBREF0 , BIBREF1 , rule-based BIBREF2 , and machine translation (MT) BIBREF3 , BIBREF4 . LS is mainly used to simplify text by substituting infrequent and difficult words with frequent and easier words. However, there are several challenges for the LS approach: a great number of transformation rules are required for reasonable coverage and should be applied based on the specific context; third, the syntax and semantic meaning of the sentence is hard to retain. Rule-based approaches use hand-crafted rules for lexical and syntactic simplification, for example, substituting difficult words in a predefined vocabulary. However, such approaches need a lot of human-involvement to manually define these rules, and it is impossible to give all possible simplification rules. MT-based approach has attracted great attention in the last several years, which addresses text simplification as a monolingual machine translation problem translating from 'ordinary' and 'simplified' sentences.
In recent years, neural Machine Translation (NMT) is a newly-proposed deep learning approach and achieves very impressive results BIBREF5 , BIBREF6 , BIBREF7 . Unlike the traditional phrased-based machine translation system which operates on small components separately, NMT system is being trained end-to-end, without the need to have external decoders, language models or phrase tables. Therefore, the existing architectures in NMT are used for text simplification BIBREF8 , BIBREF4 . However, most recent work using NMT is limited to the training data that are scarce and expensive to build. Language models trained on simplified corpora have played a central role in statistical text simplification BIBREF9 , BIBREF10 . One main reason is the amount of available simplified corpora typically far exceeds the amount of parallel data. The performance of models can be typically improved when trained on more data. Therefore, we expect simplified corpora to be especially helpful for NMT models.
In contrast to previous work, which uses the existing NMT models, we explore strategy to include simplified training corpora in the training process without changing the neural network architecture. We first propose to pair simplified training sentences with synthetic ordinary sentences during training, and treat this synthetic data as additional training data. We obtain synthetic ordinary sentences through back-translation, i.e. an automatic translation of the simplified sentence into the ordinary sentence BIBREF11 . Then, we mix the synthetic data into the original (simplified-ordinary) data to train NMT model. Experimental results on two publicly available datasets show that we can improve the text simplification quality of NMT models by mixing simplified sentences into the training set over NMT model only using the original training data.
Related Work
Automatic TS is a complicated natural language processing (NLP) task, which consists of lexical and syntactic simplification levels BIBREF12 . It has attracted much attention recently as it could make texts more accessible to wider audiences, and used as a pre-processing step, improve performances of various NLP tasks and systems BIBREF13 , BIBREF14 , BIBREF15 . Usually, hand-crafted, supervised, and unsupervised methods based on resources like English Wikipedia and Simple English Wikipedia (EW-SEW) BIBREF10 are utilized for extracting simplification rules. It is very easy to mix up the automatic TS task and the automatic summarization task BIBREF3 , BIBREF16 , BIBREF6 . TS is different from text summarization as the focus of text summarization is to reduce the length and redundant content.
At the lexical level, lexical simplification systems often substitute difficult words using more common words, which only require a large corpus of regular text to obtain word embeddings to get words similar to the complex word BIBREF1 , BIBREF9 . Biran et al. BIBREF0 adopted an unsupervised method for learning pairs of complex and simpler synonyms from a corpus consisting of Wikipedia and Simple Wikipedia. At the sentence level, a sentence simplification model was proposed by tree transformation based on statistical machine translation (SMT) BIBREF3 . Woodsend and Lapata BIBREF17 presented a data-driven model based on a quasi-synchronous grammar, a formalism that can naturally capture structural mismatches and complex rewrite operations. Wubben et al. BIBREF18 proposed a phrase-based machine translation (PBMT) model that is trained on ordinary-simplified sentence pairs. Xu et al. BIBREF19 proposed a syntax-based machine translation model using simplification-specific objective functions and features to encourage simpler output.
Compared with SMT, neural machine translation (NMT) has shown to produce state-of-the-art results BIBREF5 , BIBREF7 . The central approach of NMT is an encoder-decoder architecture implemented by recurrent neural networks, which can represent the input sequence as a vector, and then decode that vector into an output sequence. Therefore, NMT models were used for text simplification task, and achieved good results BIBREF8 , BIBREF4 , BIBREF20 . The main limitation of the aforementioned NMT models for text simplification depended on the parallel ordinary-simplified sentence pairs. Because ordinary-simplified sentence pairs are expensive and time-consuming to build, the available largest data is EW-SEW that only have 296,402 sentence pairs. The dataset is insufficiency for NMT model if we want to NMT model can obtain the best parameters. Considering simplified data plays an important role in boosting fluency for phrase-based text simplification, and we investigate the use of simplified data for text simplification. We are the first to show that we can effectively adapt neural translation models for text simplifiation with simplified corpora.
Simplified Corpora
We collected a simplified dataset from Simple English Wikipedia that are freely available, which has been previously used for many text simplification methods BIBREF0 , BIBREF10 , BIBREF3 . The simple English Wikipedia is pretty easy to understand than normal English Wikipedia. We downloaded all articles from Simple English Wikipedia. For these articles, we removed stubs, navigation pages and any article that consisted of a single sentence. We then split them into sentences with the Stanford CorNLP BIBREF21 , and deleted these sentences whose number of words are smaller than 10 or large than 40. After removing repeated sentences, we chose 600K sentences as the simplified data with 11.6M words, and the size of vocabulary is 82K.
Text Simplification using Neural Machine Translation
Our work is built on attention-based NMT BIBREF5 as an encoder-decoder network with recurrent neural networks (RNN), which simultaneously conducts dynamic alignment and generation of the target simplified sentence.
The encoder uses a bidirectional RNN that consists of forward and backward RNN. Given a source sentence INLINEFORM0 , the forward RNN and backward RNN calculate forward hidden states INLINEFORM1 and backward hidden states INLINEFORM2 , respectively. The annotation vector INLINEFORM3 is obtained by concatenating INLINEFORM4 and INLINEFORM5 .
The decoder is a RNN that predicts a target simplificated sentence with Gated Recurrent Unit (GRU) BIBREF22 . Given the previously generated target (simplified) sentence INLINEFORM0 , the probability of next target word INLINEFORM1 is DISPLAYFORM0
where INLINEFORM0 is a non-linear function, INLINEFORM1 is the embedding of INLINEFORM2 , and INLINEFORM3 is a decoding state for time step INLINEFORM4 .
State INLINEFORM0 is calculated by DISPLAYFORM0
where INLINEFORM0 is the activation function GRU.
The INLINEFORM0 is the context vector computed as a weighted annotation INLINEFORM1 , computed by DISPLAYFORM0
where the weight INLINEFORM0 is computed by DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 , INLINEFORM1 and INLINEFORM2 are weight matrices. The training objective is to maximize the likelihood of the training data. Beam search is employed for decoding.
Synthetic Simplified Sentences
We train an auxiliary system using NMT model from the simplified sentence to the ordinary sentence, which is first trained on the available parallel data. For leveraging simplified sentences to improve the quality of NMT model for text simplification, we propose to adapt the back-translation approach proposed by Sennrich et al. BIBREF11 to our scenario. More concretely, Given one sentence in simplified sentences, we use the simplified-ordinary system in translate mode with greedy decoding to translate it to the ordinary sentences, which is denoted as back-translation. This way, we obtain a synthetic parallel simplified-ordinary sentences. Both the synthetic sentences and the available parallel data are used as training data for the original NMT system.
Evaluation
We evaluate the performance of text simplification using neural machine translation on available parallel sentences and additional simplified sentences.
Dataset. We use two simplification datasets (WikiSmall and WikiLarge). WikiSmall consists of ordinary and simplified sentences from the ordinary and simple English Wikipedias, which has been used as benchmark for evaluating text simplification BIBREF17 , BIBREF18 , BIBREF8 . The training set has 89,042 sentence pairs, and the test set has 100 pairs. WikiLarge is also from Wikipedia corpus whose training set contains 296,402 sentence pairs BIBREF19 , BIBREF20 . WikiLarge includes 8 (reference) simplifications for 2,359 sentences split into 2,000 for development and 359 for testing.
Metrics. Three metrics in text simplification are chosen in this paper. BLEU BIBREF5 is one traditional machine translation metric to assess the degree to which translated simplifications differed from reference simplifications. FKGL measures the readability of the output BIBREF23 . A small FKGL represents simpler output. SARI is a recent text-simplification metric by comparing the output against the source and reference simplifications BIBREF20 .
We evaluate the output of all systems using human evaluation. The metric is denoted as Simplicity BIBREF8 . The three non-native fluent English speakers are shown reference sentences and output sentences. They are asked whether the output sentence is much simpler (+2), somewhat simpler (+1), equally (0), somewhat more difficult (-1), and much more difficult (-2) than the reference sentence.
Methods. We use OpenNMT BIBREF24 as the implementation of the NMT system for all experiments BIBREF5 . We generally follow the default settings and training procedure described by Klein et al.(2017). We replace out-of-vocabulary words with a special UNK symbol. At prediction time, we replace UNK words with the highest probability score from the attention layer. OpenNMT system used on parallel data is the baseline system. To obtain a synthetic parallel training set, we back-translate a random sample of 100K sentences from the collected simplified corpora. OpenNMT used on parallel data and synthetic data is our model. The benchmarks are run on a Intel(R) Core(TM) i7-5930K CPU@3.50GHz, 32GB Mem, trained on 1 GPU GeForce GTX 1080 (Pascal) with CUDA v. 8.0.
We choose three statistical text simplification systems. PBMT-R is a phrase-based method with a reranking post-processing step BIBREF18 . Hybrid performs sentence splitting and deletion operations based on discourse representation structures, and then simplifies sentences with PBMT-R BIBREF25 . SBMT-SARI BIBREF19 is syntax-based translation model using PPDB paraphrase database BIBREF26 and modifies tuning function (using SARI). We choose two neural text simplification systems. NMT is a basic attention-based encoder-decoder model which uses OpenNMT framework to train with two LSTM layers, hidden states of size 500 and 500 hidden units, SGD optimizer, and a dropout rate of 0.3 BIBREF8 . Dress is an encoder-decoder model coupled with a deep reinforcement learning framework, and the parameters are chosen according to the original paper BIBREF20 . For the experiments with synthetic parallel data, we back-translate a random sample of 60 000 sentences from the collected simplified sentences into ordinary sentences. Our model is trained on synthetic data and the available parallel data, denoted as NMT+synthetic.
Results. Table 1 shows the results of all models on WikiLarge dataset. We can see that our method (NMT+synthetic) can obtain higher BLEU, lower FKGL and high SARI compared with other models, except Dress on FKGL and SBMT-SARI on SARI. It verified that including synthetic data during training is very effective, and yields an improvement over our baseline NMF by 2.11 BLEU, 1.7 FKGL and 1.07 SARI. We also substantially outperform Dress, who previously reported SOTA result. The results of our human evaluation using Simplicity are also presented in Table 1. NMT on synthetic data is significantly better than PBMT-R, Dress, and SBMT-SARI on Simplicity. It indicates that our method with simplified data is effective at creating simpler output.
Results on WikiSmall dataset are shown in Table 2. We see substantial improvements (6.37 BLEU) than NMT from adding simplified training data with synthetic ordinary sentences. Compared with statistical machine translation models (PBMT-R, Hybrid, SBMT-SARI), our method (NMT+synthetic) still have better results, but slightly worse FKGL and SARI. Similar to the results in WikiLarge, the results of our human evaluation using Simplicity outperforms the other models. In conclusion, Our method produces better results comparing with the baselines, which demonstrates the effectiveness of adding simplified training data.
Conclusion
In this paper, we propose one simple method to use simplified corpora during training of NMT systems, with no changes to the network architecture. In the experiments on two datasets, we achieve substantial gains in all tasks, and new SOTA results, via back-translation of simplified sentences into the ordinary sentences, and treating this synthetic data as additional training data. Because we do not change the neural network architecture to integrate simplified corpora, our method can be easily applied to other Neural Text Simplification (NTS) systems. We expect that the effectiveness of our method not only varies with the quality of the NTS system used for back-translation, but also depends on the amount of available parallel and simplified corpora. In the paper, we have only utilized data from Wikipedia for simplified sentences. In the future, many other text sources are available and the impact of not only size, but also of domain should be investigated.
| [
"For the WikiLarge dataset, the improvement over baseline NMT is 2.11 BLEU, 1.7 FKGL and 1.07 SARI.\nFor the WikiSmall dataset, the improvement over baseline NMT is 8.37 BLEU.",
"6.37 BLEU"
] | 2,271 | qasper | en | null | 6b9d0c1e63714067fc0b54c024dc065e63f64cbaf44ab4d0 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: by how much did their model improve?
Answer: | [
"For the WikiLarge dataset, the improvement over baseline NMT is 2.11 BLEU, 1.7 FKGL and 1.07 SARI.\nFor the WikiSmall dataset, the improvement over baseline NMT is 8.37 BLEU.",
"6.37 BLEU"
] | qasper | 128 | 62 | |
how many humans evaluated the results? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Machine translation has made remarkable progress, and studies claiming it to reach a human parity are starting to appear BIBREF0. However, when evaluating translations of the whole documents rather than isolated sentences, human raters show a stronger preference for human over machine translation BIBREF1. These findings emphasize the need to shift towards context-aware machine translation both from modeling and evaluation perspective.
Most previous work on context-aware NMT assumed that either all the bilingual data is available at the document level BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10 or at least its fraction BIBREF11. But in practical scenarios, document-level parallel data is often scarce, which is one of the challenges when building a context-aware system.
We introduce an approach to context-aware machine translation using only monolingual document-level data. In our setting, a separate monolingual sequence-to-sequence model (DocRepair) is used to correct sentence-level translations of adjacent sentences. The key idea is to use monolingual data to imitate typical inconsistencies between context-agnostic translations of isolated sentences. The DocRepair model is trained to map inconsistent groups of sentences into consistent ones. The consistent groups come from the original training data; the inconsistent groups are obtained by sampling round-trip translations for each isolated sentence.
To validate the performance of our model, we use three kinds of evaluation: the BLEU score, contrastive evaluation of translation of several discourse phenomena BIBREF11, and human evaluation. We show strong improvements for all metrics.
We analyze which discourse phenomena are hard to capture using monolingual data only. Using contrastive test sets for targeted evaluation of several contextual phenomena, we compare the performance of the models trained on round-trip translations and genuine document-level parallel data. Among the four phenomena in the test sets we use (deixis, lexical cohesion, VP ellipsis and ellipsis which affects NP inflection) we find VP ellipsis to be the hardest phenomenon to be captured using round-trip translations.
Our key contributions are as follows:
we introduce the first approach to context-aware machine translation using only monolingual document-level data;
our approach shows substantial improvements in translation quality as measured by BLEU, targeted contrastive evaluation of several discourse phenomena and human evaluation;
we show which discourse phenomena are hard to capture using monolingual data only.
Our Approach: Document-level Repair
We propose a monolingual DocRepair model to correct inconsistencies between sentence-level translations of a context-agnostic MT system. It does not use any states of a trained MT model whose outputs it corrects and therefore can in principle be trained to correct translations from any black-box MT system.
The DocRepair model requires only monolingual document-level data in the target language. It is a monolingual sequence-to-sequence model that maps inconsistent groups of sentences into consistent ones. Consistent groups come from monolingual document-level data. To obtain inconsistent groups, each sentence in a group is replaced with its round-trip translation produced in isolation from context. More formally, forming a training minibatch for the DocRepair model involves the following steps (see also Figure FIGREF9):
sample several groups of sentences from the monolingual data;
for each sentence in a group, (i) translate it using a target-to-source MT model, (ii) sample a translation of this back-translated sentence in the source language using a source-to-target MT model;
using these round-trip translations of isolated sentences, form an inconsistent version of the initial groups;
use inconsistent groups as input for the DocRepair model, consistent ones as output.
At test time, the process of getting document-level translations is two-step (Figure FIGREF10):
produce translations of isolated sentences using a context-agnostic MT model;
apply the DocRepair model to a sequence of context-agnostic translations to correct inconsistencies between translations.
In the scope of the current work, the DocRepair model is the standard sequence-to-sequence Transformer. Sentences in a group are concatenated using a reserved token-separator between sentences. The Transformer is trained to correct these long inconsistent pseudo-sentences into consistent ones. The token-separator is then removed from corrected translations.
Evaluation of Contextual Phenomena
We use contrastive test sets for evaluation of discourse phenomena for English-Russian by BIBREF11. These test sets allow for testing different kinds of phenomena which, as we show, can be captured from monolingual data with varying success. In this section, we provide test sets statistics and briefly describe the tested phenomena. For more details, the reader is referred to BIBREF11.
Evaluation of Contextual Phenomena ::: Test sets
There are four test sets in the suite. Each test set contains contrastive examples. It is specifically designed to test the ability of a system to adapt to contextual information and handle the phenomenon under consideration. Each test instance consists of a true example (a sequence of sentences and their reference translation from the data) and several contrastive translations which differ from the true one only in one specific aspect. All contrastive translations are correct and plausible translations at the sentence level, and only context reveals the inconsistencies between them. The system is asked to score each candidate translation, and we compute the system accuracy as the proportion of times the true translation is preferred to the contrastive ones. Test set statistics are shown in Table TABREF15. The suites for deixis and lexical cohesion are split into development and test sets, with 500 examples from each used for validation purposes and the rest for testing. Convergence of both consistency scores on these development sets and BLEU score on a general development set are used as early stopping criteria in models training. For ellipsis, there is no dedicated development set, so we evaluate on all the ellipsis data and do not use it for development.
Evaluation of Contextual Phenomena ::: Phenomena overview
Deixis Deictic words or phrases, are referential expressions whose denotation depends on context. This includes personal deixis (“I”, “you”), place deixis (“here”, “there”), and discourse deixis, where parts of the discourse are referenced (“that's a good question”). The test set examples are all related to person deixis, specifically the T-V distinction between informal and formal you (Latin “tu” and “vos”) in the Russian translations, and test for consistency in this respect.
Ellipsis Ellipsis is the omission from a clause of one or more words that are nevertheless understood in the context of the remaining elements. In machine translation, elliptical constructions in the source language pose a problem in two situations. First, if the target language does not allow the same types of ellipsis, requiring the elided material to be predicted from context. Second, if the elided material affects the syntax of the sentence. For example, in Russian the grammatical function of a noun phrase, and thus its inflection, may depend on the elided verb, or, conversely, the verb inflection may depend on the elided subject.
There are two different test sets for ellipsis. One contains examples where a morphological form of a noun group in the last sentence can not be understood without context beyond the sentence level (“ellipsis (infl.)” in Table TABREF15). Another includes cases of verb phrase ellipsis in English, which does not exist in Russian, thus requires predicting the verb when translating into Russian (“ellipsis (VP)” in Table TABREF15).
Lexical cohesion The test set focuses on reiteration of named entities. Where several translations of a named entity are possible, a model has to prefer consistent translations over inconsistent ones.
Experimental Setup ::: Data preprocessing
We use the publicly available OpenSubtitles2018 corpus BIBREF12 for English and Russian. For a fair comparison with previous work, we train the baseline MT system on the data released by BIBREF11. Namely, our MT system is trained on 6m instances. These are sentence pairs with a relative time overlap of subtitle frames between source and target language subtitles of at least $0.9$.
We gathered 30m groups of 4 consecutive sentences as our monolingual data. We used only documents not containing groups of sentences from general development and test sets as well as from contrastive test sets. The main results we report are for the model trained on all 30m fragments.
We use the tokenization provided by the corpus and use multi-bleu.perl on lowercased data to compute BLEU score. We use beam search with a beam of 4.
Sentences were encoded using byte-pair encoding BIBREF13, with source and target vocabularies of about 32000 tokens. Translation pairs were batched together by approximate sequence length. Each training batch contained a set of translation pairs containing approximately 15000 source tokens. It has been shown that Transformer's performance depends heavily on batch size BIBREF14, and we chose a large batch size to ensure the best performance. In training context-aware models, for early stopping we use both convergence in BLEU score on the general development set and scores on the consistency development sets. After training, we average the 5 latest checkpoints.
Experimental Setup ::: Models
The baseline model, the model used for back-translation, and the DocRepair model are all Transformer base models BIBREF15. More precisely, the number of layers is $N=6$ with $h = 8$ parallel attention layers, or heads. The dimensionality of input and output is $d_{model} = 512$, and the inner-layer of a feed-forward networks has dimensionality $d_{ff}=2048$. We use regularization as described in BIBREF15.
As a second baseline, we use the two-pass CADec model BIBREF11. The first pass produces sentence-level translations. The second pass takes both the first-pass translation and representations of the context sentences as input and returns contextualized translations. CADec requires document-level parallel training data, while DocRepair only needs monolingual training data.
Experimental Setup ::: Generating round-trip translations
On the selected 6m instances we train sentence-level translation models in both directions. To create training data for DocRepair, we proceed as follows. The Russian monolingual data is first translated into English, using the Russian$\rightarrow $English model and beam search with beam size of 4. Then, we use the English$\rightarrow $Russian model to sample translations with temperature of $0{.}5$. For each sentence, we precompute 20 sampled translations and randomly choose one of them when forming a training minibatch for DocRepair. Also, in training, we replace each token in the input with a random one with the probability of $10\%$.
Experimental Setup ::: Optimizer
As in BIBREF15, we use the Adam optimizer BIBREF16, the parameters are $\beta _1 = 0{.}9$, $\beta _2 = 0{.}98$ and $\varepsilon = 10^{-9}$. We vary the learning rate over the course of training using the formula:
where $warmup\_steps = 16000$ and $scale=4$.
Results ::: General results
The BLEU scores are provided in Table TABREF24 (we evaluate translations of 4-sentence fragments). To see which part of the improvement is due to fixing agreement between sentences rather than simply sentence-level post-editing, we train the same repair model at the sentence level. Each sentence in a group is now corrected separately, then they are put back together in a group. One can see that most of the improvement comes from accounting for extra-sentential dependencies. DocRepair outperforms the baseline and CADec by 0.7 BLEU, and its sentence-level repair version by 0.5 BLEU.
Results ::: Consistency results
Scores on the phenomena test sets are provided in Table TABREF26. For deixis, lexical cohesion and ellipsis (infl.) we see substantial improvements over both the baseline and CADec. The largest improvement over CADec (22.5 percentage points) is for lexical cohesion. However, there is a drop of almost 5 percentage points for VP ellipsis. We hypothesize that this is because it is hard to learn to correct inconsistencies in translations caused by VP ellipsis relying on monolingual data alone. Figure FIGREF27(a) shows an example of inconsistency caused by VP ellipsis in English. There is no VP ellipsis in Russian, and when translating auxiliary “did” the model has to guess the main verb. Figure FIGREF27(b) shows steps of generating round-trip translations for the target side of the previous example. When translating from Russian, main verbs are unlikely to be translated as the auxiliary “do” in English, and hence the VP ellipsis is rarely present on the English side. This implies the model trained using the round-trip translations will not be exposed to many VP ellipsis examples in training. We discuss this further in Section SECREF34.
Table TABREF28 provides scores for deixis and lexical cohesion separately for different distances between sentences requiring consistency. It can be seen, that the performance of DocRepair degrades less than that of CADec when the distance between sentences requiring consistency gets larger.
Results ::: Human evaluation
We conduct a human evaluation on random 700 examples from our general test set. We picked only examples where a DocRepair translation is not a full copy of the baseline one.
The annotators were provided an original group of sentences in English and two translations: baseline context-agnostic one and the one corrected by the DocRepair model. Translations were presented in random order with no indication which model they came from. The task is to pick one of the three options: (1) the first translation is better, (2) the second translation is better, (3) the translations are of equal quality. The annotators were asked to avoid the third answer if they are able to give preference to one of the translations. No other guidelines were given.
The results are provided in Table TABREF30. In about $52\%$ of the cases annotators marked translations as having equal quality. Among the cases where one of the translations was marked better than the other, the DocRepair translation was marked better in $73\%$ of the cases. This shows a strong preference of the annotators for corrected translations over the baseline ones.
Varying Training Data
In this section, we discuss the influence of the training data chosen for document-level models. In all experiments, we used the DocRepair model.
Varying Training Data ::: The amount of training data
Table TABREF33 provides BLEU and consistency scores for the DocRepair model trained on different amount of data. We see that even when using a dataset of moderate size (e.g., 5m fragments) we can achieve performance comparable to the model trained on a large amount of data (30m fragments). Moreover, we notice that deixis scores are less sensitive to the amount of training data than lexical cohesion and ellipsis scores. The reason might be that, as we observed in our previous work BIBREF11, inconsistencies in translations due to the presence of deictic words and phrases are more frequent in this dataset than other types of inconsistencies. Also, as we show in Section SECREF7, this is the phenomenon the model learns faster in training.
Varying Training Data ::: One-way vs round-trip translations
In this section, we discuss the limitations of using only monolingual data to model inconsistencies between sentence-level translations. In Section SECREF25 we observed a drop in performance on VP ellipsis for DocRepair compared to CADec, which was trained on parallel data. We hypothesized that this is due to the differences between one-way and round-trip translations, and now we test this hypothesis. To do so, we fix the dataset and vary the way in which the input for DocRepair is generated: round-trip or one-way translations. The latter assumes that document-level data is parallel, and translations are sampled from the source side of the sentences in a group rather than from their back-translations. For parallel data, we take 1.5m parallel instances which were used for CADec training and add 1m instances from our monolingual data. For segments in the parallel part, we either sample translations from the source side or use round-trip translations. The results are provided in Table TABREF35.
The model trained on one-way translations is slightly better than the one trained on round-trip translations. As expected, VP ellipsis is the hardest phenomena to be captured using round-trip translations, and the DocRepair model trained on one-way translated data gains 6% accuracy on this test set. This shows that the DocRepair model benefits from having access to non-synthetic English data. This results in exposing DocRepair at training time to Russian translations which suffer from the same inconsistencies as the ones it will have to correct at test time.
Varying Training Data ::: Filtering: monolingual (no filtering) or parallel
Note that the scores of the DocRepair model trained on 2.5m instances randomly chosen from monolingual data (Table TABREF33) are different from the ones for the model trained on 2.5m instances combined from parallel and monolingual data (Table TABREF35). For convenience, we show these two in Table TABREF36.
The domain, the dataset these two data samples were gathered from, and the way we generated training data for DocRepair (round-trip translations) are all the same. The only difference lies in how the data was filtered. For parallel data, as in the previous work BIBREF6, we picked only sentence pairs with large relative time overlap of subtitle frames between source-language and target-language subtitles. This is necessary to ensure the quality of translation data: one needs groups of consecutive sentences in the target language where every sentence has a reliable translation.
Table TABREF36 shows that the quality of the model trained on data which came from the parallel part is worse than the one trained on monolingual data. This indicates that requiring each sentence in a group to have a reliable translation changes the distribution of the data, which might be not beneficial for translation quality and provides extra motivation for using monolingual data.
Learning Dynamics
Let us now look into how the process of DocRepair training progresses. Figure FIGREF38 shows how the BLEU scores with the reference translation and with the baseline context-agnostic translation (i.e. the input for the DocRepair model) are changing during training. First, the model quickly learns to copy baseline translations: the BLEU score with the baseline is very high. Then it gradually learns to change them, which leads to an improvement in BLEU with the reference translation and a drop in BLEU with the baseline. Importantly, the model is reluctant to make changes: the BLEU score between translations of the converged model and the baseline is 82.5. We count the number of changed sentences in every 4-sentence fragment in the test set and plot the histogram in Figure FIGREF38. In over than 20$\%$ of the cases the model has not changed base translations at all. In almost $40\%$, it modified only one sentence and left the remaining 3 sentences unchanged. The model changed more than half sentences in a group in only $14\%$ of the cases. Several examples of the DocRepair translations are shown in Figure FIGREF43.
Figure FIGREF42 shows how consistency scores are changing in training. For deixis, the model achieves the final quality quite quickly; for the rest, it needs a large number of training steps to converge.
Related Work
Our work is most closely related to two lines of research: automatic post-editing (APE) and document-level machine translation.
Related Work ::: Automatic post-editing
Our model can be regarded as an automatic post-editing system – a system designed to fix systematic MT errors that is decoupled from the main MT system. Automatic post-editing has a long history, including rule-based BIBREF17, statistical BIBREF18 and neural approaches BIBREF19, BIBREF20, BIBREF21.
In terms of architectures, modern approaches use neural sequence-to-sequence models, either multi-source architectures that consider both the original source and the baseline translation BIBREF19, BIBREF20, or monolingual repair systems, as in BIBREF21, which is concurrent work to ours. True post-editing datasets are typically small and expensive to create BIBREF22, hence synthetic training data has been created that uses original monolingual data as output for the sequence-to-sequence model, paired with an automatic back-translation BIBREF23 and/or round-trip translation as its input(s) BIBREF19, BIBREF21.
While previous work on automatic post-editing operated on the sentence level, the main novelty of this work is that our DocRepair model operates on groups of sentences and is thus able to fix consistency errors caused by the context-agnostic baseline MT system. We consider this strategy of sentence-level baseline translation and context-aware monolingual repair attractive when parallel document-level data is scarce.
For training, the DocRepair model only requires monolingual document-level data. While we create synthetic training data via round-trip translation similarly to earlier work BIBREF19, BIBREF21, note that we purposefully use sentence-level MT systems for this to create the types of consistency errors that we aim to fix with the context-aware DocRepair model. Not all types of consistency errors that we want to fix emerge from a round-trip translation, so access to parallel document-level data can be useful (Section SECREF34).
Related Work ::: Document-level NMT
Neural models of MT that go beyond the sentence-level are an active research area BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF10, BIBREF9, BIBREF11. Typically, the main MT system is modified to take additional context as its input. One limitation of these approaches is that they assume that parallel document-level training data is available.
Closest to our work are two-pass models for document-level NMT BIBREF24, BIBREF11, where a second, context-aware model takes the translation and hidden representations of the sentence-level first-pass model as its input. The second-pass model can in principle be trained on a subset of the parallel training data BIBREF11, somewhat relaxing the assumption that all training data is at the document level.
Our work is different from this previous work in two main respects. Firstly, we show that consistency can be improved with only monolingual document-level training data. Secondly, the DocRepair model is decoupled from the first-pass MT system, which improves its portability.
Conclusions
We introduce the first approach to context-aware machine translation using only monolingual document-level data. We propose a monolingual DocRepair model to correct inconsistencies between sentence-level translations. The model performs automatic post-editing on a sequence of sentence-level translations, refining translations of sentences in context of each other. Our approach results in substantial improvements in translation quality as measured by BLEU, targeted contrastive evaluation of several discourse phenomena and human evaluation. Moreover, we perform error analysis and detect which discourse phenomena are hard to capture using only monolingual document-level data. While in the current work we used text fragments of 4 sentences, in future work we would like to consider longer contexts.
Acknowledgments
We would like to thank the anonymous reviewers for their comments. The authors also thank David Talbot and Yandex Machine Translation team for helpful discussions and inspiration. Ivan Titov acknowledges support of the European Research Council (ERC StG BroadSem 678254) and the Dutch National Science Foundation (NWO VIDI 639.022.518). Rico Sennrich acknowledges support from the Swiss National Science Foundation (105212_169888), the European Union’s Horizon 2020 research and innovation programme (grant agreement no 825460), and the Royal Society (NAF\R1\180122).
| [
"Unanswerable",
"Unanswerable"
] | 3,711 | qasper | en | null | ea58638e132307bb3f2c24abeb0e2d07eaf162e3e1d12b57 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: how many humans evaluated the results?
Answer: | [
"Unanswerable",
"Unanswerable"
] | qasper | 128 | 63 | |
What is their definition of tweets going viral? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: 10pt
1.10pt
[ Characterizing Political Fake News in Twitter by its Meta-DataJulio Amador Díaz LópezAxel Oehmichen Miguel Molina-Solana( j.amador, axelfrancois.oehmichen11, mmolinas@imperial.ac.uk ) Imperial College London This article presents a preliminary approach towards characterizing political fake news on Twitter through the analysis of their meta-data. In particular, we focus on more than 1.5M tweets collected on the day of the election of Donald Trump as 45th president of the United States of America. We use the meta-data embedded within those tweets in order to look for differences between tweets containing fake news and tweets not containing them. Specifically, we perform our analysis only on tweets that went viral, by studying proxies for users' exposure to the tweets, by characterizing accounts spreading fake news, and by looking at their polarization. We found significant differences on the distribution of followers, the number of URLs on tweets, and the verification of the users.
]
Introduction
While fake news, understood as deliberately misleading pieces of information, have existed since long ago (e.g. it is not unusual to receive news falsely claiming the death of a celebrity), the term reached the mainstream, particularly so in politics, during the 2016 presidential election in the United States BIBREF0 . Since then, governments and corporations alike (e.g. Google BIBREF1 and Facebook BIBREF2 ) have begun efforts to tackle fake news as they can affect political decisions BIBREF3 . Yet, the ability to define, identify and stop fake news from spreading is limited.
Since the Obama campaign in 2008, social media has been pervasive in the political arena in the United States. Studies report that up to 62% of American adults receive their news from social media BIBREF4 . The wide use of platforms such as Twitter and Facebook has facilitated the diffusion of fake news by simplifying the process of receiving content with no significant third party filtering, fact-checking or editorial judgement. Such characteristics make these platforms suitable means for sharing news that, disguised as legit ones, try to confuse readers.
Such use and their prominent rise has been confirmed by Craig Silverman, a Canadian journalist who is a prominent figure on fake news BIBREF5 : “In the final three months of the US presidential campaign, the top-performing fake election news stories on Facebook generated more engagement than the top stories from major news outlet”.
Our current research hence departs from the assumption that social media is a conduit for fake news and asks the question of whether fake news (as spam was some years ago) can be identified, modelled and eventually blocked. In order to do so, we use a sample of more that 1.5M tweets collected on November 8th 2016 —election day in the United States— with the goal of identifying features that tweets containing fake news are likely to have. As such, our paper aims to provide a preliminary characterization of fake news in Twitter by looking into meta-data embedded in tweets. Considering meta-data as a relevant factor of analysis is in line with findings reported by Morris et al. BIBREF6 . We argue that understanding differences between tweets containing fake news and regular tweets will allow researchers to design mechanisms to block fake news in Twitter.
Specifically, our goals are: 1) compare the characteristics of tweets labelled as containing fake news to tweets labelled as not containing them, 2) characterize, through their meta-data, viral tweets containing fake news and the accounts from which they originated, and 3) determine the extent to which tweets containing fake news expressed polarized political views.
For our study, we used the number of retweets to single-out those that went viral within our sample. Tweets within that subset (viral tweets hereafter) are varied and relate to different topics. We consider that a tweet contains fake news if its text falls within any of the following categories described by Rubin et al. BIBREF7 (see next section for the details of such categories): serious fabrication, large-scale hoaxes, jokes taken at face value, slanted reporting of real facts and stories where the truth is contentious. The dataset BIBREF8 , manually labelled by an expert, has been publicly released and is available to researchers and interested parties.
From our results, the following main observations can be made:
Our findings resonate with similar work done on fake news such as the one from Allcot and Gentzkow BIBREF9 . Therefore, even if our study is a preliminary attempt at characterizing fake news on Twitter using only their meta-data, our results provide external validity to previous research. Moreover, our work not only stresses the importance of using meta-data, but also underscores which parameters may be useful to identify fake news on Twitter.
The rest of the paper is organized as follows. The next section briefly discusses where this work is located within the literature on fake news and contextualizes the type of fake news we are studying. Then, we present our hypotheses, the data, and the methodology we follow. Finally, we present our findings, conclusions of this study, and future lines of work.
Defining Fake news
Our research is connected to different strands of academic knowledge related to the phenomenon of fake news. In relation to Computer Science, a recent survey by Conroy and colleagues BIBREF10 identifies two popular approaches to single-out fake news. On the one hand, the authors pointed to linguistic approaches consisting in using text, its linguistic characteristics and machine learning techniques to automatically flag fake news. On the other, these researchers underscored the use of network approaches, which make use of network characteristics and meta-data, to identify fake news.
With respect to social sciences, efforts from psychology, political science and sociology, have been dedicated to understand why people consume and/or believe misinformation BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 . Most of these studies consistently reported that psychological biases such as priming effects and confirmation bias play an important role in people ability to discern misinformation.
In relation to the production and distribution of fake news, a recent paper in the field of Economics BIBREF9 found that most fake news sites use names that resemble those of legitimate organizations, and that sites supplying fake news tend to be short-lived. These authors also noticed that fake news items are more likely shared than legitimate articles coming from trusted sources, and they tend to exhibit a larger level of polarization.
The conceptual issue of how to define fake news is a serious and unresolved issue. As the focus of our work is not attempting to offer light on this, we will rely on work by other authors to describe what we consider as fake news. In particular, we use the categorization provided by Rubin et al. BIBREF7 . The five categories they described, together with illustrative examples from our dataset, are as follows:
Research Hypotheses
Previous works on the area (presented in the section above) suggest that there may be important determinants for the adoption and diffusion of fake news. Our hypotheses builds on them and identifies three important dimensions that may help distinguishing fake news from legit information:
Taking those three dimensions into account, we propose the following hypotheses about the features that we believe can help to identify tweets containing fake news from those not containing them. They will be later tested over our collected dataset.
Exposure.
Characterization.
Polarization.
Data and Methodology
For this study, we collected publicly available tweets using Twitter's public API. Given the nature of the data, it is important to emphasize that such tweets are subject to Twitter's terms and conditions which indicate that users consent to the collection, transfer, manipulation, storage, and disclosure of data. Therefore, we do not expect ethical, legal, or social implications from the usage of the tweets. Our data was collected using search terms related to the presidential election held in the United States on November 8th 2016. Particularly, we queried Twitter's streaming API, more precisely the filter endpoint of the streaming API, using the following hashtags and user handles: #MyVote2016, #ElectionDay, #electionnight, @realDonaldTrump and @HillaryClinton. The data collection ran for just one day (Nov 8th 2016).
One straightforward way of sharing information on Twitter is by using the retweet functionality, which enables a user to share a exact copy of a tweet with his followers. Among the reasons for retweeting, Body et al. BIBREF15 reported the will to: 1) spread tweets to a new audience, 2) to show one’s role as a listener, and 3) to agree with someone or validate the thoughts of others. As indicated, our initial interest is to characterize tweets containing fake news that went viral (as they are the most harmful ones, as they reach a wider audience), and understand how it differs from other viral tweets (that do not contain fake news). For our study, we consider that a tweet went viral if it was retweeted more than 1000 times.
Once we have the dataset of viral tweets, we eliminated duplicates (some of the tweets were collected several times because they had several handles) and an expert manually inspected the text field within the tweets to label them as containing fake news, or not containing them (according to the characterization presented before). This annotated dataset BIBREF8 is publicly available and can be freely reused.
Finally, we use the following fields within tweets (from the ones returned by Twitter's API) to compare their distributions and look for differences between viral tweets containing fake news and viral tweets not containing fake news:
In the following section, we provide graphical descriptions of the distribution of each of the identified attributes for the two sets of tweets (those labelled as containing fake news and those labelled as not containing them). Where appropriate, we normalized and/or took logarithms of the data for better representation. To gain a better understanding of the significance of those differences, we use the Kolmogorov-Smirnov test with the null hypothesis that both distributions are equal.
Results
The sample collected consisted on 1 785 855 tweets published by 848 196 different users. Within our sample, we identified 1327 tweets that went viral (retweeted more than 1000 times by the 8th of November 2016) produced by 643 users. Such small subset of viral tweets were retweeted on 290 841 occasions in the observed time-window.
The 1327 `viral' tweets were manually annotated as containing fake news or not. The annotation was carried out by a single person in order to obtain a consistent annotation throughout the dataset. Out of those 1327 tweets, we identified 136 as potentially containing fake news (according to the categories previously described), and the rest were classified as `non containing fake news'. Note that the categorization is far from being perfect given the ambiguity of fake news themselves and human judgement involved in the process of categorization. Because of this, we do not claim that this dataset can be considered a ground truth.
The following results detail characteristics of these tweets along the previously mentioned dimensions. Table TABREF23 reports the actual differences (together with their associated p-values) of the distributions of viral tweets containing fake news and viral tweets not containing them for every variable considered.
Exposure
Figure FIGREF24 shows that, in contrast to other kinds of viral tweets, those containing fake news were created more recently. As such, Twitter users were exposed to fake news related to the election for a shorter period of time.
However, in terms of retweets, Figure FIGREF25 shows no apparent difference between containing fake news or not containing them. That is confirmed by the Kolmogorov-Smirnoff test, which does not discard the hypothesis that the associated distributions are equal.
In relation to the number of favourites, users that generated at least a viral tweet containing fake news appear to have, on average, less favourites than users that do not generate them. Figure FIGREF26 shows the distribution of favourites. Despite the apparent visual differences, the difference are not statistically significant.
Finally, the number of hashtags used in viral fake news appears to be larger than those in other viral tweets. Figure FIGREF27 shows the density distribution of the number of hashtags used. However, once again, we were not able to find any statistical difference between the average number of hashtags in a viral tweet and the average number of hashtags in viral fake news.
Characterization
We found that 82 users within our sample were spreading fake news (i.e. they produced at least one tweet which was labelled as fake news). Out of those, 34 had verified accounts, and the rest were unverified. From the 48 unverified accounts, 6 have been suspended by Twitter at the date of writing, 3 tried to imitate legitimate accounts of others, and 4 accounts have been already deleted. Figure FIGREF28 shows the proportion of verified accounts to unverified accounts for viral tweets (containing fake news vs. not containing fake news). From the chart, it is clear that there is a higher chance of fake news coming from unverified accounts.
Turning to friends, accounts distributing fake news appear to have, on average, the same number of friends than those distributing tweets with no fake news. However, the density distribution of friends from the accounts (Figure FIGREF29 ) shows that there is indeed a statistically significant difference in their distributions.
If we take into consideration the number of followers, accounts generating viral tweets with fake news do have a very different distribution on this dimension, compared to those accounts generating viral tweets with no fake news (see Figure FIGREF30 ). In fact, such differences are statistically significant.
A useful representation for friends and followers is the ratio between friends/followers. Figures FIGREF31 and FIGREF32 show this representation. Notice that accounts spreading viral tweets with fake news have, on average, a larger ratio of friends/followers. The distribution of those accounts not generating fake news is more evenly distributed.
With respect to the number of mentions, Figure FIGREF33 shows that viral tweets labelled as containing fake news appear to use mentions to other users less frequently than viral tweets not containing fake news. In other words, tweets containing fake news mostly contain 1 mention, whereas other tweets tend to have two). Such differences are statistically significant.
The analysis (Figure FIGREF34 ) of the presence of media in the tweets in our dataset shows that tweets labelled as not containing fake news appear to present more media elements than those labelled as fake news. However, the difference is not statistically significant.
On the other hand, Figure FIGREF35 shows that viral tweets containing fake news appear to include more URLs to other sites than viral tweets that do not contain fake news. In fact, the difference between the two distributions is statistically significant (assuming INLINEFORM0 ).
Polarization
Finally, manual inspection of the text field of those viral tweets labelled as containing fake news shows that 117 of such tweets expressed support for Donald Trump, while only 8 supported Hillary Clinton. The remaining tweets contained fake news related to other topics, not expressing support for any of the candidates.
Discussion
As a summary, and constrained by our existing dataset, we made the following observations regarding differences between viral tweets labelled as containing fake news and viral tweets labelled as not containing them:
These findings (related to our initial hypothesis in Table TABREF44 ) clearly suggest that there are specific pieces of meta-data about tweets that may allow the identification of fake news. One such parameter is the time of exposure. Viral tweets containing fake news are shorter-lived than those containing other type of content. This notion seems to resonate with our findings showing that a number of accounts spreading fake news have already been deleted or suspended by Twitter by the time of writing. If one considers that researchers using different data have found similar results BIBREF9 , it appears that the lifetime of accounts, together with the age of the questioned viral content could be useful to identify fake news. In the light of this finding, accounts newly created should probably put under higher scrutiny than older ones. This in fact, would be a nice a-priori bias for a Bayesian classifier.
Accounts spreading fake news appear to have a larger proportion of friends/followers (i.e. they have, on average, the same number of friends but a smaller number of followers) than those spreading viral content only. Together with the fact that, on average, tweets containing fake news have more URLs than those spreading viral content, it is possible to hypothesize that, both, the ratio of friends/followers of the account producing a viral tweet and number of URLs contained in such a tweet could be useful to single-out fake news in Twitter. Not only that, but our finding related to the number of URLs is in line with intuitions behind the incentives to create fake news commonly found in the literature BIBREF9 (in particular that of obtaining revenue through click-through advertising).
Finally, it is interesting to notice that the content of viral fake news was highly polarized. This finding is also in line with those of Alcott et al. BIBREF9 . This feature suggests that textual sentiment analysis of the content of tweets (as most researchers do), together with the above mentioned parameters from meta-data, may prove useful for identifying fake news.
Conclusions
With the election of Donald Trump as President of the United States, the concept of fake news has become a broadly-known phenomenon that is getting tremendous attention from governments and media companies. We have presented a preliminary study on the meta-data of a publicly available dataset of tweets that became viral during the day of the 2016 US presidential election. Our aim is to advance the understanding of which features might be characteristic of viral tweets containing fake news in comparison with viral tweets without fake news.
We believe that the only way to automatically identify those deceitful tweets (i.e. containing fake news) is by actually understanding and modelling them. Only then, the automation of the processes of tagging and blocking these tweets can be successfully performed. In the same way that spam was fought, we anticipate fake news will suffer a similar evolution, with social platforms implementing tools to deal with them. With most works so far focusing on the actual content of the tweets, ours is a novel attempt from a different, but also complementary, angle.
Within the used dataset, we found there are differences around exposure, characteristics of accounts spreading fake news and the tone of the content. Those findings suggest that it is indeed possible to model and automatically detect fake news. We plan to replicate and validate our experiments in an extended sample of tweets (until 4 months after the US election), and tests the predictive power of the features we found relevant within our sample.
Author Disclosure Statement
No competing financial interest exist.
| [
"Viral tweets are the ones that are retweeted more than 1000 times",
"those that contain a high number of retweets"
] | 3,144 | qasper | en | null | 51cd01004f7bc29798a9671b991c5223ada1d40ccb8141e0 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What is their definition of tweets going viral?
Answer: | [
"Viral tweets are the ones that are retweeted more than 1000 times",
"those that contain a high number of retweets"
] | qasper | 128 | 64 | |
Which basic neural architecture perform best by itself? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
In the age of information dissemination without quality control, it has enabled malicious users to spread misinformation via social media and aim individual users with propaganda campaigns to achieve political and financial gains as well as advance a specific agenda. Often disinformation is complied in the two major forms: fake news and propaganda, where they differ in the sense that the propaganda is possibly built upon true information (e.g., biased, loaded language, repetition, etc.).
Prior works BIBREF0, BIBREF1, BIBREF2 in detecting propaganda have focused primarily at document level, typically labeling all articles from a propagandistic news outlet as propaganda and thus, often non-propagandistic articles from the outlet are mislabeled. To this end, EMNLP19DaSanMartino focuses on analyzing the use of propaganda and detecting specific propagandistic techniques in news articles at sentence and fragment level, respectively and thus, promotes explainable AI. For instance, the following text is a propaganda of type `slogan'.
Trump tweeted: $\underbrace{\text{`}`{\texttt {BUILD THE WALL!}"}}_{\text{slogan}}$
Shared Task: This work addresses the two tasks in propaganda detection BIBREF3 of different granularities: (1) Sentence-level Classification (SLC), a binary classification that predicts whether a sentence contains at least one propaganda technique, and (2) Fragment-level Classification (FLC), a token-level (multi-label) classification that identifies both the spans and the type of propaganda technique(s).
Contributions: (1) To address SLC, we design an ensemble of different classifiers based on Logistic Regression, CNN and BERT, and leverage transfer learning benefits using the pre-trained embeddings/models from FastText and BERT. We also employed different features such as linguistic (sentiment, readability, emotion, part-of-speech and named entity tags, etc.), layout, topics, etc. (2) To address FLC, we design a multi-task neural sequence tagger based on LSTM-CRF and linguistic features to jointly detect propagandistic fragments and its type. Moreover, we investigate performing FLC and SLC jointly in a multi-granularity network based on LSTM-CRF and BERT. (3) Our system (MIC-CIS) is ranked 3rd (out of 12 participants) and 4th (out of 25 participants) in FLC and SLC tasks, respectively.
System Description ::: Linguistic, Layout and Topical Features
Some of the propaganda techniques BIBREF3 involve word and phrases that express strong emotional implications, exaggeration, minimization, doubt, national feeling, labeling , stereotyping, etc. This inspires us in extracting different features (Table TABREF1) including the complexity of text, sentiment, emotion, lexical (POS, NER, etc.), layout, etc. To further investigate, we use topical features (e.g., document-topic proportion) BIBREF4, BIBREF5, BIBREF6 at sentence and document levels in order to determine irrelevant themes, if introduced to the issue being discussed (e.g., Red Herring).
For word and sentence representations, we use pre-trained vectors from FastText BIBREF7 and BERT BIBREF8.
System Description ::: Sentence-level Propaganda Detection
Figure FIGREF2 (left) describes the three components of our system for SLC task: features, classifiers and ensemble. The arrows from features-to-classifier indicate that we investigate linguistic, layout and topical features in the two binary classifiers: LogisticRegression and CNN. For CNN, we follow the architecture of DBLP:conf/emnlp/Kim14 for sentence-level classification, initializing the word vectors by FastText or BERT. We concatenate features in the last hidden layer before classification.
One of our strong classifiers includes BERT that has achieved state-of-the-art performance on multiple NLP benchmarks. Following DBLP:conf/naacl/DevlinCLT19, we fine-tune BERT for binary classification, initializing with a pre-trained model (i.e., BERT-base, Cased). Additionally, we apply a decision function such that a sentence is tagged as propaganda if prediction probability of the classifier is greater than a threshold ($\tau $). We relax the binary decision boundary to boost recall, similar to pankajgupta:CrossRE2019.
Ensemble of Logistic Regression, CNN and BERT: In the final component, we collect predictions (i.e., propaganda label) for each sentence from the three ($\mathcal {M}=3$) classifiers and thus, obtain $\mathcal {M}$ number of predictions for each sentence. We explore two ensemble strategies (Table TABREF1): majority-voting and relax-voting to boost precision and recall, respectively.
System Description ::: Fragment-level Propaganda Detection
Figure FIGREF2 (right) describes our system for FLC task, where we design sequence taggers BIBREF9, BIBREF10 in three modes: (1) LSTM-CRF BIBREF11 with word embeddings ($w\_e$) and character embeddings $c\_e$, token-level features ($t\_f$) such as polarity, POS, NER, etc. (2) LSTM-CRF+Multi-grain that jointly performs FLC and SLC with FastTextWordEmb and BERTSentEmb, respectively. Here, we add binary sentence classification loss to sequence tagging weighted by a factor of $\alpha $. (3) LSTM-CRF+Multi-task that performs propagandistic span/fragment detection (PFD) and FLC (fragment detection + 19-way classification).
Ensemble of Multi-grain, Multi-task LSTM-CRF with BERT: Here, we build an ensemble by considering propagandistic fragments (and its type) from each of the sequence taggers. In doing so, we first perform majority voting at the fragment level for the fragment where their spans exactly overlap. In case of non-overlapping fragments, we consider all. However, when the spans overlap (though with the same label), we consider the fragment with the largest span.
Experiments and Evaluation
Data: While the SLC task is binary, the FLC consists of 18 propaganda techniques BIBREF3. We split (80-20%) the annotated corpus into 5-folds and 3-folds for SLC and FLC tasks, respectively. The development set of each the folds is represented by dev (internal); however, the un-annotated corpus used in leaderboard comparisons by dev (external). We remove empty and single token sentences after tokenization. Experimental Setup: We use PyTorch framework for the pre-trained BERT model (Bert-base-cased), fine-tuned for SLC task. In the multi-granularity loss, we set $\alpha = 0.1$ for sentence classification based on dev (internal, fold1) scores. We use BIO tagging scheme of NER in FLC task. For CNN, we follow DBLP:conf/emnlp/Kim14 with filter-sizes of [2, 3, 4, 5, 6], 128 filters and 16 batch-size. We compute binary-F1and macro-F1 BIBREF12 in SLC and FLC, respectively on dev (internal).
Experiments and Evaluation ::: Results: Sentence-Level Propaganda
Table TABREF10 shows the scores on dev (internal and external) for SLC task. Observe that the pre-trained embeddings (FastText or BERT) outperform TF-IDF vector representation. In row r2, we apply logistic regression classifier with BERTSentEmb that leads to improved scores over FastTextSentEmb. Subsequently, we augment the sentence vector with additional features that improves F1 on dev (external), however not dev (internal). Next, we initialize CNN by FastTextWordEmb or BERTWordEmb and augment the last hidden layer (before classification) with BERTSentEmb and feature vectors, leading to gains in F1 for both the dev sets. Further, we fine-tune BERT and apply different thresholds in relaxing the decision boundary, where $\tau \ge 0.35$ is found optimal.
We choose the three different models in the ensemble: Logistic Regression, CNN and BERT on fold1 and subsequently an ensemble+ of r3, r6 and r12 from each fold1-5 (i.e., 15 models) to obtain predictions for dev (external). We investigate different ensemble schemes (r17-r19), where we observe that the relax-voting improves recall and therefore, the higher F1 (i.e., 0.673). In postprocess step, we check for repetition propaganda technique by computing cosine similarity between the current sentence and its preceding $w=10$ sentence vectors (i.e., BERTSentEmb) in the document. If the cosine-similarity is greater than $\lambda \in \lbrace .99, .95\rbrace $, then the current sentence is labeled as propaganda due to repetition. Comparing r19 and r21, we observe a gain in recall, however an overall decrease in F1 applying postprocess.
Finally, we use the configuration of r19 on the test set. The ensemble+ of (r4, r7 r12) was analyzed after test submission. Table TABREF9 (SLC) shows that our submission is ranked at 4th position.
Experiments and Evaluation ::: Results: Fragment-Level Propaganda
Table TABREF11 shows the scores on dev (internal and external) for FLC task. Observe that the features (i.e., polarity, POS and NER in row II) when introduced in LSTM-CRF improves F1. We run multi-grained LSTM-CRF without BERTSentEmb (i.e., row III) and with it (i.e., row IV), where the latter improves scores on dev (internal), however not on dev (external). Finally, we perform multi-tasking with another auxiliary task of PFD. Given the scores on dev (internal and external) using different configurations (rows I-V), it is difficult to infer the optimal configuration. Thus, we choose the two best configurations (II and IV) on dev (internal) set and build an ensemble+ of predictions (discussed in section SECREF6), leading to a boost in recall and thus an improved F1 on dev (external).
Finally, we use the ensemble+ of (II and IV) from each of the folds 1-3, i.e., $|{\mathcal {M}}|=6$ models to obtain predictions on test. Table TABREF9 (FLC) shows that our submission is ranked at 3rd position.
Conclusion and Future Work
Our system (Team: MIC-CIS) explores different neural architectures (CNN, BERT and LSTM-CRF) with linguistic, layout and topical features to address the tasks of fine-grained propaganda detection. We have demonstrated gains in performance due to the features, ensemble schemes, multi-tasking and multi-granularity architectures. Compared to the other participating systems, our submissions are ranked 3rd and 4th in FLC and SLC tasks, respectively.
In future, we would like to enrich BERT models with linguistic, layout and topical features during their fine-tuning. Further, we would also be interested in understanding and analyzing the neural network learning, i.e., extracting salient fragments (or key-phrases) in the sentence that generate propaganda, similar to pankajgupta:2018LISA in order to promote explainable AI.
| [
"BERT"
] | 1,507 | qasper | en | null | 3dd2d62c046f3b559c34003f570ed35211000500b8f0145f | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Which basic neural architecture perform best by itself?
Answer: | [
"BERT"
] | qasper | 128 | 65 | |
what is the source of the data? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Nowadays deep learning techniques outperform the other conventional methods in most of the speech-related tasks. Training robust deep neural networks for each task depends on the availability of powerful processing GPUs, as well as standard and large scale datasets. In text-independent speaker verification, large-scale datasets are available, thanks to the NIST SRE evaluations and other data collection projects such as VoxCeleb BIBREF0.
In text-dependent speaker recognition, experiments with end-to-end architectures conducted on large proprietary databases have demonstrated their superiority over traditional approaches BIBREF1. Yet, contrary to text-independent speaker recognition, text-dependent speaker recognition lacks large-scale publicly available databases. The two most well-known datasets are probably RSR2015 BIBREF2 and RedDots BIBREF3. The former contains speech data collected from 300 individuals in a controlled manner, while the latter is used primarily for evaluation rather than training, due to its small number of speakers (only 64). Motivated by this lack of large-scale dataset for text-dependent speaker verification, we chose to proceed with the collection of the DeepMine dataset, which we expect to become a standard benchmark for the task.
Apart from speaker recognition, large amounts of training data are required also for training automatic speech recognition (ASR) systems. Such datasets should not only be large in size, they should also be characterized by high variability with respect to speakers, age and dialects. While several datasets with these properties are available for languages like English, Mandarin, French, this is not the case for several other languages, such as Persian. To this end, we proceeded with collecting a large-scale dataset, suitable for building robust ASR models in Persian.
The main goal of the DeepMine project was to collect speech from at least a few thousand speakers, enabling research and development of deep learning methods. The project started at the beginning of 2017, and after designing the database and the developing Android and server applications, the data collection began in the middle of 2017. The project finished at the end of 2018 and the cleaned-up and final version of the database was released at the beginning of 2019. In BIBREF4, the running project and its data collection scenarios were described, alongside with some preliminary results and statistics. In this paper, we announce the final and cleaned-up version of the database, describe its different parts and provide various evaluation setups for each part. Finally, since the database was designed mainly for text-dependent speaker verification purposes, some baseline results are reported for this task on the official evaluation setups. Additional baseline results are also reported for Persian speech recognition. However, due to the space limitation in this paper, the baseline results are not reported for all the database parts and conditions. They will be defined and reported in the database technical documentation and in a future journal paper.
Data Collection
DeepMine is publicly available for everybody with a variety of licenses for different users. It was collected using crowdsourcing BIBREF4. The data collection was done using an Android application. Each respondent installed the application on his/her personal device and recorded several phrases in different sessions. The Android application did various checks on each utterance and if it passed all of them, the respondent was directed to the next phrase. For more information about data collection scenario, please refer to BIBREF4.
Data Collection ::: Post-Processing
In order to clean-up the database, the main post-processing step was to filter out problematic utterances. Possible problems include speaker word insertions (e.g. repeating some part of a phrase), deletions, substitutions, and involuntary disfluencies. To detect these, we implemented an alignment stage, similar to the second alignment stage in the LibriSpeech project BIBREF5. In this method, a custom decoding graph was generated for each phrase. The decoding graph allows for word skipping and word insertion in the phrase.
For text-dependent and text-prompted parts of the database, such errors are not allowed. Hence, any utterances with errors were removed from the enrollment and test lists. For the speech recognition part, a sub-part of the utterance which is correctly aligned to the corresponding transcription is kept. After the cleaning step, around 190 thousand utterances with full transcription and 10 thousand with sub-part alignment have remained in the database.
Data Collection ::: Statistics
After processing the database and removing problematic respondents and utterances, 1969 respondents remained in the database, with 1149 of them being male and 820 female. 297 of the respondents could not read English and have therefore read only the Persian prompts. About 13200 sessions were recorded by females and similarly, about 9500 sessions by males, i.e. women are over-represented in terms of sessions, even though their number is 17% smaller than that of males. Other useful statistics related to the database are shown in Table TABREF4.
The last status of the database, as well as other related and useful information about its availability can be found on its website, together with a limited number of samples.
DeepMine Database Parts
The DeepMine database consists of three parts. The first one contains fixed common phrases to perform text-dependent speaker verification. The second part consists of random sequences of words useful for text-prompted speaker verification, and the last part includes phrases with word- and phoneme-level transcription, useful for text-independent speaker verification using a random phrase (similar to Part4 of RedDots). This part can also serve for Persian ASR training. Each part is described in more details below. Table TABREF11 shows the number of unique phrases in each part of the database. For the English text-dependent part, the following phrases were selected from part1 of the RedDots database, hence the RedDots can be used as an additional training set for this part:
“My voice is my password.”
“OK Google.”
“Artificial intelligence is for real.”
“Actions speak louder than words.”
“There is no such thing as a free lunch.”
DeepMine Database Parts ::: Part1 - Text-dependent (TD)
This part contains a set of fixed phrases which are used to verify speakers in text-dependent mode. Each speaker utters 5 Persian phrases, and if the speaker can read English, 5 phrases selected from Part1 of the RedDots database are also recorded.
We have created three experimental setups with different numbers of speakers in the evaluation set. For each setup, speakers with more recording sessions are included in the evaluation set and the rest of the speakers are used for training in the background set (in the database, all background sets are basically training data). The rows in Table TABREF13 corresponds to the different experimental setups and shows the numbers of speakers in each set. Note that, for English, we have filtered the (Persian native) speakers by the ability to read English. Therefore, there are fewer speakers in each set for English than for Persian. There is a small “dev” set in each setup which can be used for parameter tuning to prevent over-tuning on the evaluation set.
For each experimental setup, we have defined several official trial lists with different numbers of enrollment utterances per trial in order to investigate the effects of having different amounts of enrollment data. All trials in one trial list have the same number of enrollment utterances (3 to 6) and only one test utterance. All enrollment utterances in a trial are taken from different consecutive sessions and the test utterance is taken from yet another session. From all the setups and conditions, the 100-spk with 3-session enrollment (3-sess) is considered as the main evaluation condition. In Table TABREF14, the number of trials for Persian 3-sess are shown for the different types of trial in the text-dependent speaker verification (SV). Note that for Imposter-Wrong (IW) trials (i.e. imposter speaker pronouncing wrong phrase), we merely create one wrong trial for each Imposter-Correct (IC) trial to limit the huge number of possible trials for this case. So, the number of trials for IC and IW cases are the same.
DeepMine Database Parts ::: Part2 - Text-prompted (TP)
For this part, in each session, 3 random sequences of Persian month names are shown to the respondent in two modes: In the first mode, the sequence consists of all 12 months, which will be used for speaker enrollment. The second mode contains a sequence of 3 month names that will be used as a test utterance. In each 8 sessions received by a respondent from the server, there are 3 enrollment phrases of all 12 months (all in just one session), and $7 \times 3$ other test phrases, containing fewer words. For a respondent who can read English, 3 random sequences of English digits are also recorded in each session. In one of the sessions, these sequences contain all digits and the remaining ones contain only 4 digits.
Similar to the text-dependent case, three experimental setups with different number of speaker in the evaluation set are defined (corresponding to the rows in Table TABREF16). However, different strategy is used for defining trials: Depending on the enrollment condition (1- to 3-sess), trials are enrolled on utterances of all words from 1 to 3 different sessions (i.e. 3 to 9 utterances). Further, we consider two conditions for test utterances: seq test utterance with only 3 or 4 words and full test utterances with all words (i.e. same words as in enrollment but in different order). From all setups an all conditions, the 100-spk with 1-session enrolment (1-sess) is considered as the main evaluation condition for the text-prompted case. In Table TABREF16, the numbers of trials (sum for both seq and full conditions) for Persian 1-sess are shown for the different types of trials in the text-prompted SV. Again, we just create one IW trial for each IC trial.
DeepMine Database Parts ::: Part3 - Text-independent (TI)
In this part, 8 Persian phrases that have already been transcribed on the phone level are displayed to the respondent. These phrases are chosen mostly from news and Persian Wikipedia. If the respondent is unable to read English, instead of 5 fixed phrases and 3 random digit strings, 8 other Persian phrases are also prompted to the respondent to have exactly 24 phrases in each recording session.
This part can be useful at least for three potential applications. First, it can be used for text-independent speaker verification. The second application of this part (same as Part4 of RedDots) is text-prompted speaker verification using random text (instead of a random sequence of words). Finally, the third application is large vocabulary speech recognition in Persian (explained in the next sub-section).
Based on the recording sessions, we created two experimental setups for speaker verification. In the first one, respondents with at least 17 recording sessions are included to the evaluation set, respondents with 16 sessions to the development and the rest of respondents to the background set (can be used as training data). In the second setup, respondents with at least 8 sessions are included to the evaluation set, respondents with 6 or 7 sessions to the development and the rest of respondents to the background set. Table TABREF18 shows numbers of speakers in each set of the database for text-independent SV case.
For text-independent SV, we have considered 4 scenarios for enrollment and 4 scenarios for test. The speaker can be enrolled using utterances from 1, 2 or 3 consecutive sessions (1sess to 3sess) or using 8 utterances from 8 different sessions. The test speech can be one utterance (1utt) for short duration scenario or all utterances in one session (1sess) for long duration case. In addition, test speech can be selected from 5 English phrases for cross-language testing (enrollment using Persian utterances and test using English utterances). From all setups, 1sess-1utt and 1sess-1sess for 438-spk set are considered as the main evaluation setups for text-independent case. Table TABREF19 shows number of trials for these setups.
For text-prompted SV with random text, the same setup as text-independent case together with corresponding utterance transcriptions can be used.
DeepMine Database Parts ::: Part3 - Speech Recognition
As explained before, Part3 of the DeepMine database can be used for Persian read speech recognition. There are only a few databases for speech recognition in Persian BIBREF6, BIBREF7. Hence, this part can at least partly address this problem and enable robust speech recognition applications in Persian. Additionally, it can be used for speaker recognition applications, such as training deep neural networks (DNNs) for extracting bottleneck features BIBREF8, or for collecting sufficient statistics using DNNs for i-vector training.
We have randomly selected 50 speakers (25 for each gender) from the all speakers in the database which have net speech (without silence parts) between 25 minutes to 50 minutes as test speakers. For each speaker, the utterances in the first 5 sessions are included to (small) test-set and the other utterances of test speakers are considered as a large-test-set. The remaining utterances of the other speakers are included in the training set. The test-set, large-test-set and train-set contain 5.9, 28.5 and 450 hours of speech respectively.
There are about 8300 utterances in Part3 which contain only Persian full names (i.e. first and family name pairs). Each phrase consists of several full names and their phoneme transcriptions were extracted automatically using a trained Grapheme-to-Phoneme (G2P). These utterances can be used to evaluate the performance of a systems for name recognition, which is usually more difficult than the normal speech recognition because of the lack of a reliable language model.
Experiments and Results
Due to the space limitation, we present results only for the Persian text-dependent speaker verification and speech recognition.
Experiments and Results ::: Speaker Verification Experiments
We conducted an experiment on text-dependent speaker verification part of the database, using the i-vector based method proposed in BIBREF9, BIBREF10 and applied it to the Persian portion of Part1. In this experiment, 20-dimensional MFCC features along with first and second derivatives are extracted from 16 kHz signals using HTK BIBREF11 with 25 ms Hamming windowed frames with 15 ms overlap.
The reported results are obtained with a 400-dimensional gender independent i-vector based system. The i-vectors are first length-normalized and are further normalized using phrase- and gender-dependent Regularized Within-Class Covariance Normalization (RWCCN) BIBREF10. Cosine distance is used to obtain speaker verification scores and phrase- and gender-dependent s-norm is used for normalizing the scores. For aligning speech frames to Gaussian components, monophone HMMs with 3 states and 8 Gaussian components in each state are used BIBREF10. We only model the phonemes which appear in the 5 Persian text-dependent phrases.
For speaker verification experiments, the results were reported in terms of Equal Error Rate (EER) and Normalized Detection Cost Function as defined for NIST SRE08 ($\mathrm {NDCF_{0.01}^{min}}$) and NIST SRE10 ($\mathrm {NDCF_{0.001}^{min}}$). As shown in Table TABREF22, in text-dependent SV there are 4 types of trials: Target-Correct and Imposter-Correct refer to trials when the pass-phrase is uttered correctly by target and imposter speakers respectively, and in same manner, Target-Wrong and Imposter-Wrong refer to trials when speakers uttered a wrong pass-phrase. In this paper, only the correct trials (i.e. Target-Correct as target trials vs Imposter-Correct as non-target trials) are considered for evaluating systems as it has been proved that these are the most challenging trials in text-dependent SV BIBREF8, BIBREF12.
Table TABREF23 shows the results of text-dependent experiments using Persian 100-spk and 3-sess setup. For filtering trials, the respondents' mobile brand and model were used in this experiment. In the table, the first two letters in the filter notation relate to the target trials and the second two letters (i.e. right side of the colon) relate for non-target trials. For target trials, the first Y means the enrolment and test utterances were recorded using a device with the same brand by the target speaker. The second Y letter means both recordings were done using exactly the same device model. Similarly, the first Y for non-target trials means that the devices of target and imposter speakers are from the same brand (i.e. manufacturer). The second Y means that, in addition to the same brand, both devices have the same model. So, the most difficult target trials are “NN”, where the speaker has used different a device at the test time. In the same manner, the most difficult non-target trials which should be rejected by the system are “YY” where the imposter speaker has used the same device model as the target speaker (note that it does not mean physically the same device because each speaker participated in the project using a personal mobile device). Hence, the similarity in the recording channel makes rejection more difficult.
The first row in Table TABREF23 shows the results for all trials. By comparing the results with the best published results on RSR2015 and RedDots BIBREF10, BIBREF8, BIBREF12, it is clear that the DeepMine database is more challenging than both RSR2015 and RedDots databases. For RSR2015, the same i-vector/HMM-based method with both RWCCN and s-norm has achieved EER less than 0.3% for both genders (Table VI in BIBREF10). The conventional Relevance MAP adaptation with HMM alignment without applying any channel-compensation techniques (i.e. without applying RWCCN and s-norm due to the lack of suitable training data) on RedDots Part1 for the male has achieved EER around 1.5% (Table XI in BIBREF10). It is worth noting that EERs for DeepMine database without any channel-compensation techniques are 2.1 and 3.7% for males and females respectively.
One interesting advantage of the DeepMine database compared to both RSR2015 and RedDots is having several target speakers with more than one mobile device. This is allows us to analyse the effects of channel compensation methods. The second row in Table TABREF23 corresponds to the most difficult trials where the target trials come from mobile devices with different models while imposter trials come from the same device models. It is clear that severe degradation was caused by this kind of channel effects (i.e. decreasing within-speaker similarities while increasing between-speaker similarities), especially for females.
The results in the third row show the condition when target speakers at the test time use exactly the same device that was used for enrollment. Comparing this row with the results in the first row proves how much improvement can be achieved when exactly the same device is used by the target speaker.
The results in the fourth row show the condition when imposter speakers also use the same device model at test time to fool the system. So, in this case, there is no device mismatch in all trials. By comparing the results with the third row, we can see how much degradation is caused if we only consider the non-target trials with the same device.
The fifth row shows similar results when the imposter speakers use device of the same brand as the target speaker but with a different model. Surprisingly, in this case, the degradation is negligible and it means that mobiles from a specific brand (manufacturer) have different recording channel properties.
The degraded female results in the sixth row as compared to the third row show the effect of using a different device model from the same brand for target trials. For males, the filters brings almost the same subsets of trials, which explains the very similar results in this case.
Looking at the first two and the last row of Table TABREF23, one can notice the significantly worse performance obtained for the female trials as compared to males. Note that these three rows include target trials where the devices used for enrollment do not necessarily match the devices used for recording test utterances. On the other hand, in rows 3 to 6, which exclude such mismatched trials, the performance for males and females is comparable. This suggest that the degraded results for females are caused by some problematic trials with device mismatch. The exact reason for this degradation is so far unclear and needs a further investigation.
In the last row of the table, the condition of the second row is relaxed: the target device should have different model possibly from the same brand and imposter device only needs to be from the same brand. In this case, as was expected, the performance degradation is smaller than in the second row.
Experiments and Results ::: Speech Recognition Experiments
In addition to speaker verification, we present several speech recognition experiments on Part3. The experiments were performed with the Kaldi toolkit BIBREF13. For training HMM-based MonoPhone model, only 20 thousands of shortest utterances are used and for other models the whole training data is used. The DNN based acoustic model is a time-delay DNN with low-rank factorized layers and skip connections without i-vector adaptation (a modified network from one of the best performing LibriSpeech recipes). The network is shown in Table TABREF25: there are 16 F-TDNN layers, with dimension 1536 and linear bottleneck layers of dimension 256. The acoustic model is trained for 10 epochs using lattice-free maximum mutual information (LF-MMI) with cross-entropy regularization BIBREF14. Re-scoring is done using a pruned trigram language model and the size of the dictionary is around 90,000 words.
Table TABREF26 shows the results in terms of word error rate (WER) for different evaluated methods. As can be seen, the created database can be used to train well performing and practically usable Persian ASR models.
Conclusions
In this paper, we have described the final version of a large speech corpus, the DeepMine database. It has been collected using crowdsourcing and, according to the best of our knowledge, it is the largest public text-dependent and text-prompted speaker verification database in two languages: Persian and English. In addition, it is the largest text-independent speaker verification evaluation database, making it suitable to robustly evaluate state-of-the-art methods on different conditions. Alongside these appealing properties, it comes with phone-level transcription, making it suitable to train deep neural network models for Persian speech recognition.
We provided several evaluation protocols for each part of the database. The protocols allow researchers to investigate the performance of different methods in various scenarios and study the effects of channels, duration and phrase text on the performance. We also provide two test sets for speech recognition: One normal test set with a few minutes of speech for each speaker and one large test set with more (30 minutes on average) speech that can be used for any speaker adaptation method.
As baseline results, we reported the performance of an i-vector/HMM based method on Persian text-dependent part. Moreover, we conducted speech recognition experiments using conventional HMM-based methods, as well as state-of-the-art deep neural network based method using Kaldi toolkit with promising performance. Text-dependent results have shown that the DeepMine database is more challenging than RSR2015 and RedDots databases.
Acknowledgments
The data collection project was mainly supported by Sharif DeepMine company. The work on the paper was supported by Czech National Science Foundation (GACR) project "NEUREM3" No. 19-26934X and the National Programme of Sustainability (NPU II) project "IT4Innovations excellence in science - LQ1602".
| [
"Android application"
] | 3,795 | qasper | en | null | d230ce079b2e4ecf5d9a987fb750dcbc319a537bdcfcc3d4 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: what is the source of the data?
Answer: | [
"Android application"
] | qasper | 128 | 66 | |
What machine learning and deep learning methods are used for RQE? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
With the availability of rich data on users' locations, profiles and search history, personalization has become the leading trend in large-scale information retrieval. However, efficiency through personalization is not yet the most suitable model when tackling domain-specific searches. This is due to several factors, such as the lexical and semantic challenges of domain-specific data that often include advanced argumentation and complex contextual information, the higher sparseness of relevant information sources, and the more pronounced lack of similarities between users' searches.
A recent study on expert search strategies among healthcare information professionals BIBREF0 showed that, for a given search task, they spend an average of 60 minutes per collection or database, 3 minutes to examine the relevance of each document, and 4 hours of total search time. When written in steps, their search strategy spans over 15 lines and can reach up to 105 lines.
With the abundance of information sources in the medical domain, consumers are more and more faced with a similar challenge, one that needs dedicated solutions that can adapt to the heterogeneity and specifics of health-related information.
Dedicated Question Answering (QA) systems are one of the viable solutions to this problem as they are designed to understand natural language questions without relying on external information on the users.
In the context of QA, the goal of Recognizing Question Entailment (RQE) is to retrieve answers to a premise question ( INLINEFORM0 ) by retrieving inferred or entailed questions, called hypothesis questions ( INLINEFORM1 ) that already have associated answers. Therefore, we define the entailment relation between two questions as: a question INLINEFORM2 entails a question INLINEFORM3 if every answer to INLINEFORM4 is also a correct answer to INLINEFORM5 BIBREF1 .
RQE is particularly relevant due to the increasing numbers of similar questions posted online BIBREF2 and its ability to solve differently the challenging issues of question understanding and answer extraction. In addition to being used to find relevant answers, these resources can also be used in training models able to recognize inference relations and similarity between questions.
Question similarity has recently attracted international challenges BIBREF3 , BIBREF4 and several research efforts proposing a wide range of approaches, including Logistic Regression, Recurrent Neural Networks (RNNs), Long Short Term Memory cells (LSTMs), and Convolutional Neural Networks (CNNs) BIBREF5 , BIBREF6 , BIBREF1 , BIBREF7 .
In this paper, we study question entailment in the medical domain and the effectiveness of the end-to-end RQE-based QA approach by evaluating the relevance of the retrieved answers. Although entailment was attempted in QA before BIBREF8 , BIBREF9 , BIBREF10 , as far as we know, we are the first to introduce and evaluate a full medical question answering approach based on question entailment for free-text questions. Our contributions are:
The next section is dedicated to related work on question answering, question similarity and entailment. In Section SECREF3 , we present two machine learning (ML) and deep learning (DL) methods for RQE and compare their performance using open-domain and clinical datasets. Section SECREF4 describes the new collection of medical question-answer pairs. In Section SECREF5 , we describe our RQE-based approach for QA. Section SECREF6 presents our evaluation of the retrieved answers and the results obtained on TREC 2017 LiveQA medical questions.
Background
In this section we define the RQE task and describe related work at the intersection of question answering, question similarity and textual inference.
Task Definition
The definition of Recognizing Question Entailment (RQE) can have a significant impact on QA results. In related work, the meaning associated with Natural Language Inference (NLI) varies among different tasks and events. For instance, Recognizing Textual Entailment (RTE) was addressed by the PASCAL challenge BIBREF12 , where the entailment relation has been assessed manually by human judges who selected relevant sentences "entailing" a set of hypotheses from a list of documents returned by different Information Retrieval (IR) methods. In another definition, the Stanford Natural Language Inference corpus SNLI BIBREF13 , used three classification labels for the relations between two sentences: entailment, neutral and contradiction. For the entailment label, the annotators who built the corpus were presented with an image and asked to write a caption “that is a definitely true description of the photo”. For the neutral label, they were asked to provide a caption “that might be a true description of the label”. They were asked for a caption that “is definitely a false description of the photo” for the contradiction label.
More recently, the multiNLI corpus BIBREF14 was shared in the scope of the RepEval 2017 shared task BIBREF15 . To build the corpus, annotators were presented with a premise text and asked to write three sentences. One novel sentence, which is “necessarily true or appropriate in the same situations as the premise,” for the entailment label, a sentence, which is “necessarily false or inappropriate whenever the premise is true,” for the contradiction label, and a last sentence, “where neither condition applies,” for the neutral label.
Whereas these NLI definitions might be suitable for the broad topic of text understanding, their relation to practical information retrieval or question answering systems is not straightforward.
In contrast, RQE has to be tailored to the question answering task. For instance, if the premise question is "looking for cold medications for a 30 yo woman", a RQE approach should be able to consider the more general (less restricted) question "looking for cold medications" as relevant, since its answers are relevant for the initial question, whereas "looking for medications for a 30 yo woman" is a useless contextualization. The entailment relation we are seeking in the QA context should include relevant and meaningful relaxations of contextual and semantic constraints (cf. Section SECREF13 ).
Related Work on Question Answering
Classical QA systems face two main challenges related to question analysis and answer extraction. Several QA approaches were proposed in the literature for the open domain BIBREF16 , BIBREF17 and the medical domain BIBREF18 , BIBREF19 , BIBREF20 . A variety of methods were developed for question analysis, focus (topic) recognition and question type identification BIBREF21 , BIBREF22 , BIBREF23 , BIBREF24 . Similarly, many different approaches tackled document or passage retrieval and answer selection and (re)ranking BIBREF25 , BIBREF26 , BIBREF27 .
An alternative approach consists in finding similar questions or FAQs that are already answered BIBREF28 , BIBREF29 . One of the earliest question answering systems based on finding similar questions and re-using the existing answers was FAQ FINDER BIBREF30 . Another system that complements the existing Q&A services of NetWellness is SimQ BIBREF2 , which allows retrieval of similar web-based consumer health questions. SimQ uses syntactic and semantic features to compute similarity between questions, and UMLS BIBREF31 as a standardized semantic knowledge source. The system achieves 72.2% precision, 78.0% recall and 75.0% F-score on NetWellness questions. However, the method was evaluated only on one question similarity dataset, and the retrieved answers were not evaluated.
The aim of the medical task at TREC 2017 LiveQA was to develop techniques for answering complex questions such as consumer health questions, as well as to identify relevant answer sources that can comply with the sensitivity of medical information retrieval.
The CMU-OAQA system BIBREF32 achieved the best performance of 0.637 average score on the medical task by using an attentional encoder-decoder model for paraphrase identification and answer ranking. The Quora question-similarity dataset was used for training. The PRNA system BIBREF33 achieved the second best performance in the medical task with 0.49 average score using Wikipedia as the first answer source and Yahoo and Google searches as secondary answer sources. Each medical question was decomposed into several subquestions. To extract the answer from the selected text passage, a bi-directional attention model trained on the SQUAD dataset was used.
Deep neural network models have been pushing the limits of performance achieved in QA related tasks using large training datasets. The results obtained by CMU-OAQA and PRNA showed that large open-domain datasets were beneficial for the medical domain. However, the best system (CMU-OAQA) relying on the same training data obtained a score of 1.139 on the LiveQA open-domain task.
While this gap in performance can be explained in part by the discrepancies between the medical test questions and the open-domain questions, it also highlights the need for larger medical datasets to support deep learning approaches in dealing with the linguistic complexity of consumer health questions and the challenge of finding correct and complete answers.
Another technique was used by ECNU-ICA team BIBREF34 based on learning question similarity via two long short-term memory (LSTM) networks applied to obtain the semantic representations of the questions. To construct a collection of similar question pairs, they searched community question answering sites such as Yahoo! and Answers.com. In contrast, the ECNU-ICA system achieved the best performance of 1.895 in the open-domain task but an average score of only 0.402 in the medical task. As the ECNU-ICA approach also relied on a neural network for question matching, this result shows that training attention-based decoder-encoder networks on the Quora dataset generalized better to the medical domain than training LSTMs on similar questions from Yahoo! and Answers.com.
The CMU-LiveMedQA team BIBREF20 designed a specific system for the medical task. Using only the provided training datasets and the assumption that each question contains only one focus, the CMU-LiveMedQA system obtained an average score of 0.353. They used a convolutional neural network (CNN) model to classify a question into a restricted set of 10 question types and crawled "relevant" online web pages to find the answers. However, the results were lower than those achieved by the systems relying on finding similar answered questions. These results support the relevance of similar question matching for the end-to-end QA task as a new way of approaching QA instead of the classical QA approaches based on Question Analysis and Answer Retrieval.
Related Work on Question Similarity and Entailment
Several efforts focused on recognizing similar questions. Jeon et al. BIBREF35 showed that a retrieval model based on translation probabilities learned from a question and answer archive can recognize semantically similar questions. Duan et al. BIBREF36 proposed a dedicated language modeling approach for question search, using question topic (user's interest) and question focus (certain aspect of the topic).
Lately, these efforts were supported by a task on Question-Question similarity introduced in the community QA challenge at SemEval (task 3B) BIBREF3 . Given a new question, the task focused on reranking all similar questions retrieved by a search engine, assuming that the answers to the similar questions will be correct answers for the new question. Different machine learning and deep learning approaches were tested in the scope of SemEval 2016 BIBREF3 and 2017 BIBREF4 task 3B. The best performing system in 2017 achieved a MAP of 47.22% using supervised Logistic Regression that combined different unsupervised similarity measures such as Cosine and Soft-Cosine BIBREF37 . The second best system achieved 46.93% MAP with a learning-to-rank method using Logistic Regression and a rich set of features including lexical and semantic features as well as embeddings generated by different neural networks (siamese, Bi-LSTM, GRU and CNNs) BIBREF38 . In the scope of this challenge, a dataset was collected from Qatar Living forum for training. We refer to this dataset as SemEval-cQA.
In another effort, an answer-based definition of RQE was proposed and tested BIBREF1 . The authors introduced a dataset of clinical questions and used a feature-based method that provided an Accuracy of 75% on consumer health questions. We will call this dataset Clinical-QE. Dos Santos et al. BIBREF5 proposed a new approach to retrieve semantically equivalent questions combining a bag-of-words representation with a distributed vector representation created by a CNN and user data collected from two Stack Exchange communities. Lei et al. BIBREF7 proposed a recurrent and convolutional model (gated convolution) to map questions to their semantic representations. The models were pre-trained within an encoder-decoder framework.
RQE Approaches and Experiments
The choice of two methods for our empirical study is motivated by the best performance achieved by Logistic Regression in question-question similarity at SemEval 2017 (best system BIBREF37 and second best system BIBREF38 ), and the high performance achieved by neural networks on larger datasets such as SNLI BIBREF13 , BIBREF39 , BIBREF40 , BIBREF41 . We first define the RQE task, then present the two approaches, and evaluate their performance on five different datasets.
Definition
In the context of QA, the goal of RQE is to retrieve answers to a new question by retrieving entailed questions with associated answers. We therefore define question entailment as:
a question INLINEFORM0 entails a question INLINEFORM1 if every answer to INLINEFORM2 is also a complete or partial answer to INLINEFORM3 .
We present below two examples of consumer health questions INLINEFORM0 and entailed questions INLINEFORM1 :
Example 1 (each answer to the entailed question B1 is a complete answer to A1):
A1: What is the latest news on tennitis, or ringing in the ear, I am 75 years old and have had ringing in the ear since my mid 5os. Thank you.
B1: What is the latest research on Tinnitus?
Example 2 (each answer to the entailed question B2 is a partial answer to A2):
A2: My mother has been diagnosed with Alzheimer's, my father is not of the greatest health either and is the main caregiver for my mother. My question is where do we start with attempting to help our parents w/ the care giving and what sort of financial options are there out there for people on fixed incomes.
B2: What resources are available for Alzheimer's caregivers?
The inclusion of partial answers in the definition of question entailment also allows efficient relaxation of the contextual constraints of the original question INLINEFORM0 to retrieve relevant answers from entailed, but less restricted, questions.
Deep Learning Model
To recognize entailment between two questions INLINEFORM0 (premise) and INLINEFORM1 (hypothesis), we adapted the neural network proposed by Bowman et al. BIBREF13 . Our DL model, presented in Figure FIGREF20 , consists of three 600d ReLU layers, with a bottom layer taking the concatenated sentence representations as input and a top layer feeding a softmax classifier. The sentence embedding model sums the Recurrent neural network (RNN) embeddings of its words. The word embeddings are first initialized with pretrained GloVe vectors. This adaptation provided the best performance in previous experiments with RQE data.
GloVe is an unsupervised learning algorithm to generate vector representations for words BIBREF42 . Training is performed on aggregated word co-occurrence statistics from a large corpus, and the resulting representations show interesting linear substructures of the word vector space. We use the pretrained common crawl version with 840B tokens and 300d vectors, which are not updated during training.
Logistic Regression Classifier
In this feature-based approach, we use Logistic Regression to classify question pairs into entailment or no-entailment. Logistic Regression achieved good results on this specific task and outperformed other statistical learning algorithms such as SVM and Naive Bayes. In a preprocessing step, we remove stop words and perform word stemming using the Porter algorithm BIBREF43 for all ( INLINEFORM0 , INLINEFORM1 ) pairs.
We use a list of nine features, selected after several experiments on RTE datasets BIBREF12 . We compute five similarity measures between the pre-processed questions and use their values as features. We use Word Overlap, the Dice coefficient based on the number of common bigrams, Cosine, Levenshtein, and the Jaccard similarities. Our feature list also includes the maximum and average values obtained with these measures and the question length ratio (length( INLINEFORM0 )/length( INLINEFORM1 )). We compute a morphosyntactic feature indicating the number of common nouns and verbs between INLINEFORM2 and INLINEFORM3 . TreeTagger BIBREF44 was used for POS tagging.
For RQE, we add an additional feature specific to the question type. We use a dictionary lookup to map triggers to the question type (e.g. Treatment, Prognosis, Inheritance). Triggers are identified for each question type based on a manual annotation of a set of medical questions (cf. Section SECREF36 ). This feature has three possible values: 2 (Perfect match between INLINEFORM0 type(s) and INLINEFORM1 type(s)), 1 (Overlap between INLINEFORM2 type(s) and INLINEFORM3 type(s)) and 0 (No common types).
Datasets Used for the RQE Study
We evaluate the RQE methods (i.e. deep learning model and logistic regression classifier) using two datasets of sentence pairs (SNLI and multiNLI), and three datasets of question pairs (Quora, Clinical-QE, and SemEval-cQA).
The Stanford Natural Language Inference corpus (SNLI) BIBREF13 contains 569,037 sentence pairs written by humans based on image captioning. The training set of the MultiNLI corpus BIBREF14 consists of 393,000 pairs of sentences from five genres of written and spoken English (e.g. Travel, Government). Two other "matched" and "mismatched" sets are also available for development (20,000 pairs). Both SNLI and multiNLI consider three types of relationships between sentences: entailment, neutral and contradiction. We converted the contradiction and neutral labels to the same non-entailment class.
The QUORA dataset of similar questions was recently published with 404,279 question pairs. We randomly selected three distinct subsets (80%/10%/10%) for training (323,423 pairs), development (40,428 pairs) and test (40,428 pairs).
The clinical-QE dataset BIBREF1 contains 8,588 question pairs and was constructed using 4,655 clinical questions asked by family doctors BIBREF45 . We randomly selected three distinct subsets (80%/10%/10%) for training (6,870 pairs), development (859 pairs) and test (859 pairs).
The question similarity dataset of SemEval 2016 Task 3B (SemEval-cQA) BIBREF3 contains 3,869 question pairs and aims to re-rank a list of related questions according to their similarity to the original question. The same dataset was used for SemEval 2017 Task 3 BIBREF4 .
To construct our test dataset, we used a publicly shared set of Consumer Health Questions (CHQs) received by the U.S. National Library of Medicine (NLM), and annotated with named entities, question types, and focus BIBREF46 , BIBREF47 . The CHQ dataset consists of 1,721 consumer information requests manually annotated with subquestions, each identified by a question type and a focus.
First, we selected automatically harvested FAQs, from U.S. National Institutes of Health (NIH) websites, that share both the same focus and the same question type with the CHQs. As FAQs are most often very short, we first assume that the CHQ entails the FAQ. Two sets of pairs were constructed: (i) positive pairs of CHQs and FAQs sharing at least one common question type and the question focus, and (ii) negative pairs corresponding to a focus mismatch or type mismatch. For each category of negative examples, we randomly selected the same number of pairs for a balanced dataset. Then, we manually validated the constructed pairs and corrected the positive and negative labels when needed. The final RQE dataset contains 850 CHQ-FAQ pairs with 405 positive and 445 negative pairs. Table TABREF26 presents examples from the five training datasets (SNLI, MultiNLI, SemEval-cQA, Clinical-QE and Quora) and the new test dataset of medical CHQ-FAQ pairs.
Results of RQE Approaches
In the first experiment, we evaluated the DL and ML methods on SNLI, multi-NLI, Quora, and Clinical-QE. For the datasets that did not have a development and test sets, we randomly selected two sets, each amounting to 10% of the data, for test and development, and used the remaining 80% for training. For MultiNLI, we used the dev1-matched set for validation and the dev2-mismatched set for testing.
Table TABREF28 presents the results of the first experiment. The DL model with GloVe word embeddings achieved better results on three datasets, with 82.80% Accuracy on SNLI, 78.52% Accuracy on MultiNLI, and 83.62% Accuracy on Quora. Logistic Regression achieved the best Accuracy of 98.60% on Clinical-RQE. We also performed a 10-fold cross-validation on the full Clinical-QE data of 8,588 question pairs, which gave 98.61% Accuracy.
In the second experiment, we used these datasets for training only and compared their performance on our test set of 850 consumer health questions. Table TABREF29 presents the results of this experiment. Logistic Regression trained on the clinical-RQE data outperformed DL models trained on all datasets, with 73.18% Accuracy.
To validate further the performance of the LR method, we evaluated it on question similarity detection. A typical approach to this task is to use an IR method to find similar question candidates, then a more sophisticated method to select and re-rank the similar questions. We followed a similar approach for this evaluation by combining the LR method with the IR baseline provided in the context of SemEval-cQA. The hybrid method combines the score provided by the Logistic Regression model and the reciprocal rank from the IR baseline using a weight-based combination:
INLINEFORM0
The weight INLINEFORM0 was set empirically through several tests on the cQA-2016 development set ( INLINEFORM1 ). Table TABREF30 presents the results on the cQA-2016 and cQA-2017 test datasets. The hybrid method (LR+IR) provided the best results on both datasets. On the 2016 test data, the LR+IR method outperformed the best system in all measures, with 80.57% Accuracy and 77.47% MAP (official system ranking measure in SemEval-cQA). On the cQA-2017 test data, the LR+IR method obtained 44.66% MAP and outperformed the cQA-2017 best system in Accuracy with 67.27%.
Discussion of RQE Results
When trained and tested on the same corpus, the DL model with GloVe embeddings gave the best results on three datasets (SNLI, MultiNLI and Quora). Logistic Regression gave the best Accuracy on the Clinical-RQE dataset with 98.60%. When tested on our test set (850 medical CHQs-FAQs pairs), Logistic Regression trained on Clinical-QE gave the best performance with 73.18% Accuracy.
The SNLI and multi-NLI models did not perform well when tested on medical RQE data. We performed additional evaluations using the RTE-1, RTE-2 and RTE-3 open-domain datasets provided by the PASCAL challenge and the results were similar. We have also tested the SemEval-cQA-2016 model and had a similar drop in performance on RQE data. This could be explained by the different types of data leading to wrong internal conceptualizations of medical terms and questions in the deep neural layers. This performance drop could also be caused by the complexity of the test consumer health questions that are often composed of several subquestions, contain contextual information, and may contain misspellings and ungrammatical sentences, which makes them more difficult to process BIBREF48 . Another aspect is the semantics of the task as discussed in Section SECREF6 . The definition of textual entailment in open-domain may not quite apply to question entailment due to the strict semantics. Also the general textual entailment definitions refer only to the premise and hypothesis, while the definition of RQE for question answering relies on the relationship between the sets of answers of the compared questions.
Building a Medical QA Collection from Trusted Resources
A RQE-based QA system requires a collection of question-answer pairs to map new user questions to the existing questions with an RQE approach, rank the retrieved questions, and present their answers to the user.
Method
To construct trusted medical question-answer pairs, we crawled websites from the National Institutes of Health (cf. Section SECREF56 ). Each web page describes a specific topic (e.g. name of a disease or a drug), and often includes synonyms of the main topic that we extracted during the crawl.
We constructed hand-crafted patterns for each website to automatically generate the question-answer pairs based on the document structure and the section titles. We also annotated each question with the associated focus (topic of the web page) as well as the question type identified with the designed patterns (cf. Section SECREF36 ).
To provide additional information about the questions that could be used for diverse IR and NLP tasks, we automatically annotated the questions with the focus, its UMLS Concept Unique Identifier (CUI) and Semantic Type. We combined two methods to recognize named entities from the titles of the crawled articles and their associated UMLS CUIs: (i) exact string matching to the UMLS Metathesaurus, and (ii) MetaMap Lite BIBREF49 . We then used the UMLS Semantic Network to retrieve the associated semantic types and groups.
Question Types
The question types were derived after the manual evaluation of 1,721 consumer health questions. Our taxonomy includes 16 types about Diseases, 20 types about Drugs and one type (Information) for the other named entities such as Procedures, Medical exams and Treatments. We describe below the considered question types and examples of associated question patterns.
Question Types about Diseases (16): Information, Research (or Clinical Trial), Causes, Treatment, Prevention, Diagnosis (Exams and Tests), Prognosis, Complications, Symptoms, Inheritance, Susceptibility, Genetic changes, Frequency, Considerations, Contact a medical professional, Support Groups.
Examples:
What research (or clinical trial) is being done for DISEASE?
What is the outlook for DISEASE?
How many people are affected by DISEASE?
When to contact a medical professional about DISEASE?
Who is at risk for DISEASE?
Where to find support for people with DISEASE?
Question Types About Drugs (20): Information, Interaction with medications, Interaction with food, Interaction with herbs and supplements, Important warning, Special instructions, Brand names, How does it work, How effective is it, Indication, Contraindication, Learn more, Side effects, Emergency or overdose, Severe reaction, Forget a dose, Dietary, Why get vaccinated, Storage and disposal, Usage, Dose.
Examples:
Are there interactions between DRUG and herbs and supplements?
What important warning or information should I know about DRUG?
Are there safety concerns or special precautions about DRUG?
What is the action of DRUG and how does it work?
Who should get DRUG and why is it prescribed?
What to do in case of a severe reaction to DRUG?
Question Type for other medical entities (e.g. Procedure, Exam, Treatment): Information.
What is Coronary Artery Bypass Surgery?
What are Liver Function Tests?
Medical Resources
We used 12 trusted websites to construct a collection of question-answer pairs. For each website, we extracted the free text of each article as well as the synonyms of the article focus (topic). These resources and their brief descriptions are provided below:
National Cancer Institute (NCI) : We extracted free text from 116 articles on various cancer types (729 QA pairs). We manually restructured the content of the articles to generate complete answers (e.g. a full answer about the treatment of all stages of a specific type of cancer). Figure FIGREF54 presents examples of QA pairs generated from a NCI article.
Genetic and Rare Diseases Information Center (GARD): This resource contains information about various aspects of genetic/rare diseases. We extracted all disease question/answer pairs from 4,278 topics (5,394 QA pairs).
Genetics Home Reference (GHR): This NLM resource contains consumer-oriented information about the effects of genetic variation on human health. We extracted 1,099 articles about diseases from this resource (5,430 QA pairs).
MedlinePlus Health Topics: This portion of MedlinePlus contains information on symptoms, causes, treatment and prevention for diseases, health conditions and wellness issues. We extracted the free texts in summary sections of 981 articles (981 QA pairs).
National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) : We extracted text from 174 health information pages on diseases studied by this institute (1,192 QA pairs).
National Institute of Neurological Disorders and Stroke (NINDS): We extracted free text from 277 information pages on neurological and stroke-related diseases from this resource (1,104 QA pairs).
NIHSeniorHealth : This website contains health and wellness information for older adults. We extracted 71 articles from this resource (769 QA pairs).
National Heart, Lung, and Blood Institute (NHLBI) : We extracted text from 135 articles on diseases, tests, procedures, and other relevant topics on disorders of heart, lung, blood, and sleep (559 QA pairs).
Centers for Disease Control and Prevention (CDC) : We extracted text from 152 articles on diseases and conditions (270 QA pairs).
MedlinePlus A.D.A.M. Medical Encyclopedia: This resource contains 4,366 articles about conditions, tests, and procedures. 17,348 QA pairs were extracted from this resource. Figure FIGREF55 presents examples of QA pairs generated from A.D.A.M encyclopedia.
MedlinePlus Drugs: We extracted free text from 1,316 articles about Drugs and generated 12,889 QA pairs.
MedlinePlus Herbs and Supplements: We extracted free text from 99 articles and generated 792 QA pairs.
The final collection contains 47,457 annotated question-answer pairs about Diseases, Drugs and other named entities (e.g. Tests) extracted from these 12 trusted resources.
The Proposed Entailment-based QA System
Our goal is to generate a ranked list of answers for a given Premise Question INLINEFORM0 by ranking the recognized Hypothesis Questions INLINEFORM1 . Based on the RQE experiments above (Section SECREF27 ), we selected Logistic Regression trained on the clinical-RQE dataset to recognize entailed questions and rank them with their classification scores.
RQE-based QA Approach
Classifying the full QA collection for each test question is not feasible for real-time applications. Therefore, we first filter the questions with an IR method to retrieve candidate questions, then classify them as entailed (or not) by the user/test question. Based on the positive results of the combination method tested on SemEval-cQA data (Section SECREF27 ), we adopted a combination method to merge the results obtained by the search engine and the RQE scores. The answers are then combined from both methods and ranked using an aggregate score. Figure FIGREF82 presents the overall architecture of the proposed QA system. We describe each module in more details next.
Finding Similar Question Candidates
For each premise question INLINEFORM0 , we use the Terrier search engine to retrieve INLINEFORM1 relevant question candidates INLINEFORM2 and then apply the RQE classifier to predict the labels for the pairs ( INLINEFORM3 , INLINEFORM4 ).
We indexed the questions of our QA collection without the associated answers. In order to improve the indexing and the performance of question retrieval, we also indexed the synonyms of the question focus and the triggers of the question type with each question. This choice allowed us to avoid the shortcomings of query expansion, including incorrect or irrelevant synonyms and the increased execution time. The synonyms of the question focus (topic) were extracted automatically from the QA collection. The triggers of each question type were defined manually in the question types taxonomy. Below are two examples of indexed questions from our QA collection, with the automatically added focus synonyms and question type triggers:
What are the treatments for Torticollis?
Focus: Torticollis. Question type: Treatment.
Added focus synonyms: "Spasmodic torticollis, Wry neck, Loxia, Cervical dystonia". Added question type triggers: "relieve, manage, cure, remedy, therapy".
What is the outlook for Legionnaire disease?
Focus: Legionnaire disease. Question Type: Prognosis.
Added focus synonyms: "Legionella pneumonia, Pontiac fever, Legionellosis". Added question type triggers: "prognosis, life expectancy".
The IR task consists of retrieving hypothesis questions INLINEFORM0 relevant to the submitted question INLINEFORM1 . As fusion of IR result has shown good performance in different tracks in TREC, we merge the results of the TF-IDF weighting function and the In-expB2 DFR model BIBREF50 .
Let INLINEFORM0 = INLINEFORM1 , INLINEFORM2 , ..., INLINEFORM3 be the set of INLINEFORM4 questions retrieved by the first IR model INLINEFORM5 and INLINEFORM6 = INLINEFORM7 , INLINEFORM8 , ..., INLINEFORM9 be the set of INLINEFORM10 questions retrieved by the second IR model INLINEFORM11 . We merge both sets by summing the scores of each retrieved question INLINEFORM12 in both INLINEFORM13 and INLINEFORM14 lists, then we re-rank the hypothesis questions INLINEFORM15 .
Combining IR and RQE Methods
The IR models and the RQE Logistic Regression model bring different perspectives to the search for relevant candidate questions. In particular, question entailment allows understanding the relations between the important terms, whereas the traditional IR methods identify the important terms, but will not notice if the relations are opposite. Moreover, some of the question types that the RQE classifier learns will not be deemed important terms by traditional IR and the most relevant questions will not be ranked at the top of the list.
Therefore, in our approach, when a question is submitted to the system, candidate questions are fetched using the IR models, then the RQE classifier is applied to filter out the non-entailed questions and re-rank the remaining candidates.
Specifically, we denote INLINEFORM0 the list of question candidates INLINEFORM1 returned by the IR system. The premise question INLINEFORM2 is then used to construct N question pairs INLINEFORM3 . The RQE classifier is then applied to filter out the question pairs that are not entailed and re-rank the remaining pairs.
More precisely, let INLINEFORM0 = INLINEFORM1 in INLINEFORM2 be the list of selected candidate questions that have a positive entailment relation with a given premise question INLINEFORM3 . We rank INLINEFORM4 by computing a hybrid score INLINEFORM5 for each candidate question INLINEFORM6 taking into account the score of the IR system INLINEFORM7 and the score of the RQE system INLINEFORM8 .
For each system INLINEFORM0 INLINEFORM1 , we normalize the associated score by dividing it by the maximum score among the INLINEFORM2 candidate questions retrieved by INLINEFORM3 for INLINEFORM4 :
INLINEFORM0
INLINEFORM0 INLINEFORM1
In our experiments, we fixed the value of INLINEFORM0 to 100. This threshold value was selected as a safe value for this task for the following reasons:
Our collection of 47,457 question-answer pairs was collected from only 12 NIH institutes and is unlikely to contain more than 100 occurrences of the same focus-type pair.
Each question was indexed with additional annotations for the question focus, its synonyms and the question type synonyms.
Evaluating RQE for Medical Question Answering
The objective of this evaluation is to study the effectiveness of RQE for Medical Question Answering, by comparing the answers retrieved by the hybrid entailment-based approach, the IR method and the other QA systems participating to the medical task at TREC 2017 LiveQA challenge (LiveQA-Med).
Evaluation Method
We developed an interface to perform the manual evaluation of the retrieved answers. Figure 5 presents the evaluation interface showing, for each test question, the top-10 answers of the evaluated QA method and the reference answer(s) used by LiveQA assessors to help judging the retrieved answers by the participating systems.
We used the test questions of the medical task at TREC-2017 LiveQA BIBREF11 . These questions are randomly selected from the consumer health questions that the NLM receives daily from all over the world. The test questions cover different medical entities and have a wide list of question types such as Comparison, Diagnosis, Ingredient, Side effects and Tapering.
For a relevant comparison, we used the same judgment scores as the LiveQA Track:
Correct and Complete Answer (4)
Correct but Incomplete (3)
Incorrect but Related (2)
Incorrect (1)
We evaluated the answers returned by the IR-based method and the hybrid QA method (IR+RQE) according to the same reference answers used in LiveQA-Med. The answers were anonymized (the method names were blinded) and presented to 3 assessors: a medical doctor (Assessor A), a medical librarian (B) and a researcher in medical informatics (C). None of the assessors participated in the development of the QA methods. Assessors B and C evaluated 1,000 answers retrieved by each of the methods (IR and IR+RQE). Assessor A evaluated 2,000 answers from both methods.
Table TABREF103 presents the inter-annotator agreement (IAA) through F1 score computed by considering one of the assessors as reference. In the first evaluation, we computed the True Positives (TP) and False Positives (FP) over all ratings and the Precision and F1 score. As there are no negative labels (only true or false positives for each category), Recall is 100%. We also computed a partial IAA by grouping the "Correct and Complete Answer" and "Correct but Incomplete" ratings (as Correct), and the "Incorrect but Related" and "Incorrect" ratings (as Incorrect). The average agreement on distinguishing the Correct and Incorrect answers is 94.33% F1 score. Therefore, we used the evaluations performed by assessor A for both methods. The official results of the TREC LiveQA track relied on one assessor per question as well.
Evaluation of the first retrieved answer
We computed the measures used by TREC LiveQA challenges BIBREF51 , BIBREF11 to evaluate the first retrieved answer for each test question:
avgScore(0-3): the average score over all questions, transferring 1-4 level grades to 0-3 scores. This is the main score used to rank LiveQA runs.
succ@i+: the number of questions with score i or above (i INLINEFORM0 {2..4}) divided by the total number of questions.
prec@i+: the number of questions with score i or above (i INLINEFORM0 {2..4}) divided by number of questions answered by the system.
Table TABREF108 presents the average scores, success and precision results. The hybrid IR+RQE QA system achieved better results than the IR-based system with 0.827 average score. It also achieved a higher score than the best results achieved in the medical challenge at LiveQA'17. Evaluating the RQE system alone is not relevant, as applying RQE on the full collection for each user question is not feasible for a real-time system because of the extended execution time.
Evaluation of the top ten answers
In this evaluation, we used Mean Average Precision (MAP) and Mean Reciprocal Rank (MRR) which are commonly used in QA to evaluate the top-10 answers for each question. We consider answers rated as “Correct and Complete Answer” or “Correct but Incomplete” as correct answers, as the test questions contain multiple subquestions while each answer in our QA collection can cover only one subquestion.
MAP is the mean of the Average Precision (AvgP) scores over all questions.
(1) INLINEFORM0
Q is the number of questions. INLINEFORM0 is the AvgP of the INLINEFORM1 question.
INLINEFORM0
K is the number of correct answers. INLINEFORM0 is the rank of INLINEFORM1 correct answer.
MRR is the average of the reciprocal ranks for each question. The reciprocal rank of a question is the multiplicative inverse of the rank of the first correct answer.
(2) INLINEFORM0
Q is the number of questions. INLINEFORM0 is the rank of the first correct answer for the INLINEFORM1 question.
Table TABREF113 presents the MAP@10 and MRR@10 of our QA methods. The IR+RQE system outperforms the IR-based QA system with 0.311 MAP@10 and 0.333 MRR@10.
Discussion of entailment-based QA for the medical domain
In our evaluation, we followed the same LiveQA guidelines with the highest possible rigor. In particular, we consulted with NIST assessors who provided us with the paraphrases of the test questions that they used to judge the answers. Our IAA on the answers rating was also high compared to related tasks, with an 88.5% F1 agreement with the exact four categories and a 94.3% agreement when reducing the categories to two: “Correct” and “Incorrect” answers. Our results show that RQE improves the overall performance and exceeds the best results in the medical LiveQA'17 challenge by a factor of 29.8%. This performance improvement is particularly interesting as:
Our answer source has only 47K question-answer pairs when LiveQA participating systems relied on much larger collections, including the World Wide Web.
Our system answered one subquestion at most when many LiveQA test questions had several subquestions.
The latter observation, (b), makes the hybrid IR+RQE approach even more promising as it gives it a large potential for the improvement of answer completeness.
The former observation, (a), provides another interesting insight: restricting the answer source to only reliable collections can actually improve the QA performance without losing coverage (i.e., our QA approach provided at least one answer to each test question and obtained the best relevance score).
In another observation, the assessors reported that many of the returned answers had a correct question type but a wrong focus, which indicates that including a focus recognition module to filter such wrong answers can improve further the QA performance in terms of precision. Another aspect that was reported is the repetition of the same (or similar) answer from different websites, which could be addressed by improving answer selection with inter-answer comparisons and removal of near-duplicates. Also, half of the LiveQA test questions are about Drugs, when only two of our resources are specialized in Drugs, among 12 sub-collections overall. Accordingly, the assessors noticed that the performance of the QA systems was better on questions about diseases than on questions about drugs, which suggests a need for extending our medical QA collection with more information about drugs and associated question types.
We also looked closely at the private websites used by the LiveQA-Med annotators to provide some of the reference answers for the test questions. For instance, the ConsumerLab website was useful to answer a question about the ingredients of a Drug (COENZYME Q10). Similarly, the eHealthMe website was used to answer a test question asking about interactions between two drugs (Phentermine and Dicyclomine) when no information was found in DailyMed. eHealthMe provides healthcare big data analysis and private research and studies including self-reported adverse drug effects by patients.
But the question remains on the extent to which such big data and other private websites could be used to automatically answer medical questions if information is otherwise unavailable. Unlike medical professionals, patients do not necessarily have the knowledge and tools to validate such information. An alternative approach could be to put limitations on medical QA systems in terms of the questions that can be answered (e.g. "What is my diagnosis for such symptoms") and build classifiers to detect such questions and warn the users about the dangers of looking for their answers online.
More generally, medical QA systems should follow some strict guidelines regarding the goal and background knowledge and resources of each system in order to protect the consumers from misleading or harmful information. Such guidelines could be based (i) on the source of the information such as health and medical information websites sponsored by the U.S. government, not-for-profit health or medical organizations, and medical university centers, or (ii) on conventions such as the code of conduct of the HON Foundation (HONcode) that addresses the reliability and usefulness of medical information on the Internet.
Our experiments show that limiting the number of answer sources with such guidelines is not only feasible, but it could also enhance the performance of the QA system from an information retrieval perspective.
Conclusion
In this paper, we carried out an empirical study of machine learning and deep learning methods for Recognizing Question Entailment in the medical domain using several datasets. We developed a RQE-based QA system to answer new medical questions using existing question-answer pairs. We built and shared a collection of 47K medical question-answer pairs. Our QA approach outperformed the best results on TREC-2017 LiveQA medical test questions. The proposed approach can be applied and adapted to open-domain as well as specific-domain QA. Deep learning models achieved interesting results on open-domain and clinical datasets, but obtained a lower performance on consumer health questions. We will continue investigating other network architectures including transfer learning, as well as creation of a large collection of consumer health questions for training to improve the performance of DL models. Future work also includes exploring integration of a Question Focus Recognition module to enhance candidate question retrieval, and expanding our question-answer collection.
Acknowledgements
We thank Halil Kilicoglu (NLM/NIH) for his help with the crawling and the manual evaluation and Sonya E. Shooshan (NLM/NIH) for her help with the judgment of the retrieved answers. We also thank Ellen Voorhees (NIST) for her valuable support with the TREC LiveQA evaluation.
We consider the case of the question number 36 in the TREC-2017 LiveQA medical test dataset:
36. congenital diaphragmatic hernia. what are the causes of congenital diaphragmatic hernia? Can cousin marriage cause this? What kind of lung disease the baby might experience life long?
This question was answered by 5 participating runs (vs. 8 runs for other questions), and all submitted answers were wrong (scores of 1 or 2). However, our IR-based QA system retrieved one excellent answer (score 4) and our hybrid IR+RQE system provided 3 excellent answers.
A) TREC 2017 LiveQA-Med Participants' Results:
B) Our IR-based QA System:
C) Our IR+RQE QA System:
| [
"Logistic Regression, neural networks"
] | 7,257 | qasper | en | null | 89a4fd3fce6114c3401790c6f9b5243fda094597657f348a | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What machine learning and deep learning methods are used for RQE?
Answer: | [
"Logistic Regression, neural networks"
] | qasper | 128 | 67 | |
What is the benchmark dataset and is its quality high? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Microblogging such as Twitter and Weibo is a popular social networking service, which allows users to post messages up to 140 characters. There are millions of active users on the platform who stay connected with friends. Unfortunately, spammers also use it as a tool to post malicious links, send unsolicited messages to legitimate users, etc. A certain amount of spammers could sway the public opinion and cause distrust of the social platform. Despite the use of rigid anti-spam rules, human-like spammers whose homepages having photos, detailed profiles etc. have emerged. Unlike previous "simple" spammers, whose tweets contain only malicious links, those "smart" spammers are more difficult to distinguish from legitimate users via content-based features alone BIBREF0 .
There is a considerable amount of previous work on spammer detection on social platforms. Researcher from Twitter Inc. BIBREF1 collect bot accounts and perform analysis on the user behavior and user profile features. Lee et al. lee2011seven use the so-called social honeypot by alluring social spammers' retweet to build a benchmark dataset, which has been extensively explored in our paper. Some researchers focus on the clustering of urls in tweets and network graph of social spammers BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , showing the power of social relationship features.As for content information modeling, BIBREF6 apply improved sparse learning methods. However, few studies have adopted topic-based features. Some researchers BIBREF7 discuss topic characteristics of spamming posts, indicating that spammers are highly likely to dwell on some certain topics such as promotion. But this may not be applicable to the current scenario of smart spammers.
In this paper, we propose an efficient feature extraction method. In this method, two new topic-based features are extracted and used to discriminate human-like spammers from legitimate users. We consider the historical tweets of each user as a document and use the Latent Dirichlet Allocation (LDA) model to compute the topic distribution for each user. Based on the calculated topic probability, two topic-based features, the Local Outlier Standard Score (LOSS) which captures the user's interests on different topics and the Global Outlier Standard Score (GOSS) which reveals the user's interests on specific topic in comparison with other users', are extracted. The two features contain both local and global information, and the combination of them can distinguish human-like spammers effectively.
To the best of our knowledge, it is the first time that features based on topic distributions are used in spammer classification. Experimental results on one public dataset and one self-collected dataset further validate that the two sets of extracted topic-based features get excellent performance on human-like spammer classification problem compared with other state-of-the-art methods. In addition, we build a Weibo dataset, which contains both legitimate users and spammers.
To summarize, our major contributions are two-fold:
In the following sections, we first propose the topic-based features extraction method in Section 2, and then introduce the two datasets in Section 3. Experimental results are discussed in Section 4, and we conclude the paper in Section 5. Future work is presented in Section 6.
Methodology
In this section, we first provide some observations we obtained after carefully exploring the social network, then the LDA model is introduced. Based on the LDA model, the ways to obtain the topic probability vector for each user and the two topic-based features are provided.
Observation
After exploring the homepages of a substantial number of spammers, we have two observations. 1) social spammers can be divided into two categories. One is content polluters, and their tweets are all about certain kinds of advertisement and campaign. The other is fake accounts, and their tweets resemble legitimate users' but it seems they are simply random copies of others to avoid being detected by anti-spam rules. 2) For legitimate users, content polluters and fake accounts, they show different patterns on topics which interest them.
Legitimate users mainly focus on limited topics which interest him. They seldom post contents unrelated to their concern.
Content polluters concentrate on certain topics.
Fake accounts focus on a wide range of topics due to random copying and retweeting of other users' tweets.
Spammers and legitimate users show different interests on some topics e.g. commercial, weather, etc.
To better illustrate our observation, Figure. 1 shows the topic distribution of spammers and legitimate users in two employed datasets(the Honeypot dataset and Weibo dataset). We can see that on both topics (topic-3 and topic-11) there exists obvious difference between the red bars and green bars, representing spammers and legitimate users. On the Honeypot dataset, spammers have a narrower shape of distribution (the outliers on the red bar tail are not counted) than that of legitimate users. This is because there are more content polluters than fake accounts. In other word, spammers in this dataset tend to concentrate on limited topics. While on the Weibo dataset, fake accounts who are interested in different topics take large proportion of spammers. Their distribution is more flat (i.e. red bars) than that of the legitimate users. Therefore we can detect spammers by means of the difference of their topic distribution patterns.
LDA model
Blei et al.blei2003latent first presented Latent Dirichlet Allocation(LDA) as an example of topic model.
Each document $i$ is deemed as a bag of words $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $ and $M$ is the number of words. Each word is attributable to one of the document's topics $Z=\left\lbrace z_{i1},z_{i2},...,z_{iK}\right\rbrace $ and $K$ is the number of topics. $\psi _{k}$ is a multinomial distribution over words for topic $k$ . $\theta _i$ is another multinomial distribution over topics for document $i$ . The smoothed generative model is illustrated in Figure. 2 . $\alpha $ and $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $0 are hyper parameter that affect scarcity of the document-topic and topic-word distributions. In this paper, $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $1 , $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $2 and $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $3 are empirically set to 0.3, 0.01 and 15. The entire content of each Twitter user is regarded as one document. We adopt Gibbs Sampling BIBREF8 to speed up the inference of LDA. Based on LDA, we can get the topic probabilities for all users in the employed dataset as: $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $4 , where $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $5 is the number of users. Each element $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $6 is a topic probability vector for the $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $7 document. $W=\left\lbrace w_{i1},w_{i2},...,w_{iM}\right\rbrace $8 is the raw topic probability vector and our features are developed on top of it.
Topic-based Features
Using the LDA model, each person in the dataset is with a topic probability vector $X_i$ . Assume $x_{ik}\in X_{i}$ denotes the likelihood that the $\emph {i}^{th}$ tweet account favors $\emph {k}^{th}$ topic in the dataset. Our topic based features can be calculated as below.
Global Outlier Standard Score measures the degree that a user's tweet content is related to a certain topic compared to the other users. Specifically, the "GOSS" score of user $i$ on topic $k$ can be calculated as Eq.( 12 ):
$$\centering \begin{array}{ll} \mu \left(x_{k}\right)=\frac{\sum _{i=1}^{n} x_{ik}}{n},\\ GOSS\left(x_{ik}\right)=\frac{x_{ik}-\mu \left(x_k\right)}{\sqrt{\underset{i}{\sum }\left(x_{ik}-\mu \left(x_{k}\right)\right)^{2}}}. \end{array}$$ (Eq. 12)
The value of $GOSS\left(x_{ik}\right)$ indicates the interesting degree of this person to the $\emph {k}^{th}$ topic. Specifically, if $GOSS\left(x_{ik}\right)$ > $GOSS\left(x_{jk}\right)$ , it means that the $\emph {i}^{th}$ person has more interest in topic $k$ than the $\emph {j}^{th}$ person. If the value $GOSS\left(x_{ik}\right)$ is extremely high or low, the $\emph {i}^{th}$ person showing extreme interest or no interest on topic $k$ which will probably be a distinctive pattern in the fowllowing classfication. Therefore, the topics interested or disliked by the $\emph {k}^{th}$0 person can be manifested by $\emph {k}^{th}$1 , from which the pattern of the interested topics with regarding to this person is found. Denote $\emph {k}^{th}$2 our first topic-based feature, and it hopefully can get good performance on spammer detection.
Local Outlier Standard Score measures the degree of interest someone shows to a certain topic by considering his own homepage content only. For instance, the "LOSS" score of account $i$ on topic $k$ can be calculated as Eq.( 13 ):
$$\centering \begin{array}{ll} \mu \left(x_{i}\right)=\frac{\sum _{k=1}^{K} x_{ik}}{K},\\ LOSS\left(x_{ik}\right)=\frac{x_{ik}-\mu \left(x_i\right)}{\sqrt{\underset{k}{\sum }\left(x_{ik}-\mu \left(x_{i}\right)\right)^{2}}}. \end{array}$$ (Eq. 13)
$\mu (x_i)$ represents the averaged interesting degree for all topics with regarding to $\emph {i}^{th}$ user and his tweet content. Similarly to $GOSS$ , the topics interested or disliked by the $\emph {i}^{th}$ person via considering his single post information can be manifested by $f_{LOSS}^{i}=[LOSS(x_{i1})\cdots LOSS(x_{iK})]$ , and $LOSS$ becomes our second topic-based features for the $\emph {i}^{th}$ person.
Dataset
We use one public dataset Social Honeypot dataset and one self-collected dataset Weibo dataset to validate the effectiveness of our proposed features.
Social Honeypot Dataset: Lee et al. lee2010devils created and deployed 60 seed social accounts on Twitter to attract spammers by reporting back what accounts interact with them. They collected 19,276 legitimate users and 22,223 spammers in their datasets along with their tweet content in 7 months. This is our first test dataset.
Our Weibo Dataset: Sina Weibo is one of the most famous social platforms in China. It has implemented many features from Twitter. The 2197 legitimate user accounts in this dataset are provided by the Tianchi Competition held by Sina Weibo. The spammers are all purchased commercially from multiple vendors on the Internet. We checked them manually and collected 802 suitable "smart" spammers accounts.
Preprocessing: Before directly performing the experiments on the employed datasets, we first delete some accounts with few posts in the two employed since the number of tweets is highly indicative of spammers. For the English Honeypot dataset, we remove stopwords, punctuations, non-ASCII words and apply stemming. For the Chinese Weibo dataset, we perform segmentation with "Jieba", a Chinese text segmentation tool. After preprocessing steps, the Weibo dataset contains 2197 legitimate users and 802 spammers, and the honeypot dataset contains 2218 legitimate users and 2947 spammers. It is worth mentioning that the Honeypot dataset has been slashed because most of the Twitter accounts only have limited number of posts, which are not enough to show their interest inclination.
Evaluation Metrics
The evaluating indicators in our model are show in 2 . We calculate precision, recall and F1-score (i.e. F1 score) as in Eq. ( 19 ). Precision is the ratio of selected accounts that are spammers. Recall is the ratio of spammers that are detected so. F1-score is the harmonic mean of precision and recall.
$$precision =\frac{TP}{TP+FP}, recall =\frac{TP}{TP+FN}\nonumber \\ F1-score = \frac{2\times precision \times recall}{precision + recall}$$ (Eq. 19)
Performance Comparisons with Baseline
Three baseline classification methods: Support Vector Machines (SVM), Adaboost, and Random Forests are adopted to evaluate our extracted features. We test each classification algorithm with scikit-learn BIBREF9 and run a 10-fold cross validation. On each dataset, the employed classifiers are trained with individual feature first, and then with the combination of the two features. From 1 , we can see that GOSS+LOSS achieves the best performance on F1-score among all others. Besides, the classification by combination of LOSS and GOSS can increase accuracy by more than 3% compared with raw topic distribution probability.
Comparison with Other Features
To compare our extracted features with previously used features for spammer detection, we use three most discriminative feature sets according to Lee et al. lee2011seven( 4 ). Two classifiers (Adaboost and SVM) are selected to conduct feature performance comparisons. Using Adaboost, our LOSS+GOSS features outperform all other features except for UFN which is 2% higher than ours with regard to precision on the Honeypot dataset. It is caused by the incorrectly classified spammers who are mostly news source after our manual check. They keep posting all kinds of news pieces covering diverse topics, which is similar to the behavior of fake accounts. However, UFN based on friendship networks is more useful for public accounts who possess large number of followers. The best recall value of our LOSS+GOSS features using SVM is up to 6% higher than the results by other feature groups. Regarding F1-score, our features outperform all other features. To further show the advantages of our proposed features, we compare our combined LOSS+GOSS with the combination of all the features from Lee et al. lee2011seven (UFN+UC+UH). It's obvious that LOSS+GOSS have a great advantage over UFN+UC+UH in terms of recall and F1-score. Moreover, by combining our LOSS+GOSS features and UFN+UC+UH features together, we obtained another 7.1% and 2.3% performance gain with regard to precision and F1-score by Adaboost. Though there is a slight decline in terms of recall. By SVM, we get comparative results on recall and F1-score but about 3.5% improvement on precision.
Conclusion
In this paper, we propose a novel feature extraction method to effectively detect "smart" spammers who post seemingly legitimate tweets and are thus difficult to identify by existing spammer classification methods. Using the LDA model, we obtain the topic probability for each Twitter user. By utilizing the topic probability result, we extract our two topic-based features: GOSS and LOSS which represent the account with global and local information. Experimental results on a public dataset and a self-built Chinese microblog dataset validate the effectiveness of the proposed features.
Future Work
In future work, the combination method of local and global information can be further improved to maximize their individual strengths. We will also apply decision theory to enhancing the performance of our proposed features. Moreover, we are also building larger datasets on both Twitter and Weibo to validate our method. Moreover, larger datasets on both Twitter and Weibo will be built to further validate our method.
| [
"Social Honeypot dataset (public) and Weibo dataset (self-collected); yes",
"Social Honeypot, which is not of high quality"
] | 2,242 | qasper | en | null | 9c46084d667b60a92a3cebfb8fa56436fa1497668ce2af56 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What is the benchmark dataset and is its quality high?
Answer: | [
"Social Honeypot dataset (public) and Weibo dataset (self-collected); yes",
"Social Honeypot, which is not of high quality"
] | qasper | 128 | 68 | |
What architecture does the decoder have? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
This paper describes our approach and results for Task 2 of the CoNLL–SIGMORPHON 2018 shared task on universal morphological reinflection BIBREF0 . The task is to generate an inflected word form given its lemma and the context in which it occurs.
Morphological (re)inflection from context is of particular relevance to the field of computational linguistics: it is compelling to estimate how well a machine-learned system can capture the morphosyntactic properties of a word given its context, and map those properties to the correct surface form for a given lemma.
There are two tracks of Task 2 of CoNLL–SIGMORPHON 2018: in Track 1 the context is given in terms of word forms, lemmas and morphosyntactic descriptions (MSD); in Track 2 only word forms are available. See Table TABREF1 for an example. Task 2 is additionally split in three settings based on data size: high, medium and low, with high-resource datasets consisting of up to 70K instances per language, and low-resource datasets consisting of only about 1K instances.
The baseline provided by the shared task organisers is a seq2seq model with attention (similar to the winning system for reinflection in CoNLL–SIGMORPHON 2016, BIBREF1 ), which receives information about context through an embedding of the two words immediately adjacent to the target form. We use this baseline implementation as a starting point and achieve the best overall accuracy of 49.87 on Task 2 by introducing three augmentations to the provided baseline system: (1) We use an LSTM to encode the entire available context; (2) We employ a multi-task learning approach with the auxiliary objective of MSD prediction; and (3) We train the auxiliary component in a multilingual fashion, over sets of two to three languages.
In analysing the performance of our system, we found that encoding the full context improves performance considerably for all languages: 11.15 percentage points on average, although it also highly increases the variance in results. Multi-task learning, paired with multilingual training and subsequent monolingual finetuning, scored highest for five out of seven languages, improving accuracy by another 9.86% on average.
System Description
Our system is a modification of the provided CoNLL–SIGMORPHON 2018 baseline system, so we begin this section with a reiteration of the baseline system architecture, followed by a description of the three augmentations we introduce.
Baseline
The CoNLL–SIGMORPHON 2018 baseline is described as follows:
The system is an encoder-decoder on character sequences. It takes a lemma as input and generates a word form. The process is conditioned on the context of the lemma [...] The baseline treats the lemma, word form and MSD of the previous and following word as context in track 1. In track 2, the baseline only considers the word forms of the previous and next word. [...] The baseline system concatenates embeddings for context word forms, lemmas and MSDs into a context vector. The baseline then computes character embeddings for each character in the input lemma. Each of these is concatenated with a copy of the context vector. The resulting sequence of vectors is encoded using an LSTM encoder. Subsequently, an LSTM decoder generates the characters in the output word form using encoder states and an attention mechanism.
To that we add a few details regarding model size and training schedule:
the number of LSTM layers is one;
embedding size, LSTM layer size and attention layer size is 100;
models are trained for 20 epochs;
on every epoch, training data is subsampled at a rate of 0.3;
LSTM dropout is applied at a rate 0.3;
context word forms are randomly dropped at a rate of 0.1;
the Adam optimiser is used, with a default learning rate of 0.001; and
trained models are evaluated on the development data (the data for the shared task comes already split in train and dev sets).
Our system
Here we compare and contrast our system to the baseline system. A diagram of our system is shown in Figure FIGREF4 .
The idea behind this modification is to provide the encoder with access to all morpho-syntactic cues present in the sentence. In contrast to the baseline, which only encodes the immediately adjacent context of a target word, we encode the entire context. All context word forms, lemmas, and MSD tags (in Track 1) are embedded in their respective high-dimensional spaces as before, and their embeddings are concatenated. However, we now reduce the entire past context to a fixed-size vector by encoding it with a forward LSTM, and we similarly represent the future context by encoding it with a backwards LSTM.
We introduce an auxiliary objective that is meant to increase the morpho-syntactic awareness of the encoder and to regularise the learning process—the task is to predict the MSD tag of the target form. MSD tag predictions are conditioned on the context encoding, as described in UID15 . Tags are generated with an LSTM one component at a time, e.g. the tag PRO;NOM;SG;1 is predicted as a sequence of four components, INLINEFORM0 PRO, NOM, SG, 1 INLINEFORM1 .
For every training instance, we backpropagate the sum of the main loss and the auxiliary loss without any weighting.
As MSD tags are only available in Track 1, this augmentation only applies to this track.
The parameters of the entire MSD (auxiliary-task) decoder are shared across languages.
Since a grouping of the languages based on language family would have left several languages in single-member groups (e.g. Russian is the sole representative of the Slavic family), we experiment with random groupings of two to three languages. Multilingual training is performed by randomly alternating between languages for every new minibatch. We do not pass any information to the auxiliary decoder as to the source language of the signal it is receiving, as we assume abstract morpho-syntactic features are shared across languages.
After 20 epochs of multilingual training, we perform 5 epochs of monolingual finetuning for each language. For this phase, we reduce the learning rate to a tenth of the original learning rate, i.e. 0.0001, to ensure that the models are indeed being finetuned rather than retrained.
We keep all hyperparameters the same as in the baseline. Training data is split 90:10 for training and validation. We train our models for 50 epochs, adding early stopping with a tolerance of five epochs of no improvement in the validation loss. We do not subsample from the training data.
We train models for 50 different random combinations of two to three languages in Track 1, and 50 monolingual models for each language in Track 2. Instead of picking the single model that performs best on the development set and thus risking to select a model that highly overfits that data, we use an ensemble of the five best models, and make the final prediction for a given target form with a majority vote over the five predictions.
Results and Discussion
Test results are listed in Table TABREF17 . Our system outperforms the baseline for all settings and languages in Track 1 and for almost all in Track 2—only in the high resource setting is our system not definitively superior to the baseline.
Interestingly, our results in the low resource setting are often higher for Track 2 than for Track 1, even though contextual information is less explicit in the Track 2 data and the multilingual multi-tasking approach does not apply to this track. We interpret this finding as an indicator that a simpler model with fewer parameters works better in a setting of limited training data. Nevertheless, we focus on the low resource setting in the analysis below due to time limitations. As our Track 1 results are still substantially higher than the baseline results, we consider this analysis valid and insightful.
Ablation Study
We analyse the incremental effect of the different features in our system, focusing on the low-resource setting in Track 1 and using development data.
Encoding the entire context with an LSTM highly increases the variance of the observed results. So we trained fifty models for each language and each architecture. Figure FIGREF23 visualises the means and standard deviations over the trained models. In addition, we visualise the average accuracy for the five best models for each language and architecture, as these are the models we use in the final ensemble prediction. Below we refer to these numbers only.
The results indicate that encoding the full context with an LSTM highly enhances the performance of the model, by 11.15% on average. This observation explains the high results we obtain also for Track 2.
Adding the auxiliary objective of MSD prediction has a variable effect: for four languages (de, en, es, and sv) the effect is positive, while for the rest it is negative. We consider this to be an issue of insufficient data for the training of the auxiliary component in the low resource setting we are working with.
We indeed see results improving drastically with the introduction of multilingual training, with multilingual results being 7.96% higher than monolingual ones on average.
We studied the five best models for each language as emerging from the multilingual training (listed in Table TABREF27 ) and found no strong linguistic patterns. The en–sv pairing seems to yield good models for these languages, which could be explained in terms of their common language family and similar morphology. The other natural pairings, however, fr–es, and de–sv, are not so frequent among the best models for these pairs of languages.
Finally, monolingual finetuning improves accuracy across the board, as one would expect, by 2.72% on average.
The final observation to be made based on this breakdown of results is that the multi-tasking approach paired with multilingual training and subsequent monolingual finetuning outperforms the other architectures for five out of seven languages: de, en, fr, ru and sv. For the other two languages in the dataset, es and fi, the difference between this approach and the approach that emerged as best for them is less than 1%. The overall improvement of the multilingual multi-tasking approach over the baseline is 18.30%.
Error analysis
Here we study the errors produced by our system on the English test set to better understand the remaining shortcomings of the approach. A small portion of the wrong predictions point to an incorrect interpretation of the morpho-syntactic conditioning of the context, e.g. the system predicted plan instead of plans in the context Our _ include raising private capital. The majority of wrong predictions, however, are nonsensical, like bomb for job, fify for fixing, and gnderrate for understand. This observation suggests that generally the system did not learn to copy the characters of lemma into inflected form, which is all it needs to do in a large number of cases. This issue could be alleviated with simple data augmentation techniques that encourage autoencoding BIBREF2 .
MSD prediction
Figure FIGREF32 summarises the average MSD-prediction accuracy for the multi-tasking experiments discussed above. Accuracy here is generally higher than on the main task, with the multilingual finetuned setup for Spanish and the monolingual setup for French scoring best: 66.59% and 65.35%, respectively. This observation illustrates the added difficulty of generating the correct surface form even when the morphosyntactic description has been identified correctly.
We observe some correlation between these numbers and accuracy on the main task: for de, en, ru and sv, the brown, pink and blue bars here pattern in the same way as the corresponding INLINEFORM0 's in Figure FIGREF23 . One notable exception to this pattern is fr where inflection gains a lot from multilingual training, while MSD prediction suffers greatly. Notice that the magnitude of change is not always the same, however, even when the general direction matches: for ru, for example, multilingual training benefits inflection much more than in benefits MSD prediction, even though the MSD decoder is the only component that is actually shared between languages. This observation illustrates the two-fold effect of multi-task training: an auxiliary task can either inform the main task through the parameters the two tasks share, or it can help the main task learning through its regularising effect.
Related Work
Our system is inspired by previous work on multi-task learning and multi-lingual learning, mainly building on two intuitions: (1) jointly learning related tasks tends to be beneficial BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 ; and (2) jointly learning related languages in an MTL-inspired framework tends to be beneficial BIBREF8 , BIBREF9 , BIBREF10 . In the context of computational morphology, multi-lingual approaches have previously been employed for morphological reinflection BIBREF2 and for paradigm completion BIBREF11 . In both of these cases, however, the available datasets covered more languages, 40 and 21, respectively, which allowed for linguistically-motivated language groupings and for parameter sharing directly on the level of characters. BIBREF10 explore parameter sharing between related languages for dependency parsing, and find that sharing is more beneficial in the case of closely related languages.
Conclusions
In this paper we described our system for the CoNLL–SIGMORPHON 2018 shared task on Universal Morphological Reinflection, Task 2, which achieved the best performance out of all systems submitted, an overall accuracy of 49.87. We showed in an ablation study that this is due to three core innovations, which extend a character-based encoder-decoder model: (1) a wide context window, encoding the entire available context; (2) multi-task learning with the auxiliary task of MSD prediction, which acts as a regulariser; (3) a multilingual approach, exploiting information across languages. In future work we aim to gain better understanding of the increase in variance of the results introduced by each of our modifications and the reasons for the varying effect of multi-task learning for different languages.
Acknowledgements
We gratefully acknowledge the support of the NVIDIA Corporation with the donation of the Titan Xp GPU used for this research.
| [
"LSTM",
"LSTM"
] | 2,289 | qasper | en | null | f755dcbd288905ec07a63f18ddee7ed22103c45e887a091e | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What architecture does the decoder have?
Answer: | [
"LSTM",
"LSTM"
] | qasper | 128 | 69 | |
Do they report results only on English data? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
With the increasing popularity of the Internet, online texts provided by social media platform (e.g. Twitter) and news media sites (e.g. Google news) have become important sources of real-world events. Therefore, it is crucial to automatically extract events from online texts.
Due to the high variety of events discussed online and the difficulty in obtaining annotated data for training, traditional template-based or supervised learning approaches for event extraction are no longer applicable in dealing with online texts. Nevertheless, newsworthy events are often discussed by many tweets or online news articles. Therefore, the same event could be mentioned by a high volume of redundant tweets or news articles. This property inspires the research community to devise clustering-based models BIBREF0 , BIBREF1 , BIBREF2 to discover new or previously unidentified events without extracting structured representations.
To extract structured representations of events such as who did what, when, where and why, Bayesian approaches have made some progress. Assuming that each document is assigned to a single event, which is modeled as a joint distribution over the named entities, the date and the location of the event, and the event-related keywords, Zhou et al. zhou2014simple proposed an unsupervised Latent Event Model (LEM) for open-domain event extraction. To address the limitation that LEM requires the number of events to be pre-set, Zhou et al. zhou2017event further proposed the Dirichlet Process Event Mixture Model (DPEMM) in which the number of events can be learned automatically from data. However, both LEM and DPEMM have two limitations: (1) they assume that all words in a document are generated from a single event which can be represented by a quadruple INLINEFORM0 entity, location, keyword, date INLINEFORM1 . However, long texts such news articles often describe multiple events which clearly violates this assumption; (2) During the inference process of both approaches, the Gibbs sampler needs to compute the conditional posterior distribution and assigns an event for each document. This is time consuming and takes long time to converge.
To deal with these limitations, in this paper, we propose the Adversarial-neural Event Model (AEM) based on adversarial training for open-domain event extraction. The principle idea is to use a generator network to learn the projection function between the document-event distribution and four event related word distributions (entity distribution, location distribution, keyword distribution and date distribution). Instead of providing an analytic approximation, AEM uses a discriminator network to discriminate between the reconstructed documents from latent events and the original input documents. This essentially helps the generator to construct a more realistic document from a random noise drawn from a Dirichlet distribution. Due to the flexibility of neural networks, the generator is capable of learning complicated nonlinear distributions. And the supervision signal provided by the discriminator will help generator to capture the event-related patterns. Furthermore, the discriminator also provides low-dimensional discriminative features which can be used to visualize documents and events.
The main contributions of the paper are summarized below:
Related Work
Our work is related to two lines of research, event extraction and Generative Adversarial Nets.
Event Extraction
Recently there has been much interest in event extraction from online texts, and approaches could be categorized as domain-specific and open-domain event extraction.
Domain-specific event extraction often focuses on the specific types of events (e.g. sports events or city events). Panem et al. panem2014structured devised a novel algorithm to extract attribute-value pairs and mapped them to manually generated schemes for extracting the natural disaster events. Similarly, to extract the city-traffic related event, Anantharam et al. anantharam2015extracting viewed the task as a sequential tagging problem and proposed an approach based on the conditional random fields. Zhang zhang2018event proposed an event extraction approach based on imitation learning, especially on inverse reinforcement learning.
Open-domain event extraction aims to extract events without limiting the specific types of events. To analyze individual messages and induce a canonical value for each event, Benson et al. benson2011event proposed an approach based on a structured graphical model. By representing an event with a binary tuple which is constituted by a named entity and a date, Ritter et al. ritter2012open employed some statistic to measure the strength of associations between a named entity and a date. The proposed system relies on a supervised labeler trained on annotated data. In BIBREF1 , Abdelhaq et al. developed a real-time event extraction system called EvenTweet, and each event is represented as a triple constituted by time, location and keywords. To extract more information, Wang el al. wang2015seeft developed a system employing the links in tweets and combing tweets with linked articles to identify events. Xia el al. xia2015new combined texts with the location information to detect the events with low spatial and temporal deviations. Zhou et al. zhou2014simple,zhou2017event represented event as a quadruple and proposed two Bayesian models to extract events from tweets.
Generative Adversarial Nets
As a neural-based generative model, Generative Adversarial Nets BIBREF3 have been extensively researched in natural language processing (NLP) community.
For text generation, the sequence generative adversarial network (SeqGAN) proposed in BIBREF4 incorporated a policy gradient strategy to optimize the generation process. Based on the policy gradient, Lin et al. lin2017adversarial proposed RankGAN to capture the rich structures of language by ranking and analyzing a collection of human-written and machine-written sentences. To overcome mode collapse when dealing with discrete data, Fedus et al. fedus2018maskgan proposed MaskGAN which used an actor-critic conditional GAN to fill in missing text conditioned on the surrounding context. Along this line, Wang et al. wang2018sentigan proposed SentiGAN to generate texts of different sentiment labels. Besides, Li et al. li2018learning improved the performance of semi-supervised text classification using adversarial training, BIBREF5 , BIBREF6 designed GAN-based models for distance supervision relation extraction.
Although various GAN based approaches have been explored for many applications, none of these approaches tackles open-domain event extraction from online texts. We propose a novel GAN-based event extraction model called AEM. Compared with the previous models, AEM has the following differences: (1) Unlike most GAN-based text generation approaches, a generator network is employed in AEM to learn the projection function between an event distribution and the event-related word distributions (entity, location, keyword, date). The learned generator captures event-related patterns rather than generating text sequence; (2) Different from LEM and DPEMM, AEM uses a generator network to capture the event-related patterns and is able to mine events from different text sources (short and long). Moreover, unlike traditional inference procedure, such as Gibbs sampling used in LEM and DPEMM, AEM could extract the events more efficiently due to the CUDA acceleration; (3) The discriminative features learned by the discriminator of AEM provide a straightforward way to visualize the extracted events.
Methodology
We describe Adversarial-neural Event Model (AEM) in this section. An event is represented as a quadruple < INLINEFORM0 >, where INLINEFORM1 stands for non-location named entities, INLINEFORM2 for a location, INLINEFORM3 for event-related keywords, INLINEFORM4 for a date, and each component in the tuple is represented by component-specific representative words.
AEM is constituted by three components: (1) The document representation module, as shown at the top of Figure FIGREF4 , defines a document representation approach which converts an input document from the online text corpus into INLINEFORM0 which captures the key event elements; (2) The generator INLINEFORM1 , as shown in the lower-left part of Figure FIGREF4 , generates a fake document INLINEFORM2 which is constituted by four multinomial distributions using an event distribution INLINEFORM3 drawn from a Dirichlet distribution as input; (3) The discriminator INLINEFORM4 , as shown in the lower-right part of Figure FIGREF4 , distinguishes the real documents from the fake ones and its output is subsequently employed as a learning signal to update the INLINEFORM5 and INLINEFORM6 . The details of each component are presented below.
Document Representation
Each document INLINEFORM0 in a given corpus INLINEFORM1 is represented as a concatenation of 4 multinomial distributions which are entity distribution ( INLINEFORM2 ), location distribution ( INLINEFORM3 ), keyword distribution ( INLINEFORM4 ) and date distribution ( INLINEFORM5 ) of the document. As four distributions are calculated in a similar way, we only describe the computation of the entity distribution below as an example.
The entity distribution INLINEFORM0 is represented by a normalized INLINEFORM1 -dimensional vector weighted by TF-IDF, and it is calculated as: INLINEFORM2
where INLINEFORM0 is the pseudo corpus constructed by removing all non-entity words from INLINEFORM1 , INLINEFORM2 is the total number of distinct entities in a corpus, INLINEFORM3 denotes the number of INLINEFORM4 -th entity appeared in document INLINEFORM5 , INLINEFORM6 represents the number of documents in the corpus, and INLINEFORM7 is the number of documents that contain INLINEFORM8 -th entity, and the obtained INLINEFORM9 denotes the relevance between INLINEFORM10 -th entity and document INLINEFORM11 .
Similarly, location distribution INLINEFORM0 , keyword distribution INLINEFORM1 and date distribution INLINEFORM2 of INLINEFORM3 could be calculated in the same way, and the dimensions of these distributions are denoted as INLINEFORM4 , INLINEFORM5 and INLINEFORM6 , respectively. Finally, each document INLINEFORM7 in the corpus is represented by a INLINEFORM8 -dimensional ( INLINEFORM9 = INLINEFORM10 + INLINEFORM11 + INLINEFORM12 + INLINEFORM13 ) vector INLINEFORM14 by concatenating four computed distributions.
Network Architecture
The generator network INLINEFORM0 is designed to learn the projection function between the document-event distribution INLINEFORM1 and the four document-level word distributions (entity distribution, location distribution, keyword distribution and date distribution).
More concretely, INLINEFORM0 consists of a INLINEFORM1 -dimensional document-event distribution layer, INLINEFORM2 -dimensional hidden layer and INLINEFORM3 -dimensional event-related word distribution layer. Here, INLINEFORM4 denotes the event number, INLINEFORM5 is the number of units in the hidden layer, INLINEFORM6 is the vocabulary size and equals to INLINEFORM7 + INLINEFORM8 + INLINEFORM9 + INLINEFORM10 . As shown in Figure FIGREF4 , INLINEFORM11 firstly employs a random document-event distribution INLINEFORM12 as an input. To model the multinomial property of the document-event distribution, INLINEFORM13 is drawn from a Dirichlet distribution parameterized with INLINEFORM14 which is formulated as: DISPLAYFORM0
where INLINEFORM0 is the hyper-parameter of the dirichlet distribution, INLINEFORM1 is the number of events which should be set in AEM, INLINEFORM2 , INLINEFORM3 represents the proportion of event INLINEFORM4 in the document and INLINEFORM5 .
Subsequently, INLINEFORM0 transforms INLINEFORM1 into a INLINEFORM2 -dimensional hidden space using a linear layer followed by layer normalization, and the transformation is defined as: DISPLAYFORM0
where INLINEFORM0 represents the weight matrix of hidden layer, and INLINEFORM1 denotes the bias term, INLINEFORM2 is the parameter of LeakyReLU activation and is set to 0.1, INLINEFORM3 and INLINEFORM4 denote the normalized hidden states and the outputs of the hidden layer, and INLINEFORM5 represents the layer normalization.
Then, to project INLINEFORM0 into four document-level event related word distributions ( INLINEFORM1 , INLINEFORM2 , INLINEFORM3 and INLINEFORM4 shown in Figure FIGREF4 ), four subnets (each contains a linear layer, a batch normalization layer and a softmax layer) are employed in INLINEFORM5 . And the exact transformation is based on the formulas below: DISPLAYFORM0
where INLINEFORM0 means softmax layer, INLINEFORM1 , INLINEFORM2 , INLINEFORM3 and INLINEFORM4 denote the weight matrices of the linear layers in subnets, INLINEFORM5 , INLINEFORM6 , INLINEFORM7 and INLINEFORM8 represent the corresponding bias terms, INLINEFORM9 , INLINEFORM10 , INLINEFORM11 and INLINEFORM12 are state vectors. INLINEFORM13 , INLINEFORM14 , INLINEFORM15 and INLINEFORM16 denote the generated entity distribution, location distribution, keyword distribution and date distribution, respectively, that correspond to the given event distribution INLINEFORM17 . And each dimension represents the relevance between corresponding entity/location/keyword/date term and the input event distribution.
Finally, four generated distributions are concatenated to represent the generated document INLINEFORM0 corresponding to the input INLINEFORM1 : DISPLAYFORM0
The discriminator network INLINEFORM0 is designed as a fully-connected network which contains an input layer, a discriminative feature layer (discriminative features are employed for event visualization) and an output layer. In AEM, INLINEFORM1 uses fake document INLINEFORM2 and real document INLINEFORM3 as input and outputs the signal INLINEFORM4 to indicate the source of the input data (lower value denotes that INLINEFORM5 is prone to predict the input data as a fake document and vice versa).
As have previously been discussed in BIBREF7 , BIBREF8 , lipschitz continuity of INLINEFORM0 network is crucial to the training of the GAN-based approaches. To ensure the lipschitz continuity of INLINEFORM1 , we employ the spectral normalization technique BIBREF9 . More concretely, for each linear layer INLINEFORM2 (bias term is omitted for simplicity) in INLINEFORM3 , the weight matrix INLINEFORM4 is normalized by INLINEFORM5 . Here, INLINEFORM6 is the spectral norm of the weight matrix INLINEFORM7 with the definition below: DISPLAYFORM0
which is equivalent to the largest singular value of INLINEFORM0 . The weight matrix INLINEFORM1 is then normalized using: DISPLAYFORM0
Obviously, the normalized weight matrix INLINEFORM0 satisfies that INLINEFORM1 and further ensures the lipschitz continuity of the INLINEFORM2 network BIBREF9 . To reduce the high cost of computing spectral norm INLINEFORM3 using singular value decomposition at each iteration, we follow BIBREF10 and employ the power iteration method to estimate INLINEFORM4 instead. With this substitution, the spectral norm can be estimated with very small additional computational time.
Objective and Training Procedure
The real document INLINEFORM0 and fake document INLINEFORM1 shown in Figure FIGREF4 could be viewed as random samples from two distributions INLINEFORM2 and INLINEFORM3 , and each of them is a joint distribution constituted by four Dirichlet distributions (corresponding to entity distribution, location distribution, keyword distribution and date distribution). The training objective of AEM is to let the distribution INLINEFORM4 (produced by INLINEFORM5 network) to approximate the real data distribution INLINEFORM6 as much as possible.
To compare the different GAN losses, Kurach kurach2018gan takes a sober view of the current state of GAN and suggests that the Jansen-Shannon divergence used in BIBREF3 performs more stable than variant objectives. Besides, Kurach also advocates that the gradient penalty (GP) regularization devised in BIBREF8 will further improve the stability of the model. Thus, the objective function of the proposed AEM is defined as: DISPLAYFORM0
where INLINEFORM0 denotes the discriminator loss, INLINEFORM1 represents the gradient penalty regularization loss, INLINEFORM2 is the gradient penalty coefficient which trade-off the two components of objective, INLINEFORM3 could be obtained by sampling uniformly along a straight line between INLINEFORM4 and INLINEFORM5 , INLINEFORM6 denotes the corresponding distribution.
The training procedure of AEM is presented in Algorithm SECREF15 , where INLINEFORM0 is the event number, INLINEFORM1 denotes the number of discriminator iterations per generator iteration, INLINEFORM2 is the batch size, INLINEFORM3 represents the learning rate, INLINEFORM4 and INLINEFORM5 are hyper-parameters of Adam BIBREF11 , INLINEFORM6 denotes INLINEFORM7 . In this paper, we set INLINEFORM8 , INLINEFORM9 , INLINEFORM10 . Moreover, INLINEFORM11 , INLINEFORM12 and INLINEFORM13 are set as 0.0002, 0.5 and 0.999.
[!h] Training procedure for AEM [1] INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3 , INLINEFORM4 , INLINEFORM5 , INLINEFORM6 the trained INLINEFORM7 and INLINEFORM8 . Initial INLINEFORM9 parameters INLINEFORM10 and INLINEFORM11 parameter INLINEFORM12 INLINEFORM13 has not converged INLINEFORM14 INLINEFORM15 Sample INLINEFORM16 , Sample a random INLINEFORM17 Sample a random number INLINEFORM18 INLINEFORM19 INLINEFORM20 INLINEFORM21 INLINEFORM22 INLINEFORM23 INLINEFORM24 Sample INLINEFORM25 noise INLINEFORM26 INLINEFORM27
Event Generation
After the model training, the generator INLINEFORM0 learns the mapping function between the document-event distribution and the document-level event-related word distributions (entity, location, keyword and date). In other words, with an event distribution INLINEFORM1 as input, INLINEFORM2 could generate the corresponding entity distribution, location distribution, keyword distribution and date distribution.
In AEM, we employ event seed INLINEFORM0 , an INLINEFORM1 -dimensional vector with one-hot encoding, to generate the event related word distributions. For example, in ten event setting, INLINEFORM2 represents the event seed of the first event. With the event seed INLINEFORM3 as input, the corresponding distributions could be generated by INLINEFORM4 based on the equation below: DISPLAYFORM0
where INLINEFORM0 , INLINEFORM1 , INLINEFORM2 and INLINEFORM3 denote the entity distribution, location distribution, keyword distribution and date distribution of the first event respectively.
Experiments
In this section, we firstly describe the datasets and baseline approaches used in our experiments and then present the experimental results.
Experimental Setup
To validate the effectiveness of AEM for extracting events from social media (e.g. Twitter) and news media sites (e.g. Google news), three datasets (FSD BIBREF12 , Twitter, and Google datasets) are employed. Details are summarized below:
FSD dataset (social media) is the first story detection dataset containing 2,499 tweets. We filter out events mentioned in less than 15 tweets since events mentioned in very few tweets are less likely to be significant. The final dataset contains 2,453 tweets annotated with 20 events.
Twitter dataset (social media) is collected from tweets published in the month of December in 2010 using Twitter streaming API. It contains 1,000 tweets annotated with 20 events.
Google dataset (news article) is a subset of GDELT Event Database INLINEFORM0 , documents are retrieved by event related words. For example, documents which contain `malaysia', `airline', `search' and `plane' are retrieved for event MH370. By combining 30 events related documents, the dataset contains 11,909 news articles.
We choose the following three models as the baselines:
K-means is a well known data clustering algorithm, we implement the algorithm using sklearn toolbox, and represent documents using bag-of-words weighted by TF-IDF.
LEM BIBREF13 is a Bayesian modeling approach for open-domain event extraction. It treats an event as a latent variable and models the generation of an event as a joint distribution of its individual event elements. We implement the algorithm with the default configuration.
DPEMM BIBREF14 is a non-parametric mixture model for event extraction. It addresses the limitation of LEM that the number of events should be known beforehand. We implement the model with the default configuration.
For social media text corpus (FSD and Twitter), a named entity tagger specifically built for Twitter is used to extract named entities including locations from tweets. A Twitter Part-of-Speech (POS) tagger BIBREF15 is used for POS tagging and only words tagged with nouns, verbs and adjectives are retained as keywords. For the Google dataset, we use the Stanford Named Entity Recognizer to identify the named entities (organization, location and person). Due to the `date' information not being provided in the Google dataset, we further divide the non-location named entities into two categories (`person' and `organization') and employ a quadruple <organization, location, person, keyword> to denote an event in news articles. We also remove common stopwords and only keep the recognized named entities and the tokens which are verbs, nouns or adjectives.
Experimental Results
To evaluate the performance of the proposed approach, we use the evaluation metrics such as precision, recall and F-measure. Precision is defined as the proportion of the correctly identified events out of the model generated events. Recall is defined as the proportion of correctly identified true events. For calculating the precision of the 4-tuple, we use following criteria:
(1) Do the entity/organization, location, date/person and keyword that we have extracted refer to the same event?
(2) If the extracted representation contains keywords, are they informative enough to tell us what happened?
Table TABREF35 shows the event extraction results on the three datasets. The statistics are obtained with the default parameter setting that INLINEFORM0 is set to 5, number of hidden units INLINEFORM1 is set to 200, and INLINEFORM2 contains three fully-connected layers. The event number INLINEFORM3 for three datasets are set to 25, 25 and 35, respectively. The examples of extracted events are shown in Table. TABREF36 .
It can be observed that K-means performs the worst over all three datasets. On the social media datasets, AEM outpoerforms both LEM and DPEMM by 6.5% and 1.7% respectively in F-measure on the FSD dataset, and 4.4% and 3.7% in F-measure on the Twitter dataset. We can also observe that apart from K-means, all the approaches perform worse on the Twitter dataset compared to FSD, possibly due to the limited size of the Twitter dataset. Moreover, on the Google dataset, the proposed AEM performs significantly better than LEM and DPEMM. It improves upon LEM by 15.5% and upon DPEMM by more than 30% in F-measure. This is because: (1) the assumption made by LEM and DPEMM that all words in a document are generated from a single event is not suitable for long text such as news articles; (2) DPEMM generates too many irrelevant events which leads to a very low precision score. Overall, we see the superior performance of AEM across all datasets, with more significant improvement on the for Google datasets (long text).
We next visualize the detected events based on the discriminative features learned by the trained INLINEFORM0 network in AEM. The t-SNE BIBREF16 visualization results on the datasets are shown in Figure FIGREF19 . For clarity, each subplot is plotted on a subset of the dataset containing ten randomly selected events. It can be observed that documents describing the same event have been grouped into the same cluster.
To further evaluate if a variation of the parameters INLINEFORM0 (the number of discriminator iterations per generator iteration), INLINEFORM1 (the number of units in hidden layer) and the structure of generator INLINEFORM2 will impact the extraction performance, additional experiments have been conducted on the Google dataset, with INLINEFORM3 set to 5, 7 and 10, INLINEFORM4 set to 100, 150 and 200, and three INLINEFORM5 structures (3, 4 and 5 layers). The comparison results on precision, recall and F-measure are shown in Figure FIGREF20 . From the results, it could be observed that AEM with the 5-layer generator performs the best and achieves 96.7% in F-measure, and the worst F-measure obtained by AEM is 85.7%. Overall, the AEM outperforms all compared approaches acorss various parameter settings, showing relatively stable performance.
Finally, we compare in Figure FIGREF37 the training time required for each model, excluding the constant time required by each model to load the data. We could observe that K-means runs fastest among all four approaches. Both LEM and DPEMM need to sample the event allocation for each document and update the relevant counts during Gibbs sampling which are time consuming. AEM only requires a fraction of the training time compared to LEM and DPEMM. Moreover, on a larger dataset such as the Google dataset, AEM appears to be far more efficient compared to LEM and DPEMM.
Conclusions and Future Work
In this paper, we have proposed a novel approach based on adversarial training to extract the structured representation of events from online text. The experimental comparison with the state-of-the-art methods shows that AEM achieves improved extraction performance, especially on long text corpora with an improvement of 15% observed in F-measure. AEM only requires a fraction of training time compared to existing Bayesian graphical modeling approaches. In future work, we will explore incorporating external knowledge (e.g. word relatedness contained in word embeddings) into the learning framework for event extraction. Besides, exploring nonparametric neural event extraction approaches and detecting the evolution of events over time from news articles are other promising future directions.
Acknowledgments
We would like to thank anonymous reviewers for their valuable comments and helpful suggestions. This work was funded by the National Key Research and Development Program of China (2016YFC1306704), the National Natural Science Foundation of China (61772132), the Natural Science Foundation of Jiangsu Province of China (BK20161430).
| [
"Unanswerable",
"Unanswerable"
] | 3,838 | qasper | en | null | db4afd55783aaf6d069c5228152492cf0804e9cf310cb238 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Do they report results only on English data?
Answer: | [
"Unanswerable",
"Unanswerable"
] | qasper | 128 | 70 | |
What is best performing model among author's submissions, what performance it had? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
In the age of information dissemination without quality control, it has enabled malicious users to spread misinformation via social media and aim individual users with propaganda campaigns to achieve political and financial gains as well as advance a specific agenda. Often disinformation is complied in the two major forms: fake news and propaganda, where they differ in the sense that the propaganda is possibly built upon true information (e.g., biased, loaded language, repetition, etc.).
Prior works BIBREF0, BIBREF1, BIBREF2 in detecting propaganda have focused primarily at document level, typically labeling all articles from a propagandistic news outlet as propaganda and thus, often non-propagandistic articles from the outlet are mislabeled. To this end, EMNLP19DaSanMartino focuses on analyzing the use of propaganda and detecting specific propagandistic techniques in news articles at sentence and fragment level, respectively and thus, promotes explainable AI. For instance, the following text is a propaganda of type `slogan'.
Trump tweeted: $\underbrace{\text{`}`{\texttt {BUILD THE WALL!}"}}_{\text{slogan}}$
Shared Task: This work addresses the two tasks in propaganda detection BIBREF3 of different granularities: (1) Sentence-level Classification (SLC), a binary classification that predicts whether a sentence contains at least one propaganda technique, and (2) Fragment-level Classification (FLC), a token-level (multi-label) classification that identifies both the spans and the type of propaganda technique(s).
Contributions: (1) To address SLC, we design an ensemble of different classifiers based on Logistic Regression, CNN and BERT, and leverage transfer learning benefits using the pre-trained embeddings/models from FastText and BERT. We also employed different features such as linguistic (sentiment, readability, emotion, part-of-speech and named entity tags, etc.), layout, topics, etc. (2) To address FLC, we design a multi-task neural sequence tagger based on LSTM-CRF and linguistic features to jointly detect propagandistic fragments and its type. Moreover, we investigate performing FLC and SLC jointly in a multi-granularity network based on LSTM-CRF and BERT. (3) Our system (MIC-CIS) is ranked 3rd (out of 12 participants) and 4th (out of 25 participants) in FLC and SLC tasks, respectively.
System Description ::: Linguistic, Layout and Topical Features
Some of the propaganda techniques BIBREF3 involve word and phrases that express strong emotional implications, exaggeration, minimization, doubt, national feeling, labeling , stereotyping, etc. This inspires us in extracting different features (Table TABREF1) including the complexity of text, sentiment, emotion, lexical (POS, NER, etc.), layout, etc. To further investigate, we use topical features (e.g., document-topic proportion) BIBREF4, BIBREF5, BIBREF6 at sentence and document levels in order to determine irrelevant themes, if introduced to the issue being discussed (e.g., Red Herring).
For word and sentence representations, we use pre-trained vectors from FastText BIBREF7 and BERT BIBREF8.
System Description ::: Sentence-level Propaganda Detection
Figure FIGREF2 (left) describes the three components of our system for SLC task: features, classifiers and ensemble. The arrows from features-to-classifier indicate that we investigate linguistic, layout and topical features in the two binary classifiers: LogisticRegression and CNN. For CNN, we follow the architecture of DBLP:conf/emnlp/Kim14 for sentence-level classification, initializing the word vectors by FastText or BERT. We concatenate features in the last hidden layer before classification.
One of our strong classifiers includes BERT that has achieved state-of-the-art performance on multiple NLP benchmarks. Following DBLP:conf/naacl/DevlinCLT19, we fine-tune BERT for binary classification, initializing with a pre-trained model (i.e., BERT-base, Cased). Additionally, we apply a decision function such that a sentence is tagged as propaganda if prediction probability of the classifier is greater than a threshold ($\tau $). We relax the binary decision boundary to boost recall, similar to pankajgupta:CrossRE2019.
Ensemble of Logistic Regression, CNN and BERT: In the final component, we collect predictions (i.e., propaganda label) for each sentence from the three ($\mathcal {M}=3$) classifiers and thus, obtain $\mathcal {M}$ number of predictions for each sentence. We explore two ensemble strategies (Table TABREF1): majority-voting and relax-voting to boost precision and recall, respectively.
System Description ::: Fragment-level Propaganda Detection
Figure FIGREF2 (right) describes our system for FLC task, where we design sequence taggers BIBREF9, BIBREF10 in three modes: (1) LSTM-CRF BIBREF11 with word embeddings ($w\_e$) and character embeddings $c\_e$, token-level features ($t\_f$) such as polarity, POS, NER, etc. (2) LSTM-CRF+Multi-grain that jointly performs FLC and SLC with FastTextWordEmb and BERTSentEmb, respectively. Here, we add binary sentence classification loss to sequence tagging weighted by a factor of $\alpha $. (3) LSTM-CRF+Multi-task that performs propagandistic span/fragment detection (PFD) and FLC (fragment detection + 19-way classification).
Ensemble of Multi-grain, Multi-task LSTM-CRF with BERT: Here, we build an ensemble by considering propagandistic fragments (and its type) from each of the sequence taggers. In doing so, we first perform majority voting at the fragment level for the fragment where their spans exactly overlap. In case of non-overlapping fragments, we consider all. However, when the spans overlap (though with the same label), we consider the fragment with the largest span.
Experiments and Evaluation
Data: While the SLC task is binary, the FLC consists of 18 propaganda techniques BIBREF3. We split (80-20%) the annotated corpus into 5-folds and 3-folds for SLC and FLC tasks, respectively. The development set of each the folds is represented by dev (internal); however, the un-annotated corpus used in leaderboard comparisons by dev (external). We remove empty and single token sentences after tokenization. Experimental Setup: We use PyTorch framework for the pre-trained BERT model (Bert-base-cased), fine-tuned for SLC task. In the multi-granularity loss, we set $\alpha = 0.1$ for sentence classification based on dev (internal, fold1) scores. We use BIO tagging scheme of NER in FLC task. For CNN, we follow DBLP:conf/emnlp/Kim14 with filter-sizes of [2, 3, 4, 5, 6], 128 filters and 16 batch-size. We compute binary-F1and macro-F1 BIBREF12 in SLC and FLC, respectively on dev (internal).
Experiments and Evaluation ::: Results: Sentence-Level Propaganda
Table TABREF10 shows the scores on dev (internal and external) for SLC task. Observe that the pre-trained embeddings (FastText or BERT) outperform TF-IDF vector representation. In row r2, we apply logistic regression classifier with BERTSentEmb that leads to improved scores over FastTextSentEmb. Subsequently, we augment the sentence vector with additional features that improves F1 on dev (external), however not dev (internal). Next, we initialize CNN by FastTextWordEmb or BERTWordEmb and augment the last hidden layer (before classification) with BERTSentEmb and feature vectors, leading to gains in F1 for both the dev sets. Further, we fine-tune BERT and apply different thresholds in relaxing the decision boundary, where $\tau \ge 0.35$ is found optimal.
We choose the three different models in the ensemble: Logistic Regression, CNN and BERT on fold1 and subsequently an ensemble+ of r3, r6 and r12 from each fold1-5 (i.e., 15 models) to obtain predictions for dev (external). We investigate different ensemble schemes (r17-r19), where we observe that the relax-voting improves recall and therefore, the higher F1 (i.e., 0.673). In postprocess step, we check for repetition propaganda technique by computing cosine similarity between the current sentence and its preceding $w=10$ sentence vectors (i.e., BERTSentEmb) in the document. If the cosine-similarity is greater than $\lambda \in \lbrace .99, .95\rbrace $, then the current sentence is labeled as propaganda due to repetition. Comparing r19 and r21, we observe a gain in recall, however an overall decrease in F1 applying postprocess.
Finally, we use the configuration of r19 on the test set. The ensemble+ of (r4, r7 r12) was analyzed after test submission. Table TABREF9 (SLC) shows that our submission is ranked at 4th position.
Experiments and Evaluation ::: Results: Fragment-Level Propaganda
Table TABREF11 shows the scores on dev (internal and external) for FLC task. Observe that the features (i.e., polarity, POS and NER in row II) when introduced in LSTM-CRF improves F1. We run multi-grained LSTM-CRF without BERTSentEmb (i.e., row III) and with it (i.e., row IV), where the latter improves scores on dev (internal), however not on dev (external). Finally, we perform multi-tasking with another auxiliary task of PFD. Given the scores on dev (internal and external) using different configurations (rows I-V), it is difficult to infer the optimal configuration. Thus, we choose the two best configurations (II and IV) on dev (internal) set and build an ensemble+ of predictions (discussed in section SECREF6), leading to a boost in recall and thus an improved F1 on dev (external).
Finally, we use the ensemble+ of (II and IV) from each of the folds 1-3, i.e., $|{\mathcal {M}}|=6$ models to obtain predictions on test. Table TABREF9 (FLC) shows that our submission is ranked at 3rd position.
Conclusion and Future Work
Our system (Team: MIC-CIS) explores different neural architectures (CNN, BERT and LSTM-CRF) with linguistic, layout and topical features to address the tasks of fine-grained propaganda detection. We have demonstrated gains in performance due to the features, ensemble schemes, multi-tasking and multi-granularity architectures. Compared to the other participating systems, our submissions are ranked 3rd and 4th in FLC and SLC tasks, respectively.
In future, we would like to enrich BERT models with linguistic, layout and topical features during their fine-tuning. Further, we would also be interested in understanding and analyzing the neural network learning, i.e., extracting salient fragments (or key-phrases) in the sentence that generate propaganda, similar to pankajgupta:2018LISA in order to promote explainable AI.
| [
"For SLC task, the \"ltuorp\" team has the best performing model (0.6323/0.6028/0.6649 for F1/P/R respectively) and for FLC task the \"newspeak\" team has the best performing model (0.2488/0.2863/0.2201 for F1/P/R respectively)."
] | 1,541 | qasper | en | null | d25cf05e9fda9bdadb01c26a57122079766a565269fac749 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What is best performing model among author's submissions, what performance it had?
Answer: | [
"For SLC task, the \"ltuorp\" team has the best performing model (0.6323/0.6028/0.6649 for F1/P/R respectively) and for FLC task the \"newspeak\" team has the best performing model (0.2488/0.2863/0.2201 for F1/P/R respectively)."
] | qasper | 128 | 71 | |
what was the baseline? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Neural machine translation (NMT) BIBREF0 , BIBREF1 , BIBREF2 has enabled end-to-end training of a translation system without needing to deal with word alignments, translation rules, and complicated decoding algorithms, which are the characteristics of phrase-based statistical machine translation (PBSMT) BIBREF3 . Although NMT can be significantly better than PBSMT in resource-rich scenarios, PBSMT performs better in low-resource scenarios BIBREF4 . Only by exploiting cross-lingual transfer learning techniques BIBREF5 , BIBREF6 , BIBREF7 , can the NMT performance approach PBSMT performance in low-resource scenarios.
However, such methods usually require an NMT model trained on a resource-rich language pair like French INLINEFORM0 English (parent), which is to be fine-tuned for a low-resource language pair like Uzbek INLINEFORM1 English (child). On the other hand, multilingual approaches BIBREF8 propose to train a single model to translate multiple language pairs. However, these approaches are effective only when the parent target or source language is relatively resource-rich like English (En). Furthermore, the parents and children models should be trained on similar domains; otherwise, one has to take into account an additional problem of domain adaptation BIBREF9 .
In this paper, we work on a linguistically distant and thus challenging language pair Japanese INLINEFORM0 Russian (Ja INLINEFORM1 Ru) which has only 12k lines of news domain parallel corpus and hence is extremely resource-poor. Furthermore, the amount of indirect in-domain parallel corpora, i.e., Ja INLINEFORM2 En and Ru INLINEFORM3 En, are also small. As we demonstrate in Section SECREF4 , this severely limits the performance of prominent low-resource techniques, such as multilingual modeling, back-translation, and pivot-based PBSMT. To remedy this, we propose a novel multistage fine-tuning method for NMT that combines multilingual modeling BIBREF8 and domain adaptation BIBREF9 .
We have addressed two important research questions (RQs) in the context of extremely low-resource machine translation (MT) and our explorations have derived rational contributions (CTs) as follows:
To the best of our knowledge, we are the first to perform such an extensive evaluation of extremely low-resource MT problem and propose a novel multilingual multistage fine-tuning approach involving multilingual modeling and domain adaptation to address it.
Our Japanese–Russian Setting
In this paper, we deal with Ja INLINEFORM0 Ru news translation. This language pair is very challenging because the languages involved have completely different writing system, phonology, morphology, grammar, and syntax. Among various domains, we experimented with translations in the news domain, considering the importance of sharing news between different language speakers. Moreover, news domain is one of the most challenging tasks, due to large presence of out-of-vocabulary (OOV) tokens and long sentences. To establish and evaluate existing methods, we also involved English as the third language. As direct parallel corpora are scarce, involving a language such as English for pivoting is quite common BIBREF10 .
There has been no clean held-out parallel data for Ja INLINEFORM0 Ru and Ja INLINEFORM1 En news translation. Therefore, we manually compiled development and test sets using News Commentary data as a source. Since the given Ja INLINEFORM2 Ru and Ja INLINEFORM3 En data share many lines in the Japanese side, we first compiled tri-text data. Then, from each line, corresponding parts across languages were manually identified, and unaligned parts were split off into a new line. Note that we have never merged two or more lines. As a result, we obtained 1,654 lines of data comprising trilingual, bilingual, and monolingual segments (mainly sentences) as summarized in Table TABREF8 . Finally, for the sake of comparability, we randomly chose 600 trilingual sentences to create a test set, and concatenated the rest of them and bilingual sentences to form development sets.
Our manually aligned development and test sets are publicly available.
Related Work
koehn-knowles:2017:NMT showed that NMT is unable to handle low-resource language pairs as opposed to PBSMT. Transfer learning approaches BIBREF5 , BIBREF6 , BIBREF7 work well when a large helping parallel corpus is available. This restricts one of the source or the target languages to be English which, in our case, is not possible. Approaches involving bi-directional NMT modeling is shown to drastically improve low-resource translation BIBREF11 . However, like most other, this work focuses on translation from and into English.
Remaining options include (a) unsupervised MT BIBREF12 , BIBREF13 , BIBREF14 , (b) parallel sentence mining from non-parallel or comparable corpora BIBREF15 , BIBREF16 , (c) generating pseudo-parallel data BIBREF17 , and (d) MT based on pivot languages BIBREF10 . The linguistic distance between Japanese and Russian makes it extremely difficult to learn bilingual knowledge, such as bilingual lexicons and bilingual word embeddings. Unsupervised MT is thus not promising yet, due to its heavy reliance on accurate bilingual word embeddings. Neither does parallel sentence mining, due to the difficulty of obtaining accurate bilingual lexicons. Pseudo-parallel data can be used to augment existing parallel corpora for training, and previous work has reported that such data generated by so-called back-translation can substantially improve the quality of NMT. However, this approach requires base MT systems that can generate somewhat accurate translations. It is thus infeasible in our scenario, because we can obtain only a weak system which is the consequence of an extremely low-resource situation. MT based on pivot languages requires large in-domain parallel corpora involving the pivot languages. This technique is thus infeasible, because the in-domain parallel corpora for Ja INLINEFORM0 En and Ru INLINEFORM1 En pairs are also extremely limited, whereas there are large parallel corpora in other domains. Section SECREF4 empirically confirms the limit of these existing approaches.
Fortunately, there are two useful transfer learning solutions using NMT: (e) multilingual modeling to incorporate multiple language pairs into a single model BIBREF8 and (f) domain adaptation to incorporate out-of-domain data BIBREF9 . In this paper, we explore a novel method involving step-wise fine-tuning to combine these two methods. By improving the translation quality in this way, we can also increase the likelihood of pseudo-parallel data being useful to further improve translation quality.
Limit of Using only In-domain Data
This section answers our first research question, [RQ1], about the translation quality that we can achieve using existing methods and in-domain parallel and monolingual data. We then use the strongest model to conduct experiments on generating and utilizing back-translated pseudo-parallel data for augmenting NMT. Our intention is to empirically identify the most effective practices as well as recognize the limitations of relying only on in-domain parallel corpora.
Data
To train MT systems among the three languages, i.e., Japanese, Russian, and English, we used all the parallel data provided by Global Voices, more specifically those available at OPUS. Table TABREF9 summarizes the size of train/development/test splits used in our experiments. The number of parallel sentences for Ja INLINEFORM0 Ru is 12k, for Ja INLINEFORM1 En is 47k, and for Ru INLINEFORM2 En is 82k. Note that the three corpora are not mutually exclusive: 9k out of 12k sentences in the Ja INLINEFORM3 Ru corpus were also included in the other two parallel corpora, associated with identical English translations. This puts a limit on the positive impact that the helping corpora can have on the translation quality.
Even when one focuses on low-resource language pairs, we often have access to larger quantities of in-domain monolingual data of each language. Such monolingual data are useful to improve quality of MT, for example, as the source of pseudo-parallel data for augmenting training data for NMT BIBREF17 and as the training data for large and smoothed language models for PBSMT BIBREF4 . Table TABREF13 summarizes the statistics on our monolingual corpora for several domains including the news domain. Note that we removed from the Global Voices monolingual corpora those sentences that are already present in the parallel corpus.
https://dumps.wikimedia.org/backup-index.html (20180501) http://www.statmt.org/wmt18/translation-task.html https://www.yomiuri.co.jp/database/glossary/ http://www.cs.jhu.edu/~kevinduh/a/multitarget-tedtalks/ http://opus.nlpl.eu/Tatoeba-v2.php
We tokenized English and Russian sentences using tokenizer.perl of Moses BIBREF3 . To tokenize Japanese sentences, we used MeCab with the IPA dictionary. After tokenization, we eliminated duplicated sentence pairs and sentences with more than 100 tokens for all the languages.
MT Methods Examined
We began with evaluating standard MT paradigms, i.e., PBSMT BIBREF3 and NMT BIBREF1 . As for PBSMT, we also examined two advanced methods: pivot-based translation relying on a helping language BIBREF10 and induction of phrase tables from monolingual data BIBREF14 .
As for NMT, we compared two types of encoder-decoder architectures: attentional RNN-based model (RNMT) BIBREF2 and the Transformer model BIBREF18 . In addition to standard uni-directional modeling, to cope with the low-resource problem, we examined two multi-directional models: bi-directional model BIBREF11 and multi-to-multi (M2M) model BIBREF8 .
After identifying the best model, we also examined the usefulness of a data augmentation method based on back-translation BIBREF17 .
First, we built a PBSMT system for each of the six translation directions. We obtained phrase tables from parallel corpus using SyMGIZA++ with the grow-diag-final heuristics for word alignment, and Moses for phrase pair extraction. Then, we trained a bi-directional MSD (monotone, swap, and discontinuous) lexicalized reordering model. We also trained three 5-gram language models, using KenLM on the following monolingual data: (1) the target side of the parallel data, (2) the concatenation of (1) and the monolingual data from Global Voices, and (3) the concatenation of (1) and all monolingual data in the news domain in Table TABREF13 .
Subsequently, using English as the pivot language, we examined the following three types of pivot-based PBSMT systems BIBREF10 , BIBREF19 for each of Ja INLINEFORM0 Ru and Ru INLINEFORM1 Ja.
2-step decoding using the source-to-English and English-to-target systems.
Obtain a new phrase table from synthetic parallel data generated by translating English side of the target–English training parallel data to the source language with the English-to-source system.
Compile a new phrase table combining those for the source-to-English and English-to-target systems.
Among these three, triangulation is the most computationally expensive method. Although we had filtered the component phrase tables using the statistical significance pruning method BIBREF20 , triangulation can generate an enormous number of phrase pairs. To reduce the computational cost during decoding and the negative effects of potentially noisy phrase pairs, we retained for each source phrase INLINEFORM0 only the INLINEFORM1 -best translations INLINEFORM2 according to the forward translation probability INLINEFORM3 calculated from the conditional probabilities in the component models as defined in utiyama:07. For each of the retained phrase pairs, we also calculated the backward translation probability, INLINEFORM4 , and lexical translation probabilities, INLINEFORM5 and INLINEFORM6 , in the same manner as INLINEFORM7 .
We also investigated the utility of recent advances in unsupervised MT. Even though we began with a publicly available implementation of unsupervised PBSMT BIBREF13 , it crashed due to unknown reasons. We therefore followed another method described in marie:usmt-unmt. Instead of short INLINEFORM0 -grams BIBREF12 , BIBREF13 , we collected a set of phrases in Japanese and Russian from respective monolingual data using the word2phrase algorithm BIBREF21 , as in marie:usmt-unmt. To reduce the complexity, we used randomly selected 10M monolingual sentences, and 300k most frequent phrases made of words among the 300k most frequent words. For each source phrase INLINEFORM1 , we selected 300-best target phrases INLINEFORM2 according to the translation probability as in D18-1549: INLINEFORM3 where INLINEFORM4 stands for a bilingual embedding of a given phrase, obtained through averaging bilingual embeddings of constituent words learned from the two monolingual data using fastText and vecmap. For each of the retained phrase pair, INLINEFORM5 was computed analogously. We also computed lexical translation probabilities relying on those learned from the given small parallel corpus.
Up to four phrase tables were jointly exploited by the multiple decoding path ability of Moses. Weights for the features were tuned using KB-MIRA BIBREF22 on the development set; we took the best weights after 15 iterations. Two hyper-parameters, namely, INLINEFORM0 for the number of pivot-based phrase pairs per source phrase and INLINEFORM1 for distortion limit, were determined by a grid search on INLINEFORM2 and INLINEFORM3 . In contrast, we used predetermined hyper-parameters for phrase table induction from monolingual data, following the convention: 200 for the dimension of word and phrase embeddings and INLINEFORM4 .
We used the open-source implementation of the RNMT and the Transformer models in tensor2tensor. A uni-directional model for each of the six translation directions was trained on the corresponding parallel corpus. Bi-directional and M2M models were realized by adding an artificial token that specifies the target language to the beginning of each source sentence and shuffling the entire training data BIBREF8 .
Table TABREF22 contains some specific hyper-parameters for our baseline NMT models. The hyper-parameters not mentioned in this table used the default values in tensor2tensor. For M2M systems, we over-sampled Ja INLINEFORM0 Ru and Ja INLINEFORM1 En training data so that their sizes match the largest Ru INLINEFORM2 En data. To reduce the number of unknown words, we used tensor2tensor's internal sub-word segmentation mechanism. Since we work in a low-resource setting, we used shared sub-word vocabularies of size 16k for the uni- and bi-directional models and 32k for the M2M models. The number of training iterations was determined by early-stopping: we evaluated our models on the development set every 1,000 updates, and stopped training if BLEU score for the development set was not improved for 10,000 updates (10 check-points). Note that the development set was created by concatenating those for the individual translation directions without any over-sampling.
Having trained the models, we averaged the last 10 check-points and decoded the test sets with a beam size of 4 and a length penalty which was tuned by a linear search on the BLEU score for the development set.
Similarly to PBSMT, we also evaluated “Cascade” and “Synthesize” methods with uni-directional NMT models.
Results
We evaluated MT models using case-sensitive and tokenized BLEU BIBREF23 on test sets, using Moses's multi-bleu.perl. Statistical significance ( INLINEFORM0 ) on the difference of BLEU scores was tested by Moses's bootstrap-hypothesis-difference-significance.pl.
Tables TABREF27 and TABREF31 show BLEU scores of all the models, except the NMT systems augmented with back-translations. Whereas some models achieved reasonable BLEU scores for Ja INLINEFORM0 En and Ru INLINEFORM1 En translation, all the results for Ja INLINEFORM2 Ru, which is our main concern, were abysmal.
Among the NMT models, Transformer models (b INLINEFORM0 ) were proven to be better than RNMT models (a INLINEFORM1 ). RNMT models could not even outperform the uni-directional PBSMT models (c1). M2M models (a3) and (b3) outperformed their corresponding uni- and bi-directional models in most cases. It is worth noting that in this extremely low-resource scenario, BLEU scores of the M2M RNMT model for the largest language pair, i.e., Ru INLINEFORM2 En, were lower than those of the uni- and bi-directional RNMT models as in TACL1081. In contrast, with the M2M Transformer model, Ru INLINEFORM3 En also benefited from multilingualism.
Standard PBSMT models (c1) achieved higher BLEU scores than uni-directional NMT models (a1) and (b1), as reported by koehn-knowles:2017:NMT, whereas they underperform the M2M Transformer NMT model (b3). As shown in Table TABREF31 , pivot-based PBSMT systems always achieved higher BLEU scores than (c1). The best model with three phrase tables, labeled “Synthesize / Triangulate / Gold,” brought visible BLEU gains with substantial reduction of OOV tokens (3047 INLINEFORM0 1180 for Ja INLINEFORM1 Ru, 4463 INLINEFORM2 1812 for Ru INLINEFORM3 Ja). However, further extension with phrase tables induced from monolingual data did not push the limit, despite their high coverage; only 336 and 677 OOV tokens were left for the two translation directions, respectively. This is due to the poor quality of the bilingual word embeddings used to extract the phrase table, as envisaged in Section SECREF3 .
None of pivot-based approaches with uni-directional NMT models could even remotely rival the M2M Transformer NMT model (b3).
Table TABREF46 shows the results of our multistage fine-tuning, where the IDs of each row refer to those described in Section SECREF41 . First of all, the final models of our multistage fine-tuning, i.e., V and VII, achieved significantly higher BLEU scores than (b3) in Table TABREF27 , a weak baseline without using any monolingual data, and #10 in Table TABREF33 , a strong baseline established with monolingual data.
The performance of the initial model (I) depends on the language pair. For Ja INLINEFORM0 Ru pair, it cannot achieve minimum level of quality since the model has never seen parallel data for this pair. The performance on Ja INLINEFORM1 En pair was much lower than the two baseline models, reflecting the crucial mismatch between training and testing domains. In contrast, Ru INLINEFORM2 En pair benefited the most and achieved surprisingly high BLEU scores. The reason might be due to the proximity of out-of-domain training data and in-domain test data.
The first fine-tuning stage significantly pushed up the translation quality for Ja INLINEFORM0 En and Ru INLINEFORM1 En pairs, in both cases with fine-tuning (II) and mixed fine-tuning (III). At this stage, both models performed only poorly for Ja INLINEFORM2 Ru pair as they have not yet seen Ja INLINEFORM3 Ru parallel data. Either model had a consistent advantage to the other.
When these models were further fine-tuned only on the in-domain Ja INLINEFORM0 Ru parallel data (IV and VI), we obtained translations of better quality than the two baselines for Ja INLINEFORM1 Ru pair. However, as a result of complete ignorance of Ja INLINEFORM2 En and Ru INLINEFORM3 En pairs, the models only produced translations of poor quality for these language pairs. In contrast, mixed fine-tuning for the second fine-tuning stage (V and VII) resulted in consistently better models than conventional fine-tuning (IV and VI), irrespective of the choice at the first stage, thanks to the gradual shift of parameters realized by in-domain Ja INLINEFORM4 En and Ru INLINEFORM5 En parallel data. Unfortunately, the translation quality for Ja INLINEFORM6 En and Ru INLINEFORM7 En pairs sometimes degraded from II and III. Nevertheless, the BLEU scores still retain the large margin against two baselines.
The last three rows in Table TABREF46 present BLEU scores obtained by the methods with fewer fine-tuning steps. The most naive model I', trained on the balanced mixture of whole five types of corpora from scratch, and the model II', obtained through a single-step conventional fine-tuning of I on all the in-domain data, achieved only BLEU scores consistently worse than VII. In contrast, when we merged our two fine-tuning steps into a single mixed fine-tuning on I, we obtained a model III' which is better for the Ja INLINEFORM0 Ru pair than VII. Nevertheless, they are still comparable to those of VII and the BLEU scores for the other two language pairs are much lower than VII. As such, we conclude that our multistage fine-tuning leads to a more robust in-domain multilingual model.
Augmentation with Back-translation
Given that the M2M Transformer NMT model (b3) achieved best results for most of the translation directions and competitive results for the rest, we further explored it through back-translation.
We examined the utility of pseudo-parallel data for all the six translation directions, unlike the work of lakew2017improving and lakew2018comparison, which concentrate only on the zero-shot language pair, and the work of W18-2710, which compares only uni- or bi-directional models. We investigated whether each translation direction in M2M models will benefit from pseudo-parallel data and if so, what kind of improvement takes place.
First, we selected sentences to be back-translated from in-domain monolingual data (Table TABREF13 ), relying on the score proposed by moore:intelligent via the following procedure.
For each language, train two 4-gram language models, using KenLM: an in-domain one on all the Global Voices data, i.e., both parallel and monolingual data, and a general-domain one on the concatenation of Global Voices, IWSLT, and Tatoeba data.
For each language, discard sentences containing OOVs according to the in-domain language model.
For each translation direction, select the INLINEFORM0 -best monolingual sentences in the news domain, according to the difference between cross-entropy scores given by the in-domain and general-domain language models.
Whereas W18-2710 exploited monolingual data much larger than parallel data, we maintained a 1:1 ratio between them BIBREF8 , setting INLINEFORM0 to the number of lines of parallel data of given language pair.
Selected monolingual sentences were then translated using the M2M Transformer NMT model (b3) to compose pseudo-parallel data. Then, the pseudo-parallel data were enlarged by over-sampling as summarized in Table TABREF32 . Finally, new NMT models were trained on the concatenation of the original parallel and pseudo-parallel data from scratch in the same manner as the previous NMT models with the same hyper-parameters.
Table TABREF33 shows the BLEU scores achieved by several reasonable combinations of six-way pseudo-parallel data. We observed that the use of all six-way pseudo-parallel data (#10) significantly improved the base model for all the translation directions, except En INLINEFORM0 Ru. A translation direction often benefited when the pseudo-parallel data for that specific direction was used.
Summary
We have evaluated an extensive variation of MT models that rely only on in-domain parallel and monolingual data. However, the resulting BLEU scores for Ja INLINEFORM2 Ru and Ru INLINEFORM3 Ja tasks do not exceed 10 BLEU points, implying the inherent limitation of the in-domain data as well as the difficulty of these translation directions.
Exploiting Large Out-of-Domain Data Involving a Helping Language
The limitation of relying only on in-domain data demonstrated in Section SECREF4 motivates us to explore other types of parallel data. As raised in our second research question, [RQ2], we considered the effective ways to exploit out-of-domain data.
According to language pair and domain, parallel data can be classified into four categories in Table TABREF40 . Among all the categories, out-of-domain data for the language pair of interest have been exploited in the domain adaptation scenarios (C INLINEFORM0 A) BIBREF9 . However, for Ja INLINEFORM1 Ru, no out-of-domain data is available. To exploit out-of-domain parallel data for Ja INLINEFORM2 En and Ru INLINEFORM3 En pairs instead, we propose a multistage fine-tuning method, which combines two types of transfer learning, i.e., domain adaptation for Ja INLINEFORM4 En and Ru INLINEFORM5 En (D INLINEFORM6 B) and multilingual transfer (B INLINEFORM7 A), relying on the M2M model examined in Section SECREF4 . We also examined the utility of fine-tuning for iteratively generating and using pseudo-parallel data.
Multistage Fine-tuning
Simply using NMT systems trained on out-of-domain data for in-domain translation is known to perform badly. In order to effectively use large-scale out-of-domain data for our extremely low-resource task, we propose to perform domain adaptation through either (a) conventional fine-tuning, where an NMT system trained on out-of-domain data is fine-tuned only on in-domain data, or (b) mixed fine-tuning BIBREF9 , where pre-trained out-of-domain NMT system is fine-tuned using a mixture of in-domain and out-of-domain data. The same options are available for transferring from Ja INLINEFORM0 En and Ru INLINEFORM1 En to Ja INLINEFORM2 Ru.
We inevitably involve two types of transfer learning, i.e., domain adaptation for Ja INLINEFORM0 En and Ru INLINEFORM1 En and multilingual transfer for Ja INLINEFORM2 Ru pair. Among several conceivable options for managing these two problems, we examined the following multistage fine-tuning.
Pre-train a multilingual model only on the Ja INLINEFORM0 En and Ru INLINEFORM1 En out-of-domain parallel data (I), where the vocabulary of the model is determined on the basis of the in-domain parallel data in the same manner as the M2M NMT models examined in Section SECREF4 .
Fine-tune the pre-trained model (I) on the in-domain Ja INLINEFORM0 En and Ru INLINEFORM1 En parallel data (fine-tuning, II) or on the mixture of in-domain and out-of-domain Ja INLINEFORM2 En and Ru INLINEFORM3 En parallel data (mixed fine-tuning, III).
Further fine-tune the models (each of II and III) for Ja INLINEFORM0 Ru on in-domain parallel data for this language pair only (fine-tuning, IV and VI) or on all the in-domain parallel data (mixed fine-tuning, V and VII).
We chose this way due to the following two reasons. First, we need to take a balance between several different parallel corpora sizes. The other reason is division of labor; we assume that solving each sub-problem one by one should enable gradual shift of parameters.
Data Selection
As an additional large-scale out-of-domain parallel data for Ja INLINEFORM0 En, we used the cleanest 1.5M sentences from the Asian Scientific Paper Excerpt Corpus (ASPEC) BIBREF24 . As for Ru INLINEFORM1 En, we used the UN and Yandex corpora released for the WMT 2018 News Translation Task. We retained Ru INLINEFORM2 En sentence pairs that contain at least one OOV token in both sides, according to the in-domain language model trained in Section SECREF34 . Table TABREF45 summarizes the statistics on the remaining out-of-domain parallel data.
Further Augmentation with Back-translation
Having obtained a better model, we examined again the utility of back-translation. More precisely, we investigated (a) whether the pseudo-parallel data generated by an improved NMT model leads to a further improvement, and (b) whether one more stage of fine-tuning on the mixture of original parallel and pseudo-parallel data will result in a model better than training a new model from scratch as examined in Section SECREF34 .
Given an NMT model, we first generated six-way pseudo-parallel data by translating monolingual data. For the sake of comparability, we used the identical monolingual sentences sampled in Section SECREF34 . Then, we further fine-tuned the given model on the mixture of the generated pseudo-parallel data and the original parallel data, following the same over-sampling procedure in Section SECREF34 . We repeated these steps five times.
Table TABREF51 shows the results. “new #10” in the second row indicates an M2M Transformer model trained from scratch on the mixture of six-way pseudo-parallel data generated by VII and the original parallel data. It achieved higher BLEU scores than #10 in Table TABREF33 thanks to the pseudo-parallel data of better quality, but underperformed the base NMT model VII. In contrast, our fine-tuned model VIII successfully surpassed VII, and one more iteration (IX) further improved BLEU scores for all translation directions, except Ru INLINEFORM0 En. Although further iterations did not necessarily gain BLEU scores, we came to a much higher plateau compared to the results in Section SECREF4 .
Conclusion
In this paper, we challenged the difficult task of Ja INLINEFORM0 Ru news domain translation in an extremely low-resource setting. We empirically confirmed the limited success of well-established solutions when restricted to in-domain data. Then, to incorporate out-of-domain data, we proposed a multilingual multistage fine-tuning approach and observed that it substantially improves Ja INLINEFORM1 Ru translation by over 3.7 BLEU points compared to a strong baseline, as summarized in Table TABREF53 . This paper contains an empirical comparison of several existing approaches and hence we hope that our paper can act as a guideline to researchers attempting to tackle extremely low-resource translation.
In the future, we plan to confirm further fine-tuning for each of specific translation directions. We will also explore the way to exploit out-of-domain pseudo-parallel data, better domain-adaptation approaches, and additional challenging language pairs.
Acknowledgments
This work was carried out when Aizhan Imankulova was taking up an internship at NICT, Japan. We would like to thank the reviewers for their insightful comments. A part of this work was conducted under the program “Promotion of Global Communications Plan: Research, Development, and Social Demonstration of Multilingual Speech Translation Technology” of the Ministry of Internal Affairs and Communications (MIC), Japan.
| [
"pivot-based translation relying on a helping language BIBREF10, nduction of phrase tables from monolingual data BIBREF14 , attentional RNN-based model (RNMT) BIBREF2, Transformer model BIBREF18, bi-directional model BIBREF11, multi-to-multi (M2M) model BIBREF8, back-translation BIBREF17",
"M2M Transformer"
] | 4,542 | qasper | en | null | f55efaaee938369fc4be8a430e531c164168be643b3cfe57 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: what was the baseline?
Answer: | [
"pivot-based translation relying on a helping language BIBREF10, nduction of phrase tables from monolingual data BIBREF14 , attentional RNN-based model (RNMT) BIBREF2, Transformer model BIBREF18, bi-directional model BIBREF11, multi-to-multi (M2M) model BIBREF8, back-translation BIBREF17",
"M2M Transformer"
] | qasper | 128 | 72 | |
What was their highest recall score? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
BioASQ is a biomedical document classification, document retrieval, and question answering competition, currently in its seventh year. We provide an overview of our submissions to semantic question answering task (7b, Phase B) of BioASQ 7 (except for 'ideal answer' test, in which we did not participate this year). In this task systems are provided with biomedical questions and are required to submit ideal and exact answers to those questions. We have used BioBERT BIBREF0 based system , see also Bidirectional Encoder Representations from Transformers(BERT) BIBREF1, and we fine tuned it for the biomedical question answering task. Our system scored near the top for factoid questions for all the batches of the challenge. More specifially, in the third test batch set, our system achieved highest ‘MRR’ score for Factoid Question Answering task. Also, for List-type question answering task our system achieved highest recall score in the fourth test batch set. Along with our detailed approach, we present the results for our submissions and also highlight identified downsides for our current approach and ways to improve them in our future experiments. In last test batch results we placed 4th for List-type questions and 3rd for Factoid-type questions.)
The QA task is organized in two phases. Phase A deals with retrieval of the relevant document, snippets, concepts, and RDF triples, and phase B deals with exact and ideal answer generations (which is a paragraph size summary of snippets). Exact answer generation is required for factoid, list, and yes/no type question.
BioASQ organizers provide the training and testing data. The training data consists of questions, gold standard documents, snippets, concepts, and ideal answers (which we did not use in this paper, but we used last year BIBREF2). The test data is split between phases A and B. The Phase A dataset consists of the questions, unique ids, question types. The Phase B dataset consists of the questions, golden standard documents, snippets, unique ids and question types. Exact answers for factoid type questions are evaluated using strict accuracy (the top answer), lenient accuracy (the top 5 answers), and MRR (Mean Reciprocal Rank) which takes into account the ranks of returned answers. Answers for the list type question are evaluated using precision, recall, and F-measure.
Related Work ::: BioAsq
Sharma et al. BIBREF3 describe a system with two stage process for factoid and list type question answering. Their system extracts relevant entities and then runs supervised classifier to rank the entities. Wiese et al. BIBREF4 propose neural network based model for Factoid and List-type question answering task. The model is based on Fast QA and predicts the answer span in the passage for a given question. The model is trained on SQuAD data set and fine tuned on the BioASQ data. Dimitriadis et al. BIBREF5 proposed two stage process for Factoid question answering task. Their system uses general purpose tools such as Metamap, BeCas to identify candidate sentences. These candidate sentences are represented in the form of features, and are then ranked by the binary classifier. Classifier is trained on candidate sentences extracted from relevant questions, snippets and correct answers from BioASQ challenge. For factoid question answering task highest ‘MRR’ achieved in the 6th edition of BioASQ competition is ‘0.4325’. Our system is a neural network model based on contextual word embeddings BIBREF1 and achieved a ‘MRR’ score ‘0.6103’ in one of the test batches for Factoid Question Answering task.
Related Work ::: A minimum background on BERT
BERT stands for "Bidirectional Encoder Representations from Transformers" BIBREF1 is a contextual word embedding model. Given a sentence as an input, contextual embedding for the words are returned. The BERT model was designed so it can be fine tuned for 11 different tasks BIBREF1, including question answering tasks. For a question answering task, question and paragraph (context) are given as an input. A BERT standard is that question text and paragraph text are separated by a separator [Sep]. BERT question-answering fine tuning involves adding softmax layer. Softmax layer takes contextual word embeddings from BERT as input and learns to identity answer span present in the paragraph (context). This process is represented in Figure FIGREF4.
BERT was originally trained to perform tasks such as language model creation using masked words and next-sentence-prediction. In other words BERT weights are learned such that context is used in building the representation of the word, not just as a loss function to help learn a context-independent representation. For detailed understanding of BERT Architecture, please refer to the original BERT paper BIBREF1.
Related Work ::: A minimum background on BERT ::: Comparison of Word Embeddings and Contextual Word Embeddings
A ‘word embedding’ is a learned representation. It is represented in the form of vector where words that have the same meaning have a similar vector representation. Consider a word embedding model 'word2vec' BIBREF6 trained on a corpus. Word embeddings generated from the model are context independent that is, word embeddings are returned regardless of where the words appear in a sentence and regardless of e.g. the sentiment of the sentence. However, contextual word embedding models like BERT also takes context of the word into consideration.
Related Work ::: Comparison of BERT and Bio-BERT
‘BERT’ and BioBERT are very similar in terms of architecture. Difference is that ‘BERT’ is pretrained on Wikipedia articles, whereas BioBERT version used in our experiments is pretrained on Wikipedia, PMC and PubMed articles. Therefore BioBERT model is expected to perform well with biomedical text, in terms of generating contextual word embeddings.
BioBERT model used in our experiments is based on BERT-Base Architecture; BERT-Base has 12 transformer Layers where as BERT-Large has 24 transformer layers. Moreover contextual word embedding vector size is 768 for BERT-Base and more for BERT-large. According to BIBREF1 Bert-Large, fine-tuned on SQuAD 1.1 question answering data BIBREF7 can achieve F1 Score of 90.9 for Question Answering task where as BERT-Base Fine-tuned on the same SQuAD question answering BIBREF7 data could achieve F1 score of 88.5. One downside of the current version BioBERT is that word-piece vocabulary is the same as that of original BERT Model, as a result word-piece vocabulary does not include biomedical jargon. Lee et al. BIBREF0 created BioBERT, using the same pre-trained BERT released by Google, and hence in the word-piece vocabulary (vocab.txt), as a result biomedical jargon is not included in word-piece vocabulary. Modifying word-piece vocabulary (vocab.txt) at this stage would loose original compatibility with ‘BERT’, hence it is left unmodified.
In our future work we would like to build pre-trained ‘BERT’ model from scratch. We would pretrain the model with biomedical corpus (PubMed, ‘PMC’) and Wikipedia. Doing so would give us scope to create word piece vocabulary to include biomedical jargon and there are chances of model performing better with biomedical jargon being included in the word piece vocabulary. We will consider this scenario in the future, or wait for the next version of BioBERT.
Experiments: Factoid Question Answering Task
For Factoid Question Answering task, we fine tuned BioBERT BIBREF0 with question answering data and added new features. Fig. FIGREF4 shows the architecture of BioBERT fine tuned for question answering tasks: Input to BioBERT is word tokenized embeddings for question and the paragraph (Context). As per the ‘BERT’ BIBREF1 standards, tokens ‘[CLS]’ and ‘[SEP]’ are appended to the tokenized input as illustrated in the figure. The resulting model has a softmax layer formed for predicting answer span indices in the given paragraph (Context). On test data, the fine tuned model generates $n$-best predictions for each question. For a question, $n$-best corresponds that $n$ answers are returned as possible answers in the decreasing order of confidence. Variable $n$ is configurable. In our paper, any further mentions of ‘answer returned by the model’ correspond to the top answer returned by the model.
Experiments: Factoid Question Answering Task ::: Setup
BioASQ provides the training data. This data is based on previous BioASQ competitions. Train data we have considered is aggregate of all train data sets till the 5th version of BioASQ competition. We cleaned the data, that is, question-answering data without answers are removed and left with a total count of ‘530’ question answers. The data is split into train and test data in the ratio of 94 to 6; that is, count of '495' for training and '35' for testing.
The original data format is converted to the BERT/BioBERT format, where BioBERT expects ‘start_index’ of the actual answer. The ‘start_index corresponds to the index of the answer text present in the paragraph/ Context. For finding ‘start_index’ we used built-in python function find(). The function returns the lowest index of the actual answer present in the context(paragraph). If the answer is not found ‘-1’ is returned as the index. The efficient way of finding start_index is, if the paragraph (Context) has multiple instances of answer text, then ‘start_index’ of the answer should be that instance of answer text whose context actually matches with what’s been asked in the question.
Example (Question, Answer and Paragraph from BIBREF8):
Question: Which drug should be used as an antidote in benzodiazepine overdose?
Answer: 'Flumazenil'
Paragraph(context):
"Flumazenil use in benzodiazepine overdose in the UK: a retrospective survey of NPIS data. OBJECTIVE: Benzodiazepine (BZD) overdose (OD) continues to cause significant morbidity and mortality in the UK. Flumazenil is an effective antidote but there is a risk of seizures, particularly in those who have co-ingested tricyclic antidepressants. A study was undertaken to examine the frequency of use, safety and efficacy of flumazenil in the management of BZD OD in the UK. METHODS: A 2-year retrospective cohort study was performed of all enquiries to the UK National Poisons Information Service involving BZD OD. RESULTS: Flumazenil was administered to 80 patients in 4504 BZD-related enquiries, 68 of whom did not have ventilatory failure or had recognised contraindications to flumazenil. Factors associated with flumazenil use were increased age, severe poisoning and ventilatory failure. Co-ingestion of tricyclic antidepressants and chronic obstructive pulmonary disease did not influence flumazenil administration. Seizure frequency in patients not treated with flumazenil was 0.3%".
Actual answer is 'Flumazenil', but there are multiple instances of word 'Flu-mazenil'. Efficient way to identify the start-index for 'Flumazenil'(answer) is to find that particular instance of the word 'Flumazenil' which matches the context for the question. In the above example 'Flumazenil' highlighted in bold is the actual instance that matches question's context. Unfortunately, we could not identify readily available tools that can achieve this goal. In our future work, we look forward to handling these scenarios effectively.
Note: The creators of 'SQuAD' BIBREF7 have handled the task of identifying answer's start_index effectively. But 'SQuAD' data set is much more general and does not include biomedical question answering data.
Experiments: Factoid Question Answering Task ::: Training and error analysis
During our training with the BioASQ data, learning rate is set to 3e-5, as mentioned in the BioBERT paper BIBREF0. We started training the model with 495 available train data and 35 test data by setting number of epochs to 50. After training with these hyper-parameters training accuracy(exact match) was 99.3%(overfitting) and testing accuracy is only 4%. In the next iteration we reduced the number of epochs to 25 then training accuracy is reduced to 98.5% and test accuracy moved to 5%. We further reduced number of epochs to 15, and the resulting training accuracy was 70% and test accuracy 15%. In the next iteration set number of epochs to 12 and achieved train accuracy of 57.7% and test accuracy 23.3%. Repeated the experiment with 11 epochs and found training accuracy to be 57.7% and test accuracy to be same 22%. In the next iteration we set number of epochs to '9' and found training accuracy of 48% and test accuracy of 15%. Hence optimum number of epochs is taken as 12 epochs.
During our error analysis we found that on test data, model tends to return text in the beginning of the context(paragraph) as the answer. On analysing train data, we found that there are '120'(out of '495') question answering data instances having start_index:0, meaning 120( 25%) question answering data has first word(s) in the context(paragraph) as the answer. We removed 70% of those instances in order to make train data more balanced. In the new train data set we are left with '411' question answering data instances. This time we got the highest test accuracy of 26% at 11 epochs. We have submitted our results for BioASQ test batch-2, got strict accuracy of 32% and our system stood in 2nd place. Initially, hyper-parameter- 'batch size' is set to ‘400’. Later it is tuned to '32'. Although accuracy(exact answer match) remained at 26%, model generated concise and better answers at batch size ‘32’, that is wrong answers are close to the expected answer in good number of cases.
Example.(from BIBREF8)
Question: Which mutated gene causes Chediak Higashi Syndrome?
Exact Answer: ‘lysosomal trafficking regulator gene’.
The answer returned by a model trained at ‘400’ batch size is ‘Autosomal-recessive complicated spastic paraplegia with a novel lysosomal trafficking regulator’, and from the one trained at ‘32’ batch size is ‘lysosomal trafficking regulator’.
In further experiments, we have fine tuned the BioBERT model with both ‘SQuAD’ dataset (version 2.0) and BioAsq train data. For training on ‘SQuAD’, hyper parameters- Learning rate and number of epochs are set to ‘3e-3’ and ‘3’ respectively as mentioned in the paper BIBREF1. Test accuracy of the model boosted to 44%. In one more experiment we trained model only on ‘SQuAD’ dataset, this time test accuracy of the model moved to 47%. The reason model did not perform up to the mark when trained with ‘SQuAD’ alongside BioASQ data could be that in formatted BioASQ data, start_index for the answer is not accurate, and affected the overall accuracy.
Our Systems and Their Performance on Factoid Questions
We have experimented with several systems and their variations, e.g. created by training with specific additional features (see next subsection). Here is their list and short descriptions. Unfortunately we did not pay attention to naming, and the systems evolved between test batches, so the overall picture can only be understood by looking at the details.
When we started the experiments our objective was to see whether BioBERT and entailment-based techniques can provide value for in the context of biomedical question answering. The answer to both questions was a yes, qualified by many examples clearly showing the limitations of both methods. Therefore we tried to address some of these limitations using feature engineering with mixed results: some clear errors got corrected and new errors got introduced, without overall improvement but convincing us that in future experiments it might be worth trying feature engineering again especially if more training data were available.
Overall we experimented with several approaches with the following aspects of the systems changing between batches, that is being absent or present:
training on BioAsq data vs. training on SQuAD
using the BioAsq snippets for context vs. using the documents from the provided URLs for context
adding or not the LAT, i.e. lexical answer type, feature (see BIBREF9, BIBREF10 and an explanation in the subsection just below).
For Yes/No questions (only) we experimented with the entailment methods.
We will discuss the performance of these models below and in Section 6. But before we do that, let us discuss a feature engineering experiment which eventually produced mixed results, but where we feel it is potentially useful in future experiments.
Our Systems and Their Performance on Factoid Questions ::: LAT Feature considered and its impact (slightly negative)
During error analysis we found that for some cases, answer being returned by the model is far away from what it is being asked in the Question.
Example: (from BIBREF8)
Question: Hy's law measures failure of which organ?
Actual Answer: ‘Liver’.
The answer returned by one of our models was ‘alanine aminotransferase’, which is an enzyme. The model returns an enzyme, when the question asked for the organ name. To address this type of errors, we decided to try the concepts of ‘Lexical Answer Type’ (LAT) and Focus Word, which was used in IBM Watson, see BIBREF11 for overview; BIBREF10 for technical details, and BIBREF9 for details on question analysis. In an example given in the last source we read:
POETS & POETRY: He was a bank clerk in the Yukon before he published "Songs of a Sourdough" in 1907.
The focus is the part of the question that is a reference to the answer. In the example above, the focus is "he".
LATs are terms in the question that indicate what type of entity is being asked for.
The headword of the focus is generally a LAT, but questions often contain additional LATs, and in the Jeopardy! domain, categories are an additional source of LATs.
(...) In the example, LATs are "he", "clerk", and "poet".
For example in the question "Which plant does oleuropein originate from?" (BIBREF8). The LAT here is ‘plant’. For the BioAsq task we did not need to explicitly distinguish between the focus and the LAT concepts. In this example, the expectation is that answer returned by the model is a plant. Thus it is conceivable that the cosine distance between contextual embedding of word 'plant' in the question and contextual embedding for the answer present in the paragraph(context) is comparatively low. As a result model learns to adjust its weights during training phase and returns answers with low cosine distance with the LAT.
We used Stanford CoreNLP BIBREF12 library to write rules for extracting lexical answer type present in the question, both 'parts of speech'(POS) and dependency parsing functionality was used. We incorporated the Lexical Answer Type into one of our systems, UNCC_QA1 in Batch 4. This system underperformed our system FACTOIDS by about 3% in the MRR measure, but corrected errors such as in the example above.
Our Systems and Their Performance on Factoid Questions ::: LAT Feature considered and its impact (slightly negative) ::: Assumptions and rules for deriving lexical answer type.
There are different question types: ‘Which’, ‘What’, ‘When’, ‘How’ etc. Each type of question is being handled differently and there are commonalities among the rules written for different question types. Question words are identified through parts of speech tags: 'WDT', 'WRB' ,'WP'. We assumed that LAT is a ‘Noun’ and follows the question word. Often it was also a subject (nsubj). This process is illustrated in Fig.FIGREF15.
LAT computation was governed by a few simple rules, e.g. when a question has multiple words that are 'Subjects’ (and ‘Noun’), a word that is in proximity to the question word is considered as ‘LAT’. These rules are different for each "Wh" word.
Namely, when the word immediately following the question word is a Noun, window size is set to ‘3’. The window size ‘3’ means we iterate through the next ‘3’ words to check if any of the word is both Noun and Subject, If so, such word is considered the ‘LAT’; else the word that is present very next to the question word is considered as the ‘LAT’.
For questions with words ‘Which’ , ‘What’, ‘When’; a Noun immediately following the question word is very often the LAT, e.g. 'enzyme' in Which enzyme is targeted by Evolocumab?. When the word immediately following the question word is not a Noun, e.g. in What is the function of the protein Magt1? the window size is set to ‘5’, and we iterate through the next ‘5’ words (if present) and search for the word that is both Noun and Subject. If present, the word is considered as the ‘LAT’; else, the Noun in close proximity to the question word and following it is returned as the ‘LAT’.
For questions with question words: ‘When’, ‘Who’, ‘Why’, the ’LAT’ is a question word itself. For the word ‘How', e.g. in How many selenoproteins are encoded in the human genome?, we look at the adjective and if we find one, we take it to be the LAT, otherwise the word 'How' is considered as the ‘LAT’.
Perhaps because of using only very simple rules, the accuracy for ‘LAT’ derivation is 75%; that is, in the remaining 25% of the cases the LAT word is identified incorrectly. Worth noting is that the overall performance the system that used LATs was slightly inferior to the system without LATs, but the types of errors changed. The training used BioBERT with the LAT feature as part of the input string. The errors it introduces usually involve finding the wrong element of the correct type e.g. wrong enzyme when two similar enzymes are described in the text, or 'neuron' when asked about a type of cell with a certain function, when the answer calls for a different cell category, adipocytes, and both are mentioned in the text. We feel with more data and additional tuning or perhaps using an ensemble model, we might be able to keep the correct answers, and improve the results on the confusing examples like the one mentioned above. Therefore if we improve our ‘LAT’ derivation logic, or have larger datasets, then perhaps the neural network techniques they will yield better results.
Our Systems and Their Performance on Factoid Questions ::: Impact of Training using BioAsq data (slightly negative)
Training on BioAsq data in our entry in Batch 1 and Batch 2 under the name QA1 showed it might lead to overfitting. This happened both with (Batch 2) and without (Batch 1) hyperparameters tuning: abysmal 18% MRR in Batch 1, and slighly better one, 40% in Batch 2 (although in Batch 2 it was overall the second best result in MRR but 16% lower than the highest score).
In Batch 3 (only), our UNCC_QA3 system was fine tuned on BioAsq and SQuAD 2.0 BIBREF7, and for data preprocessing Context paragraph is generated from relevant snippets provided in the test data. This system underperformed, by about 2% in MRR, our other entry UNCC_QA1, which was also an overall category winner for this batch. The latter was also trained on SQuAD, but not on BioAsq. We suspect that the reason could be the simplistic nature of the find() function described in Section 3.1. So, this could be an area where a better algorithm for finding the best occurrence of an entity could improve performance.
Our Systems and Their Performance on Factoid Questions ::: Impact of Using Context from URLs (negative)
In some experiments, for context in testing, we used documents for which URL pointers are provided in BioAsq. However, our system UNCC_QA3 underperformed our other system tested only on the provided snippets.
In Batch 5 the underperformance was about 6% of MRR, compared to our best system UNCC_QA1, and by 9% to the top performer.
Performance on Yes/No and List questions
Our work focused on Factoid questions. But we also have done experiments on List-type and Yes/No questions.
Performance on Yes/No and List questions ::: Entailment improves Yes/No accuracy
We started by answering always YES (in batch 2 and 3) to get the baseline performance. For batch 4 we used entailment. Our algorithm was very simple: Given a question we iterate through the candidate sentences and try to find any candidate sentence is contradicting the question (with confidence over 50%), if so 'No' is returned as answer, else 'Yes' is returned. In batch 4 this strategy produced better than the BioAsq baseline performance, and compared to our other systems, the use of entailment increased the performance by about 13% (macro F1 score). We used 'AllenNlp' BIBREF13 entailment library to find entailment of the candidate sentences with question.
Performance on Yes/No and List questions ::: For List-type the URLs have negative impact
Overall, we followed the similar strategy that's been followed for Factoid Question Answering task. We started our experiment with batch 2, where we submitted 20 best answers (with context from snippets). Starting with batch 3, we performed post processing: once models generate answer predictions (n-best predictions), we do post-processing on the predicted answers. In test batch 4, our system (called FACTOIDS) achieved highest recall score of ‘0.7033’ but low precision of 0.1119, leaving open the question of how could we have better balanced the two measures.
In the post-processing phase, we take the top ‘20’ (batch 3) and top 5 (batch 4 and 5), predicted answers, tokenize them using common separators: 'comma' , 'and', 'also', 'as well as'. Tokens with characters count more than ‘100’ are eliminated and rest of the tokens are added to the list of possible answers. BioASQ evaluation mechanism does not consider snippets with more than ‘100’ characters as a valid answer. Considering lengthy snippets in to the list of answers would reduce the mean precision score. As a final step, duplicate snippets in the answer pool are removed. For example, consider these top 3 answers predicted by the system (before post-processing):
{
"text": "dendritic cells",
"probability": 0.7554540733426441,
"start_logit": 8.466046333312988,
"end_logit": 9.536355018615723
},
{
"text": "neutrophils, macrophages and
distinct subtypes of dendritic cells",
"probability": 0.13806867348304214,
"start_logit": 6.766478538513184,
"end_logit": 9.536355018615723
},
{
"text": "macrophages and distinct subtypes of dendritic",
"probability": 0.013973475271178242,
"start_logit": 6.766478538513184,
"end_logit": 7.24576473236084
},
After execution of post-processing heuristics, the list of answers returned is as follows:
["dendritic cells"],
["neutrophils"],
["macrophages"],
["distinct subtypes of dendritic cells"]
Summary of our results
The tables below summarize all our results. They show that the performance of our systems was mixed. The simple architectures and algorithm we used worked very well only in Batch 3. However, we feel we can built a better system based on this experience. In particular we observed both the value of contextual embeddings and of feature engineering (LAT), however we failed to combine them properly.
Summary of our results ::: Factoid questions ::: Systems used in Batch 5 experiments
System description for ‘UNCC_QA1’: The system was finetuned on the SQuAD 2.0. For data preprocessing Context / paragraph was generated from relevant snippets provided in the test data.
System description for ‘QA1’ : ‘LAT’ feature was added and finetuned with SQuAD 2.0. For data preprocessing Context / paragraph was generated from relevant snippets provided in the test data.
System description for ‘UNCC_QA3’ : Fine tuning process is same as it is done for the system ‘UNCC_QA1’ in test batch-5. Difference is during data preprocessing, Context/paragraph is generated from the relevant documents for which URLS are included in the test data.
Summary of our results ::: List Questions
For List-type questions, although post processing helped in the later batches, we never managed to obtain competitive precision, although our recall was good.
Summary of our results ::: Yes/No questions
The only thing worth remembering from our performance is that using entailment can have a measurable impact (at least with respect to a weak baseline). The results (weak) are in Table 3.
Discussion, Future Experiments, and Conclusions ::: Summary:
In contrast to 2018, when we submitted BIBREF2 to BioASQ a system based on extractive summarization (and scored very high in the ideal answer category), this year we mainly targeted factoid question answering task and focused on experimenting with BioBERT. After these experiments we see the promise of BioBERT in QA tasks, but we also see its limitations. The latter we tried to address with mixed results using feature engineering. Overall these experiments allowed us to secure a best and a second best score in different test batches. Along with Factoid-type question, we also tried ‘Yes/No’ and ‘List’-type questions, and did reasonably well with our very simple approach.
For Yes/No the moral worth remembering is that reasoning has a potential to influence results, as evidenced by our adding the AllenNLP entailment BIBREF13 system increased its performance.
All our data and software is available at Github, in the previously referenced URL (end of Section 2).
Discussion, Future Experiments, and Conclusions ::: Future experiments
In the current model, we have a shallow neural network with a softmax layer for predicting answer span. Shallow networks however are not good at generalizations. In our future experiments we would like to create dense question answering neural network with a softmax layer for predicting answer span. The main idea is to get contextual word embedding for the words present in the question and paragraph (Context) and feed the contextual word embeddings retrieved from the last layer of BioBERT to the dense question answering network. The mentioned dense layered question answering neural network need to be tuned for finding right hyper parameters. An example of such architecture is shown in Fig.FIGREF30.
In one more experiment, we would like to add a better version of ‘LAT’ contextual word embedding as a feature, along with the actual contextual word embeddings for question text, and Context and feed them as input to the dense question answering neural network. By this experiment, we would like to find if ‘LAT’ feature is improving overall answer prediction accuracy. Adding ‘LAT’ feature this way instead of feeding this word piece embedding directly to the BioBERT (as we did in our above experiments) would not downgrade the quality of contextual word embeddings generated form ‘BioBERT'. Quality contextual word embeddings would lead to efficient transfer learning and chances are that it would improve the model's answer prediction accuracy.
We also see potential for incorporating domain specific inference into the task e.g. using the MedNLI dataset BIBREF14. For all types of experiments it might be worth exploring clinical BERT embeddings BIBREF15, explicitly incorporating domain knowledge (e.g. BIBREF16) and possibly deeper discourse representations (e.g. BIBREF17).
APPENDIX
In this appendix we provide additional details about the implementations.
APPENDIX ::: Systems and their descriptions:
We used several variants of our systems when experimenting with the BioASQ problems. In retrospect, it would be much easier to understand the changes if we adopted some mnemonic conventions in naming the systems. So, we apologize for the names that do not reflect the modifications, and necessitate this list.
APPENDIX ::: Systems and their descriptions: ::: Factoid Type Question Answering:
We preprocessed the test data to convert test data to BioBERT format, We generated Context/paragraph by either aggregating relevant snippets provided or by aggregating documents for which URLS are provided in the BioASQ test data.
APPENDIX ::: Systems and their descriptions: ::: System description for QA1:
We generated Context/paragraph by aggregating relevant snippets available in the test data and mapped it against the question text and question id. We ignored the content present in the documents (document URLS were provided in the original test data). The model is finetuned with BioASQ data.
data preprocessing is done in the same way as it is done for test batch-1. Model fine tuned on BioASQ data.
‘LAT’/ Focus word feature added and fine tuned with SQuAD 2.0 [reference]. For data preprocessing Context / paragraph is generated from relevant snippets provided in the test data.
APPENDIX ::: Systems and their descriptions: ::: System description for UNCC_QA_1:
System is finetuned on the SQuAD 2.0 [reference]. For data preprocessing Context / paragraph is generated from relevant snippets provided in the test data.
‘LAT’/ Focus word feature added and fine tuned with SQuAD 2.0 [reference]. For data preprocessing Context / paragraph is generated from relevant snippets provided in the test data.
The System is finetuned on the SQuAD 2.0. For data preprocessing Context / paragraph is generated from relevant snippets provided in the test data.
APPENDIX ::: Systems and their descriptions: ::: System description for UNCC_QA3:
System is finetuned on the SQuAD 2.0 [reference] and BioASQ dataset[].For data preprocessing Context / paragraph is generated from relevant snippets provided in the test data.
Fine tuning process is same as it is done for the system ‘UNCC_QA_1’ in test batch-5. Difference is during data preprocessing, Context/paragraph is generated form from the relevant documents for which URLS are included in the test data.
APPENDIX ::: Systems and their descriptions: ::: System description for UNCC_QA2:
Fine tuning process is same as for ‘UNCC_QA_1 ’. Difference is Context/paragraph is generated form from the relevant documents for which URLS are included in the test data. System ‘UNCC_QA_1’ got the highest ‘MRR’ score in the 3rd test batch set.
APPENDIX ::: Systems and their descriptions: ::: System description for FACTOIDS:
The System is finetuned on the SQuAD 2.0. For data preprocessing Context / paragraph is generated from relevant snippets provided in the test data.
APPENDIX ::: Systems and their descriptions: ::: List Type Questions:
We attempted List type questions starting from test batch ‘2’. Used similar approach that's been followed for Factoid Question answering task. For all the test batch sets, in the data pre processing phase Context/ paragraph is generated either by aggregating relevant snippets or by aggregating documents(URLS) provided in the BioASQ test data.
For test batch-2, model (System: QA1) is finetuned on BioASQ data and submitted top ‘20’ answers predicted by the model as the list of answers. system ‘QA1’ achieved low F-Measure score:‘0.0786’ in the second test batch. In the further test batches for List type questions, we finetuned the model on Squad data set [reference], implemented post processing techniques (refer section 5.2) and achieved a better F-measure score: ‘0.2862’ in the final test batch set.
In test batch-3 (Systems : ‘QA1’/’’UNCC_QA_1’/’UNCC_QA3’/’UNCC_QA2’) top 20 answers returned by the model is sent for post processing and in test batch 4 and 5 only top 5 answers are sent for post processing. System UNCC_QA2(in batch 3) for List type question answering, Context is generated from documents for which URLS are provided in the BioASQ test data. for the rest of the systems (in test batch-3) for List Type question answering task snippets present in the BioaSQ test data are used to generate context.
In test batch-4 (System : ‘FACTOIDS’/’UNCC_QA_1’/’UNCC_QA3’’) top 5 answers returned by the model is sent for post processing. In case of system ‘FACTOIDS’ snippets in the test data were used to generate context. for systems ’UNCC_QA_1’ and ’UNCC_QA3’ context is generated from the documents for which URLS are provided in the BioASQ test data.
In test batch-5 ( Systems: ‘QA1’/’UNCC_QA_1’/’UNCC_QA3’/’UNCC_QA2’ ) our approach is the same as that of test batch-4 where top 5 answers returned by the model is sent for post processing. for all the systems (in test batch-5) context is generated from the snippets provided in the BioASQ test data.
APPENDIX ::: Systems and their descriptions: ::: Yes/No Type Questions:
For the first 3 test batches, We have submitted answer ‘Yes’ to all the questions. Later, we employed ‘Sentence Entailment’ techniques(refer section 6.0) for the fourth and fifth test batch sets. Our Systems with ‘Sentence Entailment’ approach (for ‘Yes’/ ‘No’ question answering): ‘UNCC_QA_1’(test batch-4), UNCC_QA3(test batch-5).
APPENDIX ::: Additional details for Yes/No Type Questions
We used Textual Entailment in Batch 4 and 5 for ‘Yes’/‘No’ question type. The algorithm was very simple: Given a question we iterate through the candidate sentences, and look for any candidate sentences contradicting the question. If we find one 'No' is returned as answer, else 'Yes' is returned. (The confidence for contradiction was set at 50%) We used AllenNLP BIBREF13 entailment library to find entailment of the candidate sentences with question.
Flow Chart for Yes/No Question answer processing is shown in Fig.FIGREF51
APPENDIX ::: Assumptions, rules and logic flow for deriving Lexical Answer Types from questions
There are different question types, and we distinguished them based on the question words: ‘Which’, ‘What’, ‘When’, ‘How’ etc. Each type of question is being handled differently and there are commonalities among the rules written for different question types. How are question words identified? question words have parts of speech(POS): 'WDT', 'WRB', 'WP'.
Assumptions:
1) Lexical answer type (‘LAT’) or focus word is of type Noun and follows the question word.
2) The LAT word is a Subject. (This clearly not always true, but we used a very simple method). Note: ‘StanfordNLP’ dependency parsing tag for identifying subject is 'nsubj' or 'nsubjpass'.
3) When a question has multiple words that are of type Subject (and Noun), a word that is in proximity to the question word is considered as ‘LAT’.
4) For questions with question words: ‘When’, ‘Who’, ‘Why’, the ’LAT’ is a question word itself that is, ‘When’, ‘Who’, ‘Why’ respectively.
Rules and logic flow to traverse a question: The three cases below describe the logic flow of finding LATs. The figures show the grammatical structures used for this purpose.
APPENDIX ::: Assumptions, rules and logic flow for deriving Lexical Answer Types from questions ::: Case-1:
Question with question word ‘How’.
For questions with question word 'How', the adjective that follows the question word is considered as ‘LAT’ (need not follow immediately). If an adjective is absent, word 'How' is considered as ‘LAT’. When there are multiple words that are adjectives, a word in close proximity to the question word and follows it is returned as ‘LAT’. Note: The part of speech tag to identify adjectives is 'JJ'. For Other possible question words like ‘whose’. ‘LAT’/Focus word is question words itself.
Example Question: How many selenoproteins are encoded in the human genome?
APPENDIX ::: Assumptions, rules and logic flow for deriving Lexical Answer Types from questions ::: Case-2:
Questions with question words ‘Which’ , ‘What’ and all other possible question words; a 'Noun' immediately following the question word.
Example Question: Which enzyme is targeted by Evolocumab?
Here, Focus word/LAT is ‘enzyme’ which is both Noun and Subject and immediately follows the question word.
When the word immediately following the question word is a noun, the window size is set to ‘3’. This size ‘3’ means that we iterate through the next ‘3’ words (if present) to check if any of the word is both 'Noun' and 'Subject', If so, the word is considered as ‘LAT’/Focus Word. Else the word that is present very next to the question word is considered as ‘LAT’.
APPENDIX ::: Assumptions, rules and logic flow for deriving Lexical Answer Types from questions ::: Case-3:
Questions with question words ‘Which’ , ‘What’ and all other possible question words; word immediately following the question word is not a 'Noun'.
Example Question: What is the function of the protein Magt1?
Here, Focus word/LAT is ‘function ’ which is both Noun and Subject and does not immediately follow the question word.
When the very next word following the question word is not a Noun, window size is set to ‘5’. Window size ‘5’ corresponds that we iterate through the next ‘5’ words (if present) and search for the word that is both Noun and Subject. If present, the word is considered as ‘LAT’. Else, the 'Noun' close proximity to the question word and follows it is returned as ‘LAT’.
Ad we mentioned earlier, the accuracy for ‘LAT’ derivation is 75 percent. But clearly the simple logic described above can be improved, as shown in BIBREF9, BIBREF10. Whether this in turn produces improvements in this particular task is an open question.
APPENDIX ::: Proposing Future Experiments
In the current model, we have a shallow neural network with a softmax layer for predicting answer span. Shallow networks however are not good at generalizations. In our future experiments we would like to create dense question answering neural network with a softmax layer for predicting answer span. The main idea is to get contextual word embedding for the words present in the question and paragraph (Context) and feed the contextual word embeddings retrieved from the last layer of BioBERT to the dense question answering network. The mentioned dense layered question answering Neural network need to be tuned for finding right hyper parameters. An example of such architecture is shown in Fig.FIGREF30.
In another experiment we would like to only feed contextual word embeddings for Focus word/ ‘LAT’, paragraph/ Context as input to the question answering neural network. In this experiment we would neglect all embeddings for the question text except that of Focus word/ ‘LAT’. Our assumption and idea for considering focus word and neglecting remaining words in the question is that during training phase it would make more precise for the model to identify the focus of the question and map answers against the question’s focus. To validate our assumption, we would like to take sample question answering data and find the cosine distance between contextual embedding of Focus word and that of the actual answer and verify if the cosine distance is comparatively low in most of the cases.
In one more experiment, we would like to add a better version of ‘LAT’ contextual word embedding as a feature, along with the actual contextual word embeddings for question text, and Context and feed them as input to the dense question answering neural network. By this experiment, we would like to find if ‘LAT’ feature is improving overall answer prediction accuracy. Adding ‘LAT’ feature this way instead of feeding Focus word’s word piece embedding directly (as we did in our above experiments) to the BioBERT would not downgrade the quality of contextual word embeddings generated form ‘BioBERT'. Quality contextual word embeddings would lead to efficient transfer learning and chances are that it would improve the model's answer prediction accuracy.
| [
"0.7033",
"0.7033"
] | 6,810 | qasper | en | null | 6ceab4edd1d0e37d217958e7e962697124ccb6a4f449f4af | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What was their highest recall score?
Answer: | [
"0.7033",
"0.7033"
] | qasper | 128 | 73 | |
What embedding techniques are explored in the paper? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Measures of semantic similarity and relatedness quantify the degree to which two concepts are similar (e.g., INLINEFORM0 – INLINEFORM1 ) or related (e.g., INLINEFORM2 – INLINEFORM3 ). Semantic similarity can be viewed as a special case of semantic relatedness – to be similar is one of many ways that a pair of concepts may be related. The automated discovery of groups of semantically similar or related terms is critical to improving the retrieval BIBREF0 and clustering BIBREF1 of biomedical and clinical documents, and the development of biomedical terminologies and ontologies BIBREF2 .
There is a long history in using distributional methods to discover semantic similarity and relatedness (e.g., BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 ). These methods are all based on the distributional hypothesis, which holds that two terms that are distributionally similar (i.e., used in the same context) will also be semantically similar BIBREF7 , BIBREF8 . Recently word embedding techniques such as word2vec BIBREF9 have become very popular. Despite the prominent role that neural networks play in many of these approaches, at their core they remain distributional techniques that typically start with a word by word co–occurrence matrix, much like many of the more traditional approaches.
However, despite these successes distributional methods do not perform well when data is very sparse (which is common). One possible solution is to use second–order co–occurrence vectors BIBREF10 , BIBREF11 . In this approach the similarity between two words is not strictly based on their co–occurrence frequencies, but rather on the frequencies of the other words which occur with both of them (i.e., second order co–occurrences). This approach has been shown to be successful in quantifying semantic relatedness BIBREF12 , BIBREF13 . However, while more robust in the face of sparsity, second–order methods can result in significant amounts of noise, where contextual information that is overly general is included and does not contribute to quantifying the semantic relatedness between the two concepts.
Our goal then is to discover methods that automatically reduce the amount of noise in a second–order co–occurrence vector. We achieve this by incorporating pairwise semantic similarity scores derived from a taxonomy into our second–order vectors, and then using these scores to select only the most semantically similar co–occurrences (thereby reducing noise).
We evaluate our method on two datasets that have been annotated in multiple ways. One has been annotated for both similarity and relatedness, and the other has been annotated for relatedness by two different types of experts (medical doctors and medical coders). Our results show that integrating second order co–occurrences with measures of semantic similarity increases correlation with our human reference standards. We also compare our result to a number of other studies which have applied various word embedding methods to the same reference standards we have used. We find that our method often performs at a comparable or higher level than these approaches. These results suggest that our methods of integrating semantic similarity and relatedness values have the potential to improve performance of purely distributional methods.
Similarity and Relatedness Measures
This section describes the similarity and relatedness measures we integrate in our second–order co–occurrence vectors. We use two taxonomies in this study, SNOMED–CT and MeSH. SNOMED–CT (Systematized Nomenclature of Medicine Clinical Terms) is a comprehensive clinical terminology created for the electronic representation of clinical health information. MeSH (Medical Subject Headings) is a taxonomy of biomedical terms developed for indexing biomedical journal articles.
We obtain SNOMED–CT and MeSH via the Unified Medical Language System (UMLS) Metathesaurus (version 2016AA). The Metathesaurus contains approximately 2 million biomedical and clinical concepts from over 150 different terminologies that have been semi–automatically integrated into a single source. Concepts in the Metathesaurus are connected largely by two types of hierarchical relations: INLINEFORM0 / INLINEFORM1 (PAR/CHD) and INLINEFORM2 / INLINEFORM3 (RB/RN).
Similarity Measures
Measures of semantic similarity can be classified into three broad categories : path–based, feature–based and information content (IC). Path–based similarity measures use the structure of a taxonomy to measure similarity – concepts positioned close to each other are more similar than those further apart. Feature–based methods rely on set theoretic measures of overlap between features (union and intersection). The information content measures quantify the amount of information that a concept provides – more specific concepts have a higher amount of information content.
RadaMBB89 introduce the Conceptual Distance measure. This measure is simply the length of the shortest path between two concepts ( INLINEFORM0 and INLINEFORM1 ) in the MeSH hierarchy. Paths are based on broader than (RB) and narrower than (RN) relations. CaviedesC04 extends this measure to use parent (PAR) and child (CHD) relations. Our INLINEFORM2 measure is simply the reciprocal of this shortest path value (Equation EQREF3 ), so that larger values (approaching 1) indicate a high degree of similarity. DISPLAYFORM0
While the simplicity of INLINEFORM0 is appealing, it can be misleading when concepts are at different levels of specificity. Two very general concepts may have the same path length as two very specific concepts. WuP94 introduce a correction to INLINEFORM1 that incorporates the depth of the concepts, and the depth of their Least Common Subsumer (LCS). This is the most specific ancestor two concepts share. In this measure, similarity is twice the depth of the two concept's LCS divided by the product of the depths of the individual concepts (Equation EQREF4 ). Note that if there are multiple LCSs for a pair of concepts, the deepest of them is used in this measure. DISPLAYFORM0
ZhongZLY02 take a very similar approach and again scale the depth of the LCS by the sum of the depths of the two concepts (Equation EQREF5 ), where INLINEFORM0 . The value of INLINEFORM1 was set to 2 based on their recommendations. DISPLAYFORM0
PekarS02 offer another variation on INLINEFORM0 , where the shortest path of the two concepts to the LCS is used, in addition to the shortest bath between the LCS and the root of the taxonomy (Equation EQREF6 ). DISPLAYFORM0
Feature–based methods represent each concept as a set of features and then measure the overlap or sharing of features to measure similarity. In particular, each concept is represented as the set of their ancestors, and similarity is a ratio of the intersection and union of these features.
MaedcheS01 quantify the similarity between two concepts as the ratio of the intersection over their union as shown in Equation EQREF8 . DISPLAYFORM0
BatetSV11 extend this by excluding any shared features (in the numerator) as shown in Equation EQREF9 . DISPLAYFORM0
Information content is formally defined as the negative log of the probability of a concept. The effect of this is to assign rare (low probability) concepts a high measure of information content, since the underlying assumption is that more specific concepts are less frequently used than more common ones.
Resnik95 modified this notion of information content in order to use it as a similarity measure. He defines the similarity of two concepts to be the information content of their LCS (Equation EQREF11 ). DISPLAYFORM0
JiangC97, Lin98, and PirroE10 extend INLINEFORM0 by incorporating the information content of the individual concepts in various different ways. Lin98 defines the similarity between two concepts as the ratio of information content of the LCS with the sum of the individual concept's information content (Equation EQREF12 ). Note that INLINEFORM1 has the same form as INLINEFORM2 and INLINEFORM3 , and is in effect using information content as a measure of specificity (rather than depth). If there is more than one possible LCS, the LCS with the greatest IC is chosen. DISPLAYFORM0
JiangC97 define the distance between two concepts to be the sum of the information content of the two concepts minus twice the information content of the concepts' LCS. We modify this from a distance to a similarity measure by taking the reciprocal of the distance (Equation EQREF13 ). Note that the denominator of INLINEFORM0 is very similar to the numerator of INLINEFORM1 . DISPLAYFORM0
PirroE10 define the similarity between two concepts as the information content of the two concept's LCS divided by the sum of their individual information content values minus the information content of their LCS (Equation EQREF14 ). Note that INLINEFORM0 can be viewed as a set–theoretic version of INLINEFORM1 . DISPLAYFORM0
Information Content
The information content of a concept may be derived from a corpus (corpus–based) or directly from a taxonomy (intrinsic–based). In this work we focus on corpus–based techniques.
For corpus–based information content, we estimate the probability of a concept INLINEFORM0 by taking the sum of the probability of the concept INLINEFORM1 and the probability its descendants INLINEFORM2 (Equation EQREF16 ). DISPLAYFORM0
The initial probabilities of a concept ( INLINEFORM0 ) and its descendants ( INLINEFORM1 ) are obtained by dividing the number of times each concept and descendant occurs in the corpus, and dividing that by the total numbers of concepts ( INLINEFORM2 ).
Ideally the corpus from which we are estimating the probabilities of concepts will be sense–tagged. However, sense–tagging is a challenging problem in its own right, and it is not always possible to carry out reliably on larger amounts of text. In fact in this paper we did not use any sense–tagging of the corpus we derived information content from.
Instead, we estimated the probability of a concept by using the UMLSonMedline dataset. This was created by the National Library of Medicine and consists of concepts from the 2009AB UMLS and the counts of the number of times they occurred in a snapshot of Medline taken on 12 January, 2009. These counts were obtained by using the Essie Search Engine BIBREF14 which queried Medline with normalized strings from the 2009AB MRCONSO table in the UMLS. The frequency of a CUI was obtained by aggregating the frequency counts of the terms associated with the CUI to provide a rough estimate of its frequency. The information content measures then use this information to calculate the probability of a concept.
Another alternative is the use of Intrinsic Information Content. It assess the informativeness of concept based on its placement within a taxonomy by considering the number of incoming (ancestors) relative to outgoing (descendant) links BIBREF15 (Equation EQREF17 ). DISPLAYFORM0
where INLINEFORM0 are the number of descendants of concept INLINEFORM1 that are leaf nodes, INLINEFORM2 are the number of concept INLINEFORM3 's ancestors and INLINEFORM4 are the total number of leaf nodes in the taxonomy.
Relatedness Measures
Lesk86 observed that concepts that are related should share more words in their respective definitions than concepts that are less connected. He was able to perform word sense disambiguation by identifying the senses of words in a sentence with the largest number of overlaps between their definitions. An overlap is the longest sequence of one or more consecutive words that occur in both definitions. BanerjeeP03 extended this idea to WordNet, but observed that WordNet glosses are often very short, and did not contain enough information to distinguish between multiple concepts. Therefore, they created a super–gloss for each concept by adding the glosses of related concepts to the gloss of the concept itself (and then finding overlaps).
PatwardhanP06 adapted this measure to second–order co–occurrence vectors. In this approach, a vector is created for each word in a concept's definition that shows which words co–occur with it in a corpus. These word vectors are averaged to create a single co-occurrence vector for the concept. The similarity between the concepts is calculated by taking the cosine between the concepts second–order vectors. LiuMPMP12 modified and extended this measure to be used to quantify the relatedness between biomedical and clinical terms in the UMLS. The work in this paper can be seen as a further extension of PatwardhanP06 and LiuMPMP12.
Method
In this section, we describe our second–order similarity vector measure. This incorporates both contextual information using the term pair's definition and their pairwise semantic similarity scores derived from a taxonomy. There are two stages to our approach. First, a co–occurrence matrix must be constructed. Second, this matrix is used to construct a second–order co–occurrence vector for each concept in a pair of concepts to be measured for relatedness.
Co–occurrence Matrix Construction
We build an INLINEFORM0 similarity matrix using an external corpus where the rows and columns represent words within the corpus and the element contains the similarity score between the row word and column word using the similarity measures discussed above. If a word maps to more than one possible sense, we use the sense that returns the highest similarity score.
For this paper our external corpus was the NLM 2015 Medline baseline. Medline is a bibliographic database containing over 23 million citations to journal articles in the biomedical domain and is maintained by National Library of Medicine. The 2015 Medline Baseline encompasses approximately 5,600 journals starting from 1948 and contains 23,343,329 citations, of which 2,579,239 contain abstracts. In this work, we use Medline titles and abstracts from 1975 to present day. Prior to 1975, only 2% of the citations contained an abstract. We then calculate the similarity for each bigram in this dataset and include those that have a similarity score greater than a specified threshold on these experiments.
Measure Term Pairs for Relatedness
We obtain definitions for each of the two terms we wish to measure. Due to the sparsity and inconsistencies of the definitions in the UMLS, we not only use the definition of the term (CUI) but also include the definition of its related concepts. This follows the method proposed by PatwardhanP06 for general English and WordNet, and which was adapted for the UMLS and the medical domain by LiuMPMP12. In particular we add the definitions of any concepts connected via a parent (PAR), child (CHD), RB (broader than), RN (narrower than) or TERM (terms associated with CUI) relation. All of the definitions for a term are combined into a single super–gloss. At the end of this process we should have two super–glosses, one for each term to be measured for relatedness.
Next, we process each super–gloss as follows:
We extract a first–order co–occurrence vector for each term in the super–gloss from the co–occurrence matrix created previously.
We take the average of the first order co–occurrence vectors associated with the terms in a super–gloss and use that to represent the meaning of the term. This is a second–order co–occurrence vector.
After a second–order co–occurrence vector has been constructed for each term, then we calculate the cosine between these two vectors to measure the relatedness of the terms.
Data
We use two reference standards to evaluate the semantic similarity and relatedness measures . UMNSRS was annotated for both similarity and relatedness by medical residents. MiniMayoSRS was annotated for relatedness by medical doctors (MD) and medical coders (coder). In this section, we describe these data sets and describe a few of their differences.
MiniMayoSRS: The MayoSRS, developed by PakhomovPMMRC10, consists of 101 clinical term pairs whose relatedness was determined by nine medical coders and three physicians from the Mayo Clinic. The relatedness of each term pair was assessed based on a four point scale: (4.0) practically synonymous, (3.0) related, (2.0) marginally related and (1.0) unrelated. MiniMayoSRS is a subset of the MayoSRS and consists of 30 term pairs on which a higher inter–annotator agreement was achieved. The average correlation between physicians is 0.68. The average correlation between medical coders is 0.78. We evaluate our method on the mean of the physician scores, and the mean of the coders scores in this subset in the same manner as reported by PedersenPPC07.
UMNSRS: The University of Minnesota Semantic Relatedness Set (UMNSRS) was developed by PakhomovMALPM10, and consists of 725 clinical term pairs whose semantic similarity and relatedness was determined independently by four medical residents from the University of Minnesota Medical School. The similarity and relatedness of each term pair was annotated based on a continuous scale by having the resident touch a bar on a touch sensitive computer screen to indicate the degree of similarity or relatedness. The Intraclass Correlation Coefficient (ICC) for the reference standard tagged for similarity was 0.47, and 0.50 for relatedness. Therefore, as suggested by Pakhomov and colleagues,we use a subset of the ratings consisting of 401 pairs for the similarity set and 430 pairs for the relatedness set which each have an ICC of 0.73.
Experimental Framework
We conducted our experiments using the freely available open source software package UMLS::Similarity BIBREF16 version 1.47. This package takes as input two terms (or UMLS concepts) and returns their similarity or relatedness using the measures discussed in Section SECREF2 .
Correlation between the similarity measures and human judgments were estimated using Spearman's Rank Correlation ( INLINEFORM0 ). Spearman's measures the statistical dependence between two variables to assess how well the relationship between the rankings of the variables can be described using a monotonic function. We used Fisher's r-to-z transformation BIBREF17 to calculate the significance between the correlation results.
Results and Discussion
Table TABREF26 shows the Spearman's Rank Correlation between the human scores from the four reference standards and the scores from the various measures of similarity introduced in Section SECREF2 . Each class of measure is followed by the scores obtained when integrating our second order vector approach with these measures of semantic similarity.
Results Comparison
The results for UMNSRS tagged for similarity ( INLINEFORM0 ) and MiniMayoSRS tagged by coders show that all of the second-order similarity vector measures ( INLINEFORM1 ) except for INLINEFORM2 - INLINEFORM3 obtain a higher correlation than the original measures. We found that INLINEFORM4 - INLINEFORM5 and INLINEFORM6 - INLINEFORM7 obtain the highest correlations of all these results with human judgments.
For the UMNSRS dataset tagged for relatedness and MiniMayoSRS tagged by physicians (MD), the original INLINEFORM0 measure obtains a higher correlation than our measure ( INLINEFORM1 ) although the difference is not statistically significant ( INLINEFORM2 ).
In order to analyze and better understand these results, we filtered the bigram pairs used to create the initial similarity matrix based on the strength of their similarity using the INLINEFORM0 and the INLINEFORM1 measures. Note that the INLINEFORM2 measure holds to a 0 to 1 scale, while INLINEFORM3 ranges from 0 to an unspecified upper bound that is dependent on the size of the corpus from which information content is estimated. As such we use a different range of threshold values for each measure. We discuss the results of this filtering below.
Thresholding Experiments
Table TABREF29 shows the results of applying the threshold parameter on each of the reference standards using the INLINEFORM0 measure. For example, a threshold of 0 indicates that all of the bigrams were included in the similarity matrix; and a threshold of 1 indicates that only the bigram pairs with a similarity score greater than one were included.
These results show that using a threshold cutoff of 2 obtains the highest correlation for the UMNSRS dataset, and that a threshold cutoff of 4 obtains the highest correlation for the MiniMayoSRS dataset. All of the results show an increase in correlation with human judgments when incorporating a threshold cutoff over all of the original measures. The increase in the correlation for the UMNSRS tagged for similarity is statistically significant ( INLINEFORM0 ), however this is not the case for the UMNSRS tagged for relatedness nor for the MiniMayoSRS data.
Similarly, Table TABREF30 shows the results of applying the threshold parameter (T) on each of the reference standards using the INLINEFORM0 measure. Although, unlike INLINEFORM1 whose scores are greater than or equal to 0 without an upper limit, the INLINEFORM2 measure returns scores between 0 and 1 (inclusive). Therefore, here a threshold of 0 indicates that all of the bigrams were included in the similarity matrix; and a threshold of INLINEFORM3 indicates that only the bigram pairs with a similarity score greater than INLINEFORM4 were included. The results show an increase in accuracy for all of the datasets except for the MiniMayoSRS tagged for physicians. The increase in the results for the UMNSRS tagged for similarity and the MayoSRS is statistically significant ( INLINEFORM5 ). This is not the case for the UMNSRS tagged for relatedness nor the MiniMayoSRS.
Overall, these results indicate that including only those bigrams that have a sufficiently high similarity score increases the correlation results with human judgments, but what quantifies as sufficiently high varies depending on the dataset and measure.
Comparison with Previous Work
Recently, word embeddings BIBREF9 have become a popular method for measuring semantic relatedness in the biomedical domain. This is a neural network based approach that learns a representation of a word by word co–occurrence matrix. The basic idea is that the neural network learns a series of weights (the hidden layer within the neural network) that either maximizes the probability of a word given its context, referred to as the continuous bag of words (CBOW) approach, or that maximizes the probability of the context given a word, referred to as the Skip–gram approach. These approaches have been used in numerous recent papers.
muneeb2015evalutating trained both the Skip–gram and CBOW models over the PubMed Central Open Access (PMC) corpus of approximately 1.25 million articles. They evaluated the models on a subset of the UMNSRS data, removing word pairs that did not occur in their training corpus more than ten times. chiu2016how evaluated both the the Skip–gram and CBOW models over the PMC corpus and PubMed. They also evaluated the models on a subset of the UMNSRS ignoring those words that did not appear in their training corpus. Pakhomov2016corpus trained CBOW model over three different types of corpora: clinical (clinical notes from the Fairview Health System), biomedical (PMC corpus), and general English (Wikipedia). They evaluated their method using a subset of the UMNSRS restricting to single word term pairs and removing those not found within their training corpus. sajad2015domain trained the Skip–gram model over CUIs identified by MetaMap on the OHSUMED corpus, a collection of 348,566 biomedical research articles. They evaluated the method on the complete UMNSRS, MiniMayoSRS and the MayoSRS datasets; any subset information about the dataset was not explicitly stated therefore we believe a direct comparison may be possible.
In addition, a previous work very closely related to ours is a retrofitting vector method proposed by YuCBJW16 that incorporates ontological information into a vector representation by including semantically related words. In their measure, they first map a biomedical term to MeSH terms, and second build a word vector based on the documents assigned to the respective MeSH term. They then retrofit the vector by including semantically related words found in the Unified Medical Language System. They evaluate their method on the MiniMayoSRS dataset.
Table TABREF31 shows a comparison to the top correlation scores reported by each of these works on the respective datasets (or subsets) they evaluated their methods on. N refers to the number of term pairs in the dataset the authors report they evaluated their method. The table also includes our top scoring results: the integrated vector-res and vector-faith. The results show that integrating semantic similarity measures into second–order co–occurrence vectors obtains a higher or on–par correlation with human judgments as the previous works reported results with the exception of the UMNSRS rel dataset. The results reported by Pakhomov2016corpus and chiu2016how obtain a higher correlation although the results can not be directly compared because both works used different subsets of the term pairs from the UMNSRS dataset.
Conclusion and Future Work
We have presented a method for quantifying the similarity and relatedness between two terms that integrates pair–wise similarity scores into second–order vectors. The goal of this approach is two–fold. First, we restrict the context used by the vector measure to words that exist in the biomedical domain, and second, we apply larger weights to those word pairs that are more similar to each other. Our hypothesis was that this combination would reduce the amount of noise in the vectors and therefore increase their correlation with human judgments. We evaluated our method on datasets that have been manually annotated for relatedness and similarity and found evidence to support this hypothesis. In particular we discovered that guiding the creation of a second–order context vector by selecting term pairs from biomedical text based on their semantic similarity led to improved levels of correlation with human judgment.
We also explored using a threshold cutoff to include only those term pairs that obtained a sufficiently large level of similarity. We found that eliminating less similar pairs improved the overall results (to a point). In the future, we plan to explore metrics to automatically determine the threshold cutoff appropriate for a given dataset and measure. We also plan to explore additional features that can be integrated with a second–order vector measure that will reduce the noise but still provide sufficient information to quantify relatedness. We are particularly interested in approaches that learn word, phrase, and sentence embeddings from structured corpora such as literature BIBREF23 and dictionary entries BIBREF24 . Such embeddings could be integrated into a second–order vector or be used on their own.
Finally, we compared our proposed method to other distributional approaches, focusing on those that used word embeddings. Our results showed that integrating semantic similarity measures into second–order co–occurrence vectors obtains the same or higher correlation with human judgments as do various different word embedding approaches. However, a direct comparison was not possible due to variations in the subsets of the UMNSRS evaluation dataset used. In the future, we would not only like to conduct a direct comparison but also explore integrating semantic similarity into various kinds of word embeddings by training on pair–wise values of semantic similarity as well as co–occurrence statistics.
| [
"Skip–gram, CBOW",
"integrated vector-res, vector-faith, Skip–gram, CBOW"
] | 4,259 | qasper | en | null | 710f7452388fe5d22e2329697dbc4fa17d2c9a53daedafa6 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What embedding techniques are explored in the paper?
Answer: | [
"Skip–gram, CBOW",
"integrated vector-res, vector-faith, Skip–gram, CBOW"
] | qasper | 128 | 74 | |
How do they match words before reordering them? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Deep Learning approaches have achieved impressive results on various NLP tasks BIBREF0 , BIBREF1 , BIBREF2 and have become the de facto approach for any NLP task. However, these deep learning techniques have found to be less effective for low-resource languages when the available training data is very less BIBREF3 . Recently, several approaches like Multi-task learning BIBREF4 , multilingual learning BIBREF5 , semi-supervised learning BIBREF2 , BIBREF6 and transfer learning BIBREF7 , BIBREF3 have been explored by the deep learning community to overcome data sparsity in low-resource languages. Transfer learning trains a model for a parent task and fine-tunes the learned parent model weights (features) for a related child task BIBREF7 , BIBREF8 . This effectively reduces the requirement on training data for the child task as the model would have learned relevant features from the parent task data thereby, improving the performance on the child task.
Transfer learning has also been explored in the multilingual Neural Machine Translation BIBREF3 , BIBREF9 , BIBREF10 . The goal is to improve the NMT performance on the source to target language pair (child task) using an assisting source language (assisting to target translation is the parent task). Here, the parent model is trained on the assisting and target language parallel corpus and the trained weights are used to initialize the child model. The child model can now be fine-tuned on the source-target language pairs, if parallel corpus is available. The divergence between the source and the assisting language can adversely impact the benefits obtained from transfer learning. Multiple studies have shown that transfer learning works best when the languages are related BIBREF3 , BIBREF10 , BIBREF9 . Several studies have tried to address lexical divergence between the source and the target languages BIBREF10 , BIBREF11 , BIBREF12 . However, the effect of word order divergence and its mitigation has not been explored. In a practical setting, it is not uncommon to have source and assisting languages with different word order. For instance, it is possible to find parallel corpora between English and some Indian languages, but very little parallel corpora between Indian languages. Hence, it is natural to use English as an assisting language for inter-Indian language translation.
To see how word order divergence can be detrimental, let us consider the case of the standard RNN (Bi-LSTM) encoder-attention-decoder architecture BIBREF13 . The encoder generates contextual representations (annotation vectors) for each source word, which are used by the attention network to match the source words to the current decoder state. The contextual representation is word-order dependent. Hence, if the assisting and the source languages do not have similar word order the generated contextual representations will not be consistent. The attention network (and hence the decoder) sees different contextual representations for similar words in parallel sentences across different languages. This makes it difficult to transfer knowledge learned from the assisting language to the source language.
We illustrate this by visualizing the contextual representations generated by the encoder of an English to Hindi NMT system for two versions of the English input: (a) original word order (SVO) (b) word order of the source language (SOV, for Bengali). Figure FIGREF1 shows that the encoder representations obtained are very different. The attention network and the decoder now have to work with very different representations. Note that the plot below does not take into account further lexical and other divergences between source and assisting languages, since we demonstrated word order divergence with the same language on the source side.
To address this word order divergence, we propose to pre-order the assisting language sentences to match the word order of the source language. We consider an extremely resource constrained scenario, where we do not have any parallel corpus for the child task. We are limited to a bilingual dictionary for transfer information from the assisting to the source language. From our experiments, we show that there is a significant increase in the translation accuracy for the unseen source-target language pair.
Addressing Lexical Divergence
BIBREF3 explored transfer learning for NMT on low-resource languages. They studied the influence of language divergence between languages chosen for training the parent and child model, and showed that choosing similar languages for training the parent and child model leads to better improvements from transfer learning. A limitation of BIBREF3 approach is that they ignore the lexical similarity between languages and also the source language embeddings are randomly initialized. BIBREF10 , BIBREF11 , BIBREF12 take advantage of lexical similarity between languages in their work. BIBREF10 proposed to use Byte-Pair Encoding (BPE) to represent the sentences in both the parent and the child language to overcome the above limitation. They show using BPE benefits transfer learning especially when the involved languages are closely-related agglutinative languages. Similarly, BIBREF11 utilize lexical similarity between the source and assisting languages by training a character-level NMT system. BIBREF12 address lexical divergence by using bilingual embeddings and mixture of universal token embeddings. One of the languages' vocabulary, usually English vocabulary is considered as universal tokens and every word in the other languages is represented as a mixture of universal tokens. They show results on extremely low-resource languages.
Addressing Word Order Divergence
To the best of our knowledge, no work has addressed word order divergence in transfer learning for multilingual NMT. However, some work exists for other NLP tasks that could potentially address word order. For Named Entity Recognition (NER), BIBREF14 use a self-attention layer after the Bi-LSTM layer to address word-order divergence for Named Entity Recognition (NER) task. The approach does not show any significant improvements over multiple languages. A possible reason is that the divergence has to be addressed before/during construction of the contextual embeddings in the Bi-LSTM layer, and the subsequent self-attention layer does not address word-order divergence. BIBREF15 use adversarial training for cross-lingual question-question similarity ranking in community question answering. The adversarial training tries to force the encoder representations of similar sentences from different input languages to have similar representations.
Use of Pre-ordering
Pre-ordering the source language sentences to match the target language word order has been useful in addressing word-order divergence for Phrase-Based SMT BIBREF16 , BIBREF17 , BIBREF18 , BIBREF19 . Recently, BIBREF20 proposed a way to measure and reduce the divergence between the source and target languages based on morphological and syntactic properties, also termed as anisomorphism. They demonstrated that by reducing the anisomorphism between the source and target languages, consistent improvements in NMT performance were obtained. The NMT system used additional features like word forms, POS tags and dependency relations in addition to parallel corpora. On the other hand, BIBREF21 observed a drop in performance due to pre-ordering for NMT. Unlike BIBREF20 , the NMT system was trained on pre-ordered sentences and no additional features were provided to the system. Note that all these works address source-target divergence, not divergence between source languages in multilingual NMT.
Proposed Solution
Consider the task of translating from an extremely low-resource language (source) to a target language. The parallel corpus between the two languages if available may be too small to train a NMT model. Similar to existing works BIBREF3 , BIBREF10 , BIBREF12 , we use transfer learning to overcome data sparsity and train a NMT model between the source and the target languages. Specifically, the NMT model (parent model) is trained on the assisting language and target language pairs. We choose English as the assisting language in all our experiments. In our resource-scarce scenario, we have no parallel corpus for the child task. Hence, at test time, the source language sentence is translated using the parent model after performing a word-by-word translation into the assisting language.
Since the source language and the assisting language (English) have different word order, we hypothesize that it leads to inconsistencies in the contextual representations generated by the encoder for the two languages. In this paper, we propose to pre-order English sentences (assisting language sentences) to match the word-order of the source language and train the parent model on this pre-ordered corpus. In our experiments, we look at scenarios where the assisting language has SVO word order and the source language has SOV word order.
For instance, consider the English sentence Anurag will meet Thakur. One of the pre-ordering rule swaps the position of the noun phrase followed by a transitive verb with the transitive verb. The original and the resulting re-ordered parse tree will be as shown in the Table TABREF5 . Applying this reordering rule to the above sentence Anurag will meet Thakur will yield the reordered sentence Anurag Thakur will meet. Additionally, the Table TABREF5 shows the parse trees for the above sentence with and without pre-ordering.
Pre-ordering should also be beneficial for other word order divergence scenarios (e.g., SOV to SVO), but we leave verification of these additional scenarios for future work.
Experimental Setup
In this section, we describe the languages experimented with, datasets used, the network hyper-parameters used in our experiments.
Languages
We experimented with English INLINEFORM0 Hindi translation as the parent task. English is the assisting source language. Bengali, Gujarati, Marathi, Malayalam and Tamil are the primary source languages, and translation from these to Hindi constitute the child tasks. Hindi, Bengali, Gujarati and Marathi are Indo-Aryan languages, while Malayalam and Tamil are Dravidian languages. All these languages have a canonical SOV word order.
Datasets
For training English-Hindi NMT systems, we use the IITB English-Hindi parallel corpus BIBREF22 ( INLINEFORM0 sentences from the training set) and the ILCI English-Hindi parallel corpus ( INLINEFORM1 sentences). The ILCI (Indian Language Corpora Initiative) multilingual parallel corpus BIBREF23 spans multiple Indian languages from the health and tourism domains. We use the 520-sentence dev-set of the IITB parallel corpus for validation. For each child task, we use INLINEFORM2 sentences from ILCI corpus as the test set.
Network
We use OpenNMT-Torch BIBREF24 to train the NMT system. We use the standard sequence-to-sequence architecture with attention BIBREF13 . We use an encoder which contains two layers of bidirectional LSTMs with 500 neurons each. The decoder contains two LSTM layers with 500 neurons each. Input feeding approach BIBREF1 is used where the previous attention hidden state is fed as input to the decoder LSTM. We use a mini-batch of size 50 and use a dropout layer. We begin with an initial learning rate of INLINEFORM0 and decay the learning rate by a factor of INLINEFORM1 when the perplexity on validation set increases. The training is stopped when the learning rate falls below INLINEFORM2 or number of epochs=22. The English input is initialized with pre-trained embeddings trained using fastText BIBREF25 .
English vocabulary consists of INLINEFORM0 tokens appearing at least 2 times in the English training corpus. For constructing the Hindi vocabulary we considered only those tokens appearing at least 5 times in the training split resulting in a vocabulary size of INLINEFORM1 tokens. For representing English and other source languages into a common space, we translate each word in the source language into English using a bilingual dictionary (Google Translate word translation in our case). In an end-to-end solution, it would have been ideal to use bilingual embeddings or obtain word-by-word translations via bilingual embeddings BIBREF14 . But, the quality of publicly available bilingual embeddings for English-Indian languages is very low for obtaining good-quality, bilingual representations BIBREF26 , BIBREF27 . We also found that these embeddings were not useful for transfer learning.
We use the CFILT-preorder system for reordering English sentences to match the Indian language word order. It contains two re-ordering systems: (1) generic rules that apply to all Indian languages BIBREF17 , and (2) hindi-tuned rules which improve the generic rules by incorporating improvements found through an error analysis of English-Hindi reordering BIBREF28 . These Hindi-tuned rules have been found to improve reordering for many English to Indian language pairs BIBREF29 .
Results
In this section, we describe the results from our experiments on NMT task. We report the results on X-Hindi pair, where X is one of Bengali, Gujarati, Marathi, Tamil, and Malayalam. The results are presented in the Table TABREF6 . We report BLEU scores and LeBLEU scores BIBREF30 . We observe that both the pre-ordering configurations significantly improve the BLEU scores over the baseline scores. We observe larger gains when generic pre-ordering rules are used compared to the Hindi-tuned pre-ordering rules.
These results support our hypothesis that word-order divergence can limit the benefits of multilingual translation. Reducing the word order divergence can improve translation in extremely low-resource scenarios.
An analysis of the outputs revealed that pre-ordering significantly reducing the number of UNK tokens (placeholder for unknown words) in the test output (Table TABREF14 ). We hypothesize that due to word order divergence between English and Indian languages, the encoder representation generated is not consistent leading to decoder generating unknown words. However, the pre-ordered models generate better contextual representations leading to less number of unknown tokens and better translation which is also reflected in the BLEU scores.
Conclusion
In this paper, we show that handling word-order divergence between source and assisting languages is crucial for the success of multilingual NMT in an extremely low-resource setting. We show that pre-ordering the assisting language to match the word order of the source language significantly improves translation quality in an extremely low-resource setting. While the current work focused on Indian languages, we would like to validate the hypothesis on a more diverse set of languages.
| [
"Unanswerable",
"CFILT-preorder system"
] | 2,231 | qasper | en | null | d7fcca67f5487a55282f8217e4f05bc54ee67e2ddb578e73 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How do they match words before reordering them?
Answer: | [
"Unanswerable",
"CFILT-preorder system"
] | qasper | 128 | 75 | |
Does the paper explore extraction from electronic health records? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The explosion of available scientific articles in the Biomedical domain has led to the rise of Biomedical Information Extraction (BioIE). BioIE systems aim to extract information from a wide spectrum of articles including medical literature, biological literature, electronic health records, etc. that can be used by clinicians and researchers in the field. Often the outputs of BioIE systems are used to assist in the creation of databases, or to suggest new paths for research. For example, a ranked list of interacting proteins that are extracted from biomedical literature, but are not present in existing databases, can allow researchers to make informed decisions about which protein/gene to study further. Interactions between drugs are necessary for clinicians who simultaneously administer multiple drugs to their patients. A database of diseases, treatments and tests is beneficial for doctors consulting in complicated medical cases.
The main problems in BioIE are similar to those in Information Extraction:
This paper discusses, in each section, various methods that have been adopted to solve the listed problems. Each section also highlights the difficulty of Information Extraction tasks in the biomedical domain.
This paper is intended as a primer to Biomedical Information Extraction for current NLP researchers. It aims to highlight the diversity of the various techniques from Information Extraction that have been applied in the Biomedical domain. The state of biomedical text mining is reviewed regularly. For more extensive surveys, consult BIBREF0 , BIBREF1 , BIBREF2 .
Named Entity Recognition and Fact Extraction
Named Entity Recognition (NER) in the Biomedical domain usually includes recognition of entities such as proteins, genes, diseases, treatments, drugs, etc. Fact extraction involves extraction of Named Entities from a corpus, usually given a certain ontology. When compared to NER in the domain of general text, the biomedical domain has some characteristic challenges:
Some of the earliest systems were heavily dependent on hand-crafted features. The method proposed in BIBREF4 for recognition of protein names in text does not require any prepared dictionary. The work gives examples of diversity in protein names and lists multiple rules depending on simple word features as well as POS tags.
BIBREF5 adopt a machine learning approach for NER. Their NER system extracts medical problems, tests and treatments from discharge summaries and progress notes. They use a semi-Conditional Random Field (semi-CRF) BIBREF6 to output labels over all tokens in the sentence. They use a variety of token, context and sentence level features. They also use some concept mapping features using existing annotation tools, as well as Brown clustering to form 128 clusters over the unlabelled data. The dataset used is the i2b2 2010 challenge dataset. Their system achieves an F-Score of 0.85. BIBREF7 is an incremental paper on NER taggers. It uses 3 types of word-representation techniques (Brown clustering, distributional clustering, word vectors) to improve performance of the NER Conditional Random Field tagger, and achieves marginal F-Score improvements.
BIBREF8 propose a boostrapping mechanism to bootstrap biomedical ontologies using NELL BIBREF9 , which uses a coupled semi-supervised bootstrapping approach to extract facts from text, given an ontology and a small number of “seed” examples for each category. This interesting approach (called BioNELL) uses an ontology of over 100 categories. In contrast to NELL, BioNELL does not contain any relations in the ontology. BioNELL is motivated by the fact that a lot of scientific literature available online is highly reliable due to peer-review. The authors note that the algorithm used by NELL to bootstrap fails in BioNELL due to ambiguities in biomedical literature, and heavy semantic drift. One of the causes for this is that often common words such as “white”, “dad”, “arm” are used as names of genes- this can easily result in semantic drift in one iteration of the bootstrapping. In order to mitigate this, they use Pointwise Mutual Information scores for corpus level statistics, which attributes a small score to common words. In addition, in contrast to NELL, BioNELL only uses high instances as seeds in the next iteration, but adds low ranking instances to the knowledge base. Since evaluation is not possible using Mechanical Turk or a small number of experts (due to the complexity of the task), they use Freebase BIBREF10 , a knowledge base that has some biomedical concepts as well. The lexicon learned using BioNELL is used to train an NER system. The system shows a very high precision, thereby showing that BioNELL learns very few ambiguous terms.
More recently, deep learning techniques have been developed to further enhance the performance of NER systems. BIBREF11 explore recurrent neural networks for the problem of NER in biomedical text.
Relation Extraction
In Biomedical Information Extraction, Relation Extraction involves finding related entities of many different kinds. Some of these include protein-protein interactions, disease-gene relations and drug-drug interactions. Due to the explosion of available biomedical literature, it is impossible for one person to extract relevant relations from published material. Automatic extraction of relations assists in the process of database creation, by suggesting potentially related entities with links to the source article. For example, a database of drug-drug interactions is important for clinicians who administer multiple drugs simultaneously to their patients- it is imperative to know if one drug will have an adverse effect on the other. A variety of methods have been developed for relation extractions, and are often inspired by Relation Extraction in NLP tasks. These include rule-based approaches, hand-crafted patterns, feature-based and kernel machine learning methods, and more recently deep learning architectures. Relation Extraction systems over Biomedical Corpora are often affected by noisy extraction of entities, due to ambiguities in names of proteins, genes, drugs etc.
BIBREF12 was one of the first large scale Information Extraction efforts to study the feasibility of extraction of protein-protein interactions (such as “protein A activates protein B") from Biomedical text. Using 8 hand-crafted regular expressions over a fixed vocabulary, the authors were able to achieve a recall of 30% for interactions present in The Dictionary of Interacting Proteins (DIP) from abstracts in Medline. The method did not differentiate between the type of relation. The reasons for the low recall were the inconsistency in protein nomenclature, information not present in the abstract, and due to specificity of the hand-crafted patterns. On a small subset of extracted relations, they found that about 60% were true interactions between proteins not present in DIP.
BIBREF13 combine sentence level relation extraction for protein interactions with corpus level statistics. Similar to BIBREF12 , they do not consider the type of interaction between proteins- only whether they interact in the general sense of the word. They also do not differentiate between genes and their protein products (which may share the same name). They use Pointwise Mutual Information (PMI) for corpus level statistics to determine whether a pair of proteins occur together by chance or because they interact. They combine this with a confidence aggregator that takes the maximum of the confidence of the extractor over all extractions for the same protein-pair. The extraction uses a subsequence kernel based on BIBREF14 . The integrated model, that combines PMI with aggregate confidence, gives the best performance. Kernel methods have widely been studied for Relation Extraction in Biomedical Literature. Common kernels used usually exploit linguistic information by utilising kernels based on the dependency tree BIBREF15 , BIBREF16 , BIBREF17 .
BIBREF18 look at the extraction of diseases and their relevant genes. They use a dictionary from six public databases to annotate genes and diseases in Medline abstracts. In their work, the authors note that when both genes and diseases are correctly identified, they are related in 94% of the cases. The problem then reduces to filtering incorrect matches using the dictionary, which occurs due to false positives resulting from ambiguities in the names as well as ambiguities in abbreviations. To this end, they train a Max-Ent based NER classifier for the task, and get a 26% gain in precision over the unfiltered baseline, with a slight hit in recall. They use POS tags, expanded forms of abbreviations, indicators for Greek letters as well as suffixes and prefixes commonly used in biomedical terms.
BIBREF19 adopt a supervised feature-based approach for the extraction of drug-drug interaction (DDI) for the DDI-2013 dataset BIBREF20 . They partition the data in subsets depending on the syntactic features, and train a different model for each. They use lexical, syntactic and verb based features on top of shallow parse features, in addition to a hand-crafted list of trigger words to define their features. An SVM classifier is then trained on the feature vectors, with a positive label if the drug pair interacts, and negative otherwise. Their method beats other systems on the DDI-2013 dataset. Some other feature-based approaches are described in BIBREF21 , BIBREF22 .
Distant supervision methods have also been applied to relation extraction over biomedical corpora. In BIBREF23 , 10,000 neuroscience articles are distantly supervised using information from UMLS Semantic Network to classify brain-gene relations into geneExpression and otherRelation. They use lexical (bag of words, contextual) features as well as syntactic (dependency parse features). They make the “at-least one” assumption, i.e. at least one of the sentences extracted for a given entity-pair contains the relation in database. They model it as a multi-instance learning problem and adopt a graphical model similar to BIBREF24 . They test using manually annotated examples. They note that the F-score achieved are much lesser than that achieved in the general domain in BIBREF24 , and attribute to generally poorer performance of NER tools in the biomedical domain, as well as less training examples. BIBREF25 explore distant supervision methods for protein-protein interaction extraction.
More recently, deep learning methods have been applied to relation extraction in the biomedical domain. One of the main advantages of such methods over traditional feature or kernel based learning methods is that they require minimal feature engineering. In BIBREF26 , skip-gram vectors BIBREF27 are trained over 5.6Gb of unlabelled text. They use these vectors to extract protein-protein interactions by converting them into features for entities, context and the entire sentence. Using an SVM for classification, their method is able to outperform many kernel and feature based methods over a variety of datasets.
BIBREF28 follow a similar method by using word vectors trained on PubMed articles. They use it for the task of relation extraction from clinical text for entities that include problem, treatment and medical test. For a given sentence, given labelled entities, they predict the type of relation exhibited (or None) by the entity pair. These types include “treatment caused medical problem”, “test conducted to investigate medical problem”, “medical problem indicates medical problems”, etc. They use a Convolutional Neural Network (CNN) followed by feedforward neural network architecture for prediction. In addition to pre-trained word vectors as features, for each token they also add features for POS tags, distance from both the entities in the sentence, as well BIO tags for the entities. Their model performs better than a feature based SVM baseline that they train themselves.
The BioNLP'16 Shared Tasks has also introduced some Relation Extraction tasks, in particular the BB3-event subtask that involves predicting whether a “lives-in” relation holds for a Bacteria in a location. Some of the top performing models for this task are deep learning models. BIBREF29 train word embeddings with six billions words of scientific texts from PubMed. They then consider the shortest dependency path between the two entities (Bacteria and location). For each token in the path, they use word embedding features, POS type embeddings and dependency type embeddings. They train a unidirectional LSTM BIBREF30 over the dependency path, that achieves an F-Score of 52.1% on the test set.
BIBREF31 improve the performance by making modifications to the above model. Instead of using the shortest dependency path, they modify the parse tree based on some pruning strategies. They also add feature embeddings for each token to represent the distance from the entities in the shortest path. They then train a Bidirectional LSTM on the path, and obtain an F-Score of 57.1%.
The recent success of deep learning models in Biomedical Relation Extraction that require minimal feature engineering is promising. This also suggests new avenues of research in the field. An approach as in BIBREF32 can be used to combine multi-instance learning and distant supervision with a neural architecture.
Event Extraction
Event Extraction in the Biomedical domain is a task that has gained more importance recently. Event Extraction goes beyond Relation Extraction. In Biomedical Event Extraction, events generally refer to a change in the state of biological molecules such as proteins and DNA. Generally, it includes detection of targeted event types such as gene expression, regulation, localisation and transcription. Each event type in addition can have multiple arguments that need to be detected. An additional layer of complexity comes from the fact that events can also be arguments of other events, giving rise to a nested structure. This helps to capture the underlying biology better BIBREF1 . Detecting the event type often involves recognising and classifying trigger words. Often, these words are verbs such as “activates”, “inhibits”, “phosphorylation” that may indicate a single, or sometimes multiple event types. In this section, we will discuss some of the successful models for Event Extraction in some detail.
Event Extraction gained a lot of interest with the availability of an annotated corpus with the BioNLP'09 Shared Task on Event Extraction BIBREF34 . The task involves prediction of trigger words over nine event types such as expression, transcription, catabolism, binding, etc. given only annotation of named entities (proteins, genes, etc.). For each event, its class, trigger expression and arguments need to be extracted. Since the events can be arguments to other events, the final output in general is a graph representation with events and named entities as nodes, and edges that correspond to event arguments. BIBREF33 present a pipeline based method that is heavily dependent on dependency parsing. Their pipeline approach consists of three steps: trigger detection, argument detection and semantic post-processing. While the first two components are learning based systems, the last component is a rule based system. For the BioNLP'09 corpus, only 5% of the events span multiple sentences. Hence the approach does not get affected severely by considering only single sentences. It is important to note that trigger words cannot simply be reduced to a dictionary lookup. This is because a specific word may belong to multiple classes, or may not always be a trigger word for an event. For example, “activate” is found to not be a trigger word in over 70% of the cases. A multi-class SVM is trained for trigger detection on each token, using a large feature set consisting of semantic and syntactic features. It is interesting to note that the hyperparameters of this classifier are optimised based on the performance of the entire end-to-end system.
For the second component to detect arguments, labels for edges between entities must be predicted. For the BioNLP'09 Shared Task, each directed edge from one event node to another event node, or from an event node to a named entity node are classified as “theme”, “cause”, or None. The second component of the pipeline makes these predictions independently. This is also trained using a multi-class SVM which involves heavy use of syntactic features, including the shortest dependency path between the nodes. The authors note that the precision-recall choice of the first component affects the performance of the second component: since the second component is only trained on Gold examples, any error by the first component will lead to a cascading of errors. The final component, which is a semantic post-processing step, consists of rules and heuristics to correct the output of the second component. Since the edge predictions are made independently, it is possible that some event nodes do not have any edges, or have an improper combination of edges. The rule based component corrects these and applies rules to break directed cycles in the graph, and some specific heuristics for different types of events. The final model gives a cumulative F-Score of 52% on the test set, and was the best model on the task.
BIBREF35 note that previous approaches on the task suffer due to the pipeline nature and the propagation of errors. To counter this, they adopt a joint inference method based on Markov Logic Networks BIBREF36 for the same task on BioNLP'09. The Markov Logic Network jointly predicts whether each token is a trigger word, and if yes, the class it belongs to; for each dependency edge, whether it is an argument path leading to a “theme” or a “cause”. By formulating the Event Extraction problem using an MLN, the approach becomes computationally feasible and only linear in the length of the sentence. They incorporate hard constraints to encode rules such as “an argument path must have an event”, “a cause path must start with a regulation event”, etc. In addition, they also include some domain specific soft constraints as well as some linguistically-motivated context-specific soft constraints. In order to train the MLN, stochastic gradient descent was used. Certain heuristic methods are implemented in order to deal with errors due to syntactic parsing, especially ambiguities in PP-attachment and coordination. Their final system is competitive and comes very close to the system by BIBREF33 with an average F-Score of 50%. To further improve the system, they suggest leveraging additional joint-inference opportunities and integrating the syntactic parser better. Some other more recent models for Biomedical Event Extraction include BIBREF37 , BIBREF38 .
Conclusion
We have discussed some of the major problems and challenges in BioIE, and seen some of the diverse approaches adopted to solve them. Some interesting problems such as Pathway Extraction for Biological Systems BIBREF39 , BIBREF40 have not been discussed.
Biomedical Information Extraction is a challenging and exciting field for NLP researchers that demands application of state-of-the-art methods. Traditionally, there has been a dependence on hand-crafted features or heavily feature-engineered methods. However, with the advent of deep learning methods, a lot of BioIE tasks are seeing an improvement by adopting deep learning models such as Convolutional Neural Networks and LSTMs, which require minimal feature engineering. Rapid progress in developing better systems for BioIE will be extremely helpful for clinicians and researchers in the Biomedical domain.
| [
"Yes"
] | 3,035 | qasper | en | null | 901f735b1582acacb606002ab77c6c7a3fe6017d38349aa2 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Does the paper explore extraction from electronic health records?
Answer: | [
"Yes"
] | qasper | 128 | 76 | |
Who were the experts used for annotation? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Privacy policies are the documents which disclose the ways in which a company gathers, uses, shares and manages a user's data. As legal documents, they function using the principle of notice and choice BIBREF0, where companies post their policies, and theoretically, users read the policies and decide to use a company's products or services only if they find the conditions outlined in its privacy policy acceptable. Many legal jurisdictions around the world accept this framework, including the United States and the European Union BIBREF1, BIBREF2. However, the legitimacy of this framework depends upon users actually reading and understanding privacy policies to determine whether company practices are acceptable to them BIBREF3. In practice this is seldom the case BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10. This is further complicated by the highly individual and nuanced compromises that users are willing to make with their data BIBREF11, discouraging a `one-size-fits-all' approach to notice of data practices in privacy documents.
With devices constantly monitoring our environment, including our personal space and our bodies, lack of awareness of how our data is being used easily leads to problematic situations where users are outraged by information misuse, but companies insist that users have consented. The discovery of increasingly egregious uses of data by companies, such as the scandals involving Facebook and Cambridge Analytica BIBREF12, have further brought public attention to the privacy concerns of the internet and ubiquitous computing. This makes privacy a well-motivated application domain for NLP researchers, where advances in enabling users to quickly identify the privacy issues most salient to them can potentially have large real-world impact.
[1]https://play.google.com/store/apps/details?id=com.gotokeep.keep.intl [2]https://play.google.com/store/apps/details?id=com.viber.voip [3]A question might not have any supporting evidence for an answer within the privacy policy.
Motivated by this need, we contribute PrivacyQA, a corpus consisting of 1750 questions about the contents of privacy policies, paired with over 3500 expert annotations. The goal of this effort is to kickstart the development of question-answering methods for this domain, to address the (unrealistic) expectation that a large population should be reading many policies per day. In doing so, we identify several understudied challenges to our ability to answer these questions, with broad implications for systems seeking to serve users' information-seeking intent. By releasing this resource, we hope to provide an impetus to develop systems capable of language understanding in this increasingly important domain.
Related Work
Prior work has aimed to make privacy policies easier to understand. Prescriptive approaches towards communicating privacy information BIBREF21, BIBREF22, BIBREF23 have not been widely adopted by industry. Recently, there have been significant research effort devoted to understanding privacy policies by leveraging NLP techniques BIBREF24, BIBREF25, BIBREF26, BIBREF27, BIBREF28, especially by identifying specific data practices within a privacy policy. We adopt a personalized approach to understanding privacy policies, that allows users to query a document and selectively explore content salient to them. Most similar is the PolisisQA corpus BIBREF29, which examines questions users ask corporations on Twitter. Our approach differs in several ways: 1) The PrivacyQA dataset is larger, containing 10x as many questions and answers. 2) Answers are formulated by domain experts with legal training. 3) PrivacyQA includes diverse question types, including unanswerable and subjective questions.
Our work is also related to reading comprehension in the open domain, which is frequently based upon Wikipedia passages BIBREF16, BIBREF17, BIBREF15, BIBREF30 and news articles BIBREF20, BIBREF31, BIBREF32. Table.TABREF4 presents the desirable attributes our dataset shares with past approaches. This work is also tied into research in applying NLP approaches to legal documents BIBREF33, BIBREF34, BIBREF35, BIBREF36, BIBREF37, BIBREF38, BIBREF39. While privacy policies have legal implications, their intended audience consists of the general public rather than individuals with legal expertise. This arrangement is problematic because the entities that write privacy policies often have different goals than the audience. feng2015applying, tan-EtAl:2016:P16-1 examine question answering in the insurance domain, another specialized domain similar to privacy, where the intended audience is the general public.
Data Collection
We describe the data collection methodology used to construct PrivacyQA. With the goal of achieving broad coverage across application types, we collect privacy policies from 35 mobile applications representing a number of different categories in the Google Play Store. One of our goals is to include both policies from well-known applications, which are likely to have carefully-constructed privacy policies, and lesser-known applications with smaller install bases, whose policies might be considerably less sophisticated. Thus, setting 5 million installs as a threshold, we ensure each category includes applications with installs on both sides of this threshold. All policies included in the corpus are in English, and were collected before April 1, 2018, predating many companies' GDPR-focused BIBREF41 updates. We leave it to future studies BIBREF42 to look at the impact of the GDPR (e.g., to what extent GDPR requirements contribute to making it possible to provide users with more informative answers, and to what extent their disclosures continue to omit issues that matter to users).
Data Collection ::: Crowdsourced Question Elicitation
The intended audience for privacy policies consists of the general public. This informs the decision to elicit questions from crowdworkers on the contents of privacy policies. We choose not to show the contents of privacy policies to crowdworkers, a procedure motivated by a desire to avoid inadvertent biases BIBREF43, BIBREF44, BIBREF45, BIBREF46, BIBREF47, and encourage crowdworkers to ask a variety of questions beyond only asking questions based on practices described in the document.
Instead, crowdworkers are presented with public information about a mobile application available on the Google Play Store including its name, description and navigable screenshots. Figure FIGREF9 shows an example of our user interface. Crowdworkers are asked to imagine they have access to a trusted third-party privacy assistant, to whom they can ask any privacy question about a given mobile application. We use the Amazon Mechanical Turk platform and recruit crowdworkers who have been conferred “master” status and are located within the United States of America. Turkers are asked to provide five questions per mobile application, and are paid $2 per assignment, taking ~eight minutes to complete the task.
Data Collection ::: Answer Selection
To identify legally sound answers, we recruit seven experts with legal training to construct answers to Turker questions. Experts identify relevant evidence within the privacy policy, as well as provide meta-annotation on the question's relevance, subjectivity, OPP-115 category BIBREF49, and how likely any privacy policy is to contain the answer to the question asked.
Data Collection ::: Analysis
Table.TABREF17 presents aggregate statistics of the PrivacyQA dataset. 1750 questions are posed to our imaginary privacy assistant over 35 mobile applications and their associated privacy documents. As an initial step, we formulate the problem of answering user questions as an extractive sentence selection task, ignoring for now background knowledge, statistical data and legal expertise that could otherwise be brought to bear. The dataset is partitioned into a training set featuring 27 mobile applications and 1350 questions, and a test set consisting of 400 questions over 8 policy documents. This ensures that documents in training and test splits are mutually exclusive. Every question is answered by at least one expert. In addition, in order to estimate annotation reliability and provide for better evaluation, every question in the test set is answered by at least two additional experts.
Table TABREF14 describes the distribution over first words of questions posed by crowdworkers. We also observe low redundancy in the questions posed by crowdworkers over each policy, with each policy receiving ~49.94 unique questions despite crowdworkers independently posing questions. Questions are on average 8.4 words long. As declining to answer a question can be a legally sound response but is seldom practically useful, answers to questions where a minority of experts abstain to answer are filtered from the dataset. Privacy policies are ~3000 words long on average. The answers to the question asked by the users typically have ~100 words of evidence in the privacy policy document.
Data Collection ::: Analysis ::: Categories of Questions
Questions are organized under nine categories from the OPP-115 Corpus annotation scheme BIBREF49:
First Party Collection/Use: What, why and how information is collected by the service provider
Third Party Sharing/Collection: What, why and how information shared with or collected by third parties
Data Security: Protection measures for user information
Data Retention: How long user information will be stored
User Choice/Control: Control options available to users
User Access, Edit and Deletion: If/how users can access, edit or delete information
Policy Change: Informing users if policy information has been changed
International and Specific Audiences: Practices pertaining to a specific group of users
Other: General text, contact information or practices not covered by other categories.
For each question, domain experts indicate one or more relevant OPP-115 categories. We mark a category as relevant to a question if it is identified as such by at least two annotators. If no such category exists, the category is marked as `Other' if atleast one annotator has identified the `Other' category to be relevant. If neither of these conditions is satisfied, we label the question as having no agreement. The distribution of questions in the corpus across OPP-115 categories is as shown in Table.TABREF16. First party and third party related questions are the largest categories, forming nearly 66.4% of all questions asked to the privacy assistant.
Data Collection ::: Analysis ::: Answer Validation
When do experts disagree? We would like to analyze the reasons for potential disagreement on the annotation task, to ensure disagreements arise due to valid differences in opinion rather than lack of adequate specification in annotation guidelines. It is important to note that the annotators are experts rather than crowdworkers. Accordingly, their judgements can be considered valid, legally-informed opinions even when their perspectives differ. For the sake of this question we randomly sample 100 instances in the test data and analyze them for likely reasons for disagreements. We consider a disagreement to have occurred when more than one expert does not agree with the majority consensus. By disagreement we mean there is no overlap between the text identified as relevant by one expert and another.
We find that the annotators agree on the answer for 74% of the questions, even if the supporting evidence they identify is not identical i.e full overlap. They disagree on the remaining 26%. Sources of apparent disagreement correspond to situations when different experts: have differing interpretations of question intent (11%) (for example, when a user asks 'who can contact me through the app', the questions admits multiple interpretations, including seeking information about the features of the app, asking about first party collection/use of data or asking about third party collection/use of data), identify different sources of evidence for questions that ask if a practice is performed or not (4%), have differing interpretations of policy content (3%), identify a partial answer to a question in the privacy policy (2%) (for example, when the user asks `who is allowed to use the app' a majority of our annotators decline to answer, but the remaining annotators highlight partial evidence in the privacy policy which states that children under the age of 13 are not allowed to use the app), and other legitimate sources of disagreement (6%) which include personal subjective views of the annotators (for example, when the user asks `is my DNA information used in any way other than what is specified', some experts consider the boilerplate text of the privacy policy which states that it abides to practices described in the policy document as sufficient evidence to answer this question, whereas others do not).
Experimental Setup
We evaluate the ability of machine learning methods to identify relevant evidence for questions in the privacy domain. We establish baselines for the subtask of deciding on the answerability (§SECREF33) of a question, as well as the overall task of identifying evidence for questions from policies (§SECREF37). We describe aspects of the question that can render it unanswerable within the privacy domain (§SECREF41).
Experimental Setup ::: Answerability Identification Baselines
We define answerability identification as a binary classification task, evaluating model ability to predict if a question can be answered, given a question in isolation. This can serve as a prior for downstream question-answering. We describe three baselines on the answerability task, and find they considerably improve performance over a majority-class baseline.
SVM: We define 3 sets of features to characterize each question. The first is a simple bag-of-words set of features over the question (SVM-BOW), the second is bag-of-words features of the question as well as length of the question in words (SVM-BOW + LEN), and lastly we extract bag-of-words features, length of the question in words as well as part-of-speech tags for the question (SVM-BOW + LEN + POS). This results in vectors of 200, 201 and 228 dimensions respectively, which are provided to an SVM with a linear kernel.
CNN: We utilize a CNN neural encoder for answerability prediction. We use GloVe word embeddings BIBREF50, and a filter size of 5 with 64 filters to encode questions.
BERT: BERT BIBREF51 is a bidirectional transformer-based language-model BIBREF52. We fine-tune BERT-base on our binary answerability identification task with a learning rate of 2e-5 for 3 epochs, with a maximum sequence length of 128.
Experimental Setup ::: Privacy Question Answering
Our goal is to identify evidence within a privacy policy for questions asked by a user. This is framed as an answer sentence selection task, where models identify a set of evidence sentences from all candidate sentences in each policy.
Experimental Setup ::: Privacy Question Answering ::: Evaluation Metric
Our evaluation metric for answer-sentence selection is sentence-level F1, implemented similar to BIBREF30, BIBREF16. Precision and recall are implemented by measuring the overlap between predicted sentences and sets of gold-reference sentences. We report the average of the maximum F1 from each n$-$1 subset, in relation to the heldout reference.
Experimental Setup ::: Privacy Question Answering ::: Baselines
We describe baselines on this task, including a human performance baseline.
No-Answer Baseline (NA) : Most of the questions we receive are difficult to answer in a legally-sound way on the basis of information present in the privacy policy. We establish a simple baseline to quantify the effect of identifying every question as unanswerable.
Word Count Baseline : To quantify the effect of using simple lexical matching to answer the questions, we retrieve the top candidate policy sentences for each question using a word count baseline BIBREF53, which counts the number of question words that also appear in a sentence. We include the top 2, 3 and 5 candidates as baselines.
BERT: We implement two BERT-based baselines BIBREF51 for evidence identification. First, we train BERT on each query-policy sentence pair as a binary classification task to identify if the sentence is evidence for the question or not (Bert). We also experiment with a two-stage classifier, where we separately train the model on questions only to predict answerability. At inference time, if the answerable classifier predicts the question is answerable, the evidence identification classifier produces a set of candidate sentences (Bert + Unanswerable).
Human Performance: We pick each reference answer provided by an annotator, and compute the F1 with respect to the remaining references, as described in section 4.2.1. Each reference answer is treated as the prediction, and the remaining n-1 answers are treated as the gold reference. The average of the maximum F1 across all reference answers is computed as the human baseline.
Results and Discussion
The results of the answerability baselines are presented in Table TABREF31, and on answer sentence selection in Table TABREF32. We observe that bert exhibits the best performance on a binary answerability identification task. However, most baselines considerably exceed the performance of a majority-class baseline. This suggests considerable information in the question, indicating it's possible answerability within this domain.
Table.TABREF32 describes the performance of our baselines on the answer sentence selection task. The No-answer (NA) baseline performs at 28 F1, providing a lower bound on performance at this task. We observe that our best-performing baseline, Bert + Unanswerable achieves an F1 of 39.8. This suggest that bert is capable of making some progress towards answering questions in this difficult domain, while still leaving considerable headroom for improvement to reach human performance. Bert + Unanswerable performance suggests that incorporating information about answerability can help in this difficult domain. We examine this challenging phenomena of unanswerability further in Section .
Results and Discussion ::: Error Analysis
Disagreements are analyzed based on the OPP-115 categories of each question (Table.TABREF34). We compare our best performing BERT variant against the NA model and human performance. We observe significant room for improvement across all categories of questions but especially for first party, third party and data retention categories.
We analyze the performance of our strongest BERT variant, to identify classes of errors and directions for future improvement (Table.8). We observe that a majority of answerability mistakes made by the BERT model are questions which are in fact answerable, but are identified as unanswerable by BERT. We observe that BERT makes 124 such mistakes on the test set. We collect expert judgments on relevance, subjectivity , silence and information about how likely the question is to be answered from the privacy policy from our experts. We find that most of these mistakes are relevant questions. However many of them were identified as subjective by the annotators, and at least one annotator marked 19 of these questions as having no answer within the privacy policy. However, only 6 of these questions were unexpected or do not usually have an answer in privacy policies. These findings suggest that a more nuanced understanding of answerability might help improve model performance in his challenging domain.
Results and Discussion ::: What makes Questions Unanswerable?
We further ask legal experts to identify potential causes of unanswerability of questions. This analysis has considerable implications. While past work BIBREF17 has treated unanswerable questions as homogeneous, a question answering system might wish to have different treatments for different categories of `unanswerable' questions. The following factors were identified to play a role in unanswerability:
Incomprehensibility: If a question is incomprehensible to the extent that its meaning is not intelligible.
Relevance: Is this question in the scope of what could be answered by reading the privacy policy.
Ill-formedness: Is this question ambiguous or vague. An ambiguous statement will typically contain expressions that can refer to multiple potential explanations, whereas a vague statement carries a concept with an unclear or soft definition.
Silence: Other policies answer this type of question but this one does not.
Atypicality: The question is of a nature such that it is unlikely for any policy policy to have an answer to the question.
Our experts attempt to identify the different `unanswerable' factors for all 573 such questions in the corpus. 4.18% of the questions were identified as being incomprehensible (for example, `any difficulties to occupy the privacy assistant'). Amongst the comprehendable questions, 50% were identified as likely to have an answer within the privacy policy, 33.1% were identified as being privacy-related questions but not within the scope of a privacy policy (e.g., 'has Viber had any privacy breaches in the past?') and 16.9% of questions were identified as completely out-of-scope (e.g., `'will the app consume much space?'). In the questions identified as relevant, 32% were ill-formed questions that were phrased by the user in a manner considered vague or ambiguous. Of the questions that were both relevant as well as `well-formed', 95.7% of the questions were not answered by the policy in question but it was reasonable to expect that a privacy policy would contain an answer. The remaining 4.3% were described as reasonable questions, but of a nature generally not discussed in privacy policies. This suggests that the answerability of questions over privacy policies is a complex issue, and future systems should consider each of these factors when serving user's information seeking intent.
We examine a large-scale dataset of “natural” unanswerable questions BIBREF54 based on real user search engine queries to identify if similar unanswerability factors exist. It is important to note that these questions have previously been filtered, according to a criteria for bad questions defined as “(questions that are) ambiguous, incomprehensible, dependent on clear false presuppositions, opinion-seeking, or not clearly a request for factual information.” Annotators made the decision based on the content of the question without viewing the equivalent Wikipedia page. We randomly sample 100 questions from the development set which were identified as unanswerable, and find that 20% of the questions are not questions (e.g., “all I want for christmas is you mariah carey tour”). 12% of questions are unlikely to ever contain an answer on Wikipedia, corresponding closely to our atypicality category. 3% of questions are unlikely to have an answer anywhere (e.g., `what guides Santa home after he has delivered presents?'). 7% of questions are incomplete or open-ended (e.g., `the south west wind blows across nigeria between'). 3% of questions have an unresolvable coreference (e.g., `how do i get to Warsaw Missouri from here'). 4% of questions are vague, and a further 7% have unknown sources of error. 2% still contain false presuppositions (e.g., `what is the only fruit that does not have seeds?') and the remaining 42% do not have an answer within the document. This reinforces our belief that though they have been understudied in past work, any question answering system interacting with real users should expect to receive such unanticipated and unanswerable questions.
Conclusion
We present PrivacyQA, the first significant corpus of privacy policy questions and more than 3500 expert annotations of relevant answers. The goal of this work is to promote question-answering research in the specialized privacy domain, where it can have large real-world impact. Strong neural baselines on PrivacyQA achieve a performance of only 39.8 F1 on this corpus, indicating considerable room for future research. Further, we shed light on several important considerations that affect the answerability of questions. We hope this contribution leads to multidisciplinary efforts to precisely understand user intent and reconcile it with information in policy documents, from both the privacy and NLP communities.
Acknowledgements
This research was supported in part by grants from the National Science Foundation Secure and Trustworthy Computing program (CNS-1330596, CNS-1330214, CNS-15-13957, CNS-1801316, CNS-1914486, CNS-1914444) and a DARPA Brandeis grant on Personalized Privacy Assistants (FA8750-15-2-0277). The US Government is authorized to reproduce and distribute reprints for Governmental purposes not withstanding any copyright notation. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the NSF, DARPA, or the US Government. The authors would like to extend their gratitude to Elias Wright, Gian Mascioli, Kiara Pillay, Harrison Kay, Eliel Talo, Alexander Fagella and N. Cameron Russell for providing their valuable expertise and insight to this effort. The authors are also grateful to Eduard Hovy, Lorrie Cranor, Florian Schaub, Joel Reidenberg, Aditya Potukuchi and Igor Shalyminov for helpful discussions related to this work, and to the three anonymous reviewers of this draft for their constructive feedback. Finally, the authors would like to thank all crowdworkers who consented to participate in this study.
| [
"Individuals with legal training",
"Yes"
] | 3,846 | qasper | en | null | 17a2b96e126ab914e1179b4794d18513627ade82d5c35554 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Who were the experts used for annotation?
Answer: | [
"Individuals with legal training",
"Yes"
] | qasper | 128 | 77 | |
What models are used for painting embedding and what for language style transfer? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Neural networks have been successfully used to describe images with text using sequence-to-sequence models BIBREF0. However, the results are simple and dry captions which are one or two phrases long. Humans looking at a painting see more than just objects. Paintings stimulate sentiments, metaphors and stories as well. Therefore, our goal is to have a neural network describe the painting artistically in a style of choice. As a proof of concept, we present a model which generates Shakespearean prose for a given painting as shown in Figure FIGREF1. Accomplishing this task is difficult with traditional sequence to sequence models since there does not exist a large collection of Shakespearean prose which describes paintings: Shakespeare's works describes a single painting shown in Figure FIGREF3. Fortunately we have a dataset of modern English poems which describe images BIBREF1 and line-by-line modern paraphrases of Shakespeare's plays BIBREF2. Our solution is therefore to combine two separately trained models to synthesize Shakespearean prose for a given painting.
Introduction ::: Related work
A general end-to-end approach to sequence learning BIBREF3 makes minimal assumptions on the sequence structure. This model is widely used in tasks such as machine translation, text summarization, conversational modeling, and image captioning. A generative model using a deep recurrent architecture BIBREF0 has also beeen used for generating phrases describing an image. The task of synthesizing multiple lines of poetry for a given image BIBREF1 is accomplished by extracting poetic clues from images. Given the context image, the network associates image attributes with poetic descriptions using a convolutional neural net. The poem is generated using a recurrent neural net which is trained using multi-adversarial training via policy gradient.
Transforming text from modern English to Shakespearean English using text "style transfer" is challenging. An end to end approach using a sequence-to-sequence model over a parallel text corpus BIBREF2 has been proposed based on machine translation. In the absence of a parallel text corpus, generative adversarial networks (GANs) have been used, which simultaneously train two models: a generative model which captures the data distribution, and a discriminative model which evaluates the performance of the generator. Using a target domain language model as a discriminator has also been employed BIBREF4, providing richer and more stable token-level feedback during the learning process. A key challenge in both image and text style transfer is separating content from style BIBREF5, BIBREF6, BIBREF7. Cross-aligned auto-encoder models have focused on style transfer using non-parallel text BIBREF7. Recently, a fine grained model for text style transfer has been proposed BIBREF8 which controls several factors of variation in textual data by using back-translation. This allows control over multiple attributes, such as gender and sentiment, and fine-grained control over the trade-off between content and style.
Methods
We use a total three datasets: two datasets for generating an English poem from an image, and Shakespeare plays and their English translations for text style transfer.
We train a model for generating poems from images based on two datasets BIBREF1. The first dataset consists of image and poem pairs, namely a multi-modal poem dataset (MultiM-Poem), and the second dataset is a large poem corpus, namely a uni-modal poem dataset (UniM-Poem). The image and poem pairs are extended by adding the nearest three neighbor poems from the poem corpus without redundancy, and an extended image and poem pair dataset is constructed and denoted as MultiM-Poem(Ex)BIBREF1.
We use a collection of line-by-line modern paraphrases for 16 of Shakespeare’s plays BIBREF2, for training a style transfer network from English poems to Shakespearean prose. We use 18,395 sentences from the training data split. We keep 1,218 sentences in the validation data set and 1,462 sentences in our test set.
Methods ::: Image To Poem Actor-Critic Model
For generating a poem from images we use an existing actor-critic architecture BIBREF1. This involves 3 parallel CNNs: an object CNN, sentiment CNN, and scene CNN, for feature extraction. These features are combined with a skip-thought model which provides poetic clues, which are then fed into a sequence-to-sequence model trained by policy gradient with 2 discriminator networks for rewards. This as a whole forms a pipeline that takes in an image and outputs a poem as shown on the top left of Figure FIGREF4. A CNN-RNN generative model acts as an agent. The parameters of this agent define a policy whose execution determines which word is selected as an action. When the agent selects all words in a poem, it receives a reward. Two discriminative networks, shown on the top right of Figure FIGREF4, are defined to serve as rewards concerning whether the generated poem properly describes the input image and whether the generated poem is poetic. The goal of the poem generation model is to generate a sequence of words as a poem for an image to maximize the expected return.
Methods ::: Shakespearizing Poetic Captions
For Shakespearizing modern English texts, we experimented with various types of sequence to sequence models. Since the size of the parallel translation data available is small, we leverage a dictionary providing a mapping between Shakespearean words and modern English words to retrofit pre-trained word embeddings. Incorporating this extra information improves the translation task. The large number of shared word types between the source and target sentences indicates that sharing the representation between them is beneficial.
Methods ::: Shakespearizing Poetic Captions ::: Seq2Seq with Attention
We use a sequence-to-sequence model which consists of a single layer unidrectional LSTM encoder and a single layer LSTM decoder and pre-trained retrofitted word embeddings shared between source and target sentences. We experimented with two different types of attention: global attention BIBREF9, in which the model makes use of the output from the encoder and decoder for the current time step only, and Bahdanau attention BIBREF10, where computing attention requires the output of the decoder from the prior time step. We found that global attention performs better in practice for our task of text style transfer.
Methods ::: Shakespearizing Poetic Captions ::: Seq2Seq with a Pointer Network
Since a pair of corresponding Shakespeare and modern English sentences have significant vocabulary overlap we extend the sequence-to-sequence model mentioned above using pointer networks BIBREF11 that provide location based attention and have been used to enable copying of tokens directly from the input. Moreover, there are lot of proper nouns and rare words which might not be predicted by a vanilla sequence to sequence model.
Methods ::: Shakespearizing Poetic Captions ::: Prediction
For both seq2seq models, we use the attention matrices returned at each decoder time step during inference, to compute the next word in the translated sequence if the decoder output at the current time step is the UNK token. We replace the UNKs in the target output with the highest aligned, maximum attention, source word. The seq2seq model with global attention gives the best results with an average target BLEU score of 29.65 on the style transfer dataset, compared with an average target BLEU score of 26.97 using the seq2seq model with pointer networks.
Results
We perform a qualitative analysis of the Shakespearean prose generated for the input paintings. We conducted a survey, in which we presented famous paintings including those shown in Figures FIGREF1 and FIGREF10 and the corresponding Shakespearean prose generated by the model, and asked 32 students to rate them on the basis of content, creativity and similarity to Shakespearean style on a Likert scale of 1-5. Figure FIGREF12 shows the result of our human evaluation.
The average content score across the paintings is 3.7 which demonstrates that the prose generated is relevant to the painting. The average creativity score is 3.9 which demonstrates that the model captures more than basic objects in the painting successfully using poetic clues in the scene. The average style score is 3.9 which demonstrates that the prose generated is perceived to be in the style of Shakespeare.
We also perform a quantitative analysis of style transfer by generating BLEU scores for the model output using the style transfer dataset. The variation of the BLEU scores with the source sentence lengths is shown in Figure FIGREF11. As expected, the BLEU scores decrease with increase in source sentence lengths.
Results ::: Implementation
All models were trained on Google Colab with a single GPU using Python 3.6 and Tensorflow 2.0. The number of hidden units for the encoder and decoder is 1,576 and 256 for seq2seq with global attention and seq2seq with pointer networks respectively. Adam optimizer was used with the default learning rate of 0.001. The model was trained for 25 epochs. We use pre-trained retrofitted word embeddings of dimension 192.
Results ::: Limitations
Since we do not have an end-to-end dataset, the generated English poem may not work well with Shakespeare style transfer as shown in Figure FIGREF12 for "Starry Night" with a low average content score. This happens when the style transfer dataset does not have similar words in the training set of sentences. A solution would be to expand the style transfer dataset, for a better representation of the poem data.
Results ::: Conclusions and Future Work
In conclusion, combining two pipelines with an intermediate representation works well in practice. We observe that a CNN-RNN based image-to-poem net combined with a seq2seq model with parallel text corpus for text style transfer synthesizes Shakespeare-style prose for a given painting. For the seq2seq model used, we observe that it performs better in practice using global attention as compared with local attention. We make our models and code publicly available BIBREF12. In future work we would like to experiment with GANs in the absence of non-parallel datasets, so that we can use varied styles for text style transfer. We would also like to experiment with cross aligned auto-encoders, which form a latent content representation, to efficiently separate style and content.
| [
"generating a poem from images we use an existing actor-critic architecture, various types of sequence to sequence models"
] | 1,653 | qasper | en | null | 6eec5b0ef38f64f6a545677f730b4ea1c980db6d380de966 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What models are used for painting embedding and what for language style transfer?
Answer: | [
"generating a poem from images we use an existing actor-critic architecture, various types of sequence to sequence models"
] | qasper | 128 | 78 | |
On top of BERT does the RNN layer work better or the transformer layer? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Bidirectional Encoder Representations from Transformers (BERT) is a novel Transformer BIBREF0 model, which recently achieved state-of-the-art performance in several language understanding tasks, such as question answering, natural language inference, semantic similarity, sentiment analysis, and others BIBREF1. While well-suited to dealing with relatively short sequences, Transformers suffer from a major issue that hinders their applicability in classification of long sequences, i.e. they are able to consume only a limited context of symbols as their input BIBREF2.
There are several natural language (NLP) processing tasks that involve such long sequences. Of particular interest are topic identification of spoken conversations BIBREF3, BIBREF4, BIBREF5 and call center customer satisfaction prediction BIBREF6, BIBREF7, BIBREF8, BIBREF9. Call center conversations, while usually quite short and to the point, often involve agents trying to solve very complex issues that the customers experience, resulting in some calls taking even an hour or more. For speech analytics purposes, these calls are typically transcribed using an automatic speech recognition (ASR) system, and processed in textual representations further down the NLP pipeline. These transcripts sometimes exceed the length of 5000 words. Furthermore, temporal information might play an important role in tasks like CSAT. For example, a customer may be angry at the beginning of the call, but after her issue is resolved, she would be very satisfied with the way it was handled. Therefore, simple bag of words models, or any model that does not include temporal dependencies between the inputs, may not be well-suited to handle this category of tasks. This motivates us to employ model such as BERT in this task.
In this paper, we propose a method that builds upon BERT's architecture. We split the input text sequence into shorter segments in order to obtain a representation for each of them using BERT. Then, we use either a recurrent LSTM BIBREF10 network, or another Transformer, to perform the actual classification. We call these techniques Recurrence over BERT (RoBERT) and Transformer over BERT (ToBERT). Given that these models introduce a hierarchy of representations (segment-wise and document-wise), we refer to them as Hierarchical Transformers. To the best of our knowledge, no attempt has been done before to use the Transformer architecture for classification of such long sequences.
Our novel contributions are:
Two extensions - RoBERT and ToBERT - to the BERT model, which enable its application in classification of long texts by performing segmentation and using another layer on top of the segment representations.
State-of-the-art results on the Fisher topic classification task.
Significant improvement on the CSAT prediction task over the MS-CNN model.
Related work
Several dimensionality reduction algorithms such as RBM, autoencoders, subspace multinomial models (SMM) are used to obtain a low dimensional representation of documents from a simple BOW representation and then classify it using a simple linear classifiers BIBREF11, BIBREF12, BIBREF13, BIBREF4. In BIBREF14 hierarchical attention networks are used for document classification. They evaluate their model on several datasets with average number of words around 150. Character-level CNN are explored in BIBREF15 but it is prohibitive for very long documents. In BIBREF16, dataset collected from arXiv papers is used for classification. For classification, they sample random blocks of words and use them together for classification instead of using full document which may work well as arXiv papers are usually coherent and well written on a well defined topic. Their method may not work well on spoken conversations as random block of words usually do not represent topic of full conversation.
Several researchers addressed the problem of predicting customer satisfaction BIBREF6, BIBREF7, BIBREF8, BIBREF9. In most of these works, logistic regression, SVM, CNN are applied on different kinds of representations.
In BIBREF17, authors use BERT for document classification but the average document length is less than BERT maximum length 512. TransformerXL BIBREF2 is an extension to the Transformer architecture that allows it to better deal with long inputs for the language modelling task. It relies on the auto-regressive property of the model, which is not the case in our tasks.
Method ::: BERT
Because our work builds heavily upon BERT, we provide a brief summary of its features. BERT is built upon the Transformer architecture BIBREF0, which uses self-attention, feed-forward layers, residual connections and layer normalization as the main building blocks. It has two pre-training objectives:
Masked language modelling - some of the words in a sentence are being masked and the model has to predict them based on the context (note the difference from the typical autoregressive language model training objective);
Next sentence prediction - given two input sequences, decide whether the second one is the next sentence or not.
BERT has been shown to beat the state-of-the-art performance on 11 tasks with no modifications to the model architecture, besides adding a task-specific output layer BIBREF1. We follow same procedure suggested in BIBREF1 for our tasks. Fig. FIGREF8 shows the BERT model for classification. We obtain two kinds of representation from BERT: pooled output from last transformer block, denoted by H, and posterior probabilities, denoted by P. There are two variants of BERT - BERT-Base and BERT-Large. In this work we are using BERT-Base for faster training and experimentation, however, our methods are applicable to BERT-Large as well. BERT-Base and BERT-Large are different in model parameters such as number of transformer blocks, number of self-attention heads. Total number of parameters in BERT-Base are 110M and 340M in BERT-Large.
BERT suffers from major limitations in terms of handling long sequences. Firstly, the self-attention layer has a quadratic complexity $O(n^2)$ in terms of the sequence length $n$ BIBREF0. Secondly, BERT uses a learned positional embeddings scheme BIBREF1, which means that it won't likely be able to generalize to positions beyond those seen in the training data.
To investigate the effect of fine-tuning BERT on task performance, we use either the pre-trained BERT weights, or the weights from a BERT fine-tuned on the task-specific dataset on a segment-level (i.e. we preserve the original label but fine-tune on each segment separately instead of on the whole text sequence). We compare these results to using the fine-tuned segment-level BERT predictions directly as inputs to the next layer.
Method ::: Recurrence over BERT
Given that BERT is limited to a particular input length, we split the input sequence into segments of a fixed size with overlap. For each of these segments, we obtain H or P from BERT model. We then stack these segment-level representations into a sequence, which serves as input to a small (100-dimensional) LSTM layer. Its output serves as a document embedding. Finally, we use two fully connected layers with ReLU (30-dimensional) and softmax (the same dimensionality as the number of classes) activations to obtain the final predictions.
With this approach, we overcome BERT's computational complexity, reducing it to $O(n/k * k^2) = O(nk)$ for RoBERT, with $k$ denoting the segment size (the LSTM component has negligible linear complexity $O(k)$). The positional embeddings are also no longer an issue.
Method ::: Transformer over BERT
Given that Transformers' edge over recurrent networks is their ability to effectively capture long distance relationships between words in a sequence BIBREF0, we experiment with replacing the LSTM recurrent layer in favor of a small Transformer model (2 layers of transformer building block containing self-attention, fully connected, etc.). To investigate if preserving the information about the input sequence order is important, we also build a variant of ToBERT which learns positional embeddings at the segment-level representations (but is limited to sequences of length seen during the training).
ToBERT's computational complexity $O(\frac{n^2}{k^2})$ is asymptotically inferior to RoBERT, as the top-level Transformer model again suffers from quadratic complexity in the number of segments. However, in practice this number is much smaller than the input sequence length (${\frac{n}{k}} << n$), so we haven't observed performance or memory issues with our datasets.
Experiments
We evaluated our models on 3 different datasets:
CSAT dataset for CSAT prediction, consisting of spoken transcripts (automatic via ASR).
20 newsgroups for topic identification task, consisting of written text;
Fisher Phase 1 corpus for topic identification task, consisting of spoken transcripts (manual);
Experiments ::: CSAT
CSAT dataset consists of US English telephone speech from call centers. For each call in this dataset, customers participated in that call gave a rating on his experience with agent. Originally, this dataset has labels rated on a scale 1-9 with 9 being extremely satisfied and 1 being extremely dissatisfied. Fig. FIGREF16 shows the histogram of ratings for our dataset. As the distribution is skewed towards extremes, we choose to do binary classification with ratings above 4.5 as satisfied and below 4.5 as dissatisfied. Quantization of ratings also helped us to create a balanced dataset. This dataset contains 4331 calls and we split them into 3 sets for our experiments: 2866 calls for training, 362 calls for validation and, finally, 1103 calls for testing.
We obtained the transcripts by employing an ASR system. The ASR system uses TDNN-LSTM acoustic model trained on Fisher and Switchboard datasets with lattice-free maximum mutual information criterion BIBREF18. The word error rates using four-gram language models were 9.2% and 17.3% respectively on Switchboard and CallHome portions of Eval2000 dataset.
Experiments ::: 20 newsgroups
20 newsgroups data set is one of the frequently used datasets in the text processing community for text classification and text clustering. This data set contains approximately 20,000 English documents from 20 topics to be identified, with 11314 documents for training and 7532 for testing. In this work, we used only 90% of documents for training and the remaining 10% for validation. For fair comparison with other publications, we used 53160 words vocabulary set available in the datasets website.
Experiments ::: Fisher
Fisher Phase 1 US English corpus is often used for automatic speech recognition in speech community. In this work, we used it for topic identification as in BIBREF3. The documents are 10-minute long telephone conversations between two people discussing a given topic. We used same training and test splits as BIBREF3 in which 1374 and 1372 documents are used for training and testing respectively. For validation of our model, we used 10% of training dataset and the remaining 90% was used for actual model training. The number of topics in this data set is 40.
Experiments ::: Dataset Statistics
Table TABREF22 shows statistics of our datasets. It can be observed that average length of Fisher is much higher than 20 newsgroups and CSAT. Cumulative distribution of document lengths for each dataset is shown in Fig. FIGREF21. It can be observed that almost all of the documents in Fisher dataset have length more than 1000 words. For CSAT, more than 50% of the documents have length greater than 500 and for 20newsgroups only 10% of the documents have length greater than 500. Note that, for CSAT and 20newsgroups, there are few documents with length more than 5000.
Experiments ::: Architecture and Training Details
In this work, we split document into segments of 200 tokens with a shift of 50 tokens to extract features from BERT model. For RoBERT, LSTM model is trained to minimize cross-entropy loss with Adam optimizer BIBREF19. The initial learning rate is set to $0.001$ and is reduced by a factor of $0.95$ if validation loss does not decrease for 3-epochs. For ToBERT, the Transformer is trained with the default BERT version of Adam optimizer BIBREF1 with an initial learning rate of $5e$-5. We report accuracy in all of our experiments. We chose a model with the best validation accuracy to calculate accuracy on the test set. To accomodate for non-determinism of some TensorFlow GPU operations, we report accuracy averaged over 5 runs.
Results
Table TABREF25 presents results using pre-trained BERT features. We extracted features from the pooled output of final transformer block as these were shown to be working well for most of the tasks BIBREF1. The features extracted from a pre-trained BERT model without any fine-tuning lead to a sub-par performance. However, We also notice that ToBERT model exploited the pre-trained BERT features better than RoBERT. It also converged faster than RoBERT. Table TABREF26 shows results using features extracted after fine-tuning BERT model with our datasets. Significant improvements can be observed compared to using pre-trained BERT features. Also, it can be noticed that ToBERT outperforms RoBERT on Fisher and 20newsgroups dataset by 13.63% and 0.81% respectively. On CSAT, ToBERT performs slightly worse than RoBERT but it is not statistically significant as this dataset is small.
Table TABREF27 presents results using fine-tuned BERT predictions instead of the pooled output from final transformer block. For each document, having obtained segment-wise predictions we can obtain final prediction for the whole document in three ways:
Compute the average of all segment-wise predictions and find the most probable class;
Find the most frequently predicted class;
Train a classification model.
It can be observed from Table TABREF27 that a simple averaging operation or taking most frequent predicted class works competitively for CSAT and 20newsgroups but not for the Fisher dataset. We believe the improvements from using RoBERT or ToBERT, compared to simple averaging or most frequent operations, are proportional to the fraction of long documents in the dataset. CSAT and 20newsgroups have (on average) significantly shorter documents than Fisher, as seen in Fig. FIGREF21. Also, significant improvements for Fisher could be because of less confident predictions from BERT model as this dataset has 40 classes. Fig. FIGREF31 presents the comparison of average voting and ToBERT for various document length ranges for Fisher dataset. We used fine-tuned BERT segment-level predictions (P) for this analysis. It can be observed that ToBERT outperforms average voting in every interval. To the best of our knowledge, this is a state-of-the-art result reported on the Fisher dataset.
Table TABREF32 presents the effect of position embeddings on the model performance. It can be observed that position embeddings did not significantly affect the model performance for Fisher and 20newsgroups, but they helped slightly in CSAT prediction (an absolute improvement of 0.64% F1-score). We think that this is explained by the fact that Fisher and 20newsgroups are topic identification tasks, and the topic does not change much throughout these documents. However, CSAT may vary during the call, and in some cases a naive assumption that the sequential nature of the transcripts is irrelevant may lead to wrong conclusions.
Table TABREF33 compares our results with previous works. It can be seen that our model ToBERT outperforms CNN based experiments by significant margin on CSAT and Fisher datasets. For CSAT dataset, we used multi-scale CNN (MS-CNN) as the baseline, given its strong results on Fisher and 20newsgroups. The setup was replicated from BIBREF5 for comparison. We also see that our result on 20 newsgroups is 0.6% worse than the state-of-the-art.
Conclusions
In this paper, we presented two methods for long documents using BERT model: RoBERT and ToBERT. We evaluated our experiments on two classification tasks - customer satisfaction prediction and topic identification - using 3 datasets: CSAT, 20newsgroups and Fisher. We observed that ToBERT outperforms RoBERT on pre-trained BERT features and fine-tuned BERT features for all our tasks. Also, we noticed that fine-tuned BERT performs better than pre-trained BERT. We have shown that both RoBERT and ToBERT improved the simple baselines of taking an average (or the most frequent) of segment-wise predictions for long documents to obtain final prediction. Position embeddings did not significantly affect our models performance, but slightly improved the accuracy on the CSAT task. We obtained the best results on Fisher dataset and good improvements for CSAT task compared to the CNN baseline. It is interesting to note that the longer the average input in a given task, the bigger improvement we observe w.r.t. the baseline for that task. Our results confirm that both RoBERT and ToBERT can be used for long sequences with competitive performance and quick fine-tuning procedure. For future work, we shall focus on training models on long documents directly (i.e. in an end-to-end manner).
| [
"Transformer over BERT (ToBERT)",
"The transformer layer"
] | 2,655 | qasper | en | null | d696cd0b146419d53d72071d8b8949fc562ea4addb87eb96 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: On top of BERT does the RNN layer work better or the transformer layer?
Answer: | [
"Transformer over BERT (ToBERT)",
"The transformer layer"
] | qasper | 128 | 79 | |
Do the authors hypothesize that humans' robustness to noise is due to their general knowledge? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Machine Reading Comprehension (MRC), as the name suggests, requires a machine to read a passage and answer its relevant questions. Since the answer to each question is supposed to stem from the corresponding passage, a common MRC solution is to develop a neural-network-based MRC model that predicts an answer span (i.e. the answer start position and the answer end position) from the passage of each given passage-question pair. To facilitate the explorations and innovations in this area, many MRC datasets have been established, such as SQuAD BIBREF0 , MS MARCO BIBREF1 , and TriviaQA BIBREF2 . Consequently, many pioneering MRC models have been proposed, such as BiDAF BIBREF3 , R-NET BIBREF4 , and QANet BIBREF5 . According to the leader board of SQuAD, the state-of-the-art MRC models have achieved the same performance as human beings. However, does this imply that they have possessed the same reading comprehension ability as human beings?
OF COURSE NOT. There is a huge gap between MRC models and human beings, which is mainly reflected in the hunger for data and the robustness to noise. On the one hand, developing MRC models requires a large amount of training examples (i.e. the passage-question pairs labeled with answer spans), while human beings can achieve good performance on evaluation examples (i.e. the passage-question pairs to address) without training examples. On the other hand, BIBREF6 revealed that intentionally injected noise (e.g. misleading sentences) in evaluation examples causes the performance of MRC models to drop significantly, while human beings are far less likely to suffer from this. The reason for these phenomena, we believe, is that MRC models can only utilize the knowledge contained in each given passage-question pair, but in addition to this, human beings can also utilize general knowledge. A typical category of general knowledge is inter-word semantic connections. As shown in Table TABREF1 , such general knowledge is essential to the reading comprehension ability of human beings.
A promising strategy to bridge the gap mentioned above is to integrate the neural networks of MRC models with the general knowledge of human beings. To this end, it is necessary to solve two problems: extracting general knowledge from passage-question pairs and utilizing the extracted general knowledge in the prediction of answer spans. The first problem can be solved with knowledge bases, which store general knowledge in structured forms. A broad variety of knowledge bases are available, such as WordNet BIBREF7 storing semantic knowledge, ConceptNet BIBREF8 storing commonsense knowledge, and Freebase BIBREF9 storing factoid knowledge. In this paper, we limit the scope of general knowledge to inter-word semantic connections, and thus use WordNet as our knowledge base. The existing way to solve the second problem is to encode general knowledge in vector space so that the encoding results can be used to enhance the lexical or contextual representations of words BIBREF10 , BIBREF11 . However, this is an implicit way to utilize general knowledge, since in this way we can neither understand nor control the functioning of general knowledge. In this paper, we discard the existing implicit way and instead explore an explicit (i.e. understandable and controllable) way to utilize general knowledge.
The contribution of this paper is two-fold. On the one hand, we propose a data enrichment method, which uses WordNet to extract inter-word semantic connections as general knowledge from each given passage-question pair. On the other hand, we propose an end-to-end MRC model named as Knowledge Aided Reader (KAR), which explicitly uses the above extracted general knowledge to assist its attention mechanisms. Based on the data enrichment method, KAR is comparable in performance with the state-of-the-art MRC models, and significantly more robust to noise than them. When only a subset ( INLINEFORM0 – INLINEFORM1 ) of the training examples are available, KAR outperforms the state-of-the-art MRC models by a large margin, and is still reasonably robust to noise.
Data Enrichment Method
In this section, we elaborate a WordNet-based data enrichment method, which is aimed at extracting inter-word semantic connections from each passage-question pair in our MRC dataset. The extraction is performed in a controllable manner, and the extracted results are provided as general knowledge to our MRC model.
Semantic Relation Chain
WordNet is a lexical database of English, where words are organized into synsets according to their senses. A synset is a set of words expressing the same sense so that a word having multiple senses belongs to multiple synsets, with each synset corresponding to a sense. Synsets are further related to each other through semantic relations. According to the WordNet interface provided by NLTK BIBREF12 , there are totally sixteen types of semantic relations (e.g. hypernyms, hyponyms, holonyms, meronyms, attributes, etc.). Based on synset and semantic relation, we define a new concept: semantic relation chain. A semantic relation chain is a concatenated sequence of semantic relations, which links a synset to another synset. For example, the synset “keratin.n.01” is related to the synset “feather.n.01” through the semantic relation “substance holonym”, the synset “feather.n.01” is related to the synset “bird.n.01” through the semantic relation “part holonym”, and the synset “bird.n.01” is related to the synset “parrot.n.01” through the semantic relation “hyponym”, thus “substance holonym INLINEFORM0 part holonym INLINEFORM1 hyponym” is a semantic relation chain, which links the synset “keratin.n.01” to the synset “parrot.n.01”. We name each semantic relation in a semantic relation chain as a hop, therefore the above semantic relation chain is a 3-hop chain. By the way, each single semantic relation is equivalent to a 1-hop chain.
Inter-word Semantic Connection
The key problem in the data enrichment method is determining whether a word is semantically connected to another word. If so, we say that there exists an inter-word semantic connection between them. To solve this problem, we define another new concept: the extended synsets of a word. Given a word INLINEFORM0 , whose synsets are represented as a set INLINEFORM1 , we use another set INLINEFORM2 to represent its extended synsets, which includes all the synsets that are in INLINEFORM3 or that can be linked to from INLINEFORM4 through semantic relation chains. Theoretically, if there is no limitation on semantic relation chains, INLINEFORM5 will include all the synsets in WordNet, which is meaningless in most situations. Therefore, we use a hyper-parameter INLINEFORM6 to represent the permitted maximum hop count of semantic relation chains. That is to say, only the chains having no more than INLINEFORM7 hops can be used to construct INLINEFORM8 so that INLINEFORM9 becomes a function of INLINEFORM10 : INLINEFORM11 (if INLINEFORM12 , we will have INLINEFORM13 ). Based on the above statements, we formulate a heuristic rule for determining inter-word semantic connections: a word INLINEFORM14 is semantically connected to another word INLINEFORM15 if and only if INLINEFORM16 .
General Knowledge Extraction
Given a passage-question pair, the inter-word semantic connections that connect any word to any passage word are regarded as the general knowledge we need to extract. Considering the requirements of our MRC model, we only extract the positional information of such inter-word semantic connections. Specifically, for each word INLINEFORM0 , we extract a set INLINEFORM1 , which includes the positions of the passage words that INLINEFORM2 is semantically connected to (if INLINEFORM3 itself is a passage word, we will exclude its own position from INLINEFORM4 ). We can control the amount of the extracted results by setting the hyper-parameter INLINEFORM5 : if we set INLINEFORM6 to 0, inter-word semantic connections will only exist between synonyms; if we increase INLINEFORM7 , inter-word semantic connections will exist between more words. That is to say, by increasing INLINEFORM8 within a certain range, we can usually extract more inter-word semantic connections from a passage-question pair, and thus can provide the MRC model with more general knowledge. However, due to the complexity and diversity of natural languages, only a part of the extracted results can serve as useful general knowledge, while the rest of them are useless for the prediction of answer spans, and the proportion of the useless part always rises when INLINEFORM9 is set larger. Therefore we set INLINEFORM10 through cross validation (i.e. according to the performance of the MRC model on the development examples).
Knowledge Aided Reader
In this section, we elaborate our MRC model: Knowledge Aided Reader (KAR). The key components of most existing MRC models are their attention mechanisms BIBREF13 , which are aimed at fusing the associated representations of each given passage-question pair. These attention mechanisms generally fall into two categories: the first one, which we name as mutual attention, is aimed at fusing the question representations into the passage representations so as to obtain the question-aware passage representations; the second one, which we name as self attention, is aimed at fusing the question-aware passage representations into themselves so as to obtain the final passage representations. Although KAR is equipped with both categories, its most remarkable feature is that it explicitly uses the general knowledge extracted by the data enrichment method to assist its attention mechanisms. Therefore we separately name the attention mechanisms of KAR as knowledge aided mutual attention and knowledge aided self attention.
Task Definition
Given a passage INLINEFORM0 and a relevant question INLINEFORM1 , the task is to predict an answer span INLINEFORM2 , where INLINEFORM3 , so that the resulting subsequence INLINEFORM4 from INLINEFORM5 is an answer to INLINEFORM6 .
Overall Architecture
As shown in Figure FIGREF7 , KAR is an end-to-end MRC model consisting of five layers:
Lexicon Embedding Layer. This layer maps the words to the lexicon embeddings. The lexicon embedding of each word is composed of its word embedding and character embedding. For each word, we use the pre-trained GloVe BIBREF14 word vector as its word embedding, and obtain its character embedding with a Convolutional Neural Network (CNN) BIBREF15 . For both the passage and the question, we pass the concatenation of the word embeddings and the character embeddings through a shared dense layer with ReLU activation, whose output dimensionality is INLINEFORM0 . Therefore we obtain the passage lexicon embeddings INLINEFORM1 and the question lexicon embeddings INLINEFORM2 .
Context Embedding Layer. This layer maps the lexicon embeddings to the context embeddings. For both the passage and the question, we process the lexicon embeddings (i.e. INLINEFORM0 for the passage and INLINEFORM1 for the question) with a shared bidirectional LSTM (BiLSTM) BIBREF16 , whose hidden state dimensionality is INLINEFORM2 . By concatenating the forward LSTM outputs and the backward LSTM outputs, we obtain the passage context embeddings INLINEFORM3 and the question context embeddings INLINEFORM4 .
Coarse Memory Layer. This layer maps the context embeddings to the coarse memories. First we use knowledge aided mutual attention (introduced later) to fuse INLINEFORM0 into INLINEFORM1 , the outputs of which are represented as INLINEFORM2 . Then we process INLINEFORM3 with a BiLSTM, whose hidden state dimensionality is INLINEFORM4 . By concatenating the forward LSTM outputs and the backward LSTM outputs, we obtain the coarse memories INLINEFORM5 , which are the question-aware passage representations.
Refined Memory Layer. This layer maps the coarse memories to the refined memories. First we use knowledge aided self attention (introduced later) to fuse INLINEFORM0 into themselves, the outputs of which are represented as INLINEFORM1 . Then we process INLINEFORM2 with a BiLSTM, whose hidden state dimensionality is INLINEFORM3 . By concatenating the forward LSTM outputs and the backward LSTM outputs, we obtain the refined memories INLINEFORM4 , which are the final passage representations.
Answer Span Prediction Layer. This layer predicts the answer start position and the answer end position based on the above layers. First we obtain the answer start position distribution INLINEFORM0 : INLINEFORM1 INLINEFORM2
where INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 are trainable parameters; INLINEFORM3 represents the refined memory of each passage word INLINEFORM4 (i.e. the INLINEFORM5 -th column in INLINEFORM6 ); INLINEFORM7 represents the question summary obtained by performing an attention pooling over INLINEFORM8 . Then we obtain the answer end position distribution INLINEFORM9 : INLINEFORM10 INLINEFORM11
where INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 are trainable parameters; INLINEFORM3 represents vector concatenation. Finally we construct an answer span prediction matrix INLINEFORM4 , where INLINEFORM5 represents the upper triangular matrix of a matrix INLINEFORM6 . Therefore, for the training, we minimize INLINEFORM7 on each training example whose labeled answer span is INLINEFORM8 ; for the inference, we separately take the row index and column index of the maximum element in INLINEFORM9 as INLINEFORM10 and INLINEFORM11 .
Knowledge Aided Mutual Attention
As a part of the coarse memory layer, knowledge aided mutual attention is aimed at fusing the question context embeddings INLINEFORM0 into the passage context embeddings INLINEFORM1 , where the key problem is to calculate the similarity between each passage context embedding INLINEFORM2 (i.e. the INLINEFORM3 -th column in INLINEFORM4 ) and each question context embedding INLINEFORM5 (i.e. the INLINEFORM6 -th column in INLINEFORM7 ). To solve this problem, BIBREF3 proposed a similarity function: INLINEFORM8
where INLINEFORM0 is a trainable parameter; INLINEFORM1 represents element-wise multiplication. This similarity function has also been adopted by several other works BIBREF17 , BIBREF5 . However, since context embeddings contain high-level information, we believe that introducing the pre-extracted general knowledge into the calculation of such similarities will make the results more reasonable. Therefore we modify the above similarity function to the following form: INLINEFORM2
where INLINEFORM0 represents the enhanced context embedding of a word INLINEFORM1 . We use the pre-extracted general knowledge to construct the enhanced context embeddings. Specifically, for each word INLINEFORM2 , whose context embedding is INLINEFORM3 , to construct its enhanced context embedding INLINEFORM4 , first recall that we have extracted a set INLINEFORM5 , which includes the positions of the passage words that INLINEFORM6 is semantically connected to, thus by gathering the columns in INLINEFORM7 whose indexes are given by INLINEFORM8 , we obtain the matching context embeddings INLINEFORM9 . Then by constructing a INLINEFORM10 -attended summary of INLINEFORM11 , we obtain the matching vector INLINEFORM12 (if INLINEFORM13 , which makes INLINEFORM14 , we will set INLINEFORM15 ): INLINEFORM16 INLINEFORM17
where INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 are trainable parameters; INLINEFORM3 represents the INLINEFORM4 -th column in INLINEFORM5 . Finally we pass the concatenation of INLINEFORM6 and INLINEFORM7 through a dense layer with ReLU activation, whose output dimensionality is INLINEFORM8 . Therefore we obtain the enhanced context embedding INLINEFORM9 .
Based on the modified similarity function and the enhanced context embeddings, to perform knowledge aided mutual attention, first we construct a knowledge aided similarity matrix INLINEFORM0 , where each element INLINEFORM1 . Then following BIBREF5 , we construct the passage-attended question summaries INLINEFORM2 and the question-attended passage summaries INLINEFORM3 : INLINEFORM4 INLINEFORM5
where INLINEFORM0 represents softmax along the row dimension and INLINEFORM1 along the column dimension. Finally following BIBREF17 , we pass the concatenation of INLINEFORM2 , INLINEFORM3 , INLINEFORM4 , and INLINEFORM5 through a dense layer with ReLU activation, whose output dimensionality is INLINEFORM6 . Therefore we obtain the outputs INLINEFORM7 .
Knowledge Aided Self Attention
As a part of the refined memory layer, knowledge aided self attention is aimed at fusing the coarse memories INLINEFORM0 into themselves. If we simply follow the self attentions of other works BIBREF4 , BIBREF18 , BIBREF19 , BIBREF17 , then for each passage word INLINEFORM1 , we should fuse its coarse memory INLINEFORM2 (i.e. the INLINEFORM3 -th column in INLINEFORM4 ) with the coarse memories of all the other passage words. However, we believe that this is both unnecessary and distracting, since each passage word has nothing to do with many of the other passage words. Thus we use the pre-extracted general knowledge to guarantee that the fusion of coarse memories for each passage word will only involve a precise subset of the other passage words. Specifically, for each passage word INLINEFORM5 , whose coarse memory is INLINEFORM6 , to perform the fusion of coarse memories, first recall that we have extracted a set INLINEFORM7 , which includes the positions of the other passage words that INLINEFORM8 is semantically connected to, thus by gathering the columns in INLINEFORM9 whose indexes are given by INLINEFORM10 , we obtain the matching coarse memories INLINEFORM11 . Then by constructing a INLINEFORM12 -attended summary of INLINEFORM13 , we obtain the matching vector INLINEFORM14 (if INLINEFORM15 , which makes INLINEFORM16 , we will set INLINEFORM17 ): INLINEFORM18 INLINEFORM19
where INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 are trainable parameters. Finally we pass the concatenation of INLINEFORM3 and INLINEFORM4 through a dense layer with ReLU activation, whose output dimensionality is INLINEFORM5 . Therefore we obtain the fusion result INLINEFORM6 , and further the outputs INLINEFORM7 .
Related Works
Attention Mechanisms. Besides those mentioned above, other interesting attention mechanisms include performing multi-round alignment to avoid the problems of attention redundancy and attention deficiency BIBREF20 , and using mutual attention as a skip-connector to densely connect pairwise layers BIBREF21 .
Data Augmentation. It is proved that properly augmenting training examples can improve the performance of MRC models. For example, BIBREF22 trained a generative model to generate questions based on unlabeled text, which substantially boosted their performance; BIBREF5 trained a back-and-forth translation model to paraphrase training examples, which brought them a significant performance gain.
Multi-step Reasoning. Inspired by the fact that human beings are capable of understanding complex documents by reading them over and over again, multi-step reasoning was proposed to better deal with difficult MRC tasks. For example, BIBREF23 used reinforcement learning to dynamically determine the number of reasoning steps; BIBREF19 fixed the number of reasoning steps, but used stochastic dropout in the output layer to avoid step bias.
Linguistic Embeddings. It is both easy and effective to incorporate linguistic embeddings into the input layer of MRC models. For example, BIBREF24 and BIBREF19 used POS embeddings and NER embeddings to construct their input embeddings; BIBREF25 used structural embeddings based on parsing trees to constructed their input embeddings.
Transfer Learning. Several recent breakthroughs in MRC benefit from feature-based transfer learning BIBREF26 , BIBREF27 and fine-tuning-based transfer learning BIBREF28 , BIBREF29 , which are based on certain word-level or sentence-level models pre-trained on large external corpora in certain supervised or unsupervised manners.
Experimental Settings
MRC Dataset. The MRC dataset used in this paper is SQuAD 1.1, which contains over INLINEFORM0 passage-question pairs and has been randomly partitioned into three parts: a training set ( INLINEFORM1 ), a development set ( INLINEFORM2 ), and a test set ( INLINEFORM3 ). Besides, we also use two of its adversarial sets, namely AddSent and AddOneSent BIBREF6 , to evaluate the robustness to noise of MRC models. The passages in the adversarial sets contain misleading sentences, which are aimed at distracting MRC models. Specifically, each passage in AddSent contains several sentences that are similar to the question but not contradictory to the answer, while each passage in AddOneSent contains a human-approved random sentence that may be unrelated to the passage.
Implementation Details. We tokenize the MRC dataset with spaCy 2.0.13 BIBREF30 , manipulate WordNet 3.0 with NLTK 3.3, and implement KAR with TensorFlow 1.11.0 BIBREF31 . For the data enrichment method, we set the hyper-parameter INLINEFORM0 to 3. For the dense layers and the BiLSTMs, we set the dimensionality unit INLINEFORM1 to 600. For model optimization, we apply the Adam BIBREF32 optimizer with a learning rate of INLINEFORM2 and a mini-batch size of 32. For model evaluation, we use Exact Match (EM) and F1 score as evaluation metrics. To avoid overfitting, we apply dropout BIBREF33 to the dense layers and the BiLSTMs with a dropout rate of INLINEFORM3 . To boost the performance, we apply exponential moving average with a decay rate of INLINEFORM4 .
Model Comparison in both Performance and the Robustness to Noise
We compare KAR with other MRC models in both performance and the robustness to noise. Specifically, we not only evaluate the performance of KAR on the development set and the test set, but also do this on the adversarial sets. As for the comparative objects, we only consider the single MRC models that rank in the top 20 on the SQuAD 1.1 leader board and have reported their performance on the adversarial sets. There are totally five such comparative objects, which can be considered as representatives of the state-of-the-art MRC models. As shown in Table TABREF12 , on the development set and the test set, the performance of KAR is on par with that of the state-of-the-art MRC models; on the adversarial sets, KAR outperforms the state-of-the-art MRC models by a large margin. That is to say, KAR is comparable in performance with the state-of-the-art MRC models, and significantly more robust to noise than them.
To verify the effectiveness of general knowledge, we first study the relationship between the amount of general knowledge and the performance of KAR. As shown in Table TABREF13 , by increasing INLINEFORM0 from 0 to 5 in the data enrichment method, the amount of general knowledge rises monotonically, but the performance of KAR first rises until INLINEFORM1 reaches 3 and then drops down. Then we conduct an ablation study by replacing the knowledge aided attention mechanisms with the mutual attention proposed by BIBREF3 and the self attention proposed by BIBREF4 separately, and find that the F1 score of KAR drops by INLINEFORM2 on the development set, INLINEFORM3 on AddSent, and INLINEFORM4 on AddOneSent. Finally we find that after only one epoch of training, KAR already achieves an EM of INLINEFORM5 and an F1 score of INLINEFORM6 on the development set, which is even better than the final performance of several strong baselines, such as DCN (EM / F1: INLINEFORM7 / INLINEFORM8 ) BIBREF36 and BiDAF (EM / F1: INLINEFORM9 / INLINEFORM10 ) BIBREF3 . The above empirical findings imply that general knowledge indeed plays an effective role in KAR.
To demonstrate the advantage of our explicit way to utilize general knowledge over the existing implicit way, we compare the performance of KAR with that reported by BIBREF10 , which used an encoding-based method to utilize the general knowledge dynamically retrieved from Wikipedia and ConceptNet. Since their best model only achieved an EM of INLINEFORM0 and an F1 score of INLINEFORM1 on the development set, which is much lower than the performance of KAR, we have good reason to believe that our explicit way works better than the existing implicit way.
Model Comparison in the Hunger for Data
We compare KAR with other MRC models in the hunger for data. Specifically, instead of using all the training examples, we produce several training subsets (i.e. subsets of the training examples) so as to study the relationship between the proportion of the available training examples and the performance. We produce each training subset by sampling a specific number of questions from all the questions relevant to each passage. By separately sampling 1, 2, 3, and 4 questions on each passage, we obtain four training subsets, which separately contain INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , and INLINEFORM3 of the training examples. As shown in Figure FIGREF15 , with KAR, SAN (re-implemented), and QANet (re-implemented without data augmentation) trained on these training subsets, we evaluate their performance on the development set, and find that KAR performs much better than SAN and QANet. As shown in Figure FIGREF16 and Figure FIGREF17 , with the above KAR, SAN, and QANet trained on the same training subsets, we also evaluate their performance on the adversarial sets, and still find that KAR performs much better than SAN and QANet. That is to say, when only a subset of the training examples are available, KAR outperforms the state-of-the-art MRC models by a large margin, and is still reasonably robust to noise.
Analysis
According to the experimental results, KAR is not only comparable in performance with the state-of-the-art MRC models, but also superior to them in terms of both the hunger for data and the robustness to noise. The reasons for these achievements, we believe, are as follows:
Conclusion
In this paper, we innovatively integrate the neural networks of MRC models with the general knowledge of human beings. Specifically, inter-word semantic connections are first extracted from each given passage-question pair by a WordNet-based data enrichment method, and then provided as general knowledge to an end-to-end MRC model named as Knowledge Aided Reader (KAR), which explicitly uses the general knowledge to assist its attention mechanisms. Experimental results show that KAR is not only comparable in performance with the state-of-the-art MRC models, but also superior to them in terms of both the hunger for data and the robustness to noise. In the future, we plan to use some larger knowledge bases, such as ConceptNet and Freebase, to improve the quality and scope of the general knowledge.
Acknowledgments
This work is partially supported by a research donation from iFLYTEK Co., Ltd., Hefei, China, and a discovery grant from Natural Sciences and Engineering Research Council (NSERC) of Canada.
| [
"Yes",
"Yes"
] | 4,127 | qasper | en | null | 04174ea6c9849bfa946452cdbc20fc0c4ff9d27596099dcc | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Do the authors hypothesize that humans' robustness to noise is due to their general knowledge?
Answer: | [
"Yes",
"Yes"
] | qasper | 128 | 80 | |
What cyberbulling topics did they address? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Cyberbullying has been defined by the National Crime Prevention Council as the use of the Internet, cell phones or other devices to send or post text or images intended to hurt or embarrass another person. Various studies have estimated that between to 10% to 40% of internet users are victims of cyberbullying BIBREF0 . Effects of cyberbullying can range from temporary anxiety to suicide BIBREF1 . Many high profile incidents have emphasized the prevalence of cyberbullying on social media. Most recently in October 2017, a Swedish model Arvida Byström was cyberbullied to the extent of receiving rape threats after she appeared in an advertisement with hairy legs.
Detection of cyberbullying in social media is a challenging task. Definition of what constitutes cyberbullying is quite subjective. For example, frequent use of swear words might be considered as bullying by the general population. However, for teen oriented social media platforms such as Formspring, this does not necessarily mean bullying (Table TABREF9 ). Across multiple SMPs, cyberbullies attack victims on different topics such as race, religion, and gender. Depending on the topic of cyberbullying, vocabulary and perceived meaning of words vary significantly across SMPs. For example, in our experiments we found that for word `fat', the most similar words as per Twitter dataset are `female' and `woman' (Table TABREF23 ). However, other two datasets do not show such particular bias against women. This platform specific semantic similarity between words is a key aspect of cyberbullying detection across SMPs. Style of communication varies significantly across SMPs. For example, Twitter posts are short and lack anonymity. Whereas posts on Q&A oriented SMPs are long and have option of anonymity (Table TABREF7 ). Fast evolving words and hashtags in social media make it difficult to detect cyberbullying using swear word list based simple filtering approaches. The option of anonymity in certain social networks also makes it harder to identify cyberbullying as profile and history of the bully might not be available.
Past works on cyberbullying detection have at least one of the following three bottlenecks. First (Bottleneck B1), they target only one particular social media platform. How these methods perform across other SMPs is unknown. Second (Bottleneck B2), they address only one topic of cyberbullying such as racism, and sexism. Depending on the topic, vocabulary and nature of cyberbullying changes. These models are not flexible in accommodating changes in the definition of cyberbullying. Third (Bottleneck B3), they rely on carefully handcrafted features such as swear word list and POS tagging. However, these handcrafted features are not robust against variations in writing style. In contrast to existing bottlenecks, this work targets three different types of social networks (Formspring: a Q&A forum, Twitter: microblogging, and Wikipedia: collaborative knowledge repository) for three topics of cyberbullying (personal attack, racism, and sexism) without doing any explicit feature engineering by developing deep learning based models along with transfer learning.
We experimented with diverse traditional machine learning models (logistic regression, support vector machine, random forest, naive Bayes) and deep neural network models (CNN, LSTM, BLSTM, BLSTM with Attention) using variety of representation methods for words (bag of character n-gram, bag of word unigram, GloVe embeddings, SSWE embeddings). Summary of our findings and research contributions is as follows.
Datasets
Please refer to Table TABREF7 for summary of datasets used. We performed experiments using large, diverse, manually annotated, and publicly available datasets for cyberbullying detection in social media. We cover three different types of social networks: teen oriented Q&A forum (Formspring), large microblogging platform (Twitter), and collaborative knowledge repository (Wikipedia talk pages). Each dataset addresses a different topic of cyberbullying. Twitter dataset contains examples of racism and sexism. Wikipedia dataset contains examples of personal attack. However, Formspring dataset is not specifically about any single topic. All three datasets have the problem of class imbalance where posts labeled as cyberbullying are in the minority as compared to neutral posts. Variation in the number of posts across datasets also affects vocabulary size that represents the number of distinct words encountered in the dataset. We measure the size of a post in terms of the number of words in the post. For each dataset, there are only a few posts with large size. We truncate such large posts to the size of post ranked at 95 percentile in that dataset. For example, in Wikipedia dataset, the largest post has 2846 words. However, size of post ranked at 95 percentile in that dataset is only 231. Any post larger than size 231 in Wikipedia dataset will be truncated by considering only first 231 words. This truncation affects only a small minority of posts in each dataset. However, it is required for efficiently training various models in our experiments. Details of each dataset are as follows.
Formspring BIBREF2 : It was a question and answer based website where users could openly invite others to ask and answer questions. The dataset includes 12K annotated question and answer pairs. Each post is manually labeled by three workers. Among these pairs, 825 were labeled as containing cyberbullying content by at least two Amazon Mechanical turk workers.
Twitter BIBREF3 : This dataset includes 16K annotated tweets. The authors bootstrapped the corpus collection, by performing an initial manual search of common slurs and terms used pertaining to religious, sexual, gender, and ethnic minorities. Of the 16K tweets, 3117 are labeled as sexist, 1937 as racist, and the remaining are marked as neither sexist nor racist.
Wikipedia BIBREF4 : For each page in Wikipedia, a corresponding talk page maintains the history of discussion among users who participated in its editing. This data set includes over 100k labeled discussion comments from English Wikipedia's talk pages. Each comment was labeled by 10 annotators via Crowdflower on whether it contains a personal attack. There are total 13590 comments labeled as personal attack.
Use of Swear Words and Anonymity
Please refer to Table TABREF9 . We use the following short forms in this section: B=Bullying, S=Swearing, A=Anonymous. Some of the values for Twitter dataset are undefined as Twitter does not allow anonymous postings. Use of swear words has been repeatedly linked to cyberbullying. However, preliminary analysis of datasets reveals that depending on swear word usage can neither lead to high precision nor high recall for cyberbullying detection. Swear word list based methods will have low precision as P(B INLINEFORM0 S) is not close to 1. In fact, for teen oriented social network Formspring, 78% of the swearing posts are non-bullying. Swear words based filtering will be irritating to the users in such SMPs where swear words are used casually. Swear word list based methods will also have a low recall as P(S INLINEFORM1 B) is not close to 1. For Twitter dataset, 82% of bullying posts do not use any swear words. Such passive-aggressive cyberbullying will go undetected with swear word list based methods. Anonymity is another clue that is used for detecting cyberbullying as bully might prefer to hide its identity. Anonymity definitely leads to increased use of swear words (P(S INLINEFORM2 A) INLINEFORM3 P(S)) and cyberbullying (P(B INLINEFORM4 A) INLINEFORM5 P(B), and P(B INLINEFORM6 A&S)) INLINEFORM7 P(B)). However, significant fraction of anonymous posts are non-bullying (P(B INLINEFORM8 A) not close to 1) and many of bullying posts are not anonymous (P(A INLINEFORM9 B) not close to 1). Further, anonymity might not be allowed by many SMPs such as Twitter.
Related Work
Cyberbullying is recognized as a phenomenon at least since 2003 BIBREF5 . Use of social media exploded with launching of multiple platforms such as Wikipedia (2001), MySpace (2003), Orkut (2004), Facebook (2004), and Twitter (2005). By 2006, researchers had pointed that cyberbullying was as serious phenomenon as offline bullying BIBREF6 . However, automatic detection of cyberbullying was addressed only since 2009 BIBREF7 . As a research topic, cyberbullying detection is a text classification problem. Most of the existing works fit in the following template: get training dataset from single SMP, engineer variety of features with certain style of cyberbullying as the target, apply a few traditional machine learning methods, and evaluate success in terms of measures such as F1 score and accuracy. These works heavily rely on handcrafted features such as use of swear words. These methods tend to have low precision for cyberbullying detection as handcrafted features are not robust against variations in bullying style across SMPs and bullying topics. Only recently, deep learning has been applied for cyberbullying detection BIBREF8 . Table TABREF27 summarizes important related work.
Deep Neural Network (DNN) Based Models
We experimented with four DNN based models for cyberbullying detection: CNN, LSTM, BLSTM, and BLSTM with attention. These models are listed in the increasing complexity of their neural architecture and amount of information used by these models. Please refer to Figure 1 for general architecture that we have used across four models. Various models differ only in the Neural Architecture layer while having identical rest of the layers. CNNs are providing state-of-the-results on extracting contextual feature for classification tasks in images, videos, audios, and text. Recently, CNNs were used for sentiment classification BIBREF9 . Long Short Term Memory networks are a special kind of RNN, capable of learning long-term dependencies. Their ability to use their internal memory to process arbitrary sequences of inputs has been found to be effective for text classification BIBREF10 . Bidirectional LSTMs BIBREF11 further increase the amount of input information available to the network by encoding information in both forward and backward direction. By using two directions, input information from both the past and future of the current time frame can be used. Attention mechanisms allow for a more direct dependence between the state of the model at different points in time. Importantly, attention mechanism lets the model learn what to attend to based on the input sentence and what it has produced so far.
The embedding layer processes a fixed size sequence of words. Each word is represented as a real-valued vector, also known as word embeddings. We have experimented with three methods for initializing word embeddings: random, GloVe BIBREF12 , and SSWE BIBREF13 . During the training, model improves upon the initial word embeddings to learn task specific word embeddings. We have observed that these task specific word embeddings capture the SMP specific and topic specific style of cyberbullying. Using GloVe vectors over random vector initialization has been reported to improve performance for some NLP tasks. Most of the word embedding methods such as GloVe, consider only syntactic context of the word while ignoring the sentiment conveyed by the text. SSWE method overcomes this problem by incorporating the text sentiment as one of the parameters for word embedding generation. We experimented with various dimension size for word embeddings. Experimental results reported here are with dimension size as 50. There was no significant variation in results with dimension size ranging from 30 to 200.
To avoid overfitting, we used two dropout layers, one before the neural architecture layer and one after, with dropout rates of 0.25 and 0.5 respectively. Fully connected layer is a dense output layer with the number of neurons equal to the number of classes, followed by softmax layer that provides softmax activation. All our models are trained using backpropagation. The optimizer used for training is Adam and the loss function is categorical cross-entropy. Besides learning the network weights, these methods also learn task-specific word embeddings tuned towards the bullying labels (See Section SECREF21 ). Our code is available at: https://github.com/sweta20/Detecting-Cyberbullying-Across-SMPs.
Experiments
Existing works have heavily relied on traditional machine learning models for cyberbullying detection. However, they do not study the performance of these models across multiple SMPs. We experimented with four models: logistic regression (LR), support vector machine (SVM), random forest (RF), and naive Bayes (NB), as these are used in previous works (Table TABREF27 ). We used two data representation methods: character n-gram and word unigram. Past work in the domain of detecting abusive language have showed that simple n-gram features are more powerful than linguistic and syntactic features, hand-engineered lexicons, and word and paragraph embeddings BIBREF14 . As compared to DNN models, performance of all four traditional machine learning models was significantly lower. Please refer to Table TABREF11 .
All DNN models reported here were implemented using Keras. We pre-process the data, subjecting it to standard operations of removal of stop words, punctuation marks and lowercasing, before annotating it to assigning respective labels to each comment. For each trained model, we report its performance after doing five-fold cross-validation. We use following short forms.
Effect of Oversampling Bullying Instances
The training datasets had a major problem of class imbalance with posts marked as bullying in the minority. As a result, all models were biased towards labeling the posts as non-bullying. To remove this bias, we oversampled the data from bullying class thrice. That is, we replicated bullying posts thrice in the training data. This significantly improved the performance of all DNN models with major leap in all three evaluation measures. Table TABREF17 shows the effect of oversampling for a variety of word embedding methods with BLSTM Attention as the detection model. Results for other models are similar BIBREF15 . We can notice that oversampled datasets (F+, T+, W+) have far better performance than their counterparts (F, T, W respectively). Oversampling particularly helps the smallest dataset Formspring where number of training instances for bullying class is quite small (825) as compared to other two datasets (about 5K and 13K). We also experimented with varying the replication rate for bullying posts BIBREF15 . However, we observed that for bullying posts, replication rate of three is good enough.
Choice of Initial Word Embeddings and Model
Initial word embeddings decide data representation for DNN models. However during the training, DNN models modify these initial word embeddings to learn task specific word embeddings. We have experimented with three methods to initialize word embeddings. Please refer to Table TABREF19 . This table shows the effect of varying initial word embeddings for multiple DNN models across datasets. We can notice that initial word embeddings do not have a significant effect on cyberbullying detection when oversampling of bullying posts is done (rows corresponding to F+, T+, W+). In the absence of oversampling (rows corresponding to F, T W), there is a gap in performance of simplest (CNN) and most complex (BLSTM with attention) models. However, this gap goes on reducing with the increase in the size of datasets.
Table TABREF20 compares the performance of four DNN models for three evaluation measures while using SSWE as the initial word embeddings. We have noticed that most of the time LSTM performs weaker than other three models. However, performance gap in the other three models is not significant.
Task Specific Word Embeddings
DNN models learn word embeddings over the training data. These learned embeddings across multiple datasets show the difference in nature and style of bullying across cyberbullying topics and SMPs. Here we report results for BLSTM with attention model. Results for other models are similar. We first verify that important words for each topic of cyberbullying form clusters in the learned embeddings. To enable the visualization of grouping, we reduced dimensionality with t-SNE BIBREF16 , a well-known technique for dimensionality reduction particularly well suited for visualization of high dimensional datasets. Please refer to Table TABREF22 . This table shows important clusters observed in t-SNE projection of learned word embeddings. Each cluster shows that words most relevant to a particular topic of bullying form cluster.
We also observed changes in the meanings of the words across topics of cyberbullying. Table TABREF23 shows most similar words for a given query word for two datasets. Twitter dataset which is heavy on sexism and racism, considers word slave as similar to targets of racism and sexism. However, Wikipedia dataset that is about personal attacks does not show such bias.
Transfer Learning
We used transfer learning to check if the knowledge gained by DNN models on one dataset can be used to improve cyberbullying detection performance on other datasets. We report results where BLSTM with attention is used as the DNN model. Results for other models are similar BIBREF15 . We experimented with following three flavors of transfer learning.
Complete Transfer Learning (TL1): In this flavor, a model trained on one dataset was directly used to detect cyberbullying in other datasets without any extra training. TL1 resulted in significantly low recall indicating that three datasets have different nature of cyberbullying with low overlap (Table TABREF25 ). However precision was relatively higher for TL1, indicating that DNN models are cautious in labeling a post as bully (Table TABREF25 ). TL1 also helps to measure similarity in nature of cyberbullying across three datasets. We can observe that bullying nature in Formspring and Wikipedia datasets is more similar to each other than the Twitter dataset. This can be inferred from the fact that with TL1, cyberbullying detection performance for Formspring dataset is higher when base model is Wikipedia (precision =0.51 and recall=0.66)as compared to Twitter as the base model (precision=0.38 and recall=0.04). Similarly, for Wikipedia dataset, Formspring acts as a better base model than Twitter while using TL1 flavor of transfer learning. Nature of SMP might be a factor behind this similarity in nature of cyberbullying. Both Formspring and Wikipedia are task oriented social networks (Q&A and collaborative knowledge repository respectively) that allow anonymity and larger posts. Whereas communication on Twitter is short, free of anonymity and not oriented towards a particular task.
Feature Level Transfer Learning (TL2): In this flavor, a model was trained on one dataset and only learned word embeddings were transferred to another dataset for training a new model. As compared to TL1, recall score improved dramatically with TL2 (Table TABREF25 ). Improvement in precision was also significant (Table TABREF25 ). These improvements indicate that learned word embeddings are an essential part of knowledge transfer across datasets for cyberbullying detection.
Model Level Transfer Learning (TL3): In this flavor, a model was trained on one dataset and learned word embeddings, as well as network weights, were transferred to another dataset for training a new model. TL3 does not result in any significant improvement over TL2. This lack of improvement indicates that transfer of network weights is not essential for cyberbullying detection and learned word embeddings is the key knowledge gained by the DNN models.
DNN based models coupled with transfer learning beat the best-known results for all three datasets. Previous best F1 scores for Wikipedia BIBREF4 and Twitter BIBREF8 datasets were 0.68 and 0.93 respectively. We achieve F1 scores of 0.94 for both these datasets using BLSTM with attention and feature level transfer learning (Table TABREF25 ). For Formspring dataset, authors have not reported F1 score. Their method has accuracy score of 78.5% BIBREF2 . We achieve F1 score of 0.95 with accuracy score of 98% for the same dataset.
Conclusion and Future Work
We have shown that DNN models can be used for cyberbullying detection on various topics across multiple SMPs using three datasets and four DNN models. These models coupled with transfer learning beat state of the art results for all three datasets. These models can be further improved with extra data such as information about the profile and social graph of users. Most of the current datasets do not provide any information about the severity of bullying. If such fine-grained information is made available, then cyberbullying detection models can be further improved to take a variety of actions depending on the perceived seriousness of the posts.
| [
"personal attack, racism, and sexism",
"racism, sexism, personal attack, not specifically about any single topic"
] | 3,244 | qasper | en | null | b805e336d2e8cce895100cfde3a536e632ddd5296ddece21 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What cyberbulling topics did they address?
Answer: | [
"personal attack, racism, and sexism",
"racism, sexism, personal attack, not specifically about any single topic"
] | qasper | 128 | 81 | |
How do they obtain the new context represetation? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Relation classification is the task of assigning sentences with two marked entities to a predefined set of relations. The sentence “We poured the <e1>milk</e1> into the <e2>pumpkin mixture</e2>.”, for example, expresses the relation Entity-Destination(e1,e2). While early research mostly focused on support vector machines or maximum entropy classifiers BIBREF0 , BIBREF1 , recent research showed performance improvements by applying neural networks (NNs) BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 on the benchmark data from SemEval 2010 shared task 8 BIBREF8 .
This study investigates two different types of NNs: recurrent neural networks (RNNs) and convolutional neural networks (CNNs) as well as their combination. We make the following contributions:
(1) We propose extended middle context, a new context representation for CNNs for relation classification. The extended middle context uses all parts of the sentence (the relation arguments, left of the relation arguments, between the arguments, right of the arguments) and pays special attention to the middle part.
(2) We present connectionist bi-directional RNN models which are especially suited for sentence classification tasks since they combine all intermediate hidden layers for their final decision. Furthermore, the ranking loss function is introduced for the RNN model optimization which has not been investigated in the literature for relation classification before.
(3) Finally, we combine CNNs and RNNs using a simple voting scheme and achieve new state-of-the-art results on the SemEval 2010 benchmark dataset.
Related Work
In 2010, manually annotated data for relation classification was released in the context of a SemEval shared task BIBREF8 . Shared task participants used, i.a., support vector machines or maximum entropy classifiers BIBREF0 , BIBREF1 . Recently, their results on this data set were outperformed by applying NNs BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 .
zeng2014 built a CNN based only on the context between the relation arguments and extended it with several lexical features. kim2014 and others used convolutional filters of different sizes for CNNs. nguyen applied this to relation classification and obtained improvements over single filter sizes. deSantos2015 replaced the softmax layer of the CNN with a ranking layer. They showed improvements and published the best result so far on the SemEval dataset, to our knowledge.
socher used another NN architecture for relation classification: recursive neural networks that built recursive sentence representations based on syntactic parsing. In contrast, zhang investigated a temporal structured RNN with only words as input. They used a bi-directional model with a pooling layer on top.
Convolutional Neural Networks (CNN)
CNNs perform a discrete convolution on an input matrix with a set of different filters. For NLP tasks, the input matrix represents a sentence: Each column of the matrix stores the word embedding of the corresponding word. By applying a filter with a width of, e.g., three columns, three neighboring words (trigram) are convolved. Afterwards, the results of the convolution are pooled. Following collobertWeston, we perform max-pooling which extracts the maximum value for each filter and, thus, the most informative n-gram for the following steps. Finally, the resulting values are concatenated and used for classifying the relation expressed in the sentence.
Input: Extended Middle Context
One of our contributions is a new input representation especially designed for relation classification. The contexts are split into three disjoint regions based on the two relation arguments: the left context, the middle context and the right context. Since in most cases the middle context contains the most relevant information for the relation, we want to focus on it but not ignore the other regions completely. Hence, we propose to use two contexts: (1) a combination of the left context, the left entity and the middle context; and (2) a combination of the middle context, the right entity and the right context. Due to the repetition of the middle context, we force the network to pay special attention to it. The two contexts are processed by two independent convolutional and max-pooling layers. After pooling, the results are concatenated to form the sentence representation. Figure FIGREF3 depicts this procedure. It shows an examplary sentence: “He had chest pain and <e1>headaches</e1> from <e2>mold</e2> in the bedroom.” If we only considered the middle context “from”, the network might be tempted to predict a relation like Entity-Origin(e1,e2). However, by also taking the left and right context into account, the model can detect the relation Cause-Effect(e2,e1). While this could also be achieved by integrating the whole context into the model, using the whole context can have disadvantages for longer sentences: The max pooling step can easily choose a value from a part of the sentence which is far away from the mention of the relation. With splitting the context into two parts, we reduce this danger. Repeating the middle context increases the chance for the max pooling step to pick a value from the middle context.
Convolutional Layer
Following previous work (e.g., BIBREF5 , BIBREF6 ), we use 2D filters spanning all embedding dimensions. After convolution, a max pooling operation is applied that stores only the highest activation of each filter. We apply filters with different window sizes 2-5 (multi-windows) as in BIBREF5 , i.e. spanning a different number of input words.
Recurrent Neural Networks (RNN)
Traditional RNNs consist of an input vector, a history vector and an output vector. Based on the representation of the current input word and the previous history vector, a new history is computed. Then, an output is predicted (e.g., using a softmax layer). In contrast to most traditional RNN architectures, we use the RNN for sentence modeling, i.e., we predict an output vector only after processing the whole sentence and not after each word. Training is performed using backpropagation through time BIBREF9 which unfolds the recurrent computations of the history vector for a certain number of time steps. To avoid exploding gradients, we use gradient clipping with a threshold of 10 BIBREF10 .
Input of the RNNs
Initial experiments showed that using trigrams as input instead of single words led to superior results. Hence, at timestep INLINEFORM0 we do not only give word INLINEFORM1 to the model but the trigram INLINEFORM2 by concatenating the corresponding word embeddings.
Connectionist Bi-directional RNNs
Especially for relation classification, the processing of the relation arguments might be easier with knowledge of the succeeding words. Therefore in bi-directional RNNs, not only a history vector of word INLINEFORM0 is regarded but also a future vector. This leads to the following conditioned probability for the history INLINEFORM1 at time step INLINEFORM2 : DISPLAYFORM0
Thus, the network can be split into three parts: a forward pass which processes the original sentence word by word (Equation EQREF6 ); a backward pass which processes the reversed sentence word by word (Equation ); and a combination of both (Equation ). All three parts are trained jointly. This is also depicted in Figure FIGREF7 .
Combining forward and backward pass by adding their hidden layer is similar to BIBREF7 . We, however, also add a connection to the previous combined hidden layer with weight INLINEFORM0 to be able to include all intermediate hidden layers into the final decision of the network (see Equation ). We call this “connectionist bi-directional RNN”.
In our experiments, we compare this RNN with uni-directional RNNs and bi-directional RNNs without additional hidden layer connections.
Word Representations
Words are represented by concatenated vectors: a word embedding and a position feature vector.
Pretrained word embeddings. In this study, we used the word2vec toolkit BIBREF11 to train embeddings on an English Wikipedia from May 2014. We only considered words appearing more than 100 times and added a special PADDING token for convolution. This results in an embedding training text of about 485,000 terms and INLINEFORM0 tokens. During model training, the embeddings are updated.
Position features. We incorporate randomly initialized position embeddings similar to zeng2014, nguyen and deSantos2015. In our RNN experiments, we investigate different possibilities of integrating position information: position embeddings, position embeddings with entity presence flags (flags indicating whether the current word is one of the relation arguments), and position indicators BIBREF7 .
Objective Function: Ranking Loss
Ranking. We applied the ranking loss function proposed in deSantos2015 to train our models. It maximizes the distance between the true label INLINEFORM0 and the best competitive label INLINEFORM1 given a data point INLINEFORM2 . The objective function is DISPLAYFORM0
with INLINEFORM0 and INLINEFORM1 being the scores for the classes INLINEFORM2 and INLINEFORM3 respectively. The parameter INLINEFORM4 controls the penalization of the prediction errors and INLINEFORM5 and INLINEFORM6 are margins for the correct and incorrect classes. Following deSantos2015, we set INLINEFORM7 . We do not learn a pattern for the class Other but increase its difference to the best competitive label by using only the second summand in Equation EQREF10 during training.
Experiments and Results
We used the relation classification dataset of the SemEval 2010 task 8 BIBREF8 . It consists of sentences which have been manually labeled with 19 relations (9 directed relations and one artificial class Other). 8,000 sentences have been distributed as training set and 2,717 sentences served as test set. For evaluation, we applied the official scoring script and report the macro F1 score which also served as the official result of the shared task.
RNN and CNN models were implemented with theano BIBREF12 , BIBREF13 . For all our models, we use L2 regularization with a weight of 0.0001. For CNN training, we use mini batches of 25 training examples while we perform stochastic gradient descent for the RNN. The initial learning rates are 0.2 for the CNN and 0.01 for the RNN. We train the models for 10 (CNN) and 50 (RNN) epochs without early stopping. As activation function, we apply tanh for the CNN and capped ReLU for the RNN. For tuning the hyperparameters, we split the training data into two parts: 6.5k (training) and 1.5k (development) sentences. We also tuned the learning rate schedule on dev.
Beside of training single models, we also report ensemble results for which we combined the presented single models with a voting process.
Performance of CNNs
As a baseline system, we implemented a CNN similar to the one described by zeng2014. It consists of a standard convolutional layer with filters with only one window size, followed by a softmax layer. As input it uses the middle context. In contrast to zeng2014, our CNN does not have an additional fully connected hidden layer. Therefore, we increased the number of convolutional filters to 1200 to keep the number of parameters comparable. With this, we obtain a baseline result of 73.0. After including 5 dimensional position features, the performance was improved to 78.6 (comparable to 78.9 as reported by zeng2014 without linguistic features).
In the next step, we investigate how this result changes if we successively add further features to our CNN: multi-windows for convolution (window sizes: 2,3,4,5 and 300 feature maps each), ranking layer instead of softmax and our proposed extended middle context. Table TABREF12 shows the results. Note that all numbers are produced by CNNs with a comparable number of parameters. We also report F1 for increasing the word embedding dimensionality from 50 to 400. The position embedding dimensionality is 5 in combination with 50 dimensional word embeddings and 35 with 400 dimensional word embeddings. Our results show that especially the ranking layer and the embedding size have an important impact on the performance.
Performance of RNNs
As a baseline for the RNN models, we apply a uni-directional RNN which predicts the relation after processing the whole sentence. With this model, we achieve an F1 score of 61.2 on the SemEval test set.
Afterwards, we investigate the impact of different position features on the performance of uni-directional RNNs (position embeddings, position embeddings concatenated with a flag indicating whether the current word is an entity or not, and position indicators BIBREF7 ). The results indicate that position indicators (i.e. artificial words that indicate the entity presence) perform the best on the SemEval data. We achieve an F1 score of 73.4 with them. However, the difference to using position embeddings with entity flags is not statistically significant.
Similar to our CNN experiments, we successively vary the RNN models by using bi-directionality, by adding connections between the hidden layers (“connectionist”), by applying ranking instead of softmax to predict the relation and by increasing the word embedding dimension to 400.
The results in Table TABREF14 show that all of these variations lead to statistically significant improvements. Especially the additional hidden layer connections and the integration of the ranking layer have a large impact on the performance.
Combination of CNNs and RNNs
Finally, we combine our CNN and RNN models using a voting process. For each sentence in the test set, we apply several CNN and RNN models presented in Tables TABREF12 and TABREF14 and predict the class with the most votes. In case of a tie, we pick one of the most frequent classes randomly. The combination achieves an F1 score of 84.9 which is better than the performance of the two NN types alone. It, thus, confirms our assumption that the networks provide complementary information: while the RNN computes a weighted combination of all words in the sentence, the CNN extracts the most informative n-grams for the relation and only considers their resulting activations.
Comparison with State of the Art
Table TABREF16 shows the results of our models ER-CNN (extended ranking CNN) and R-RNN (ranking RNN) in the context of other state-of-the-art models. Our proposed models obtain state-of-the-art results on the SemEval 2010 task 8 data set without making use of any linguistic features.
Conclusion
In this paper, we investigated different features and architectural choices for convolutional and recurrent neural networks for relation classification without using any linguistic features. For convolutional neural networks, we presented a new context representation for relation classification. Furthermore, we introduced connectionist recurrent neural networks for sentence classification tasks and performed the first experiments with ranking recurrent neural networks. Finally, we showed that even a simple combination of convolutional and recurrent neural networks improved results. With our neural models, we achieved new state-of-the-art results on the SemEval 2010 task 8 benchmark data.
Acknowledgments
Heike Adel is a recipient of the Google European Doctoral Fellowship in Natural Language Processing and this research is supported by this fellowship.
This research was also supported by Deutsche Forschungsgemeinschaft: grant SCHU 2246/4-2.
| [
"They use two independent convolutional and max-pooling layers on (1) a combination of the left context, the left entity and the middle context; and (2) a combination of the middle context, the right entity and the right context. They concatenated the two results after pooling to get the new context representation.... | 2,435 | qasper | en | null | 06baf6c4ae0f97a3131aa06bfa2b4878bd99ef96659cbf9c | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How do they obtain the new context represetation?
Answer: | [
"They use two independent convolutional and max-pooling layers on (1) a combination of the left context, the left entity and the middle context; and (2) a combination of the middle context, the right entity and the right context. They concatenated the two results after pooling to get the new context representation.... | qasper | 128 | 82 | |
How many different types of entities exist in the dataset? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Named Entity Recognition (NER) is a foremost NLP task to label each atomic elements of a sentence into specific categories like "PERSON", "LOCATION", "ORGANIZATION" and othersBIBREF0. There has been an extensive NER research on English, German, Dutch and Spanish language BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5, and notable research on low resource South Asian languages like HindiBIBREF6, IndonesianBIBREF7 and other Indian languages (Kannada, Malayalam, Tamil and Telugu)BIBREF8. However, there has been no study on developing neural NER for Nepali language. In this paper, we propose a neural based Nepali NER using latest state-of-the-art architecture based on grapheme-level which doesn't require any hand-crafted features and no data pre-processing.
Recent neural architecture like BIBREF1 is used to relax the need to hand-craft the features and need to use part-of-speech tag to determine the category of the entity. However, this architecture have been studied for languages like English, and German and not been applied to languages like Nepali which is a low resource language i.e limited data set to train the model. Traditional methods like Hidden Markov Model (HMM) with rule based approachesBIBREF9,BIBREF10, and Support Vector Machine (SVM) with manual feature-engineeringBIBREF11 have been applied but they perform poor compared to neural. However, there has been no research in Nepali NER using neural network. Therefore, we created the named entity annotated dataset partly with the help of Dataturk to train a neural model. The texts used for this dataset are collected from various daily news sources from Nepal around the year 2015-2016.
Following are our contributions:
We present a novel Named Entity Recognizer (NER) for Nepali language. To best of our knowledge we are the first to propose neural based Nepali NER.
As there are not good quality dataset to train NER we release a dataset to support future research
We perform empirical evaluation of our model with state-of-the-art models with relative improvement of upto 10%
In this paper, we present works similar to ours in Section SECREF2. We describe our approach and dataset statistics in Section SECREF3 and SECREF4, followed by our experiments, evaluation and discussion in Section SECREF5, SECREF6, and SECREF7. We conclude with our observations in Section SECREF8.
To facilitate further research our code and dataset will be made available at github.com/link-yet-to-be-updated
Related Work
There has been a handful of research on Nepali NER task based on approaches like Support Vector Machine and gazetteer listBIBREF11 and Hidden Markov Model and gazetteer listBIBREF9,BIBREF10.
BIBREF11 uses SVM along with features like first word, word length, digit features and gazetteer (person, organization, location, middle name, verb, designation and others). It uses one vs rest classification model to classify each word into different entity classes. However, it does not the take context word into account while training the model. Similarly, BIBREF9 and BIBREF10 uses Hidden Markov Model with n-gram technique for extracting POS-tags. POS-tags with common noun, proper noun or combination of both are combined together, then uses gazetteer list as look-up table to identify the named entities.
Researchers have shown that the neural networks like CNNBIBREF12, RNNBIBREF13, LSTMBIBREF14, GRUBIBREF15 can capture the semantic knowledge of language better with the help of pre-trained embbeddings like word2vecBIBREF16, gloveBIBREF17 or fasttextBIBREF18.
Similar approaches has been applied to many South Asian languages like HindiBIBREF6, IndonesianBIBREF7, BengaliBIBREF19 and In this paper, we present the neural network architecture for NER task in Nepali language, which doesn't require any manual feature engineering nor any data pre-processing during training. First we are comparing BiLSTMBIBREF14, BiLSTM+CNNBIBREF20, BiLSTM+CRFBIBREF1, BiLSTM+CNN+CRFBIBREF2 models with CNN modelBIBREF0 and Stanford CRF modelBIBREF21. Secondly, we show the comparison between models trained on general word embeddings, word embedding + character-level embedding, word embedding + part-of-speech(POS) one-hot encoding and word embedding + grapheme clustered or sub-word embeddingBIBREF22. The experiments were performed on the dataset that we created and on the dataset received from ILPRL lab. Our extensive study shows that augmenting word embedding with character or grapheme-level representation and POS one-hot encoding vector yields better results compared to using general word embedding alone.
Approach
In this section, we describe our approach in building our model. This model is partly inspired from multiple models BIBREF20,BIBREF1, andBIBREF2
Approach ::: Bidirectional LSTM
We used Bi-directional LSTM to capture the word representation in forward as well as reverse direction of a sentence. Generally, LSTMs take inputs from left (past) of the sentence and computes the hidden state. However, it is proven beneficialBIBREF23 to use bi-directional LSTM, where, hidden states are computed based from right (future) of sentence and both of these hidden states are concatenated to produce the final output as $h_t$=[$\overrightarrow{h_t}$;$\overleftarrow{h_t}$], where $\overrightarrow{h_t}$, $\overleftarrow{h_t}$ = hidden state computed in forward and backward direction respectively.
Approach ::: Features ::: Word embeddings
We have used Word2Vec BIBREF16, GloVe BIBREF17 and FastText BIBREF18 word vectors of 300 dimensions. These vectors were trained on the corpus obtained from Nepali National Corpus. This pre-lemmatized corpus consists of 14 million words from books, web-texts and news papers. This corpus was mixed with the texts from the dataset before training CBOW and skip-gram version of word2vec using gensim libraryBIBREF24. This trained model consists of vectors for 72782 unique words.
Light pre-processing was performed on the corpus before training it. For example, invalid characters or characters other than Devanagari were removed but punctuation and numbers were not removed. We set the window context at 10 and the rare words whose count is below 5 are dropped. These word embeddings were not frozen during the training session because fine-tuning word embedding help achieve better performance compared to frozen oneBIBREF20.
We have used fasttext embeddings in particular because of its sub-word representation ability, which is very useful in highly inflectional language as shown in Table TABREF25. We have trained the word embedding in such a way that the sub-word size remains between 1 and 4. We particularly chose this size because in Nepali language a single letter can also be a word, for example e, t, C, r, l, n, u and a single character (grapheme) or sub-word can be formed after mixture of dependent vowel signs with consonant letters for example, C + O + = CO, here three different consonant letters form a single sub-word.
The two-dimensional visualization of an example word npAl is shown in FIGREF14. Principal Component Analysis (PCA) technique was used to generate this visualization which helps use to analyze the nearest neighbor words of a given sample word. 84 and 104 nearest neighbors were observed using word2vec and fasttext embedding respectively on the same corpus.
Approach ::: Features ::: Character-level embeddings
BIBREF20 and BIBREF2 successfully presented that the character-level embeddings, extracted using CNN, when combined with word embeddings enhances the NER model performance significantly, as it is able to capture morphological features of a word. Figure FIGREF7 shows the grapheme-level CNN used in our model, where inputs to CNN are graphemes. Character-level CNN is also built in similar fashion, except the inputs are characters. Grapheme or Character -level embeddings are randomly initialized from [0,1] with real values with uniform distribution of dimension 30.
Approach ::: Features ::: Grapheme-level embeddings
Grapheme is atomic meaningful unit in writing system of any languages. Since, Nepali language is highly morphologically inflectional, we compared grapheme-level representation with character-level representation to evaluate its effect. For example, in character-level embedding, each character of a word npAl results into n + + p + A + l has its own embedding. However, in grapheme level, a word npAl is clustered into graphemes, resulting into n + pA + l. Here, each grapheme has its own embedding. This grapheme-level embedding results good scores on par with character-level embedding in highly inflectional languages like Nepali, because graphemes also capture syntactic information similar to characters. We created grapheme clusters using uniseg package which is helpful in unicode text segmentations.
Approach ::: Features ::: Part-of-speech (POS) one hot encoding
We created one-hot encoded vector of POS tags and then concatenated with pre-trained word embeddings before passing it to BiLSTM network. A sample of data is shown in figure FIGREF13.
Dataset Statistics ::: OurNepali dataset
Since, we there was no publicly available standard Nepali NER dataset and did not receive any dataset from the previous researchers, we had to create our own dataset. This dataset contains the sentences collected from daily newspaper of the year 2015-2016. This dataset has three major classes Person (PER), Location (LOC) and Organization (ORG). Pre-processing was performed on the text before creation of the dataset, for example all punctuations and numbers besides ',', '-', '|' and '.' were removed. Currently, the dataset is in standard CoNLL-2003 IO formatBIBREF25.
Since, this dataset is not lemmatized originally, we lemmatized only the post-positions like Ek, kO, l, mA, m, my, jF, sg, aEG which are just the few examples among 299 post positions in Nepali language. We obtained these post-positions from sanjaalcorps and added few more to match our dataset. We will be releasing this list in our github repository. We found out that lemmatizing the post-positions boosted the F1 score by almost 10%.
In order to label our dataset with POS-tags, we first created POS annotated dataset of 6946 sentences and 16225 unique words extracted from POS-tagged Nepali National Corpus and trained a BiLSTM model with 95.14% accuracy which was used to create POS-tags for our dataset.
The dataset released in our github repository contains each word in newline with space separated POS-tags and Entity-tags. The sentences are separated by empty newline. A sample sentence from the dataset is presented in table FIGREF13.
Dataset Statistics ::: ILPRL dataset
After much time, we received the dataset from Bal Krishna Bal, ILPRL, KU. This dataset follows standard CoNLL-2003 IOB formatBIBREF25 with POS tags. This dataset is prepared by ILPRL Lab, KU and KEIV Technologies. Few corrections like correcting the NER tags had to be made on the dataset. The statistics of both the dataset is presented in table TABREF23.
Table TABREF24 presents the total entities (PER, LOC, ORG and MISC) from both of the dataset used in our experiments. The dataset is divided into three parts with 64%, 16% and 20% of the total dataset into training set, development set and test set respectively.
Experiments
In this section, we present the details about training our neural network. The neural network architecture are implemented using PyTorch framework BIBREF26. The training is performed on a single Nvidia Tesla P100 SXM2. We first run our experiment on BiLSTM, BiLSTM-CNN, BiLSTM-CRF BiLSTM-CNN-CRF using the hyper-parameters mentioned in Table TABREF30. The training and evaluation was done on sentence-level. The RNN variants are initialized randomly from $(-\sqrt{k},\sqrt{k})$ where $k=\frac{1}{hidden\_size}$.
First we loaded our dataset and built vocabulary using torchtext library. This eased our process of data loading using its SequenceTaggingDataset class. We trained our model with shuffled training set using Adam optimizer with hyper-parameters mentioned in table TABREF30. All our models were trained on single layer of LSTM network. We found out Adam was giving better performance and faster convergence compared to Stochastic Gradient Descent (SGD). We chose those hyper-parameters after many ablation studies. The dropout of 0.5 is applied after LSTM layer.
For CNN, we used 30 different filters of sizes 3, 4 and 5. The embeddings of each character or grapheme involved in a given word, were passed through the pipeline of Convolution, Rectified Linear Unit and Max-Pooling. The resulting vectors were concatenated and applied dropout of 0.5 before passing into linear layer to obtain the embedding size of 30 for the given word. This resulting embedding is concatenated with word embeddings, which is again concatenated with one-hot POS vector.
Experiments ::: Tagging Scheme
Currently, for our experiments we trained our model on IO (Inside, Outside) format for both the dataset, hence the dataset does not contain any B-type annotation unlike in BIO (Beginning, Inside, Outside) scheme.
Experiments ::: Early Stopping
We used simple early stopping technique where if the validation loss does not decrease after 10 epochs, the training was stopped, else the training will run upto 100 epochs. In our experience, training usually stops around 30-50 epochs.
Experiments ::: Hyper-parameters Tuning
We ran our experiment looking for the best hyper-parameters by changing learning rate from (0,1, 0.01, 0.001, 0.0001), weight decay from [$10^{-1}$, $10^{-2}$, $10^{-3}$, $10^{-4}$, $10^{-5}$, $10^{-6}$, $10^{-7}$], batch size from [1, 2, 4, 8, 16, 32, 64, 128], hidden size from [8, 16, 32, 64, 128, 256, 512 1024]. Table TABREF30 shows all other hyper-parameter used in our experiment for both of the dataset.
Experiments ::: Effect of Dropout
Figure FIGREF31 shows how we end up choosing 0.5 as dropout rate. When the dropout layer was not used, the F1 score are at the lowest. As, we slowly increase the dropout rate, the F1 score also gradually increases, however after dropout rate = 0.5, the F1 score starts falling down. Therefore, we have chosen 0.5 as dropout rate for all other experiments performed.
Evaluation
In this section, we present the details regarding evaluation and comparison of our models with other baselines.
Table TABREF25 shows the study of various embeddings and comparison among each other in OurNepali dataset. Here, raw dataset represents such dataset where post-positions are not lemmatized. We can observe that pre-trained embeddings significantly improves the score compared to randomly initialized embedding. We can deduce that Skip Gram models perform better compared CBOW models for word2vec and fasttext. Here, fastText_Pretrained represents the embedding readily available in fastText website, while other embeddings are trained on the Nepali National Corpus as mentioned in sub-section SECREF11. From this table TABREF25, we can clearly observe that model using fastText_Skip Gram embeddings outperforms all other models.
Table TABREF35 shows the model architecture comparison between all the models experimented. The features used for Stanford CRF classifier are words, letter n-grams of upto length 6, previous word and next word. This model is trained till the current function value is less than $1\mathrm {e}{-2}$. The hyper-parameters of neural network experiments are set as shown in table TABREF30. Since, word embedding of character-level and grapheme-level is random, their scores are near.
All models are evaluated using CoNLL-2003 evaluation scriptBIBREF25 to calculate entity-wise precision, recall and f1 score.
Discussion
In this paper we present that we can exploit the power of neural network to train the model to perform downstream NLP tasks like Named Entity Recognition even in Nepali language. We showed that the word vectors learned through fasttext skip gram model performs better than other word embedding because of its capability to represent sub-word and this is particularly important to capture morphological structure of words and sentences in highly inflectional like Nepali. This concept can come handy in other Devanagari languages as well because the written scripts have similar syntactical structure.
We also found out that stemming post-positions can help a lot in improving model performance because of inflectional characteristics of Nepali language. So when we separate out its inflections or morphemes, we can minimize the variations of same word which gives its root word a stronger word vector representations compared to its inflected versions.
We can clearly imply from tables TABREF23, TABREF24, and TABREF35 that we need more data to get better results because OurNepali dataset volume is almost ten times bigger compared to ILPRL dataset in terms of entities.
Conclusion and Future work
In this paper, we proposed a novel NER for Nepali language and achieved relative improvement of upto 10% and studies different factors effecting the performance of the NER for Nepali language.
We also present a neural architecture BiLSTM+CNN(grapheme-level) which turns out to be performing on par with BiLSTM+CNN(character-level) under the same configuration. We believe this will not only help Nepali language but also other languages falling under the umbrellas of Devanagari languages. Our model BiLSTM+CNN(grapheme-level) and BiLSTM+CNN(G)+POS outperforms all other model experimented in OurNepali and ILPRL dataset respectively.
Since this is the first named entity recognition research in Nepal language using neural network, there are many rooms for improvement. We believe initializing the grapheme-level embedding with fasttext embeddings might help boosting the performance, rather than randomly initializing it. In future, we plan to apply other latest techniques like BERT, ELMo and FLAIR to study its effect on low-resource language like Nepali. We also plan to improve the model using cross-lingual or multi-lingual parameter sharing techniques by jointly training with other Devanagari languages like Hindi and Bengali.
Finally, we would like to contribute our dataset to Nepali NLP community to move forward the research going on in language understanding domain. We believe there should be special committee to create and maintain such dataset for Nepali NLP and organize various competitions which would elevate the NLP research in Nepal.
Some of the future works are listed below:
Proper initialization of grapheme level embedding from fasttext embeddings.
Apply robust POS-tagger for Nepali dataset
Lemmatize the OurNepali dataset with robust and efficient lemmatizer
Improve Nepali language score with cross-lingual learning techniques
Create more dataset using Wikipedia/Wikidata framework
Acknowledgments
The authors of this paper would like to express sincere thanks to Bal Krishna Bal, Kathmandu University Professor for providing us the POS-tagged Nepali NER data.
| [
"OurNepali contains 3 different types of entities, ILPRL contains 4 different types of entities",
"three"
] | 2,851 | qasper | en | null | ac6538e9c173ba7b453d6bf62480d56eb3761b5f6f73c328 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How many different types of entities exist in the dataset?
Answer: | [
"OurNepali contains 3 different types of entities, ILPRL contains 4 different types of entities",
"three"
] | qasper | 128 | 83 | |
How much higher quality is the resulting annotated data? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Assembling training corpora of annotated natural language examples in specialized domains such as biomedicine poses considerable challenges. Experts with the requisite domain knowledge to perform high-quality annotation tend to be expensive, while lay annotators may not have the necessary knowledge to provide high-quality annotations. A practical approach for collecting a sufficiently large corpus would be to use crowdsourcing platforms like Amazon Mechanical Turk (MTurk). However, crowd workers in general are likely to provide noisy annotations BIBREF0 , BIBREF1 , BIBREF2 , an issue exacerbated by the technical nature of specialized content. Some of this noise may reflect worker quality and can be modeled BIBREF0 , BIBREF1 , BIBREF3 , BIBREF4 , but for some instances lay people may simply lack the domain knowledge to provide useful annotation.
In this paper we report experiments on the EBM-NLP corpus comprising crowdsourced annotations of medical literature BIBREF5 . We operationalize the concept of annotation difficulty and show how it can be exploited during training to improve information extraction models. We then obtain expert annotations for the abstracts predicted to be most difficult, as well as for a similar number of randomly selected abstracts. The annotation of highly specialized data and the use of lay and expert annotators allow us to examine the following key questions related to lay and expert annotations in specialized domains:
Can we predict item difficulty? We define a training instance as difficult if a lay annotator or an automated model disagree on its labeling. We show that difficulty can be predicted, and that it is distinct from inter-annotator agreement. Further, such predictions can be used during training to improve information extraction models.
Are there systematic differences between expert and lay annotations? We observe decidedly lower agreement between lay workers as compared to domain experts. Lay annotations have high precision but low recall with respect to expert annotations in the new data that we collected. More generally, we expect lay annotations to be lower quality, which may translate to lower precision, recall, or both, compared to expert annotations. Can one rely solely on lay annotations? Reasonable models can be trained using lay annotations alone, but similar performance can be achieved using markedly less expert data. This suggests that the optimal ratio of expert to crowd annotations for specialized tasks will depend on the cost and availability of domain experts. Expert annotations are preferable whenever its collection is practical. But in real-world settings, a combination of expert and lay annotations is better than using lay data alone.
Does it matter what data is annotated by experts? We demonstrate that a system trained on combined data achieves better predictive performance when experts annotate difficult examples rather than instances selected at i.i.d. random.
Our contributions in this work are summarized as follows. We define a task difficulty prediction task and show how this is related to, but distinct from, inter-worker agreement. We introduce a new model for difficulty prediction combining learned representations induced via a pre-trained `universal' sentence encoder BIBREF6 , and a sentence encoder learned from scratch for this task. We show that predicting annotation difficulty can be used to improve the task routing and model performance for a biomedical information extraction task. Our results open up a new direction for ensuring corpus quality. We believe that item difficulty prediction will likely be useful in other, non-specialized tasks as well, and that the most effective data collection in specialized domains requires research addressing the fundamental questions we examine here.
Related Work
Crowdsourcing annotation is now a well-studied problem BIBREF7 , BIBREF0 , BIBREF1 , BIBREF2 . Due to the noise inherent in such annotations, there have also been considerable efforts to develop aggregation models that minimize noise BIBREF0 , BIBREF1 , BIBREF3 , BIBREF4 .
There are also several surveys of crowdsourcing in biomedicine specifically BIBREF8 , BIBREF9 , BIBREF10 . Some work in this space has contrasted model performance achieved using expert vs. crowd annotated training data BIBREF11 , BIBREF12 , BIBREF13 . Dumitrache et al. Dumitrache:2018:CGT:3232718.3152889 concluded that performance is similar under these supervision types, finding no clear advantage from using expert annotators. This differs from our findings, perhaps owing to differences in design. The experts we used already hold advanced medical degrees, for instance, while those in prior work were medical students. Furthermore, the task considered here would appear to be of greater difficulty: even a system trained on $\sim $ 5k instances performs reasonably, but far from perfect. By contrast, in some of the prior work where experts and crowd annotations were deemed equivalent, a classifier trained on 300 examples can achieve very high accuracy BIBREF12 .
More relevant to this paper, prior work has investigated methods for `task routing' in active learning scenarios in which supervision is provided by heterogeneous labelers with varying levels of expertise BIBREF14 , BIBREF15 , BIBREF16 , BIBREF17 , BIBREF14 . The related question of whether effort is better spent collecting additional annotations for already labeled (but potentially noisily so) examples or novel instances has also been addressed BIBREF18 . What distinguishes the work here is our focus on providing an operational definition of instance difficulty, showing that this can be predicted, and then using this to inform task routing.
Application Domain
Our specific application concerns annotating abstracts of articles that describe the conduct and results of randomized controlled trials (RCTs). Experimentation in this domain has become easy with the recent release of the EBM-NLP BIBREF5 corpus, which includes a reasonably large training dataset annotated via crowdsourcing, and a modest test set labeled by individuals with advanced medical training. More specifically, the training set comprises 4,741 medical article abstracts with crowdsourced annotations indicating snippets (sequences) that describe the Participants (p), Interventions (i), and Outcome (o) elements of the respective RCT, and the test set is composed of 191 abstracts with p, i, o sequence annotations from three medical experts.
Table 1 shows an example of difficult and easy examples according to our definition of difficulty. The underlined text demarcates the (consensus) reference label provided by domain experts. In the difficult examples, crowd workers marked text distinct from these reference annotations; whereas in the easy cases they reproduced them with reasonable fidelity. The difficult sentences usually exhibit complicated structure and feature jargon.
An abstract may contain some `easy' and some `difficult' sentences. We thus perform our analysis at the sentence level. We split abstracts into sentences using spaCy. We excluded sentences that comprise fewer than two tokens, as these are likely an artifact of errors in sentence splitting. In total, this resulted in 57,505 and 2,428 sentences in the train and test set abstracts, respectively.
Quantifying Task Difficulty
The test set includes annotations from both crowd workers and domain experts. We treat the latter as ground truth and then define the difficulty of sentences in terms of the observed agreement between expert and lay annotators. Formally, for annotation task $t$ and instance $i$ :
$$\text{Difficulty}_{ti} = \frac{\sum _{j=1}^n{f(\text{label}_{ij}, y_i})}{n}$$ (Eq. 3)
where $f$ is a scoring function that measures the quality of the label from worker $j$ for sentence $i$ , as compared to a ground truth annotation, $y_i$ . The difficulty score of sentence $i$ is taken as an average over the scores for all $n$ layworkers. We use Spearmans' correlation coefficient as a scoring function. Specifically, for each sentence we create two vectors comprising counts of how many times each token was annotated by crowd and expert workers, respectively, and calculate the correlation between these. Sentences with no labels are treated as maximally easy; those with only either crowd worker or expert label(s) are assumed maximally difficult.
The training set contains only crowdsourced annotations. To label the training data, we use a 10-fold validation like setting. We iteratively retrain the LSTM-CRF-Pattern sequence tagger of Patel et al. patel2018syntactic on 9 folds of the training data and use that trained model to predict labels for the 10th. In this way we obtain predictions on the full training set. We then use predicted spans as proxy `ground truth' annotations to calculate the difficulty score of sentences as described above; we normalize these to the [ $0, 1$ ] interval. We validate this approximation by comparing the proxy scores against reference scores over the test set, the Pearson's correlation coefficients are 0.57 for Population, 0.71 for Intervention and 0.68 for Outcome.
There exist many sentences that contain neither manual nor predicted annotations. We treat these as maximally easy sentences (with difficulty scores of 0). Such sentences comprise 51%, 42% and 36% for Population, Interventions and Outcomes data respectively, indicating that it is easier to identify sentences that have no Population spans, but harder to identify sentences that have no Interventions or Outcomes spans. This is intuitive as descriptions of the latter two tend to be more technical and dense with medical jargon.
We show the distribution of the automatically labeled scores for sentences that do contain spans in Figure 1 . The mean of the Population (p) sentence scores is significantly lower than that for other types of sentences (i and o), again indicating that they are easier on average to annotate. This aligns with a previous finding that annotating Interventions and Outcomes is more difficult than annotating Participants BIBREF5 .
Many sentences contain spans tagged by the LSTM-CRF-Pattern model, but missed by all crowd workers, resulting in a maximally difficult score (1). Inspection of such sentences revealed that some are truly difficult examples, but others are tagging model errors. In either case, such sentences have confused workers and/or the model, and so we retain them all as `difficult' sentences.
Content describing the p, i and o, respectively, is quite different. As such, one sentence usually contains (at most) only one of these three content types. We thus treat difficulty prediction for the respective label types as separate tasks.
Difficulty is not Worker Agreement
Our definition of difficulty is derived from agreement between expert and crowd annotations for the test data, and agreement between a predictive model and crowd annotations in the training data. It is reasonable to ask if these measures are related to inter-annotator agreement, a metric often used in language technology research to identify ambiguous or difficult items. Here we explicitly verify that our definition of difficulty only weakly correlates with inter-annotator agreement.
We calculate inter-worker agreement between crowd and expert annotators using Spearman's correlation coefficient. As shown in Table 2 , average agreement between domain experts are considerably higher than agreements between crowd workers for all three label types. This is a clear indication that the crowd annotations are noisier.
Furthermore, we compare the correlation between inter-annotator agreement and difficulty scores in the training data. Given that the majority of sentences do not contain a PICO span, we only include in these calculations those that contain a reference label. Pearson's r are 0.34, 0.30 and 0.31 for p, i and o, respectively, confirming that inter-worker agreement and our proposed difficulty score are quite distinct.
Predicting Annotation Difficulty
We treat difficulty prediction as a regression problem, and propose and evaluate neural model variants for the task. We first train RNN BIBREF19 and CNN BIBREF20 models.
We also use the universal sentence encoder (USE) BIBREF6 to induce sentence representations, and train a model using these as features. Following BIBREF6 , we then experiment with an ensemble model that combines the `universal' and task-specific representations to predict annotation difficulty. We expect these universal embeddings to capture general, high-level semantics, and the task specific representations to capture more granular information. Figure 2 depicts the model architecture. Sentences are fed into both the universal sentence encoder and, separately, a task specific neural encoder, yielding two representations. We concatenate these and pass the combined vector to the regression layer.
Experimental Setup and Results
We trained models for each label type separately. Word embeddings were initialized to 300d GloVe vectors BIBREF21 trained on common crawl data; these are fine-tuned during training. We used the Adam optimizer BIBREF22 with learning rate and decay set to 0.001 and 0.99, respectively. We used batch sizes of 16.
We used the large version of the universal sentence encoder with a transformer BIBREF23 . We did not update the pretrained sentence encoder parameters during training. All hyperparamaters for all models (including hidden layers, hidden sizes, and dropout) were tuned using Vizier BIBREF24 via 10-fold cross validation on the training set maximizing for F1.
As a baseline, we also trained a linear Support-Vector Regression BIBREF25 model on $n$ -gram features ( $n$ ranges from 1 to 3).
Table 3 reports Pearson correlation coefficients between the predictions with each of the neural models and the ground truth difficulty scores. Rows 1-4 correspond to individual models, and row 5 reports the ensemble performance. Columns correspond to label type. Results from all models outperform the baseline SVR model: Pearson's correlation coefficients range from 0.550 to 0.622. The regression correlations are the lowest.
The RNN model realizes the strongest performance among the stand-alone (non-ensemble) models, outperforming variants that exploit CNN and USE representations. Combining the RNN and USE further improves results. We hypothesize that this is due to complementary sentence information encoded in universal representations.
For all models, correlations for Intervention and Outcomes are higher than for Population, which is expected given the difficulty distributions in Figure 1 . In these, the sentences are more uniformly distributed, with a fair number of difficult and easier sentences. By contrast, in Population there are a greater number of easy sentences and considerably fewer difficult sentences, which makes the difficulty ranking task particularly challenging.
Better IE with Difficulty Prediction
We next present experiments in which we attempt to use the predicted difficulty during training to improve models for information extraction of descriptions of Population, Interventions and Outcomes from medical article abstracts. We investigate two uses: (1) simply removing the most difficult sentences from the training set, and, (2) re-weighting the most difficult sentences.
We again use LSTM-CRF-Pattern as the base model and experimenting on the EBM-NLP corpus BIBREF5 . This is trained on either (1) the training set with difficult sentences removed, or (2) the full training set but with instances re-weighted in proportion to their predicted difficulty score. Following BIBREF5 , we use the Adam optimizer with learning rate of 0.001, decay 0.9, batch size 20 and dropout 0.5. We use pretrained 200d GloVe vectors BIBREF21 to initialize word embeddings, and use 100d hidden char representations. Each word is thus represented with 300 dimensions in total. The hidden size is 100 for the LSTM in the character representation component, and 200 for the LSTM in the information extraction component. We train for 15 epochs, saving parameters that achieve the best F1 score on a nested development set.
Removing Difficult Examples
We first evaluate changes in performance induced by training the sequence labeling model using less data by removing difficult sentences prior to training. The hypothesis here is that these difficult instances are likely to introduce more noise than signal. We used a cross-fold approach to predict sentence difficulties, training on 9/10ths of the data and scoring the remaining 1/10th at a time. We then sorted sentences by predicted difficulty scores, and experimented with removing increasing numbers of these (in order of difficulty) prior to training the LSTM-CRF-Pattern model.
Figure 3 shows the results achieved by the LSTM-CRF-Pattern model after discarding increasing amounts of the training data: the $x$ and $y$ axes correspond to the the percentage of data removed and F1 scores, respectively. We contrast removing sentences predicted to be difficult with removing them (a) randomly (i.i.d.), and, (b) in inverse order of predicted inter-annotator agreement. The agreement prediction model is trained exactly the same like difficult prediction model, with simply changing the difficult score to annotation agreement. F1 scores actually improve (marginally) when we remove the most difficult sentences, up until we drop 4% of the data for Population and Interventions, and 6% for Outcomes. Removing training points at i.i.d. random degrades performance, as expected. Removing sentences in order of disagreement seems to have similar effect as removing them by difficulty score when removing small amount of the data, but the F1 scores drop much faster when removing more data. These findings indicate that sentences predicted to be difficult are indeed noisy, to the extent that they do not seem to provide the model useful signal.
Re-weighting by Difficulty
We showed above that removing a small number of the most difficult sentences does not harm, and in fact modestly improves, medical IE model performance. However, using the available data we are unable to test if this will be useful in practice, as we would need additional data to determine how many difficult sentences should be dropped.
We instead explore an alternative, practical means of exploiting difficulty predictions: we re-weight sentences during training inversely to their predicted difficulty. Formally, we weight sentence $i$ with difficulty scores above $\tau $ according to: $1-a\cdot (d_i-\tau )/(1-\tau )$ , where $d_i$ is the difficulty score for sentence $i$ , and $a$ is a parameter codifying the minimum weight value. We set $\tau $ to 0.8 so as to only re-weight sentences with difficulty in the top 20th percentile, and we set $a$ to 0.5. The re-weighting is equivalent to down-sampling the difficult sentences. LSTM-CRF-Pattern is our base model.
Table 4 reports the precision, recall and F1 achieved both with and without sentence re-weighting. Re-weighting improves all metrics modestly but consistently. All F1 differences are statistically significant under a sign test ( $p<0.01$ ). The model with best precision is different for Patient, Intervention and Outcome labels. However re-weighting by difficulty does consistently yield the best recall for all three extraction types, with the most notable improvement for i and o, where recall improved by 10 percentage points. This performance increase translated to improvements in F1 across all types, as compared to the base model and to re-weighting by agreement.
Involving Expert Annotators
The preceding experiments demonstrate that re-weighting difficult sentences annotated by the crowd generally improves the extraction models. Presumably the performance is influenced by the annotation quality.
We now examine the possibility that the higher quality and more consistent annotations of domain experts on the difficult instances will benefit the extraction model. This simulates an annotation strategy in which we route difficult instances to domain experts and easier ones to crowd annotators. We also contrast the value of difficult data to that of an i.i.d. random sample of the same size, both annotated by experts.
Expert annotations of Random and Difficult Instances
We re-annotate by experts a subset of most difficult instances and the same number of random instances. As collecting annotations from experts is slow and expensive, we only re-annotate the difficult instances for the interventions extraction task. We re-annotate the abstracts which cover the sentences with predicted difficulty scores in the top 5 percentile. We rank the abstracts from the training set by the count of difficult sentences, and re-annotate the abstracts that contain the most difficult sentences. Constrained by time and budget, we select only 2000 abstracts for re-annotation; 1000 of these are top-ranked, and 1000 are randomly sampled. This re-annotation cost $3,000. We have released the new annotation data at: https://github.com/bepnye/EBM-NLP.
Following BIBREF5 , we recruited five medical experts via Up-work with advanced medical training and strong technical reading/writing skills. The expert annotator were asked to read the entire abstract and highlight, using the BRAT toolkit BIBREF26 , all spans describing medical Interventions. Each abstract is only annotated by one expert. We examined 30 re-annotated abstracts to ensure the annotation quality before hiring the annotator.
Table 5 presents the results of LSTM-CRF-Pattern model trained on the reannotated difficult subset and the random subset. The first two rows show the results for models trained with expert annotations. The model trained on random data has a slightly better F1 than that trained on the same amount of difficult data. The model trained on random data has higher precision but lower recall.
Rows 3 and 4 list the results for models trained on the same data but with crowd annotation. Models trained with expert-annotated data are clearly superior to those trained with crowd labels with respect to F1, indicating that the experts produced higher quality annotations. For crowdsourced annotations, training the model with data sampled at i.i.d. random achieves 2% higher F1 than when difficult instances are used. When expert annotations are used, this difference is less than 1%. This trend in performance may be explained by differences in annotation quality: the randomly sampled set was more consistently annotated by both experts and crowd because the difficult set is harder. However, in both cases expert annotations are better, with a bigger difference between the expert and crowd models on the difficult set.
The last row is the model trained on all 5k abstracts with crowd annotations. Its F1 score is lower than either expert model trained on only 20% of data, suggesting that expert annotations should be collected whenever possible. Again the crowd model on complete data has higher precision than expert models but its recall is much lower.
Routing To Experts or Crowd
So far a system was trained on one type of data, either labeled by crowd or experts. We now examine the performance of a system trained on data that was routed to either experts or crowd annotators depending on their predicted difficult. Given the results presented so far mixing annotators may be beneficial given their respective trade-offs of precision and recall. We use the annotations from experts for an abstract if it exists otherwise use crowd annotations. The results are presented in Table 6 .
Rows 1 and 2 repeat the performance of the models trained on difficult subset and random subset with expert annotations only respectively. The third row is the model trained by combining difficult and random subsets with expert annotations. There are around 250 abstracts in the overlap of these two sets, so there are total 1.75k abstracts used for training the D+R model. Rows 4 to 6 are the models trained on all 5k abstracts with mixed annotations, where Other means the rest of the abstracts with crowd annotation only.
The results show adding more training data with crowd annotation still improves at least 1 point F1 score in all three extraction tasks. The improvement when the difficult subset with expert annotations is mixed with the remaining crowd annotation is 3.5 F1 score, much larger than when a random set of expert annotations are added. The model trained with re-annotating the difficult subset (D+Other) also outperforms the model with re-annotating the random subset (R+Other) by 2 points in F1. The model trained with re-annotating both of difficult and random subsets (D+R+Other), however, achieves only marginally higher F1 than the model trained with the re-annotated difficult subset (D+Other). In sum, the results clearly indicate that mixing expert and crowd annotations leads to better models than using solely crowd data, and better than using expert data alone. More importantly, there is greater gain in performance when instances are routed according to difficulty, as compared to randomly selecting the data for expert annotators. These findings align with our motivating hypothesis that annotation quality for difficult instances is important for final model performance. They also indicate that mixing annotations from expert and crowd could be an effective way to achieve acceptable model performance given a limited budget.
How Many Expert Annotations?
We established that crowd annotation are still useful in supplementing expert annotations for medical IE. Obtaining expert annotations for the one thousand most difficult instances greatly improved the model performance. However the choice of how many difficult instances to annotate was an uninformed choice. Here we check if less expert data would have yielded similar gains. Future work will need to address how best to choose this parameter for a routing system.
We simulate a routing scenario in which we send consecutive batches of the most difficult examples to the experts for annotation. We track changes in performance as we increase the number of most-difficult-articles sent to domain experts. As shown in Figure 4 , adding expert annotations for difficult articles consistently increases F1 scores. The performance gain is mostly from increased recall; the precision changes only a bit with higher quality annotation. This observation implies that crowd workers often fail to mark target tokens, but do not tend to produce large numbers of false positives. We suspect such failures to identify relevant spans/tokens are due to insufficient domain knowledge possessed by crowd workers.
The F1 score achieved after re-annotating the 600 most-difficult articles reaches 68.1%, which is close to the performance when re-annotating 1000 random articles. This demonstrates the effectiveness of recognizing difficult instances. The trend when we use up all expert data is still upward, so adding even more expert data is likely to further improve performance. Unfortunately we exhausted our budget and were not able to obtain additional expert annotations. It is likely that as the size of the expert annotations increases, the value of crowd annotations will diminish. This investigation is left for future work.
Conclusions
We have introduced the task of predicting annotation difficulty for biomedical information extraction (IE). We trained neural models using different learned representations to score texts in terms of their difficulty. Results from all models were strong with Pearson’s correlation coefficients higher than 0.45 in almost all evaluations, indicating the feasibility of this task. An ensemble model combining universal and task specific feature sentence vectors yielded the best results.
Experiments on biomedical IE tasks show that removing up to $\sim $ 10% of the sentences predicted to be most difficult did not decrease model performance, and that re-weighting sentences inversely to their difficulty score during training improves predictive performance. Simulations in which difficult examples are routed to experts and other instances to crowd annotators yields the best results, outperforming the strategy of randomly selecting data for expert annotation, and substantially improving upon the approach of relying exclusively on crowd annotations. In future work, routing strategies based on instance difficulty could be further investigated for budget-quality trade-off.
Acknowledgements
This work has been partially supported by NSF1748771 grant. Wallace was support in part by NIH/NLM R01LM012086.
| [
"improvement when the difficult subset with expert annotations is mixed with the remaining crowd annotation is 3.5 F1 score, much larger than when a random set of expert annotations are added"
] | 4,399 | qasper | en | null | 8f650cd8021e553a4106d292af9908f29595f88fc39385b2 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How much higher quality is the resulting annotated data?
Answer: | [
"improvement when the difficult subset with expert annotations is mixed with the remaining crowd annotation is 3.5 F1 score, much larger than when a random set of expert annotations are added"
] | qasper | 128 | 84 | |
How big is imbalance in analyzed corpora? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
In recent years, gender has become a hot topic within the political, societal and research spheres. Numerous studies have been conducted in order to evaluate the presence of women in media, often revealing their under-representation, such as the Global Media Monitoring Project BIBREF0. In the French context, the CSA BIBREF1 produces a report on gender representation in media on a yearly basis. The 2017 report shows that women represent 40% of French media speakers, with a significant drop during high-audience hours (6:00-8:00pm) reaching a value of only 29%. Another large scale study confirmed this trend with an automatic analysis of gender in French audiovisuals streams, highlighting a huge variation across type of shows BIBREF2.
Besides the social impact of gender representation, broadcast recordings are also a valuable source of data for the speech processing community. Indeed, automatic speech recognition (ASR) systems require large amount of annotated speech data to be efficiently trained, which leaves us facing the emerging concern about the fact that "AI artifacts tend to reflect the goals, knowledge and experience of their creators" BIBREF3. Since we know that women are under-represented in media and that the AI discipline has retained a male-oriented focus BIBREF4, we can legitimately wonder about the impact of using such data as a training set for ASR technologies. This concern is strengthened by the recent works uncovering gender bias in several natural language processing (NLP) tools such as BIBREF5, BIBREF6, BIBREF7, BIBREF8.
In this paper, we first highlight the importance of TV and radio broadcast as a source of data for ASR, and the potential impact it can have. We then perform a statistical analysis of gender representation in a data set composed of four state-of-the-art corpora of French broadcast, widely used within the speech community. Finally we question the impact of such a representation on the systems developed on this data, through the perspective of an ASR system.
From gender representation in data to gender bias in AI ::: On the importance of data
The ever growing use of machine learning in science has been enabled by several progresses among which the exponential growth of data available. The quality of a system now depends mostly on the quality and quantity of the data it has been trained on. If it does not discard the importance of an appropriate architecture, it reaffirms the fact that rich and large corpora are a valuable resource. Corpora are research contributions which do not only allow to save and observe certain phenomena or validate a hypothesis or model, but are also a mandatory part of the technology development. This trend is notably observable within the NLP field, where industrial technologies, such as Apple, Amazon or Google vocal assistants now reach high performance level partly due to the amount of data possessed by these companies BIBREF9.
Surprisingly, as data is said to be “the new oil", few data sets are available for ASR systems. The best known are corpora like TIMIT BIBREF10, Switchboard BIBREF11 or Fisher BIBREF12 which date back to the early 1990s. The scarceness of available corpora is justified by the fact that gathering and annotating audio data is costly both in terms of money and time. Telephone conversations and broadcast recordings have been the primary source of spontaneous speech used. Out of all the 130 audio resources proposed by LDC to train automatic speech recognition systems in English, approximately 14% of them are based on broadcast news and conversation. For French speech technologies, four corpora containing radio and TV broadcast are the most widely used: ESTER1 BIBREF13, ESTER2 BIBREF14, ETAPE BIBREF15 and REPERE BIBREF16. These four corpora have been built alongside evaluation campaigns and are still, to our knowledge, the largest French ones of their type available to date.
From gender representation in data to gender bias in AI ::: From data to bias
The gender issue has returned to the forefront of the media scene in recent years and with the emergence of AI technologies in our daily lives, gender bias has become a scientific topic that researchers are just beginning to address. Several studies revealed the existence of gender bias in AI technologies such as face recognition (GenderShades BIBREF17), NLP (word embeddings BIBREF5 and semantics BIBREF6) and machine translation (BIBREF18, BIBREF7). The impact of the training data used within these deep-learning algorithms is therefore questioned.
Bias can be found at different levels as pointed out by BIBREF19. BIBREF20 defines bias as a skew that produces a type of harm. She distinguishes two types of harms that are allocation harm and representation harm. The allocation harm occurs when a system is performing better or worse for a certain group while representational harm contributes to the perpetuation of stereotypes. Both types of harm are the results of bias in machine learning that often comes from the data systems are trained on. Disparities in representation in our social structures is captured and reflected by the training data, through statistical patterns. The GenderShades study is a striking example of what data disparity and lack of representation can produce: the authors tested several gender recognition modules used by facial recognition tools and found difference in error-rate as high as 34 percentage points between recognition of white male and black female faces. The scarce presence of women and colored people in training set resulted in bias in performance towards these two categories, with a strong intersectional bias. As written by BIBREF21 "A data set may have many millions of pieces of data, but this does not mean it is random or representative. To make statistical claims about a data set, we need to know where data is coming from; it is similarly important to know and account for the weaknesses in that data." (p.668).
Regarding ASR technology, little work has explored the presence of gender bias within the systems and no consensus has been reached. BIBREF22 found that speech recognizers perform better on female voice on a broadcast news and telephone corpus. They proposed several explanations to this observation, such as the larger presence of non-professional male speech in the broadcast data, implying a less prepared speech for these speakers or a more normative language and standard pronunciation for women linked to the traditional role of women in language acquisition and education. The same trend was observed by BIBREF23. More recently, BIBREF24 discovered a gender bias within YouTube's automatic captioning system but this bias was not observed in a second study evaluating Bing Speech system and YouTube Automatic Captions on a larger data set BIBREF8. However race and dialect bias were found. General American speakers and white speakers had the lowest error rate for both systems. If the better performance on General American speakers could be explained by the fact that they are all voice professionals, producing clear and articulated speech, but no explanation is provided for biases towards non-white speakers.
Gender bias in ASR technology is still an open research question as no clear answer has been reached so far. It seems that many parameters are to take into account to achieve a general agreement. As we established the importance of TV and radio broadcast as a source of data for ASR, and the potential impact it can have, the following content of this paper is structured as this: we first describe statistically the gender representation of a data set composed of four state-of-the-art corpora of French broadcast, widely used within the speech community, introducing the notion of speaker's role to refine our analysis in terms of voice professionalism. We then question the impact of such a representation on a ASR system trained on these data. BIBREF25
Methodology
This section is organized as follows: we first present the data we are working on. In a second time we explain how we proceed to describe the gender representation in our corpus and introduce the notion of speaker's role. The third subsection introduces the ASR system and metrics used to evaluate gender bias in performance.
Methodology ::: Data presentation
Our data consists of two sets used to train and evaluate our automatic speech recognition system. Four major evaluation campaigns have enabled the creation of wide corpora of French broadcast speech: ESTER1 BIBREF13, ESTER2 BIBREF14, ETAPE BIBREF15 and REPERE BIBREF16. These four collections contain radio and/or TV broadcasts aired between 1998 and 2013 which are used by most academic researchers in ASR. Show duration varies between 10min and an hour. As years went by and speech processing research was progressing, the difficulty of the tasks augmented and the content of these evaluation corpora changed. ESTER1 and ESTER2 mainly contain prepared speech such as broadcast news, whereas ETAPE and REPERE consists also of debates and entertainment shows, spontaneous speech introducing more difficulty in its recognition.
Our training set contains 27,085 speech utterances produced by 2,506 speakers, accounting for approximately 100 hours of speech. Our evaluation set contains 74,064 speech utterances produced by 1,268 speakers for a total of 70 hours of speech. Training data by show, medium and speech type is summarized in Table and evaluation data in Table . Evaluation data has a higher variety of shows with both prepared (P) and spontaneous (S) speech type (accented speech from African radio broadcast is also included in the evaluation set).
Methodology ::: Methodology for descriptive analysis of gender representation in training data
We first describe the gender representation in training data. Gender representation is measured in terms of number of speakers, number of utterances (or speech turns), and turn lengths (descriptive statistics are given in Section SECREF16). Each speech turn was mapped to its speaker in order to associate it with a gender.
As pointed out by the CSA report BIBREF1, women presence tends to be marginal within the high-audience hours, showing that women are represented but less than men and within certain given conditions. It is clear that a small number of speakers is responsible for a large number of speech turns. Most of these speakers are journalists, politicians, presenters and such, who are representative of a show. Therefore, we introduce the notion of speaker's role to refine our exploration of gender disparity, following studies which quantified women's presence in terms of role. Within our work, we define the notion of speaker role by two criteria specifying the speaker's on-air presence, namely the number of speech turns and the cumulative duration of his or her speaking time in a show. Based on the available speech transcriptions and meta-data, we compute for each speaker the number of speech turns uttered as well as their total length. We then use the following criteria to define speaker's role: a speaker is considered as speaking often (respectively seldom) if he/she accumulates a total of turns higher (respectively lower) than 1% of the total number of speech turns in a given show. The same process is applied to identify speakers talking for a long period from those who do not. We end up with two salient roles called Anchors and Punctual speakers:
the Anchor speakers (A) are above the threshold of 1% for both criteria, meaning they are intervening often and for a long time thus holding an important place in interaction;
the Punctual speakers (PS) on the contrary are below the threshold of 1% for both the total number of turns and the total speech time.
These roles are defined at the show level. They could be roughly assimilated to the categorization “host/guest” in radio and TV shows. Anchors could be described as professional speakers, producing mostly prepared speech, whereas Punctual speakers are more likely to be “everyday people". The concept of speaker's role makes sense at both sociological and technical levels. An Anchor speaker is more likely to be known from the audience (society), but he or she will also likely have a professional (clear) way of speaking (as mentioned by BIBREF22 and BIBREF8), as well as a high number of utterances, augmenting the amount of data available for a given gender category.
Methodology ::: Gender bias evaluation procedure of an ASR system performance ::: ASR system
The ASR system used in this work is described in BIBREF25. It uses the KALDI toolkit BIBREF26, following a standard Kaldi recipe. The acoustic model is based on a hybrid HMM-DNN architecture and trained on the data summarized in Table . Acoustic training data correspond to 100h of non-spontaneous speech type (mostly broadcast news) coming from both radio and TV shows. A 5-gram language model is trained from several French corpora (3,323M words in total) using SRILM toolkit BIBREF27. The pronunciation model is developed using the lexical resource BDLEX BIBREF28 as well as automatic grapheme-to-phoneme (G2P) transcription to find pronunciation variants of our vocabulary (limited to 80K). It is important to re-specify here, for further analysis, that our Kaldi pipeline follows speaker adaptive training (SAT) where we train and decode using speaker adapted features (fMLLR-adapted features) in per-speaker mode. It is well known that speaker adaptation acts as an effective procedure to reduce mismatch between training and evaluation conditions BIBREF29, BIBREF26.
Methodology ::: Gender bias evaluation procedure of an ASR system performance ::: Evaluation
Word Error Rate (WER) is a common metric to evaluate ASR performance. It is measured as the sum of errors (insertions, deletions and substitutions) divided by the total number of words in the reference transcription. As we are investigating the impact on performance of speaker's gender and role, we computed the WER for each speaker at the episode (show occurrence) level. Analyzing at such granularity allows us to avoid large WER variation that could be observed at utterance level (especially for short speech turns) but also makes possible to get several WER values for a given speaker, one for each occurrence of a show in which he/she appears on. Speaker's gender was provided by the meta-data and role was obtained using the criteria from Section SECREF6 computed for each show. This enables us to analyze our results across gender and role categories which was done using Wilcoxon rank sum tests also called Mann-Whitney U test (with $\alpha $= 0.001) BIBREF30. The choice of a Wilcoxon rank sum test and not the commonly used t-test is motivated by the non-normality of our data.
Results ::: Descriptive analysis of gender representation in training data ::: Gender representation
As expected, we observe a disparity in terms of gender representation in our data (see Table ). Women represent 33.16% of the speakers, confirming the figures given by the GMMP report BIBREF0. However, it is worth noticing that women account for only 22.57% of the total speech time, which leads us to conclude that women also speak less than men.
Results ::: Descriptive analysis of gender representation in training data ::: Speaker's role representation
Table presents roles' representation in training data and shows that despite the small number of Anchor speakers in our data (3.79%), they nevertheless concentrate 35.71 % of the total speech time.
Results ::: Descriptive analysis of gender representation in training data ::: Role and gender interaction
When crossing both parameters, we can observe that the gender distribution is not constant throughout roles. Women represent 29.47% of the speakers within the Anchor category, even less than among the Punctual speakers. Their percentage of speech is also smaller. When calculating the average speech time uttered by a female Anchor, we obtain a value of 15.9 min against 25.2 min for a male Anchor, which suggests that even within the Anchor category men tend to speak more. This confirms the existence of gender disparities within French media. It corroborates with the analysis of the CSA BIBREF1, which shows that women were less present during high-audience hours. Our study shows that they are also less present in important roles. These results legitimate our initial questioning on the impact of gender balance on ASR performance trained on broadcast recordings.
Results ::: Performance (WER) analysis on evaluation data ::: Impact of gender on WER
As explained in Section SECREF13, WER is the sum of errors divided by the number of words in the transcription reference. The higher the WER, the poorer the system performance. Our 70h evaluation data contains a large amount of spontaneous speech and is very challenging for the ASR system trained on prepared speech: we observe an overall average WER of 42.9% for women and 34.3% for men. This difference of WER between men and women is statistically significant (med(M) = 25%; med(F) = 29%; U = 709040; p-value < 0.001).
However, when observing gender differences across shows, no clear trend can be identified, as shown in Figure FIGREF21. For shows like Africa1 Infos or La Place du Village, we find an average WER lower for women than for men, while the trend is reversed for shows such as Un Temps de Pauchon or Le Masque et la Plume. The disparity of the results depending on the show leads us to believe that other factors may be entangled within the observed phenomenon.
Results ::: Performance (WER) analysis on evaluation data ::: Impact of role on WER
Speaker's role seems to have an impact on WER: we obtain an average WER of 30.8% for the Anchor speakers and 42.23% for the Punctual speakers. This difference is statistically significant with a p-value smaller than $10^{-14}$ (med(A) = 21%; med(P) = 31%; U = 540,430; p-value < 0.001) .
Results ::: Performance (WER) analysis on evaluation data ::: Role and gender interaction
Figure FIGREF25 presents the WER distribution (WER being obtained for each speaker in a show occurrence) according to the speaker's role and gender. It is worth noticing that the gender difference is only significant within the Punctual speakers group. The average WER is of 49.04% for the women and 38.56% for the men with a p-value smaller than $10^{-6}$ (med(F) = 39%; med(M) = 29%; U = 251,450; p-value < 0.001), whereas it is just a trend between male and female Anchors (med(F) = 21%; med(M) = 21%; U = 116,230; p-value = 0.173). This could be explained by the quantity of data available per speaker.
Results ::: Performance (WER) analysis on evaluation data ::: Speech type as a third entangled factor?
In order to try to explain the observed variation in our results depending on shows and gender (Figure FIGREF21), we add the notion of speech type to shed some light on our results. BIBREF22 and BIBREF24 suggested that the speaker professionalism, associated with clear and hyper-articulated speech could be an explaining factor for better performance.
Based on our categorization in prepared speech (mostly news reports) and spontaneous speech (mostly debates and entertainment shows), we cross this parameter in our performance analysis. As shown on Figure FIGREF26, these results confirm the inherent challenge of spontaneous speech compared to prepared speech. WER scores are similar between men and women when considering prepared speech (med(F) = 18%; med(M) = 21%; U = 217,160; p-value = 0.005) whereas they are worse for women (61.29%) than for men (46.51%) with p-value smaller than $10^{-14}$ for the spontaneous speech type (med(F) = 61%; med(M) = 37%; U = 153,580; p-value < 0.001).
Discussion
We find a clear disparity in terms of women presence and speech quantity in French media. Our data being recorded between 1998 and 2013, we can expect this disparity to be smaller on more recent broadcast recordings, especially since the French government displays efforts toward parity in media representation. One can also argue that even if our analysis was conducted on a large amount of data it does not reach the exhaustiveness of large-scale studies such as the one of BIBREF2. Nonetheless it does not affect the relevance of our findings, because if real-world gender representation might be more balanced today, these corpora are still used as training data for AI systems.
The performance difference across gender we observed corroborates (on a larger quantity and variety of language data produced by more than 2400 speakers) the results obtained by BIBREF24 on isolated words recognition. However the following study on read speech does not replicate these results. Yet a performance degradation is observed across dialect and race BIBREF8. BIBREF22 found lower WER for women than men on broadcast news and conversational telephone speech for both English and French. The authors suggest that gender stereotypes associated with women role in education and language acquisition induce a more normative elocution. We observed that the higher the degree of normativity of speech the smaller the gender difference. No significant gender bias is observed for prepared speech nor within the Anchor category. Even if we do not find similar results with lower WER for women than men, we obtained a median WER smaller for women on prepared speech and equal to the male median WER for the Anchor speakers.
Another explanation could be the use of adaptation within the pipeline. Most broadcast programs transcription systems have a speaker adaptation step within their decoding pipeline, which is the case for our system. An Anchor speaker intervening more often would have a larger quantity of data to realize such adaptation of the acoustic model. On the contrary, Punctual speakers who appear scarcely in the data are not provided with the same amount of adaptation data. Hence we can hypothesize that gender performance difference observed for Punctual speakers is due to the fact that female speech is further from the (initial non-adapted) acoustic model as it was trained on unbalanced data (as shown in Table ). Considering that Punctual speakers represent 92.78% of the speakers, this explains why gender difference is significant over our entire data set. A way to confirm our hypothesis would be to reproduce our analysis on WER values obtained without using speaker adapted features at the decoding step.
When decoding prepared speech (hence similar to the training data), no significant difference is found in WER between men and women, revealing that the speaker adaptation step could be sufficient to reach same performance for both genders. But when decoding more spontaneous speech, there is a mismatch with the initial acoustic model (trained on prepared speech). Consequently, the speaker adaptation step might not be enough to recover good ASR performance, especially for women for whom less adaptation data is available (see Section 4.2.3).
Conclusion
This paper has investigated gender bias in ASR performance through the following research questions: i) what is the proportion of men and women in French radio and TV media data ? ii) what is the impact of the observed disparity on ASR performance ? iii) is this as simple as a problem of gender proportion in the training data or are other factors entangled ? Our contributions are the following:
Descriptive analysis of the broadcast data used to train our ASR system confirms the already known disparity, where 65% of the speakers are men, speaking more than 75% of the time.
When investigating WER scores according to gender, speaker's role and speech type, huge variations are observed. We conclude that gender is clearly a factor of variation in ASR performance, with a WER increase of 24% for women compared to men, exhibiting a clear gender bias.
Gender bias varies across speaker's role and speech spontaneity level. Performance for Punctual speakers respectively spontaneous speech seems to reinforce this gender bias with a WER increase of 27.2% respectively 31.8% between male and female speakers.
We found that an ASR system trained on unbalanced data regarding gender produces gender bias performance. Therefore, in order to create fair systems it is necessary to take into account the representation problems in society that are going to be encapsulated in the data. Understanding how women under-representation in broadcast data can lead to bias in ASR performances is the key to prevent re-implementing and reinforcing discrimination already existing in our societies. This is in line with the concept of “Fairness by Design" proposed by BIBREF31.
Gender, race, religion, nationality are all characteristics that we deem unfair to classify on, and these ethical standpoints needs to be taken into account in systems' design. Characteristics that are not considered as relevant in a given task can be encapsulated in data nonetheless, and lead to bias performance. Being aware of the demographic skews our data set might contain is a first step to track the life cycle of a training data set and a necessary step to control the tools we develop.
| [
"Women represent 33.16% of the speakers"
] | 4,055 | qasper | en | null | 07484380079dcec47eb573ab62eaff39a1a44b32006ec74e | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How big is imbalance in analyzed corpora?
Answer: | [
"Women represent 33.16% of the speakers"
] | qasper | 128 | 85 | |
What dataset does this approach achieve state of the art results on? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Data
We build and test our MMT models on the Multi30K dataset BIBREF21 . Each image in Multi30K contains one English (EN) description taken from Flickr30K BIBREF22 and human translations into German (DE), French (FR) and Czech BIBREF23 , BIBREF24 , BIBREF25 . The dataset contains 29,000 instances for training, 1,014 for development, and 1,000 for test. We only experiment with German and French, which are languages for which we have in-house expertise for the type of analysis we present. In addition to the official Multi30K test set (test 2016), we also use the test set from the latest WMT evaluation competition, test 2018 BIBREF25 .
Degradation of source
In addition to using the Multi30K dataset as is (standard setup), we probe the ability of our models to address the three linguistic phenomena where additional context has been proved important (Section ): ambiguities, gender-neutral words and noisy input. In a controlled experiment where we aim to remove the influence of frequency biases, we degrade the source sentences by masking words through three strategies to replace words by a placeholder: random source words, ambiguous source words and gender unmarked source words. The procedure is applied to the train, validation and test sets. For the resulting dataset generated for each setting, we compare models having access to text-only context versus additional text and multimodal contexts. We seek to get insights into the contribution of each type of context to address each type of degradation.
In this setting (RND) we simulate erroneous source words by randomly dropping source content words. We first tag the entire source sentences using the spacy toolkit BIBREF26 and then drop nouns, verbs, adjectives and adverbs and replace these with a default BLANK token. By focusing on content words, we differ from previous work that suggests that neural machine translation is robust to non-content word noise in the source BIBREF27 .
In this setting (AMB), we rely on the MLT dataset BIBREF11 which provides a list of source words with multiple translations in the Multi30k training set. We replace ambiguous words with the BLANK token in the source language, which results in two language-specific datasets.
In this setting (PERS), we use the Flickr Entities dataset BIBREF28 to identify all the words that were annotated by humans as corresponding to the category person. We then replace such source words with the BLANK token.
The statistics of the resulting datasets for the three degradation strategies are shown in Table TABREF10 . We note that RND and PERS are the same for language pairs as the degradation only depends on the source side, while for AMB the words replaced depend on the target language.
Models
Based on the models described in Section we experiment with eight variants: (a) baseline transformer model (base); (b) base with AIC (base+sum); (c) base with AIF using spacial (base+att) or object based (base+obj) image features; (d) standard deliberation model (del); (e) deliberation models enriched with image information: del+sum, del+att and del+obj.
Training
In all cases, we optimise our models with cross entropy loss. For deliberation network models, we first train the standard transformer model until convergence, and use it to initialise the encoder and first-pass decoder. For each of the training samples, we follow BIBREF19 and obtain a set of 10-best samples from the first pass decoder, with a beam search of size 10. We use these as the first-pass decoder samples. We use Adam as optimiser BIBREF29 and train the model until convergence.
Results
In this section we present results of our experiments, first in the original dataset without any source degradation (Section SECREF18 ) and then in the setup with various source degradation strategies (Section SECREF25 ).
Standard setup
Table TABREF14 shows the results of our main experiments on the 2016 and 2018 test sets for French and German. We use Meteor BIBREF31 as the main metric, as in the WMT tasks BIBREF25 . We compare our transformer baseline to transformer models enriched with image information, as well as to the deliberation models, with or without image information.
We first note that our multimodal models achieve the state of the art performance for transformer networks (constrained models) on the English-German dataset, as compared to BIBREF30 . Second, our deliberation models lead to significant improvements over this baseline across test sets (average INLINEFORM0 , INLINEFORM1 ).
Transformer-based models enriched with image information (base+sum, base+att and base+obj), on the other hand, show no major improvements with respect to the base performance. This is also the case for deliberation models with image information (del+sum, del+att, del+obj), which do not show significant improvement over the vanilla deliberation performance (del).
However, as it has been shown in the WMT shared tasks on MMT BIBREF23 , BIBREF24 , BIBREF25 , automatic metrics often fail to capture nuances in translation quality, such as, the ones we expect the visual modality to help with, which – according to human perception – lead to better translations. To test this assumption in our settings, we performed human evaluation involving professional translators and native speakers of both French and German (three annotators).
The annotators were asked to rank randomly selected test samples according to how well they convey the meaning of the source, given the image (50 samples per language pair per annotator). For each source segment, the annotator was shown the outputs of three systems: base+att, the current MMT state-of-the-art BIBREF30 , del and del+obj. A rank could be assigned from 1 to 3, allowing ties BIBREF32 . Annotators could assign zero rank to all translations if they were judged incomprehensible. Following the common practice in WMT BIBREF32 , each system was then assigned a score which reflects the proportion of times it was judged to be better or equal other systems.
Table TABREF19 shows the human evaluation results. They are consistent with the automatic evaluation results when it comes to the preference of humans towards the deliberation-based setups, but show a more positive outlook regarding the addition of visual information (del+obj over del) for French.
Manual inspection of translations suggests that deliberation setups tend to improve both the grammaticality and adequacy of the first pass outputs. For German, the most common modifications performed by the second-pass decoder are substitutions of adjectives and verbs (for test 2016, 15% and 12% respectively, of all the edit distance operations). Changes to adjectives are mainly grammatical, changes to verbs are contextual (e.g., changing laufen to rennen, both verbs mean run, but the second refers to running very fast). For French, 15% of all the changes are substitutions of nouns (for test 2016). These are again very contextual. For example, the French word travailleur (worker) is replaced by ouvrier (manual worker) in the contexts where tools, machinery or buildings are mentioned. For our analysis we used again spacy.
The information on detected objects is particularly helpful for specific adequacy issues. Figure FIGREF15 demonstrates some such cases. In the first case, the base+att model misses the translation of race car: the German word Rennen translates only the word race. del introduces the word car (Auto) into the translation. Finally, del+obj correctly translates the expression race car (Rennwagen) by exploiting the object information. For French, del translates the source part in a body of water, missing from the base+att translation. del+obj additionally translated the word paddling according to the detected object Paddle.
Source degradation setup
Results of our source degradation experiments are shown in Table TABREF20 . A first observation is that – as with the standard setup – the performance of our deliberation models is overall better than that of the base models. The results of the multimodal models differ for German and French. For German, del+obj is the most successful configuration and shows statistically significant improvements over base for all setups. Moreover, for RND and AMB, it shows statistically significant improvements over del. However, especially for RND and AMB, del and del+sum are either the same or slightly worse than base.
For French, all the deliberation models show statistically significant improvements over base (average INLINEFORM0 , INLINEFORM1 ), but the image information added to del only improve scores significantly for test 2018 RND.
This difference in performances for French and German is potentially related to the need of more significant restructurings while translating from English into German. This is where a more complex del+obj architecture is more helpful. This is especially true for RND and AMB setups where blanked words could also be verbs, the part-of-speech most influenced by word order differences between English and German (see the decreasing complexity of translations for del and del+obj for the example (c) in Figure FIGREF21 ).
To get an insight into the contribution of different contexts to the resolution of blanks, we performed manual analysis of examples coming from the English-German base, del and del+obj setups (50 random examples per setup), where we count correctly translated blanks per system.
The results are shown in Table TABREF27 . As expected, they show that the RND and AMB blanks are more difficult to resolve (at most 40% resolved as compared to 61% for PERS). Translations of the majority of those blanks tend to be guessed by the textual context alone (especially for verbs). Image information is more helpful for PERS: we observe an increase of 10% in resolved blanks for del+obj as compared to del. However, for PERS the textual context is still enough in the majority of the cases: models tend to associate men with sports or women with cooking and are usually right (see Figure FIGREF21 example (c)).
The cases where image helps seem to be those with rather generic contexts: see Figure FIGREF21 (b) where enjoying a summer day is not associated with any particular gender and make other models choose homme (man) or femme (woman), and only base+obj chooses enfant (child) (the option closest to the reference).
In some cases detected objects are inaccurate or not precise enough to be helpful (e.g., when an object Person is detected) and can even harm correct translations.
Conclusions
We have proposed a novel approach to multimodal machine translation which makes better use of context, both textual and visual. Our results show that further exploring textual context through deliberation networks already leads to better results than the previous state of the art. Adding visual information, and in particular structural representations of this information, proved beneficial when input text contains noise and the language pair requires substantial restructuring from source to target. Our findings suggest that the combination of a deliberation approach and information from additional modalities is a promising direction for machine translation that is robust to noisy input. Our code and pre-processing scripts are available at https://github.com/ImperialNLP/MMT-Delib.
Acknowledgments
The authors thank the anonymous reviewers for their useful feedback. This work was supported by the MultiMT (H2020 ERC Starting Grant No. 678017) and MMVC (Newton Fund Institutional Links Grant, ID 352343575) projects. We also thank the annotators for their valuable help.
| [
"the English-German dataset"
] | 1,833 | qasper | en | null | 65e11f98300887fc3b0d3912db406f8d8825d3f61c7a12a2 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What dataset does this approach achieve state of the art results on?
Answer: | [
"the English-German dataset"
] | qasper | 128 | 86 | |
What are strong baselines model is compared to? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Chinese word segmentation (CWS) is a task for Chinese natural language process to delimit word boundary. CWS is a basic and essential task for Chinese which is written without explicit word delimiters and different from alphabetical languages like English. BIBREF0 treats Chinese word segmentation (CWS) as a sequence labeling task with character position tags, which is followed by BIBREF1, BIBREF2, BIBREF3. Traditional CWS models depend on the design of features heavily which effects the performance of model. To minimize the effort in feature engineering, some CWS models BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10, BIBREF11 are developed following neural network architecture for sequence labeling tasks BIBREF12. Neural CWS models perform strong ability of feature representation, employing unigram and bigram character embedding as input and approach good performance.
The CWS task is often modeled as one graph model based on a scoring model that means it is composed of two parts, one part is an encoder which is used to generate the representation of characters from the input sequence, the other part is a decoder which performs segmentation according to the encoder scoring. Table TABREF1 summarizes typical CWS models according to their decoding ways for both traditional and neural models. Markov models such as BIBREF13 and BIBREF4 depend on the maximum entropy model or maximum entropy Markov model both with a Viterbi decoder. Besides, conditional random field (CRF) or Semi-CRF for sequence labeling has been used for both traditional and neural models though with different representations BIBREF2, BIBREF15, BIBREF10, BIBREF17, BIBREF18. Generally speaking, the major difference between traditional and neural network models is about the way to represent input sentences.
Recent works about neural CWS which focus on benchmark dataset, namely SIGHAN Bakeoff BIBREF21, may be put into the following three categories roughly.
Encoder. Practice in various natural language processing tasks has been shown that effective representation is essential to the performance improvement. Thus for better CWS, it is crucial to encode the input character, word or sentence into effective representation. Table TABREF2 summarizes regular feature sets for typical CWS models including ours as well. The building blocks that encoders use include recurrent neural network (RNN) and convolutional neural network (CNN), and long-term memory network (LSTM).
Graph model. As CWS is a kind of structure learning task, the graph model determines which type of decoder should be adopted for segmentation, also it may limit the capability of defining feature, as shown in Table 2, not all graph models can support the word features. Thus recent work focused on finding more general or flexible graph model to make model learn the representation of segmentation more effective as BIBREF9, BIBREF11.
External data and pre-trained embedding. Whereas both encoder and graph model are about exploring a way to get better performance only by improving the model strength itself. Using external resource such as pre-trained embeddings or language representation is an alternative for the same purpose BIBREF22, BIBREF23. SIGHAN Bakeoff defines two types of evaluation settings, closed test limits all the data for learning should not be beyond the given training set, while open test does not take this limitation BIBREF21. In this work, we will focus on the closed test setting by finding a better model design for further CWS performance improvement.
Shown in Table TABREF1, different decoders have particular decoding algorithms to match the respective CWS models. Markov models and CRF-based models often use Viterbi decoders with polynomial time complexity. In general graph model, search space may be too large for model to search. Thus it forces graph models to use an approximate beam search strategy. Beam search algorithm has a kind low-order polynomial time complexity. Especially, when beam width $b$=1, the beam search algorithm will reduce to greedy algorithm with a better time complexity $O(Mn)$ against the general beam search time complexity $O(Mnb^2)$, where $n$ is the number of units in one sentences, $M$ is a constant representing the model complexity. Greedy decoding algorithm can bring the fastest speed of decoding while it is not easy to guarantee the precision of decoding when the encoder is not strong enough.
In this paper, we focus on more effective encoder design which is capable of offering fast and accurate Chinese word segmentation with only unigram feature and greedy decoding. Our proposed encoder will only consist of attention mechanisms as building blocks but nothing else. Motivated by the Transformer BIBREF24 and its strength of capturing long-range dependencies of input sentences, we use a self-attention network to generate the representation of input which makes the model encode sentences at once without feeding input iteratively. Considering the weakness of the Transformer to model relative and absolute position information directly BIBREF25 and the importance of localness information, position information and directional information for CWS, we further improve the architecture of standard multi-head self-attention of the Transformer with a directional Gaussian mask and get a variant called Gaussian-masked directional multi-head attention. Based on the newly improved attention mechanism, we expand the encoder of the Transformer to capture different directional information. With our powerful encoder, our model uses only simple unigram features to generate representation of sentences.
For decoder which directly performs the segmentation, we use the bi-affinal attention scorer, which has been used in dependency parsing BIBREF26 and semantic role labeling BIBREF27, to implement greedy decoding on finding the boundaries of words. In our proposed model, greedy decoding ensures a fast segmentation while powerful encoder design ensures a good enough segmentation performance even working with greedy decoder together. Our model will be strictly evaluated on benchmark datasets from SIGHAN Bakeoff shared task on CWS in terms of closed test setting, and the experimental results show that our proposed model achieves new state-of-the-art.
The technical contributions of this paper can be summarized as follows.
We propose a CWS model with only attention structure. The encoder and decoder are both based on attention structure.
With a powerful enough encoder, we for the first time show that unigram (character) featues can help yield strong performance instead of diverse $n$-gram (character and word) features in most of previous work.
To capture the representation of localness information and directional information, we propose a variant of directional multi-head self-attention to further enhance the state-of-the-art Transformer encoder.
Models
The CWS task is often modelled as one graph model based on an encoder-based scoring model. The model for CWS task is composed of an encoder to represent the input and a decoder based on the encoder to perform actual segmentation. Figure FIGREF6 is the architecture of our model. The model feeds sentence into encoder. Embedding captures the vector $e=(e_1,...,e_n)$ of the input character sequences of $c=(c_1,...,c_n)$. The encoder maps vector sequences of $ {e}=(e_1,..,e_n)$ to two sequences of vector which are $ {v^b}=(v_1^b,...,v_n^b)$ and ${v^f}=(v_1^f,...v_n^f)$ as the representation of sentences. With $v^b$ and $v^f$, the bi-affinal scorer calculates the probability of each segmentation gaps and predicts the word boundaries of input. Similar as the Transformer, the encoder is an attention network with stacked self-attention and point-wise, fully connected layers while our encoder includes three independent directional encoders.
Models ::: Encoder Stacks
In the Transformer, the encoder is composed of a stack of N identical layers and each layer has one multi-head self-attention layer and one position-wise fully connected feed-forward layer. One residual connection is around two sub-layers and followed by layer normalization BIBREF24. This architecture provides the Transformer a good ability to generate representation of sentence.
With the variant of multi-head self-attention, we design a Gaussian-masked directional encoder to capture representation of different directions to improve the ability of capturing the localness information and position information for the importance of adjacent characters. One unidirectional encoder can capture information of one particular direction.
For CWS tasks, one gap of characters, which is from a word boundary, can divide one sequence into two parts, one part in front of the gap and one part in the rear of it. The forward encoder and backward encoder are used to capture information of two directions which correspond to two parts divided by the gap.
One central encoder is paralleled with forward and backward encoders to capture the information of entire sentences. The central encoder is a special directional encoder for forward and backward information of sentences. The central encoder can fuse the information and enable the encoder to capture the global information.
The encoder outputs one forward information and one backward information of each positions. The representation of sentence generated by center encoder will be added to these information directly:
where $v^{b}=(v^b_1,...,v^b_n)$ is the backward information, $v^{f}=(v^f_1,...,v^f_n)$ is the forward information, $r^{b}=(r^b_1,...,r^b_n)$ is the output of backward encoder, $r^{c}=(r^c_1,...,r^c_n)$ is the output of center encoder and $r^{f}=(r^f_1,...,r^f_n)$ is the output of forward encoder.
Models ::: Gaussian-Masked Directional Multi-Head Attention
Similar as scaled dot-product attention BIBREF24, Gaussian-masked directional attention can be described as a function to map queries and key-value pairs to the representation of input. Here queries, keys and values are all vectors. Standard scaled dot-product attention is calculated by dotting query $Q$ with all keys $K$, dividing each values by $\sqrt{d_k}$, where $\sqrt{d_k}$ is the dimension of keys, and apply a softmax function to generate the weights in the attention:
Different from scaled dot-product attention, Gaussian-masked directional attention expects to pay attention to the adjacent characters of each positions and cast the localness relationship between characters as a fix Gaussian weight for attention. We assume that the Gaussian weight only relys on the distance between characters.
Firstly we introduce the Gaussian weight matrix $G$ which presents the localness relationship between each two characters:
where $g_{ij}$ is the Gaussian weight between character $i$ and $j$, $dis_{ij}$ is the distance between character $i$ and $j$, $\Phi (x)$ is the cumulative distribution function of Gaussian, $\sigma $ is the standard deviation of Gaussian function and it is a hyperparameter in our method. Equation (DISPLAY_FORM13) can ensure the Gaussian weight equals 1 when $dis_{ij}$ is 0. The larger distance between charactersis, the smaller the weight is, which makes one character can affect its adjacent characters more compared with other characters.
To combine the Gaussian weight to the self-attention, we produce the Hadamard product of Gaussian weight matrix $G$ and the score matrix produced by $Q{K^{T}}$
where $AG$ is the Gaussian-masked attention. It ensures that the relationship between two characters with long distances is weaker than adjacent characters.
The scaled dot-product attention models the relationship between two characters without regard to their distances in one sequence. For CWS task, the weight between adjacent characters should be more important while it is hard for self-attention to achieve the effect explicitly because the self-attention cannot get the order of sentences directly. The Gaussian-masked attention adjusts the weight between characters and their adjacent character to a larger value which stands for the effect of adjacent characters.
For forward and backward encoder, the self-attention sublayer needs to use a triangular matrix mask to let the self-attention focus on different weights:
where $pos_i$ is the position of character $c_i$. The triangular matrix for forward and backward encode are:
$\left[ \begin{matrix} 1 & 0 & 0 & \cdots &0\\ 1 & 1 & 0 & \cdots &0\\ 1 & 1 & 1 & \cdots &0\\ \vdots &\vdots &\vdots &\ddots &\vdots \\ 1 & 1 & 1 & \cdots & 1\\ \end{matrix} \right]$ $\left[ \begin{matrix} 1 & 1 & 1 & \cdots &1 \\ 0 & 1 & 1 & \cdots &1 \\ 0 & 0& 1 & \cdots &1 \\ \vdots &\vdots &\vdots &\ddots &\vdots \\ 0 & 0 & 0 & \cdots & 1\\ \end{matrix}\right]$
Similar as BIBREF24, we use multi-head attention to capture information from different dimension positions as Figure FIGREF16 and get Gaussian-masked directional multi-head attention. With multi-head attention architecture, the representation of input can be captured by
where $MH$ is the Gaussian-masked multi-head attention, ${W_i^q, W_i^k,W_i^v} \in \mathbb {R}^{d_k \times d_h}$ is the parameter matrices to generate heads, $d_k$ is the dimension of model and $d_h$ is the dimension of one head.
Models ::: Bi-affinal Attention Scorer
Regarding word boundaries as gaps between any adjacent words converts the character labeling task to the gap labeling task. Different from character labeling task, gap labeling task requires information of two adjacent characters. The relationship between adjacent characters can be represented as the type of gap. The characteristic of word boundaries makes bi-affine attention an appropriate scorer for CWS task.
Bi-affinal attention scorer is the component that we use to label the gap. Bi-affinal attention is developed from bilinear attention which has been used in dependency parsing BIBREF26 and SRL BIBREF27. The distribution of labels in a labeling task is often uneven which makes the output layer often include a fixed bias term for the prior probability of different labels BIBREF27. Bi-affine attention uses bias terms to alleviate the burden of the fixed bias term and get the prior probability which makes it different from bilinear attention. The distribution of the gap is uneven that is similar as other labeling task which fits bi-affine.
Bi-affinal attention scorer labels the target depending on information of independent unit and the joint information of two units. In bi-affinal attention, the score $s_{ij}$ of characters $c_i$ and $c_j$ $(i < j)$ is calculated by:
where $v_i^f$ is the forward information of $c_i$ and $v_i^b$ is the backward information of $c_j$. In Equation (DISPLAY_FORM21), $W$, $U$ and $b$ are all parameters that can be updated in training. $W$ is a matrix with shape $(d_i \times N\times d_j)$ and $U$ is a $(N\times (d_i + d_j))$ matrix where $d_i$ is the dimension of vector $v_i^f$ and $N$ is the number of labels.
In our model, the biaffine scorer uses the forward information of character in front of the gap and the backward information of the character behind the gap to distinguish the position of characters. Figure FIGREF22 is an example of labeling gap. The method of using biaffine scorer ensures that the boundaries of words can be determined by adjacent characters with different directional information. The score vector of the gap is formed by the probability of being a boundary of word. Further, the model generates all boundaries using activation function in a greedy decoding way.
Experiments ::: Experimental Settings ::: Data
We train and evaluate our model on datasets from SIGHAN Bakeoff 2005 BIBREF21 which has four datasets, PKU, MSR, AS and CITYU. Table TABREF23 shows the statistics of train data. We use F-score to evaluate CWS models. To train model with pre-trained embeddings in AS and CITYU, we use OpenCC to transfer data from traditional Chinese to simplified Chinese.
Experiments ::: Experimental Settings ::: Pre-trained Embedding
We only use unigram feature so we only trained character embeddings. Our pre-trained embedding are pre-trained on Chinese Wikipedia corpus by word2vec BIBREF29 toolkit. The corpus used for pre-trained embedding is all transferred to simplified Chinese and not segmented. On closed test, we use embeddings initialized randomly.
Experiments ::: Experimental Settings ::: Hyperparameters
For different datasets, we use two kinds of hyperparameters which are presented in Table TABREF24. We use hyperparameters in Table TABREF24 for small corpora (PKU and CITYU) and normal corpora (MSR and AS). We set the standard deviation of Gaussian function in Equation (DISPLAY_FORM13) to 2. Each training batch contains sentences with at most 4096 tokens.
Experiments ::: Experimental Settings ::: Optimizer
To train our model, we use the Adam BIBREF30 optimizer with $\beta _1=0.9$, $\beta _2=0.98$ and $\epsilon =10^{-9}$. The learning rate schedule is the same as BIBREF24:
where $d$ is the dimension of embeddings, $step$ is the step number of training and $warmup_step$ is the step number of warmup. When the number of steps is smaller than the step of warmup, the learning rate increases linearly and then decreases.
Experiments ::: Hardware and Implements
We trained our models on a single CPU (Intel i7-5960X) with an nVidia 1080 Ti GPU. We implement our model in Python with Pytorch 1.0.
Experiments ::: Results
Tables TABREF25 and TABREF26 reports the performance of recent models and ours in terms of closed test setting. Without the assistance of unsupervised segmentation features userd in BIBREF20, our model outperforms all the other models in MSR and AS except BIBREF18 and get comparable performance in PKU and CITYU. Note that all the other models for this comparison adopt various $n$-gram features while only our model takes unigram ones.
With unsupervised segmentation features introduced by BIBREF20, our model gets a higher result. Specially, the results in MSR and AS achieve new state-of-the-art and approaching previous state-of-the-art in CITYU and PKU. The unsupervised segmentation features are derived from the given training dataset, thus using them does not violate the rule of closed test of SIGHAN Bakeoff.
Table TABREF36 compares our model and recent neural models in terms of open test setting in which any external resources, especially pre-trained embeddings or language models can be used. In MSR and AS, our model gets a comparable result while our results in CITYU and PKU are not remarkable.
However, it is well known that it is always hard to compare models when using open test setting, especially with pre-trained embedding. Not all models may use the same method and data to pre-train. Though pre-trained embedding or language model can improve the performance, the performance improvement itself may be from multiple sources. It often that there is a success of pre-trained embedding to improve the performance, while it cannot prove that the model is better.
Compared with other LSTM models, our model performs better in AS and MSR than in CITYU and PKU. Considering the scale of different corpora, we believe that the size of corpus affects our model and the larger size is, the better model performs. For small corpus, the model tends to be overfitting.
Tables TABREF25 and TABREF26 also show the decoding time in different datasets. Our model finishes the segmentation with the least decoding time in all four datasets, thanks to the architecture of model which only takes attention mechanism as basic block.
Related Work ::: Chinese Word Segmentation
CWS is a task for Chinese natural language process to delimit word boundary. BIBREF0 for the first time formulize CWS as a sequence labeling task. BIBREF3 show that different character tag sets can make essential impact for CWS. BIBREF2 use CRFs as a model for CWS, achieving new state-of-the-art. Works of statistical CWS has built the basis for neural CWS.
Neural word segmentation has been widely used to minimize the efforts in feature engineering which was important in statistical CWS. BIBREF4 introduce the neural model with sliding-window based sequence labeling. BIBREF6 propose a gated recursive neural network (GRNN) for CWS to incorporate complicated combination of contextual character and n-gram features. BIBREF7 use LSTM to learn long distance information. BIBREF9 propose a neural framework that eliminates context windows and utilize complete segmentation history. BIBREF33 explore a joint model that performs segmentation, POS-Tagging and chunking simultaneously. BIBREF34 propose a feature-enriched neural model for joint CWS and part-of-speech tagging. BIBREF35 present a joint model to enhance the segmentation of Chinese microtext by performing CWS and informal word detection simultaneously. BIBREF17 propose a character-based convolutional neural model to capture $n$-gram features automatically and an effective approach to incorporate word embeddings. BIBREF11 improve the model in BIBREF9 and propose a greedy neural word segmenter with balanced word and character embedding inputs. BIBREF23 propose a novel neural network model to incorporate unlabeled and partially-labeled data. BIBREF36 propose two methods that extend the Bi-LSTM to perform incorporating dictionaries into neural networks for CWS. BIBREF37 propose Switch-LSTMs to segment words and provided a more flexible solution for multi-criteria CWS which is easy to transfer the learned knowledge to new criteria.
Related Work ::: Transformer
Transformer BIBREF24 is an attention-based neural machine translation model. The Transformer is one kind of self-attention networks (SANs) which is proposed in BIBREF38. Encoder of the Transformer consists of one self-attention layer and a position-wise feed-forward layer. Decoder of the Transformer contains one self-attention layer, one encoder-decoder attention layer and one position-wise feed-forward layer. The Transformer uses residual connections around the sublayers and then followed by a layer normalization layer.
Scaled dot-product attention is the key component in the Transformer. The input of attention contains queries, keys, and values of input sequences. The attention is generated using queries and keys like Equation (DISPLAY_FORM11). Structure of scaled dot-product attention allows the self-attention layer generate the representation of sentences at once and contain the information of the sentence which is different from RNN that process characters of sentences one by one. Standard self-attention is similar as Gaussian-masked direction attention while it does not have directional mask and gaussian mask. BIBREF24 also propose multi-head attention which is better to generate representation of sentence by dividing queries, keys and values to different heads and get information from different subspaces.
Conclusion
In this paper, we propose an attention mechanism only based Chinese word segmentation model. Our model uses self-attention from the Transformer encoder to take sequence input and bi-affine attention scorer to predict the label of gaps. To improve the ability of capturing the localness and directional information of self-attention based encoder, we propose a variant of self-attention called Gaussian-masked directional multi-head attention to replace the standard self-attention. We also extend the Transformer encoder to capture directional features. Our model uses only unigram features instead of multiple $n$-gram features in previous work. Our model is evaluated on standard benchmark dataset, SIGHAN Bakeoff 2005, which shows not only our model performs segmentation faster than any previous models but also gives new higher or comparable segmentation performance against previous state-of-the-art models.
| [
"Baseline models are:\n- Chen et al., 2015a\n- Chen et al., 2015b\n- Liu et al., 2016\n- Cai and Zhao, 2016\n- Cai et al., 2017\n- Zhou et al., 2017\n- Ma et al., 2018\n- Wang et al., 2019"
] | 3,629 | qasper | en | null | 4f6f6dfa672ed697a94d1d1ee528e50645f01f568707b0b5 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What are strong baselines model is compared to?
Answer: | [
"Baseline models are:\n- Chen et al., 2015a\n- Chen et al., 2015b\n- Liu et al., 2016\n- Cai and Zhao, 2016\n- Cai et al., 2017\n- Zhou et al., 2017\n- Ma et al., 2018\n- Wang et al., 2019"
] | qasper | 128 | 87 | |
What type of classifiers are used? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Event detection on microblogging platforms such as Twitter aims to detect events preemptively. A main task in event detection is detecting events of predetermined types BIBREF0, such as concerts or controversial events based on microposts matching specific event descriptions. This task has extensive applications ranging from cyber security BIBREF1, BIBREF2 to political elections BIBREF3 or public health BIBREF4, BIBREF5. Due to the high ambiguity and inconsistency of the terms used in microposts, event detection is generally performed though statistical machine learning models, which require a labeled dataset for model training. Data labeling is, however, a long, laborious, and usually costly process. For the case of micropost classification, though positive labels can be collected (e.g., using specific hashtags, or event-related date-time information), there is no straightforward way to generate negative labels useful for model training. To tackle this lack of negative labels and the significant manual efforts in data labeling, BIBREF1 (BIBREF1, BIBREF3) introduced a weak supervision based learning approach, which uses only positively labeled data, accompanied by unlabeled examples by filtering microposts that contain a certain keyword indicative of the event type under consideration (e.g., `hack' for cyber security). Another key technique in this context is expectation regularization BIBREF6, BIBREF7, BIBREF1. Here, the estimated proportion of relevant microposts in an unlabeled dataset containing a keyword is given as a keyword-specific expectation. This expectation is used in the regularization term of the model's objective function to constrain the posterior distribution of the model predictions. By doing so, the model is trained with an expectation on its prediction for microposts that contain the keyword. Such a method, however, suffers from two key problems:
Due to the unpredictability of event occurrences and the constantly changing dynamics of users' posting frequency BIBREF8, estimating the expectation associated with a keyword is a challenging task, even for domain experts;
The performance of the event detection model is constrained by the informativeness of the keyword used for model training. As of now, we lack a principled method for discovering new keywords and improve the model performance.
To address the above issues, we advocate a human-AI loop approach for discovering informative keywords and estimating their expectations reliably. Our approach iteratively leverages 1) crowd workers for estimating keyword-specific expectations, and 2) the disagreement between the model and the crowd for discovering new informative keywords. More specifically, at each iteration after we obtain a keyword-specific expectation from the crowd, we train the model using expectation regularization and select those keyword-related microposts for which the model's prediction disagrees the most with the crowd's expectation; such microposts are then presented to the crowd to identify new keywords that best explain the disagreement. By doing so, our approach identifies new keywords which convey more relevant information with respect to existing ones, thus effectively boosting model performance. By exploiting the disagreement between the model and the crowd, our approach can make efficient use of the crowd, which is of critical importance in a human-in-the-loop context BIBREF9, BIBREF10. An additional advantage of our approach is that by obtaining new keywords that improve model performance over time, we are able to gain insight into how the model learns for specific event detection tasks. Such an advantage is particularly useful for event detection using complex models, e.g., deep neural networks, which are intrinsically hard to understand BIBREF11, BIBREF12. An additional challenge in involving crowd workers is that their contributions are not fully reliable BIBREF13. In the crowdsourcing literature, this problem is usually tackled with probabilistic latent variable models BIBREF14, BIBREF15, BIBREF16, which are used to perform truth inference by aggregating a redundant set of crowd contributions. Our human-AI loop approach improves the inference of keyword expectation by aggregating contributions not only from the crowd but also from the model. This, however, comes with its own challenge as the model's predictions are further dependent on the results of expectation inference, which is used for model training. To address this problem, we introduce a unified probabilistic model that seamlessly integrates expectation inference and model training, thereby allowing the former to benefit from the latter while resolving the inter-dependency between the two.
To the best of our knowledge, we are the first to propose a human-AI loop approach that iteratively improves machine learning models for event detection. In summary, our work makes the following key contributions:
A novel human-AI loop approach for micropost event detection that jointly discovers informative keywords and estimates their expectation;
A unified probabilistic model that infers keyword expectation and simultaneously performs model training;
An extensive empirical evaluation of our approach on multiple real-world datasets demonstrating that our approach significantly improves the state of the art by an average of 24.3% AUC.
The rest of this paper is organized as follows. First, we present our human-AI loop approach in Section SECREF2. Subsequently, we introduce our proposed probabilistic model in Section SECREF3. The experimental setup and results are presented in Section SECREF4. Finally, we briefly cover related work in Section SECREF5 before concluding our work in Section SECREF6.
The Human-AI Loop Approach
Given a set of labeled and unlabeled microposts, our goal is to extract informative keywords and estimate their expectations in order to train a machine learning model. To achieve this goal, our proposed human-AI loop approach comprises two crowdsourcing tasks, i.e., micropost classification followed by keyword discovery, and a unified probabilistic model for expectation inference and model training. Figure FIGREF6 presents an overview of our approach. Next, we describe our approach from a process-centric perspective.
Following previous studies BIBREF1, BIBREF17, BIBREF2, we collect a set of unlabeled microposts $\mathcal {U}$ from a microblogging platform and post-filter, using an initial (set of) keyword(s), those microposts that are potentially relevant to an event category. Then, we collect a set of event-related microposts (i.e., positively labeled microposts) $\mathcal {L}$, post-filtering with a list of seed events. $\mathcal {U}$ and $\mathcal {L}$ are used together to train a discriminative model (e.g., a deep neural network) for classifying the relevance of microposts to an event. We denote the target model as $p_\theta (y|x)$, where $\theta $ is the model parameter to be learned and $y$ is the label of an arbitrary micropost, represented by a bag-of-words vector $x$. Our approach iterates several times $t=\lbrace 1, 2, \ldots \rbrace $ until the performance of the target model converges. Each iteration starts from the initial keyword(s) or the new keyword(s) discovered in the previous iteration. Given such a keyword, denoted by $w^{(t)}$, the iteration starts by sampling microposts containing the keyword from $\mathcal {U}$, followed by dynamically creating micropost classification tasks and publishing them on a crowdsourcing platform.
Micropost Classification. The micropost classification task requires crowd workers to label the selected microposts into two classes: event-related and non event-related. In particular, workers are given instructions and examples to differentiate event-instance related microposts and general event-category related microposts. Consider, for example, the following microposts in the context of Cyber attack events, both containing the keyword `hack':
Credit firm Equifax says 143m Americans' social security numbers exposed in hack
This micropost describes an instance of a cyber attack event that the target model should identify. This is, therefore, an event-instance related micropost and should be considered as a positive example. Contrast this with the following example:
Companies need to step their cyber security up
This micropost, though related to cyber security in general, does not mention an instance of a cyber attack event, and is of no interest to us for event detection. This is an example of a general event-category related micropost and should be considered as a negative example.
In this task, each selected micropost is labeled by multiple crowd workers. The annotations are passed to our probabilistic model for expectation inference and model training.
Expectation Inference & Model Training. Our probabilistic model takes crowd-contributed labels and the model trained in the previous iteration as input. As output, it generates a keyword-specific expectation, denoted as $e^{(t)}$, and an improved version of the micropost classification model, denoted as $p_{\theta ^{(t)}}(y|x)$. The details of our probabilistic model are given in Section SECREF3.
Keyword Discovery. The keyword discovery task aims at discovering a new keyword (or a set of keywords) that is most informative for model training with respect to existing keywords. To this end, we first apply the current model $p_{\theta ^{(t)}}(y|x)$ on the unlabeled microposts $\mathcal {U}$. For those that contain the keyword $w^{(t)}$, we calculate the disagreement between the model predictions and the keyword-specific expectation $e^{(t)}$:
and select the ones with the highest disagreement for keyword discovery. These selected microposts are supposed to contain information that can explain the disagreement between the model prediction and keyword-specific expectation, and can thus provide information that is most different from the existing set of keywords for model training.
For instance, our study shows that the expectation for the keyword `hack' is 0.20, which means only 20% of the initial set of microposts retrieved with the keyword are event-related. A micropost selected with the highest disagreement (Eq. DISPLAY_FORM7), whose likelihood of being event-related as predicted by the model is $99.9\%$, is shown as an example below:
RT @xxx: Hong Kong securities brokers hit by cyber attacks, may face more: regulator #cyber #security #hacking https://t.co/rC1s9CB
This micropost contains keywords that can better indicate the relevance to a cyber security event than the initial keyword `hack', e.g., `securities', `hit', and `attack'.
Note that when the keyword-specific expectation $e^{(t)}$ in Equation DISPLAY_FORM7 is high, the selected microposts will be the ones that contain keywords indicating the irrelevance of the microposts to an event category. Such keywords are also useful for model training as they help improve the model's ability to identify irrelevant microposts.
To identify new keywords in the selected microposts, we again leverage crowdsourcing, as humans are typically better than machines at providing specific explanations BIBREF18, BIBREF19. In the crowdsourcing task, workers are first asked to find those microposts where the model predictions are deemed correct. Then, from those microposts, workers are asked to find the keyword that best indicates the class of the microposts as predicted by the model. The keyword most frequently identified by the workers is then used as the initial keyword for the following iteration. In case multiple keywords are selected, e.g., the top-$N$ frequent ones, workers will be asked to perform $N$ micropost classification tasks for each keyword in the next iteration, and the model training will be performed on multiple keyword-specific expectations.
Unified Probabilistic Model
This section introduces our probabilistic model that infers keyword expectation and trains the target model simultaneously. We start by formalizing the problem and introducing our model, before describing the model learning method.
Problem Formalization. We consider the problem at iteration $t$ where the corresponding keyword is $w^{(t)}$. In the current iteration, let $\mathcal {U}^{(t)} \subset \mathcal {U}$ denote the set of all microposts containing the keyword and $\mathcal {M}^{(t)}= \lbrace x_{m}\rbrace _{m=1}^M\subset \mathcal {U}^{(t)}$ be the randomly selected subset of $M$ microposts labeled by $N$ crowd workers $\mathcal {C} = \lbrace c_n\rbrace _{n=1}^N$. The annotations form a matrix $\mathbf {A}\in \mathbb {R}^{M\times N}$ where $\mathbf {A}_{mn}$ is the label for the micropost $x_m$ contributed by crowd worker $c_n$. Our goal is to infer the keyword-specific expectation $e^{(t)}$ and train the target model by learning the model parameter $\theta ^{(t)}$. An additional parameter of our probabilistic model is the reliability of crowd workers, which is essential when involving crowdsourcing. Following Dawid and Skene BIBREF14, BIBREF16, we represent the annotation reliability of worker $c_n$ by a latent confusion matrix $\pi ^{(n)}$, where the $rs$-th element $\pi _{rs}^{(n)}$ denotes the probability of $c_n$ labeling a micropost as class $r$ given the true class $s$.
Unified Probabilistic Model ::: Expectation as Model Posterior
First, we introduce an expectation regularization technique for the weakly supervised learning of the target model $p_{\theta ^{(t)}}(y|x)$. In this setting, the objective function of the target model is composed of two parts, corresponding to the labeled microposts $\mathcal {L}$ and the unlabeled ones $\mathcal {U}$.
The former part aims at maximizing the likelihood of the labeled microposts:
where we assume that $\theta $ is generated from a prior distribution (e.g., Laplacian or Gaussian) parameterized by $\sigma $.
To leverage unlabeled data for model training, we make use of the expectations of existing keywords, i.e., {($w^{(1)}$, $e^{(1)}$), ..., ($w^{(t-1)}$, $e^{(t-1)}$), ($w^{(t)}$, $e^{(t)}$)} (Note that $e^{(t)}$ is inferred), as a regularization term to constrain model training. To do so, we first give the model's expectation for each keyword $w^{(k)}$ ($1\le k\le t$) as follows:
which denotes the empirical expectation of the model’s posterior predictions on the unlabeled microposts $\mathcal {U}^{(k)}$ containing keyword $w^{(k)}$. Expectation regularization can then be formulated as the regularization of the distance between the Bernoulli distribution parameterized by the model's expectation and the expectation of the existing keyword:
where $D_{KL}[\cdot \Vert \cdot ]$ denotes the KL-divergence between the Bernoulli distributions $Ber(e^{(k)})$ and $Ber(\mathbb {E}_{x\sim \mathcal {U}^{(k)}}(y))$, and $\lambda $ controls the strength of expectation regularization.
Unified Probabilistic Model ::: Expectation as Class Prior
To learn the keyword-specific expectation $e^{(t)}$ and the crowd worker reliability $\pi ^{(n)}$ ($1\le n\le N$), we model the likelihood of the crowd-contributed labels $\mathbf {A}$ as a function of these parameters. In this context, we view the expectation as the class prior, thus performing expectation inference as the learning of the class prior. By doing so, we connect expectation inference with model training.
Specifically, we model the likelihood of an arbitrary crowd-contributed label $\mathbf {A}_{mn}$ as a mixture of multinomials where the prior is the keyword-specific expectation $e^{(t)}$:
where $e_s^{(t)}$ is the probability of the ground truth label being $s$ given the keyword-specific expectation as the class prior; $K$ is the set of possible ground truth labels (binary in our context); and $r=\mathbf {A}_{mn}$ is the crowd-contributed label. Then, for an individual micropost $x_m$, the likelihood of crowd-contributed labels $\mathbf {A}_{m:}$ is given by:
Therefore, the objective function for maximizing the likelihood of the entire annotation matrix $\mathbf {A}$ can be described as:
Unified Probabilistic Model ::: Unified Probabilistic Model
Integrating model training with expectation inference, the overall objective function of our proposed model is given by:
Figure FIGREF18 depicts a graphical representation of our model, which combines the target model for training (on the left) with the generative model for crowd-contributed labels (on the right) through a keyword-specific expectation.
Model Learning. Due to the unknown ground truth labels of crowd-annotated microposts ($y_m$ in Figure FIGREF18), we resort to expectation maximization for model learning. The learning algorithm iteratively takes two steps: the E-step and the M-step. The E-step infers the ground truth labels given the current model parameters. The M-step updates the model parameters, including the crowd reliability parameters $\pi ^{(n)}$ ($1\le n\le N$), the keyword-specific expectation $e^{(t)}$, and the parameter of the target model $\theta ^{(t)}$. The E-step and the crowd parameter update in the M-step are similar to the Dawid-Skene model BIBREF14. The keyword expectation is inferred by taking into account both the crowd-contributed labels and the model prediction:
The parameter of the target model is updated by gradient descent. For example, when the target model to be trained is a deep neural network, we use back-propagation with gradient descent to update the weight matrices.
Experiments and Results
This section presents our experimental setup and results for evaluating our approach. We aim at answering the following questions:
[noitemsep,leftmargin=*]
Q1: How effectively does our proposed human-AI loop approach enhance the state-of-the-art machine learning models for event detection?
Q2: How well does our keyword discovery method work compare to existing keyword expansion methods?
Q3: How effective is our approach using crowdsourcing at obtaining new keywords compared with an approach labelling microposts for model training under the same cost?
Q4: How much benefit does our unified probabilistic model bring compared to methods that do not take crowd reliability into account?
Experiments and Results ::: Experimental Setup
Datasets. We perform our experiments with two predetermined event categories: cyber security (CyberAttack) and death of politicians (PoliticianDeath). These event categories are chosen as they are representative of important event types that are of interest to many governments and companies. The need to create our own dataset was motivated by the lack of public datasets for event detection on microposts. The few available datasets do not suit our requirements. For example, the publicly available Events-2012 Twitter dataset BIBREF20 contains generic event descriptions such as Politics, Sports, Culture etc. Our work targets more specific event categories BIBREF21. Following previous studies BIBREF1, we collect event-related microposts from Twitter using 11 and 8 seed events (see Section SECREF2) for CyberAttack and PoliticianDeath, respectively. Unlabeled microposts are collected by using the keyword `hack' for CyberAttack, while for PoliticianDeath, we use a set of keywords related to `politician' and `death' (such as `bureaucrat', `dead' etc.) For each dataset, we randomly select 500 tweets from the unlabeled subset and manually label them for evaluation. Table TABREF25 shows key statistics from our two datasets.
Comparison Methods. To demonstrate the generality of our approach on different event detection models, we consider Logistic Regression (LR) BIBREF1 and Multilayer Perceptron (MLP) BIBREF2 as the target models. As the goal of our experiments is to demonstrate the effectiveness of our approach as a new model training technique, we use these widely used models. Also, we note that in our case other neural network models with more complex network architectures for event detection, such as the bi-directional LSTM BIBREF17, turn out to be less effective than a simple feedforward network. For both LR and MLP, we evaluate our proposed human-AI loop approach for keyword discovery and expectation estimation by comparing against the weakly supervised learning method proposed by BIBREF1 (BIBREF1) and BIBREF17 (BIBREF17) where only one initial keyword is used with an expectation estimated by an individual expert.
Parameter Settings. We empirically set optimal parameters based on a held-out validation set that contains 20% of the test data. These include the hyperparamters of the target model, those of our proposed probabilistic model, and the parameters used for training the target model. We explore MLP with 1, 2 and 3 hidden layers and apply a grid search in 32, 64, 128, 256, 512 for the dimension of the embeddings and that of the hidden layers. For the coefficient of expectation regularization, we follow BIBREF6 (BIBREF6) and set it to $\lambda =10 \times $ #labeled examples. For model training, we use the Adam BIBREF22 optimization algorithm for both models.
Evaluation. Following BIBREF1 (BIBREF1) and BIBREF3 (BIBREF3), we use accuracy and area under the precision-recall curve (AUC) metrics to measure the performance of our proposed approach. We note that due to the imbalance in our datasets (20% positive microposts in CyberAttack and 27% in PoliticianDeath), accuracy is dominated by negative examples; AUC, in comparison, better characterizes the discriminative power of the model.
Crowdsourcing. We chose Level 3 workers on the Figure-Eight crowdsourcing platform for our experiments. The inter-annotator agreement in micropost classification is taken into account through the EM algorithm. For keyword discovery, we filter keywords based on the frequency of the keyword being selected by the crowd. In terms of cost-effectiveness, our approach is motivated from the fact that crowdsourced data annotation can be expensive, and is thus designed with minimal crowd involvement. For each iteration, we selected 50 tweets for keyword discovery and 50 tweets for micropost classification per keyword. For a dataset with 80k tweets (e.g., CyberAttack), our approach only requires to manually inspect 800 tweets (for 8 keywords), which is only 1% of the entire dataset.
Experiments and Results ::: Results of our Human-AI Loop (Q1)
Table TABREF26 reports the evaluation of our approach on both the CyberAttack and PoliticianDeath event categories. Our approach is configured such that each iteration starts with 1 new keyword discovered in the previous iteration.
Our approach improves LR by 5.17% (Accuracy) and 18.38% (AUC), and MLP by 10.71% (Accuracy) and 30.27% (AUC) on average. Such significant improvements clearly demonstrate that our approach is effective at improving model performance. We observe that the target models generally converge between the 7th and 9th iteration on both datasets when performance is measured by AUC. The performance can slightly degrade when the models are further trained for more iterations on both datasets. This is likely due to the fact that over time, the newly discovered keywords entail lower novel information for model training. For instance, for the CyberAttack dataset the new keyword in the 9th iteration `election' frequently co-occurs with the keyword `russia' in the 5th iteration (in microposts that connect Russian hackers with US elections), thus bringing limited new information for improving the model performance. As a side remark, we note that the models converge faster when performance is measured by accuracy. Such a comparison result confirms the difference between the metrics and shows the necessity for more keywords to discriminate event-related microposts from non event-related ones.
Experiments and Results ::: Comparative Results on Keyword Discovery (Q2)
Figure FIGREF31 shows the evaluation of our approach when discovering new informative keywords for model training (see Section SECREF2: Keyword Discovery). We compare our human-AI collaborative way of discovering new keywords against a query expansion (QE) approach BIBREF23, BIBREF24 that leverages word embeddings to find similar words in the latent semantic space. Specifically, we use pre-trained word embeddings based on a large Google News dataset for query expansion. For instance, the top keywords resulting from QE for `politician' are, `deputy',`ministry',`secretary', and `minister'. For each of these keywords, we use the crowd to label a set of tweets and obtain a corresponding expectation.
We observe that our approach consistently outperforms QE by an average of $4.62\%$ and $52.58\%$ AUC on CyberAttack and PoliticianDeath, respectively. The large gap between the performance improvements for the two datasets is mainly due to the fact that microposts that are relevant for PoliticianDeath are semantically more complex than those for CyberAttack, as they encode noun-verb relationship (e.g., “the king of ... died ...”) rather than a simple verb (e.g., “... hacked.”) for the CyberAttack microposts. QE only finds synonyms of existing keywords related to either `politician' or `death', however cannot find a meaningful keyword that fully characterizes the death of a politician. For instance, QE finds the keywords `kill' and `murder', which are semantically close to `death' but are not specifically relevant to the death of a politician. Unlike QE, our approach identifies keywords that go beyond mere synonyms and that are more directly related to the end task, i.e., discriminating event-related microposts from non related ones. Examples are `demise' and `condolence'. As a remark, we note that in Figure FIGREF31(b), the increase in QE performance on PoliticianDeath is due to the keywords `deputy' and `minister', which happen to be highly indicative of the death of a politician in our dataset; these keywords are also identified by our approach.
Experiments and Results ::: Cost-Effectiveness Results (Q3)
To demonstrate the cost-effectiveness of using crowdsourcing for obtaining new keywords and consequently, their expectations, we compare the performance of our approach with an approach using crowdsourcing to only label microposts for model training at the same cost. Specifically, we conducted an additional crowdsourcing experiment where the same cost used for keyword discovery in our approach is used to label additional microposts for model training. These newly labeled microposts are used with the microposts labeled in the micropost classification task of our approach (see Section SECREF2: Micropost Classification) and the expectation of the initial keyword to train the model for comparison. The model trained in this way increases AUC by 0.87% for CyberAttack, and by 1.06% for PoliticianDeath; in comparison, our proposed approach increases AUC by 33.42% for PoliticianDeath and by 15.23% for CyberAttack over the baseline presented by BIBREF1). These results show that using crowdsourcing for keyword discovery is significantly more cost-effective than simply using crowdsourcing to get additional labels when training the model.
Experiments and Results ::: Expectation Inference Results (Q4)
To investigate the effectiveness of our expectation inference method, we compare it against a majority voting approach, a strong baseline in truth inference BIBREF16. Figure FIGREF36 shows the result of this evaluation. We observe that our approach results in better models for both CyberAttack and PoliticianDeath. Our manual investigation reveals that workers' annotations are of high reliability, which explains the relatively good performance of majority voting. Despite limited margin for improvement, our method of expectation inference improves the performance of majority voting by $0.4\%$ and $1.19\%$ AUC on CyberAttack and PoliticianDeath, respectively.
Related Work
Event Detection. The techniques for event extraction from microblogging platforms can be classified according to their domain specificity and their detection method BIBREF0. Early works mainly focus on open domain event detection BIBREF25, BIBREF26, BIBREF27. Our work falls into the category of domain-specific event detection BIBREF21, which has drawn increasing attention due to its relevance for various applications such as cyber security BIBREF1, BIBREF2 and public health BIBREF4, BIBREF5. In terms of technique, our proposed detection method is related to the recently proposed weakly supervised learning methods BIBREF1, BIBREF17, BIBREF3. This comes in contrast with fully-supervised learning methods, which are often limited by the size of the training data (e.g., a few hundred examples) BIBREF28, BIBREF29.
Human-in-the-Loop Approaches. Our work extends weakly supervised learning methods by involving humans in the loop BIBREF13. Existing human-in-the-loop approaches mainly leverage crowds to label individual data instances BIBREF9, BIBREF10 or to debug the training data BIBREF30, BIBREF31 or components BIBREF32, BIBREF33, BIBREF34 of a machine learning system. Unlike these works, we leverage crowd workers to label sampled microposts in order to obtain keyword-specific expectations, which can then be generalized to help classify microposts containing the same keyword, thus amplifying the utility of the crowd. Our work is further connected to the topic of interpretability and transparency of machine learning models BIBREF11, BIBREF35, BIBREF12, for which humans are increasingly involved, for instance for post-hoc evaluations of the model's interpretability. In contrast, our approach directly solicits informative keywords from the crowd for model training, thereby providing human-understandable explanations for the improved model.
Conclusion
In this paper, we presented a new human-AI loop approach for keyword discovery and expectation estimation to better train event detection models. Our approach takes advantage of the disagreement between the crowd and the model to discover informative keywords and leverages the joint power of the crowd and the model in expectation inference. We evaluated our approach on real-world datasets and showed that it significantly outperforms the state of the art and that it is particularly useful for detecting events where relevant microposts are semantically complex, e.g., the death of a politician. As future work, we plan to parallelize the crowdsourcing tasks and optimize our pipeline in order to use our event detection approach in real-time.
Acknowledgements
This project has received funding from the Swiss National Science Foundation (grant #407540_167320 Tighten-it-All) and from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement 683253/GraphInt).
| [
"probabilistic model",
"Logistic Regression, Multilayer Perceptron"
] | 4,475 | qasper | en | null | 4a244628cbffa02d2240d412aeeed45c53fec95e66595a00 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What type of classifiers are used?
Answer: | [
"probabilistic model",
"Logistic Regression, Multilayer Perceptron"
] | qasper | 128 | 88 | |
Which toolkits do they use? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
As social media, specially Twitter, takes on an influential role in presidential elections in the U.S., natural language processing of political tweets BIBREF0 has the potential to help with nowcasting and forecasting of election results as well as identifying the main issues with a candidate – tasks of much interest to journalists, political scientists, and campaign organizers BIBREF1. As a methodology to obtain training data for a machine learning system that analyzes political tweets, BIBREF2 devised a crowdsourcing scheme with variable crowdworker numbers based on the difficulty of the annotation task. They provided a dataset of tweets where the sentiments towards political candidates were labeled both by experts in political communication and by crowdworkers who were likely not domain experts. BIBREF2 revealed that crowdworkers can match expert performance relatively accurately and in a budget-efficient manner. Given this result, the authors envisioned future work in which groundtruth labels would be crowdsourced for a large number of tweets and then used to design an automated NLP tool for political tweet analysis.
The question we address here is: How accurate are existing NLP tools for political tweet analysis? These tools would provide a baseline performance that any new machine learning system for political tweet analysis would compete against. We here explore whether existing NLP systems can answer the questions "What sentiment?" and "Towards whom?" accurately for the dataset of political tweets provided by BIBREF2. In our analysis, we include NLP tools with publicly-available APIs, even if the tools were not specifically designed for short texts like tweets, and, in particular, political tweets.
Our experiments reveal that the task of entity-level sentiment analysis is difficult for existing tools to answer accurately while the recognition of the entity, here, which politician, was easier.
NLP Toolkits
NLP toolkits typically have the following capabilities: tokenization, part-of-speech (PoS) tagging, chunking, named entity recognition and sentiment analysis. In a study by BIBREF3, it is shown that the well-known NLP toolkits NLTK BIBREF4, Stanford CoreNLP BIBREF5, and TwitterNLP BIBREF6 have tokenization, PoS tagging and NER modules in their pipelines. There are two main approaches for NER: (1) rule-based and (2) statistical or machine learning based. The most ubiquitous algorithms for sequence tagging use Hidden Markov Models BIBREF7, Maximum Entropy Markov Models BIBREF7, BIBREF8, or Conditional Random Fields BIBREF9. Recent works BIBREF10, BIBREF11 have used recurrent neural networks with attention modules for NER.
Sentiment detection tools like SentiStrength BIBREF12 and TensiStrength BIBREF13 are rule-based tools, relying on various dictionaries of emoticons, slangs, idioms, and ironic phrases, and set of rules that can detect the sentiment of a sentence overall or a targeted sentiment. Given a list of keywords, TensiStrength (similar to SentiStrength) reports the sentiment towards selected entities in a sentence, based on five levels of relaxation and five levels of stress.
Among commercial NLP toolkits (e.g., BIBREF14, BIBREF15, BIBREF16), we selected BIBREF17 and BIBREF18 for our experiments, which, to the best of our knowledge, are the only publicly accessible commercial APIs for the task of entity-level sentiment analysis that is agnostic to the text domain. We also report results of TensiStrength BIBREF13, TwitterNLP BIBREF6, BIBREF19, CogComp-NLP BIBREF20, and Stanford NLP NER BIBREF21.
Dataset and Analysis Methodology
We used the 1,000-tweet dataset by BIBREF2 that contains the named-entities labels and entity-level sentiments for each of the four 2016 presidential primary candidates Bernie Sanders, Donald Trump, Hillary Clinton, and Ted Cruz, provided by crowdworkers, and by experts in political communication, whose labels are considered groundtruth. The crowdworkers were located in the US and hired on the BIBREF22 platform. For the task of entity-level sentiment analysis, a 3-scale rating of "negative," "neutral," and "positive" was used by the annotators.
BIBREF2 proposed a decision tree approach for computing the number of crowdworkers who should analyze a tweet based on the difficulty of the task. Tweets are labeled by 2, 3, 5, or 7 workers based on the difficulty of the task and the level of disagreement between the crowdworkers. The model computes the number of workers based on how long a tweet is, the presence of a link in a tweet, and the number of present sarcasm signals. Sarcasm is often used in political tweets and causes disagreement between the crowdworkers. The tweets that are deemed to be sarcastic by the decision tree model, are expected to be more difficult to annotate, and hence are allocated more crowdworkers to work on.
We conducted two sets of experiments. In the first set, we used BIBREF23, BIBREF17, and BIBREF18, for entity-level sentiment analysis; in the second set, BIBREF17, BIBREF19, BIBREF24, BIBREF25, and BIBREF26, BIBREF18 for named-entity recognition.
In the experiments that we conducted with TwitterNLP for named-entity recognition, we worked with the default values of the model. Furthermore, we selected the 3-class Stanford NER model, which uses the classes “person,” “organization,” and “location” because it resulted in higher accuracy compared to the 7-class model. For CogComp-NLP NER we used Ontonotes 5.0 NER model BIBREF27. For spaCy NER we used the `en_core_web_lg' model.
We report the experimental results for our two tasks in terms of the correct classification rate (CCR). For sentiment analysis, we have a three-class problem (positive, negative, and neutral), where the classes are mutually exclusive. The CCR, averaged for a set of tweets, is defined to be the number of correctly-predicted sentiments over the number of groundtruth sentiments in these tweets. For NER, we consider that each tweet may reference up to four candidates, i.e., targeted entities. The CCR, averaged for a set of tweets, is the number of correctly predicted entities (candidates) over the number of groundtruth entities (candidates) in this set.
Results and Discussion
The dataset of 1,000 randomly selected tweets contains more than twice as many tweets about Trump than about the other candidates. In the named-entity recognition experiment, the average CCR of crowdworkers was 98.6%, while the CCR of the automated systems ranged from 77.2% to 96.7%. For four of the automated systems, detecting the entity Trump was more difficult than the other entities (e.g., spaCy 72.7% for the entity Trump vs. above 91% for the other entities). An example of incorrect NER is shown in Figure FIGREF1 top. The difficulties the automated tools had in NER may be explained by the fact that the tools were not trained on tweets, except for TwitterNLP, which was not in active development when the data was created BIBREF1.
In the sentiment analysis experiments, we found that a tweet may contain multiple sentiments. The groundtruth labels contain 210 positive sentiments, 521 neutral sentiments, and 305 negative sentiments to the candidates. We measured the CCR, across all tweets, to be 31.7% for Rosette Text Analytics, 43.2% for Google Cloud, 44.2% for TensiStrength, and 74.7% for the crowdworkers. This means the difference between the performance of the tools and the crowdworkers is significant – more than 30 percent points.
Crowdworkers correctly identified 62% of the neutral, 85% of the positive, and 92% of the negative sentiments. Google Cloud correctly identified 88% of the neutral sentiments, but only 3% of the positive, and 19% of the negative sentiments. TensiStrength correctly identified 87.2% of the neutral sentiments, but 10.5% of the positive, and 8.1% of the negative sentiments. Rosette Text Analytics correctly identified 22.7% of neutral sentiments, 38.1% of negative sentiments and 40.9% of positive sentiments. The lowest and highest CCR pertains to tweets about Trump and Sanders for both Google Cloud and TensiStrength, Trump and Clinton for Rosette Text Analytics, and Clinton and Cruz for crowdworkers. An example of incorrect ELS analysis is shown in Figure FIGREF1 bottom.
Conclusions and Future Work
Our results show that existing NLP systems cannot accurately perform sentiment analysis of political tweets in the dataset we experimented with. Labeling by humans, even non-expert crowdworkers, yields accuracy results that are well above the results of existing automated NLP systems. In future work we will therefore use a crowdworker-labeled dataset to train a new machine-learning based NLP system for tweet analysis. We will ensure that the training data is balanced among classes. Our plan is to use state-of-the-art deep neural networks and compare their performance for entity-level sentiment analysis of political tweets.
Acknowledgments
Partial support of this work by the Hariri Institute for Computing and Computational Science & Engineering at Boston University (to L.G.) and a Google Faculty Research Award (to M.B. and L.G.) is gratefully acknowledged. Additionally, we would like to thank Daniel Khashabi for his help in running the CogComp-NLP Python API and Mike Thelwal for his help with TensiStrength. We are also grateful to the Stanford NLP group for clarifying some of the questions we had with regards to the Stanford NER tool.
| [
"BIBREF17, BIBREF18, TensiStrength BIBREF13, TwitterNLP BIBREF6, BIBREF19, CogComp-NLP BIBREF20, Stanford NLP NER BIBREF21",
"BIBREF23, BIBREF17, BIBREF18, BIBREF19, BIBREF24, BIBREF25, BIBREF26"
] | 1,452 | qasper | en | null | 4f1d5a7b6c42269c845959f8a8d7ef95cb7e0c122c9a721c | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Which toolkits do they use?
Answer: | [
"BIBREF17, BIBREF18, TensiStrength BIBREF13, TwitterNLP BIBREF6, BIBREF19, CogComp-NLP BIBREF20, Stanford NLP NER BIBREF21",
"BIBREF23, BIBREF17, BIBREF18, BIBREF19, BIBREF24, BIBREF25, BIBREF26"
] | qasper | 128 | 89 | |
On what datasets are experiments performed? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Question Generation (QG) is the task of automatically creating questions from a range of inputs, such as natural language text BIBREF0, knowledge base BIBREF1 and image BIBREF2. QG is an increasingly important area in NLP with various application scenarios such as intelligence tutor systems, open-domain chatbots and question answering dataset construction. In this paper, we focus on question generation from reading comprehension materials like SQuAD BIBREF3. As shown in Figure FIGREF1, given a sentence in the reading comprehension paragraph and the text fragment (i.e., the answer) that we want to ask about, we aim to generate a question that is asked about the specified answer.
Question generation for reading comprehension is firstly formalized as a declarative-to-interrogative sentence transformation problem with predefined rules or templates BIBREF4, BIBREF0. With the rise of neural models, Du2017LearningTA propose to model this task under the sequence-to-sequence (Seq2Seq) learning framework BIBREF5 with attention mechanism BIBREF6. However, question generation is a one-to-many sequence generation problem, i.e., several aspects can be asked given a sentence. Zhou2017NeuralQG propose the answer-aware question generation setting which assumes the answer, a contiguous span inside the input sentence, is already known before question generation. To capture answer-relevant words in the sentence, they adopt a BIO tagging scheme to incorporate the answer position embedding in Seq2Seq learning. Furthermore, Sun2018AnswerfocusedAP propose that tokens close to the answer fragments are more likely to be answer-relevant. Therefore, they explicitly encode the relative distance between sentence words and the answer via position embedding and position-aware attention.
Although existing proximity-based answer-aware approaches achieve reasonable performance, we argue that such intuition may not apply to all cases especially for sentences with complex structure. For example, Figure FIGREF1 shows such an example where those approaches fail. This sentence contains a few facts and due to the parenthesis (i.e. “the area's coldest month”), some facts intertwine: “The daily mean temperature in January is 0.3$^\circ $C” and “January is the area's coldest month”. From the question generated by a proximity-based answer-aware baseline, we find that it wrongly uses the word “coldest” but misses the correct word “mean” because “coldest” has a shorter distance to the answer “0.3$^\circ $C”.
In summary, their intuition that “the neighboring words of the answer are more likely to be answer-relevant and have a higher chance to be used in the question” is not reliable. To quantitatively show this drawback of these models, we implement the approach proposed by Sun2018AnswerfocusedAP and analyze its performance under different relative distances between the answer and other non-stop sentence words that also appear in the ground truth question. The results are shown in Table TABREF2. We find that the performance drops at most 36% when the relative distance increases from “$0\sim 10$” to “$>10$”. In other words, when the useful context is located far away from the answer, current proximity-based answer-aware approaches will become less effective, since they overly emphasize neighboring words of the answer.
To address this issue, we extract the structured answer-relevant relations from sentences and propose a method to jointly model such structured relation and the unstructured sentence for question generation. The structured answer-relevant relation is likely to be to the point context and thus can help keep the generated question to the point. For example, Figure FIGREF1 shows our framework can extract the right answer-relevant relation (“The daily mean temperature in January”, “is”, “32.6$^\circ $F (0.3$^\circ $C)”) among multiple facts. With the help of such structured information, our model is less likely to be confused by sentences with a complex structure. Specifically, we firstly extract multiple relations with an off-the-shelf Open Information Extraction (OpenIE) toolbox BIBREF7, then we select the relation that is most relevant to the answer with carefully designed heuristic rules.
Nevertheless, it is challenging to train a model to effectively utilize both the unstructured sentence and the structured answer-relevant relation because both of them could be noisy: the unstructured sentence may contain multiple facts which are irrelevant to the target question, while the limitation of the OpenIE tool may produce less accurate extracted relations. To explore their advantages simultaneously and avoid the drawbacks, we design a gated attention mechanism and a dual copy mechanism based on the encoder-decoder framework, where the former learns to control the information flow between the unstructured and structured inputs, while the latter learns to copy words from two sources to maintain the informativeness and faithfulness of generated questions.
In the evaluations on the SQuAD dataset, our system achieves significant and consistent improvement as compared to all baseline methods. In particular, we demonstrate that the improvement is more significant with a larger relative distance between the answer and other non-stop sentence words that also appear in the ground truth question. Furthermore, our model is capable of generating diverse questions for a single sentence-answer pair where the sentence conveys multiple relations of its answer fragment.
Framework Description
In this section, we first introduce the task definition and our protocol to extract structured answer-relevant relations. Then we formalize the task under the encoder-decoder framework with gated attention and dual copy mechanism.
Framework Description ::: Problem Definition
We formalize our task as an answer-aware Question Generation (QG) problem BIBREF8, which assumes answer phrases are given before generating questions. Moreover, answer phrases are shown as text fragments in passages. Formally, given the sentence $S$, the answer $A$, and the answer-relevant relation $M$, the task of QG aims to find the best question $\overline{Q}$ such that,
where $A$ is a contiguous span inside $S$.
Framework Description ::: Answer-relevant Relation Extraction
We utilize an off-the-shelf toolbox of OpenIE to the derive structured answer-relevant relations from sentences as to the point contexts. Relations extracted by OpenIE can be represented either in a triple format or in an n-ary format with several secondary arguments, and we employ the latter to keep the extractions as informative as possible and avoid extracting too many similar relations in different granularities from one sentence. We join all arguments in the extracted n-ary relation into a sequence as our to the point context. Figure FIGREF5 shows n-ary relations extracted from OpenIE. As we can see, OpenIE extracts multiple relations for complex sentences. Here we select the most informative relation according to three criteria in the order of descending importance: (1) having the maximal number of overlapped tokens between the answer and the relation; (2) being assigned the highest confidence score by OpenIE; (3) containing maximum non-stop words. As shown in Figure FIGREF5, our criteria can select answer-relevant relations (waved in Figure FIGREF5), which is especially useful for sentences with extraneous information. In rare cases, OpenIE cannot extract any relation, we treat the sentence itself as the to the point context.
Table TABREF8 shows some statistics to verify the intuition that the extracted relations can serve as more to the point context. We find that the tokens in relations are 61% more likely to be used in the target question than the tokens in sentences, and thus they are more to the point. On the other hand, on average the sentences contain one more question token than the relations (1.86 v.s. 2.87). Therefore, it is still necessary to take the original sentence into account to generate a more accurate question.
Framework Description ::: Our Proposed Model ::: Overview.
As shown in Figure FIGREF10, our framework consists offour components (1) Sentence Encoder and Relation Encoder, (2) Decoder, (3) Gated Attention Mechanism and (4) Dual Copy Mechanism. The sentence encoder and relation encoder encode the unstructured sentence and the structured answer-relevant relation, respectively. To select and combine the source information from the two encoders, a gated attention mechanism is employed to jointly attend both contextualized information sources, and a dual copy mechanism copies words from either the sentence or the relation.
Framework Description ::: Our Proposed Model ::: Answer-aware Encoder.
We employ two encoders to integrate information from the unstructured sentence $S$ and the answer-relevant relation $M$ separately. Sentence encoder takes in feature-enriched embeddings including word embeddings $\mathbf {w}$, linguistic embeddings $\mathbf {l}$ and answer position embeddings $\mathbf {a}$. We follow BIBREF9 to transform POS and NER tags into continuous representation ($\mathbf {l}^p$ and $\mathbf {l}^n$) and adopt a BIO labelling scheme to derive the answer position embedding (B: the first token of the answer, I: tokens within the answer fragment except the first one, O: tokens outside of the answer fragment). For each word $w_i$ in the sentence $S$, we simply concatenate all features as input: $\mathbf {x}_i^s= [\mathbf {w}_i; \mathbf {l}^p_i; \mathbf {l}^n_i; \mathbf {a}_i]$. Here $[\mathbf {a};\mathbf {b}]$ denotes the concatenation of vectors $\mathbf {a}$ and $\mathbf {b}$.
We use bidirectional LSTMs to encode the sentence $(\mathbf {x}_1^s, \mathbf {x}_2^s, ..., \mathbf {x}_n^s)$ to get a contextualized representation for each token:
where $\overrightarrow{\mathbf {h}}^{s}_i$ and $\overleftarrow{\mathbf {h}}^{s}_i$ are the hidden states at the $i$-th time step of the forward and the backward LSTMs. The output state of the sentence encoder is the concatenation of forward and backward hidden states: $\mathbf {h}^{s}_i=[\overrightarrow{\mathbf {h}}^{s}_i;\overleftarrow{\mathbf {h}}^{s}_i]$. The contextualized representation of the sentence is $(\mathbf {h}^{s}_1, \mathbf {h}^{s}_2, ..., \mathbf {h}^{s}_n)$.
For the relation encoder, we firstly join all items in the n-ary relation $M$ into a sequence. Then we only take answer position embedding as an extra feature for the sequence: $\mathbf {x}_i^m= [\mathbf {w}_i; \mathbf {a}_i]$. Similarly, we take another bidirectional LSTMs to encode the relation sequence and derive the corresponding contextualized representation $(\mathbf {h}^{m}_1, \mathbf {h}^{m}_2, ..., \mathbf {h}^{m}_n)$.
Framework Description ::: Our Proposed Model ::: Decoder.
We use an LSTM as the decoder to generate the question. The decoder predicts the word probability distribution at each decoding timestep to generate the question. At the t-th timestep, it reads the word embedding $\mathbf {w}_{t}$ and the hidden state $\mathbf {u}_{t-1}$ of the previous timestep to generate the current hidden state:
Framework Description ::: Our Proposed Model ::: Gated Attention Mechanism.
We design a gated attention mechanism to jointly attend the sentence representation and the relation representation. For sentence representation $(\mathbf {h}^{s}_1, \mathbf {h}^{s}_2, ..., \mathbf {h}^{s}_n)$, we employ the Luong2015EffectiveAT's attention mechanism to obtain the sentence context vector $\mathbf {c}^s_t$,
where $\mathbf {W}_a$ is a trainable weight. Similarly, we obtain the vector $\mathbf {c}^m_t$ from the relation representation $(\mathbf {h}^{m}_1, \mathbf {h}^{m}_2, ..., \mathbf {h}^{m}_n)$. To jointly model the sentence and the relation, a gating mechanism is designed to control the information flow from two sources:
where $\odot $ represents element-wise dot production and $\mathbf {W}_g, \mathbf {W}_h$ are trainable weights. Finally, the predicted probability distribution over the vocabulary $V$ is computed as:
where $\mathbf {W}_V$ and $\mathbf {b}_V$ are parameters.
Framework Description ::: Our Proposed Model ::: Dual Copy Mechanism.
To deal with the rare and unknown words, the decoder applies the pointing method BIBREF10, BIBREF11, BIBREF12 to allow copying a token from the input sentence at the $t$-th decoding step. We reuse the attention score $\mathbf {\alpha }_{t}^s$ and $\mathbf {\alpha }_{t}^m$ to derive the copy probability over two source inputs:
Different from the standard pointing method, we design a dual copy mechanism to copy from two sources with two gates. The first gate is designed for determining copy tokens from two sources of inputs or generate next word from $P_V$, which is computed as $g^v_t = \text{sigmoid}(\mathbf {w}^v_g \tilde{\mathbf {h}}_t + b^v_g)$. The second gate takes charge of selecting the source (sentence or relation) to copy from, which is computed as $g^c_t = \text{sigmoid}(\mathbf {w}^c_g [\mathbf {c}_t^s;\mathbf {c}_t^m] + b^c_g)$. Finally, we combine all probabilities $P_V$, $P_S$ and $P_M$ through two soft gates $g^v_t$ and $g^c_t$. The probability of predicting $w$ as the $t$-th token of the question is:
Framework Description ::: Our Proposed Model ::: Training and Inference.
Given the answer $A$, sentence $S$ and relation $M$, the training objective is to minimize the negative log-likelihood with regard to all parameters:
where $\mathcal {\lbrace }Q\rbrace $ is the set of all training instances, $\theta $ denotes model parameters and $\text{log} P(Q|A,S,M;\theta )$ is the conditional log-likelihood of $Q$.
In testing, our model targets to generate a question $Q$ by maximizing:
Experimental Setting ::: Dataset & Metrics
We conduct experiments on the SQuAD dataset BIBREF3. It contains 536 Wikipedia articles and 100k crowd-sourced question-answer pairs. The questions are written by crowd-workers and the answers are spans of tokens in the articles. We employ two different data splits by following Zhou2017NeuralQG and Du2017LearningTA . In Zhou2017NeuralQG, the original SQuAD development set is evenly divided into dev and test sets, while Du2017LearningTA treats SQuAD development set as its development set and splits original SQuAD training set into a training set and a test set. We also filter out questions which do not have any overlapped non-stop words with the corresponding sentences and perform some preprocessing steps, such as tokenization and sentence splitting. The data statistics are given in Table TABREF27.
We evaluate with all commonly-used metrics in question generation BIBREF13: BLEU-1 (B1), BLEU-2 (B2), BLEU-3 (B3), BLEU-4 (B4) BIBREF17, METEOR (MET) BIBREF18 and ROUGE-L (R-L) BIBREF19. We use the evaluation script released by Chen2015MicrosoftCC.
Experimental Setting ::: Baseline Models
We compare with the following models.
[leftmargin=*]
s2s BIBREF13 proposes an attention-based sequence-to-sequence neural network for question generation.
NQG++ BIBREF9 takes the answer position feature and linguistic features into consideration and equips the Seq2Seq model with copy mechanism.
M2S+cp BIBREF14 conducts multi-perspective matching between the answer and the sentence to derive an answer-aware sentence representation for question generation.
s2s+MP+GSA BIBREF8 introduces a gated self-attention into the encoder and a maxout pointer mechanism into the decoder. We report their sentence-level results for a fair comparison.
Hybrid BIBREF15 is a hybrid model which considers the answer embedding for the question word generation and the position of context words for modeling the relative distance between the context words and the answer.
ASs2s BIBREF16 replaces the answer in the sentence with a special token to avoid its appearance in the generated questions.
Experimental Setting ::: Implementation Details
We take the most frequent 20k words as our vocabulary and use the GloVe word embeddings BIBREF20 for initialization. The embedding dimensions for POS, NER, answer position are set to 20. We use two-layer LSTMs in both encoder and decoder, and the LSTMs hidden unit size is set to 600.
We use dropout BIBREF21 with the probability $p=0.3$. All trainable parameters, except word embeddings, are randomly initialized with the Xavier uniform in $(-0.1, 0.1)$ BIBREF22. For optimization in the training, we use SGD as the optimizer with a minibatch size of 64 and an initial learning rate of 1.0. We train the model for 15 epochs and start halving the learning rate after the 8th epoch. We set the gradient norm upper bound to 3 during the training.
We adopt the teacher-forcing for the training. In the testing, we select the model with the lowest perplexity and beam search with size 3 is employed for generating questions. All hyper-parameters and models are selected on the validation dataset.
Results and Analysis ::: Main Results
Table TABREF30 shows automatic evaluation results for our model and baselines (copied from their papers). Our proposed model which combines structured answer-relevant relations and unstructured sentences achieves significant improvements over proximity-based answer-aware models BIBREF9, BIBREF15 on both dataset splits. Presumably, our structured answer-relevant relation is a generalization of the context explored by the proximity-based methods because they can only capture short dependencies around answer fragments while our extractions can capture both short and long dependencies given the answer fragments. Moreover, our proposed framework is a general one to jointly leverage structured relations and unstructured sentences. All compared baseline models which only consider unstructured sentences can be further enhanced under our framework.
Recall that existing proximity-based answer-aware models perform poorly when the distance between the answer fragment and other non-stop sentence words that also appear in the ground truth question is large (Table TABREF2). Here we investigate whether our proposed model using the structured answer-relevant relations can alleviate this issue or not, by conducting experiments for our model under the same setting as in Table TABREF2. The broken-down performances by different relative distances are shown in Table TABREF40. We find that our proposed model outperforms Hybrid (our re-implemented version for this experiment) on all ranges of relative distances, which shows that the structured answer-relevant relations can capture both short and long term answer-relevant dependencies of the answer in sentences. Furthermore, comparing the performance difference between Hybrid and our model, we find the improvements become more significant when the distance increases from “$0\sim 10$” to “$>10$”. One reason is that our model can extract relations with distant dependencies to the answer, which greatly helps our model ignore the extraneous information. Proximity-based answer-aware models may overly emphasize the neighboring words of answers and become less effective as the useful context becomes further away from the answer in the complex sentences. In fact, the breakdown intervals in Table TABREF40 naturally bound its sentence length, say for “$>10$”, the sentences in this group must be longer than 10. Thus, the length variances in these two intervals could be significant. To further validate whether our model can extract long term dependency words. We rerun the analysis of Table TABREF40 only for long sentences (length $>$ 20) of each interval. The improvement percentages over Hybrid are shown in Table TABREF40, which become more significant when the distance increases from “$0\sim 10$” to “$>10$”.
Results and Analysis ::: Case Study
Figure FIGREF42 provides example questions generated by crowd-workers (ground truth questions), the baseline Hybrid BIBREF15, and our model. In the first case, there are two subsequences in the input and the answer has no relation with the second subsequence. However, we see that the baseline model prediction copies irrelevant words “The New York Times” while our model can avoid using the extraneous subsequence “The New York Times noted ...” with the help of the structured answer-relevant relation. Compared with the ground truth question, our model cannot capture the cross-sentence information like “her fifth album”, where the techniques in paragraph-level QG models BIBREF8 may help. In the second case, as discussed in Section SECREF1, this sentence contains a few facts and some facts intertwine. We find that our model can capture distant answer-relevant dependencies such as “mean temperature” while the proximity-based baseline model wrongly takes neighboring words of the answer like “coldest” in the generated question.
Results and Analysis ::: Diverse Question Generation
Another interesting observation is that for the same answer-sentence pair, our model can generate diverse questions by taking different answer-relevant relations as input. Such capability improves the interpretability of our model because the model is given not only what to be asked (i.e., the answer) but also the related fact (i.e., the answer-relevant relation) to be covered in the question. In contrast, proximity-based answer-aware models can only generate one question given the sentence-answer pair regardless of how many answer-relevant relations in the sentence. We think such capability can also validate our motivation: questions should be generated according to the answer-aware relations instead of neighboring words of answer fragments. Figure FIGREF45 show two examples of diverse question generation. In the first case, the answer fragment `Hugh L. Dryden' is the appositive to `NASA Deputy Administrator' but the subject to the following tokens `announced the Apollo program ...'. Our framework can extract these two answer-relevant relations, and by feeding them to our model separately, we can receive two questions asking different relations with regard to the answer.
Related Work
The topic of question generation, initially motivated for educational purposes, is tackled by designing many complex rules for specific question types BIBREF4, BIBREF23. Heilman2010GoodQS improve rule-based question generation by introducing a statistical ranking model. First, they remove extraneous information in the sentence to transform it into a simpler one, which can be transformed easily into a succinct question with predefined sets of general rules. Then they adopt an overgenerate-and-rank approach to select the best candidate considering several features.
With the rise of dominant neural sequence-to-sequence learning models BIBREF5, Du2017LearningTA frame question generation as a sequence-to-sequence learning problem. Compared with rule-based approaches, neural models BIBREF24 can generate more fluent and grammatical questions. However, question generation is a one-to-many sequence generation problem, i.e., several aspects can be asked given a sentence, which confuses the model during train and prevents concrete automatic evaluation. To tackle this issue, Zhou2017NeuralQG propose the answer-aware question generation setting which assumes the answer is already known and acts as a contiguous span inside the input sentence. They adopt a BIO tagging scheme to incorporate the answer position information as learned embedding features in Seq2Seq learning. Song2018LeveragingCI explicitly model the information between answer and sentence with a multi-perspective matching model. Kim2019ImprovingNQ also focus on the answer information and proposed an answer-separated Seq2Seq model by masking the answer with special tokens. All answer-aware neural models treat question generation as a one-to-one mapping problem, but existing models perform poorly for sentences with a complex structure (as shown in Table TABREF2).
Our work is inspired by the process of extraneous information removing in BIBREF0, BIBREF25. Different from Heilman2010GoodQS which directly use the simplified sentence for generation and cao2018faithful which only consider aggregate two sources of information via gated attention in summarization, we propose to combine the structured answer-relevant relation and the original sentence. Factoid question generation from structured text is initially investigated by Serban2016GeneratingFQ, but our focus here is leveraging structured inputs to help question generation over unstructured sentences. Our proposed model can take advantage of unstructured sentences and structured answer-relevant relations to maintain informativeness and faithfulness of generated questions. The proposed model can also be generalized in other conditional sequence generation tasks which require multiple sources of inputs, e.g., distractor generation for multiple choice questions BIBREF26.
Conclusions and Future Work
In this paper, we propose a question generation system which combines unstructured sentences and structured answer-relevant relations for generation. The unstructured sentences maintain the informativeness of generated questions while structured answer-relevant relations keep the faithfulness of questions. Extensive experiments demonstrate that our proposed model achieves state-of-the-art performance across several metrics. Furthermore, our model can generate diverse questions with different structured answer-relevant relations. For future work, there are some interesting dimensions to explore, such as difficulty levels BIBREF27, paragraph-level information BIBREF8 and conversational question generation BIBREF28.
Acknowledgments
This work is supported by the Research Grants Council of the Hong Kong Special Administrative Region, China (No. CUHK 14208815 and No. CUHK 14210717 of the General Research Fund). We would like to thank the anonymous reviewers for their comments. We would also like to thank Department of Computer Science and Engineering, The Chinese University of Hong Kong for the conference grant support.
| [
"SQuAD",
"SQuAD"
] | 3,757 | qasper | en | null | 5f7af98db66df4388108e26cde4781423ca2580bb48de4fa | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: On what datasets are experiments performed?
Answer: | [
"SQuAD",
"SQuAD"
] | qasper | 128 | 90 | |
what are the existing approaches? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Users of photo-sharing websites such as Flickr often provide short textual descriptions in the form of tags to help others find the images. With the availability of GPS systems in current electronic devices such as smartphones, latitude and longitude coordinates are nowadays commonly made available as well. The tags associated with such georeferenced photos often describe the location where these photos were taken, and Flickr can thus be regarded as a source of environmental information. The use of Flickr for modelling urban environments has already received considerable attention. For instance, various approaches have been proposed for modelling urban regions BIBREF0 , and for identifying points-of-interest BIBREF1 and itineraries BIBREF2 , BIBREF3 . However, the usefulness of Flickr for characterizing the natural environment, which is the focus of this paper, is less well-understood.
Many recent studies have highlighted that Flickr tags capture valuable ecological information, which can be used as a complementary source to more traditional sources. To date, however, ecologists have mostly used social media to conduct manual evaluations of image content with little automated exploitation of the associated tags BIBREF4 , BIBREF5 , BIBREF6 . One recent exception is BIBREF7 , where bag-of-words representations derived from Flickr tags were found to give promising result for predicting a range of different environemental phenomena.
Our main hypothesis in this paper is that by using vector space embeddings instead of bag-of-words representations, the ecological information which is implicitly captured by Flickr tags can be utilized in a more effective way. Vector space embeddings are representations in which the objects from a given domain are encoded using relatively low-dimensional vectors. They have proven useful in natural language processing, especially for encoding word meaning BIBREF8 , BIBREF9 , and in machine learning more generally. In this paper, we are interested in the use of such representations for modelling geographic locations. Our main motivation for using vector space embeddings is that they allow us to integrate the textual information we get from Flickr with available structured information in a very natural way. To this end, we rely on an adaptation of the GloVe word embedding model BIBREF9 , but rather than learning word vectors, we learn vectors representing locations. Similar to how the representation of a word in GloVe is determined by the context words surrounding it, the representation of a location in our model is determined by the tags of the photos that have been taken near that location. To incorporate numerical features from structured environmental datasets (e.g. average temperature), we associate with each such feature a linear mapping that can be used to predict that feature from a given location vector. This is inspired by the fact that salient properties of a given domain can often be modelled as directions in vector space embeddings BIBREF10 , BIBREF11 , BIBREF12 . Finally, evidence from categorical datasets (e.g. land cover types) is taken into account by requiring that locations belonging to the same category are represented using similar vectors, similar to how semantic types are sometimes modelled in the context of knowledge graph embedding BIBREF13 .
While our point-of-departure is a standard word embedding model, we found that the off-the-shelf GloVe model performed surprisingly poorly, meaning that a number of modifications are needed to achieve good results. Our main findings are as follows. First, given that the number of tags associated with a given location can be quite small, it is important to apply some kind of spatial smoothing, i.e. the importance of a given tag for a given location should not only depend on the occurrences of the tag at that location, but also on its occurrences at nearby locations. To this end, we use a formulation which is based on spatially smoothed version of pointwise mutual information. Second, given the wide diversity in the kind of information that is covered by Flickr tags, we find that term selection is in some cases critical to obtain vector spaces that capture the relevant aspects of geographic locations. For instance, many tags on Flickr refer to photography related terms, which we would normally not want to affect the vector representation of a given location. Finally, even with these modifications, vector space embeddings learned from Flickr tags alone are sometimes outperformed by bag-of-words representations. However, our vector space embeddings lead to substantially better predictions in cases where structured (scientific) information is also taken into account. In this sense, the main value of using vector space embeddings in this context is not so much about abstracting away from specific tag usages, but rather about the fact that such representations allow us to integrate numerical and categorical features in a much more natural way than is possible with bag-of-words representations.
The remainder of this paper is organized as follows. In the next section, we provide a discussion of existing work. Section SECREF3 then presents our model for embedding geographic locations from Flickr tags and structured data. Next, in Section SECREF4 we provide a detailed discussion about the experimental results. Finally, Section SECREF5 summarizes our conclusions.
Vector space embeddings
The use of low-dimensional vector space embeddings for representing objects has already proven effective in a large number of applications, including natural language processing (NLP), image processing, and pattern recognition. In the context of NLP, the most prominent example is that of word embeddings, which represent word meaning using vectors of typically around 300 dimensions. A large number of different methods for learning such word embeddings have already been proposed, including Skip-gram and the Continuous Bag-of-Words (CBOW) model BIBREF8 , GloVe BIBREF9 , and fastText BIBREF14 . They have been applied effectively in many downstream NLP tasks such as sentiment analysis BIBREF15 , part of speech tagging BIBREF16 , BIBREF17 , and text classification BIBREF18 , BIBREF19 . The model we consider in this paper builds on GloVe, which was designed to capture linear regularities of word-word co-occurrence. In GloVe, there are two word vectors INLINEFORM0 and INLINEFORM1 for each word in the vocabulary, which are learned by minimizing the following objective: DISPLAYFORM0
where INLINEFORM0 is the number of times that word INLINEFORM1 appears in the context of word INLINEFORM2 , INLINEFORM3 is the vocabulary size, INLINEFORM4 is the target word bias, INLINEFORM5 is the context word bias. The weighting function INLINEFORM6 is used to limit the impact of rare terms. It is defined as 1 if INLINEFORM7 and as INLINEFORM8 otherwise, where INLINEFORM9 is usually fixed to 100 and INLINEFORM10 to 0.75. Intuitively, the target word vectors INLINEFORM11 correspond to the actual word representations which we would like to find, while the context word vectors INLINEFORM12 model how occurrences of INLINEFORM13 in the context of a given word INLINEFORM14 affect the representation of this latter word. In this paper we will use a similar model, which will however be aimed at learning location vectors instead of the target word vectors.
Beyond word embeddings, various methods have been proposed for learning vector space representations from structured data such as knowledge graphs BIBREF20 , BIBREF21 , BIBREF22 , social networks BIBREF23 , BIBREF24 and taxonomies BIBREF25 , BIBREF26 . The idea of combining a word embedding model with structured information has also been explored by several authors, for example to improve the word embeddings based on information coming from knowledge graphs BIBREF27 , BIBREF28 . Along similar lines, various lexicons have been used to obtain word embeddings that are better suited at modelling sentiment BIBREF15 and antonymy BIBREF29 , among others. The method proposed by BIBREF30 imposes the condition that words that belong to the same semantic category are closer together than words from different categories, which is somewhat similar in spirit to how we will model categorical datasets in our model.
Embeddings for geographic information
The problem of representing geographic locations using embeddings has also attracted some attention. An early example is BIBREF31 , which used principal component analysis and stacked autoencoders to learn low-dimensional vector representations of city neighbourhoods based on census data. They use these representations to predict attributes such as crime, which is not included in the given census data, and find that in most of the considered evaluation tasks, the low-dimensional vector representations lead to more faithful predictions than the original high-dimensional census data.
Some existing works combine word embedding models with geographic coordinates. For example, in BIBREF32 an approach is proposed to learn word embeddings based on the assumption that words which tend to be used in the same geographic locations are likely to be similar. Note that their aim is dual to our aim in this paper: while they use geographic location to learn word vectors, we use textual descriptions to learn vectors representing geographic locations.
Several methods also use word embedding models to learn representations of Points-of-Interest (POIs) that can be used for predicting user visits BIBREF33 , BIBREF34 , BIBREF35 . These works use the machinery of existing word embedding models to learn POI representations, intuitively by letting sequences of POI visits by a user play the role of sequences of words in a sentence. In other words, despite the use of word embedding models, many of these approaches do not actually consider any textual information. For example, in BIBREF34 the Skip-gram model is utilized to create a global pattern of users' POIs. Each location was treated as a word and the other locations visited before or after were treated as context words. They then use a pair-wise ranking loss BIBREF36 which takes into account the user's location visit frequency to personalize the location recommendations. The methods of BIBREF34 were extended in BIBREF35 to use a temporal embedding and to take more account of geographic context, in particular the distances between preferred and non-preferred neighboring POIs, to create a “geographically hierarchical pairwise preference ranking model”. Similarly, in BIBREF37 the CBOW model was trained with POI data. They ordered POIs spatially within the traffic-based zones of urban areas. The ordering was used to generate characteristic vectors of POI types. Zone vectors represented by averaging the vectors of the POIs contained in them, were then used as features to predict land use types. In the CrossMap method BIBREF38 they learned embeddings for spatio-temporal hotspots obtained from social media data of locations, times and text. In one form of embedding, intended to enable reconstruction of records, neighbourhood relations in space and time were encoded by averaging hotspots in a target location's spatial and temporal neighborhoods. They also proposed a graph-based embedding method with nodes of location, time and text. The concatenation of the location, time and text vectors were then used as features to predict peoples' activities in urban environments. Finally, in BIBREF39 , a method is proposed that uses the Skip-gram model to represent POI types, based on the intuition that the vector representing a given POI type should be predictive of the POI types that found near places of that type.
Our work is different from these studies, as our focus is on representing locations based on a given text description of that location (in the form of Flickr tags), along with numerical and categorical features from scientific datasets.
Analyzing Flickr tags
Many studies have focused on analyzing Flickr tags to extract useful information in domains such as linguistics BIBREF40 , geography BIBREF0 , BIBREF41 , and ecology BIBREF42 , BIBREF7 , BIBREF43 . Most closely related to our work, BIBREF7 found that the tags of georeferenced Flickr photos can effectively supplement traditional scientific environmental data in tasks such as predicting climate features, land cover, species occurrence, and human assessments of scenicness. To encode locations, they simply combine a bag-of-words representation of geographically nearby tags with a feature vector that encodes associated structured scientific data. They found that the predictive value of Flickr tags is roughly on a par with that of the scientific datasets, and that combining both types of information leads to significantly better results than using either of them alone. As we show in this paper, however, their straightforward way of combining both information sources, by concatenating the two types of feature vectors, is far from optimal.
Despite the proven importance of Flickr tags, the problem of embedding Flickr tags has so far received very limited attention. To the best of our knowledge, BIBREF44 is the only work that generated embeddings for Flickr tags. However, their focus was on learning embeddings that capture word meaning (being evaluated on word similarity tasks), whereas we use such embeddings as part of our method for representing locations.
Model Description
In this section, we introduce our embedding model, which combines Flickr tags and structured scientific information to represent a set of locations INLINEFORM0 . The proposed model has the following form: DISPLAYFORM0
where INLINEFORM0 and INLINEFORM1 are parameters to control the importance of each component in the model. Component INLINEFORM2 will be used to constrain the representation of the locations based on their textual description (i.e. Flickr tags), INLINEFORM3 will be used to constrain the representation of the locations based on their numerical features, and INLINEFORM4 will impose the constraint that locations belonging to the same category should be close together in the space. We will discuss each of these components in more detail in the following sections.
Tag Based Location Embedding
Many of the tags associated with Flickr photos describe characteristics of the places where these photos were taken BIBREF45 , BIBREF46 , BIBREF47 . For example, tags may correspond to place names (e.g. Brussels, England, Scandinavia), landmarks (e.g. Eiffel Tower, Empire State Building) or land cover types (e.g. mountain, forest, beach). To allow us to build location models using such tags, we collected the tags and meta-data of 70 million Flickr photos with coordinates in Europe (which is the region our experiments will focus on), all of which were uploaded to Flickr before the end of September 2015. In this section we first explain how tags can be weighted to obtain bag-of-words representations of locations from Flickr. Subsequently we describe a tag selection method, which will allow us to specialize the embedding depending on which aspects of the considered locations are of interest, after which we discuss the actual embedding model.
Tag weighting. Let INLINEFORM0 be a set of geographic locations, each characterized by latitude and longitude coordinates. To generate a bag-of-words representation of a given location, we have to weight the relevance of each tag to that location. To this end, we have followed the weighting scheme from BIBREF7 , which combines a Gaussian kernel (to model spatial proximity) with Positive Pointwise Mutual Information (PPMI) BIBREF48 , BIBREF49 .
Let us write INLINEFORM0 for the set of users who have assigned tag INLINEFORM1 to a photo with coordinates near INLINEFORM2 . To assess how relevant INLINEFORM3 is to the location INLINEFORM4 , the number of times INLINEFORM5 occurs in photos near INLINEFORM6 is clearly an important criterion. However, rather than simply counting the number of occurrences within some fixed radius, we use a Gaussian kernel to weight the tag occurrences according to their distance from that location: INLINEFORM7
where the threshold INLINEFORM0 is assumed to be fixed, INLINEFORM1 is the location of a Flickr photo, INLINEFORM2 is the Haversine distance, and we will assume that the bandwidth parameter INLINEFORM3 is set to INLINEFORM4 . A tag occurrence is counted only once for all photos by the same user at the same location, which is important to reduce the impact of bulk uploading. The value INLINEFORM5 reflects how frequent tag INLINEFORM6 is near location INLINEFORM7 , but it does not yet take into account the total number of tag occurrences near INLINEFORM8 , nor how popular the tag INLINEFORM9 is overall. To measure how strongly tag INLINEFORM10 is associated with location INLINEFORM11 , we use PPMI, which is a commonly used measure of association in natural language processing. However, rather than estimating PPMI scores from term frequencies, we will use the INLINEFORM12 values instead: INLINEFORM13
where: INLINEFORM0
with INLINEFORM0 the set of all tags, and INLINEFORM1 the set of locations.
Tag selection. Inspired by BIBREF50 , we use a term selection method in order to focus on the tags that are most important for the tasks that we want to consider and reduce the impact of tags that might relate only to a given individual or a group of users. In particular, we obtained good results with a method based on Kullback-Leibler (KL) divergence, which is based on BIBREF51 . Let INLINEFORM0 be a set of (mutually exclusive) properties of locations in which we are interested (e.g. land cover categories). For the ease of presentation, we will identify INLINEFORM1 with the set of locations that have the corresponding property. Then, we select tags from INLINEFORM2 that maximize the following score: INLINEFORM3
where INLINEFORM0 is the probability that a photo with tag INLINEFORM1 has a location near INLINEFORM2 and INLINEFORM3 is the probability that an arbitrary tag occurrence is assigned to a photo near a location in INLINEFORM4 . Since INLINEFORM5 often has to be estimated from a small number of tag occurrences, it is estimated using Bayesian smoothing: INLINEFORM6
where INLINEFORM0 is a parameter controlling the amount of smoothing, which will be tuned in the experiments. On the other hand, for INLINEFORM1 we can simply use a maximum likelihood estimation: INLINEFORM2
Location embedding. We now want to find a vector INLINEFORM0 for each location INLINEFORM1 such that similar locations are represented using similar vectors. To achieve this, we use a close variant of the GloVe model, where tag occurrences are treated as context words of geographic locations. In particular, with each location INLINEFORM2 we associate a vector INLINEFORM3 and with each tag INLINEFORM4 we associate a vector INLINEFORM5 and a bias term INLINEFORM6 , and consider the following objective (which in our full model ( EQREF7 ) will be combined with components that are derived from the structured information): INLINEFORM7
Note how tags play the role of the context words in the GloVe model, while instead of learning target word vectors we now learn location vectors. In contrast to GloVe, our objective does not directly refer to co-occurrence statistics, but instead uses the INLINEFORM0 scores. One important consequence of this is that we can also consider pairs INLINEFORM1 for which INLINEFORM2 does not occur in INLINEFORM3 at all; such pairs are usually called negative examples. While they cannot be used in the standard GloVe model, some authors have already reported that introducing negative examples in variants of GloVe can lead to an improvement BIBREF52 . In practice, evaluating the full objective above would not be computationally feasible, as we may need to consider millions of locations and millions of tags. Therefore, rather than considering all tags in INLINEFORM4 for the inner summation, we only consider those tags that appear at least once near location INLINEFORM5 together with a sample of negative examples.
Structured Environmental Data
There is a wide variety of structured data that can be used to describe locations. In this work, we have restricted ourselves to the same datasets as BIBREF7 . These include nine (real-valued) numerical features, which are latitude, longitude, elevation, population, and five climate related features (avg. temperature, avg. precipitation, avg. solar radiation, avg. wind speed, and avg. water vapor pressure). In addition, 180 categorical features were used, which are CORINE land cover classes at level 1 (5 classes), level 2 (15 classes) and level 3 (44 classes) and 116 soil types (SoilGrids). Note that each location should belong to exactly 4 categories: one CORINE class at each of the three levels and a soil type.
Numerical features. Numerical features can be treated similarly to the tag occurrences, i.e. we will assume that the value of a given numerical feature can be predicted from the location vectors using a linear mapping. In particular, for each numerical feature INLINEFORM0 we consider a vector INLINEFORM1 and a bias term INLINEFORM2 , and the following objective: INLINEFORM3
where we write INLINEFORM0 for set of all numerical features and INLINEFORM1 is the value of feature INLINEFORM2 for location INLINEFORM3 , after z-score normalization.
Categorical features. To take into account the categorical features, we impose the constraint that locations belonging to the same category should be close together in the space. To formalize this, we represent each category type INLINEFORM0 as a vector INLINEFORM1 , and consider the following objective: INLINEFORM2
Evaluation Tasks
We will use the method from BIBREF7 as our main baseline. This will allow us to directly evaluate the effectiveness of embeddings for the considered problem, since we have used the same structured datasets and same tag weighting scheme. For this reason, we will also follow their evaluation methodology. In particular, we will consider three evaluation tasks:
Predicting the distribution of 100 species across Europe, using the European network of nature protected sites Natura 2000 dataset as ground truth. For each of these species, a binary classification problem is considered. The set of locations INLINEFORM0 is defined as the 26,425 distinct sites occurring in the dataset.
Predicting soil type, again each time treating the task as a binary classification problem, using the same set of locations INLINEFORM0 as in the species distribution experiments. For these experiments, none of the soil type features are used for generating the embeddings.
Predicting CORINE land cover classes at levels 1, 2 and level 3, each time treating the task as a binary classification problem, using the same set of locations INLINEFORM0 as in the species distribution experiments. For these experiments, none of the CORINE features are used for generating the embeddings.
In addition, we will also consider the following regression tasks:
Predicting 5 climate related features: the average precipitation, temperature, solar radiation, water vapor pressure, and wind speed. We again use the same set of locations INLINEFORM0 as for species distribution in this experiment. None of the climate features is used for constructing the embeddings for this experiment.
Predicting people's subjective opinions of landscape beauty in Britain, using the crowdsourced dataset from the ScenicOrNot website as ground truth. The set INLINEFORM0 is chosen as the set of locations of 191 605 rated locations from the ScenicOrNot dataset for which at least one georeferenced Flickr photo exists within a 1 km radius.
Experimental Setup
In all experiments, we use Support Vector Machines (SVMs) for classification problems and Support Vector Regression (SVR) for regression problems to make predictions from our representations of geographic locations. In both cases, we used the SVM INLINEFORM0 implementation BIBREF53 . For each experiment, the set of locations INLINEFORM1 was split into two-thirds for training, one-sixth for testing, and one-sixth for tuning the parameters. All embedding models are learned with Adagrad using 30 iterations. The number of dimensions is chosen for each experiment from INLINEFORM2 based on the tuning data. For the parameters of our model in Equation EQREF7 , we considered values of INLINEFORM3 from {0.1, 0.01, 0.001, 0.0001} and values of INLINEFORM4 from {1, 10, 100, 1000, 10 000, 100 000}. To compute KL divergence, we need to determine a set of classes INLINEFORM5 for each experiment. For classification problems, we can simply consider the given categories, but for the regression problems we need to define such classes by discretizing the numerical values. For the scenicness experiments, we considered scores 3 and 7 as cut-off points, leading to three classes (i.e. less than 3, between 3 and 7, and above 7). Similarly, for each climate related features, we consider two cut-off values for discretization: 5 and 15 for average temperature, 50 and 100 for average precipitation, 10 000 and 17 000 for average solar radiation, 0.7 and 1 for average water vapor pressure, and 3 and 5 for wind speed. The smoothing parameter INLINEFORM6 was selected among INLINEFORM7 based on the tuning data. In all experiments where term selection is used, we select the top 100 000 tags. We fixed the radius INLINEFORM8 at 1km when counting the number of tag occurrences. Finally, we set the number of negative examples as 10 times the number of positive examples for each location, but with a cap at 1000 negative examples in each region for computational reasons. We tune all parameters with respect to the F1 score for the classification tasks, and Spearman INLINEFORM9 for the regression tasks.
Variants and Baseline Methods
We will refer to our model as EGEL (Embedding GEographic Locations), and will consider the following variants. EGEL-Tags only uses the information from the Flickr tags (i.e. component INLINEFORM0 ), without using any negative examples and without feature selection. EGEL-Tags+NS is similar to EGEL-Tags but with the addition of negative examples. EGEL-KL(Tags+NS) additionally considers term selection. EGEL-All is our full method, i.e. it additionally uses the structured information. We also consider the following baselines. BOW-Tags represents locations using a bag-of-words representation, using the same tag weighting as the embedding model. BOW-KL(Tags) uses the same representation but after term selection, using the same KL-based method as the embedding model. BOW-All combines the bag-of-words representation with the structured information, encoded as proposed in BIBREF7 . GloVe uses the objective from the original GloVe model for learning location vectors, i.e. this variant differs from EGEL-Tags in that instead of INLINEFORM1 we use the number of co-occurrences of tag INLINEFORM2 near location INLINEFORM3 , measured as INLINEFORM4 .
Results and Discussion
We present our results for the binary classification tasks in Tables TABREF23 – TABREF24 in terms of average precision, average recall and macro average F1 score. The results of the regression tasks are reported in Tables TABREF25 and TABREF29 in terms of the mean absolute error between the predicted and actual scores, as well as the Spearman INLINEFORM0 correlation between the rankings induced by both sets of scores. It can be clearly seen from the results that our proposed method (EGEL-All) can effectively integrate Flickr tags with the available structured information. It outperforms the baselines for all the considered tasks. Furthermore, note that the PPMI-based weighting in EGEL-Tags consistently outperforms GloVe and that both the addition of negative examples and term selection lead to further improvements. The use of term selection leads to particularly substantial improvements for the regression problems.
While our experimental results confirm the usefulness of embeddings for predicting environmental features, this is only consistently the case for the variants that use both the tags and the structured datasets. In particular, comparing BOW-Tags with EGEL-Tags, we sometimes see that the former achieves the best results. While this might seem surprising, it is in accordance with the findings in BIBREF54 , BIBREF38 , among others, where it was also found that bag-of-words representations can sometimes lead to surprisingly effective baselines. Interestingly, we note that in all cases where EGEL-KL(Tags+NS) performs worse than BOW-Tags, we also find that BOW-KL(Tags) performs worse than BOW-Tags. This suggests that for these tasks there is a very large variation in the kind of tags that can inform the prediction model, possibly including e.g. user-specific tags. Some of the information captured by such highly specific but rare tags is likely to be lost in the embedding.
To further analyze the difference in performance between BoW representations and embeddings, Figure TABREF29 compares the performance of the GloVe model with the bag-of-words model for predicting place scenicness, as a function of the number of tag occurrences at the considered locations. What is clearly noticeable in Figure TABREF29 is that GloVe performs better than the bag-of-words model for large corpora and worse for smaller corpora. This issue has been alleviated in our embedding method by the addition of negative examples.
Conclusions
In this paper, we have proposed a model to learn geographic location embeddings using Flickr tags, numerical environmental features, and categorical information. The experimental results show that our model can integrate Flickr tags with structured information in a more effective way than existing methods, leading to substantial improvements over baseline methods on various prediction tasks about the natural environment.
Acknowledgments
Shelan Jeawak has been sponsored by HCED Iraq. Steven Schockaert has been supported by ERC Starting Grant 637277.
| [
"BOW-Tags, BOW-KL(Tags), BOW-All, GloVe"
] | 4,658 | qasper | en | null | a80067307d72e349f14a1f6765d914acf0b43764afe77ab7 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: what are the existing approaches?
Answer: | [
"BOW-Tags, BOW-KL(Tags), BOW-All, GloVe"
] | qasper | 128 | 91 | |
Do they use attention? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Background
Teaching machine to read and comprehend a given passage/paragraph and answer its corresponding questions is a challenging task. It is also one of the long-term goals of natural language understanding, and has important applications in e.g., building intelligent agents for conversation and customer service support. In a real world setting, it is necessary to judge whether the given questions are answerable given the available knowledge, and then generate correct answers for the ones which are able to infer an answer in the passage or an empty answer (as an unanswerable question) otherwise.
In comparison with many existing MRC systems BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 , which extract answers by finding a sub-string in the passages/paragraphs, we propose a model that not only extracts answers but also predicts whether such an answer should exist. Using a multi-task learning approach (c.f. BIBREF5 ), we extend the Stochastic Answer Network (SAN) BIBREF1 for MRC answer span detector to include a classifier that whether the question is unanswerable. The unanswerable classifier is a pair-wise classification model BIBREF6 which predicts a label indicating whether the given pair of a passage and a question is unanswerable. The two models share the same lower layer to save the number of parameters, and separate the top layers for different tasks (the span detector and binary classifier).
Our model is pretty simple and intuitive, yet efficient. Without relying on the large pre-trained language models (ELMo) BIBREF7 , the proposed model achieves competitive results to the state-of-the-art on Stanford Question Answering Dataset (SQuAD) 2.0.
The contribution of this work is summarized as follows. First, we propose a simple yet efficient model for MRC that handles unanswerable questions and is optimized jointly. Second, our model achieves competitive results on SQuAD v2.0.
Model
The Machine Reading Comprehension is a task which takes a question INLINEFORM0 and a passage/paragraph INLINEFORM1 as inputs, and aims to find an answer span INLINEFORM2 in INLINEFORM3 . We assume that if the question is answerable, the answer INLINEFORM4 exists in INLINEFORM5 as a contiguous text string; otherwise, INLINEFORM6 is an empty string indicating an unanswerable question. Note that to handle the unanswerable questions, we manually append a dumpy text string NULL at the end of each corresponding passage/paragraph. Formally, the answer is formulated as INLINEFORM7 . In case of unanswerable questions, INLINEFORM8 points to the last token of the passage.
Our model is a variation of SAN BIBREF1 , as shown in Figure FIGREF2 . The main difference is the additional binary classifier added in the model justifying whether the question is unanswerable. Roughly, the model includes two different layers: the shared layer and task specific layer. The shared layer is almost identical to the lower layers of SAN, which has a lexicon encoding layer, a contextual layer and a memory generation layer. On top of it, there are different answer modules for different tasks. We employ the SAN answer module for the span detector and a one-layer feed forward neural network for the binary classification task. It can also be viewed as a multi-task learning BIBREF8 , BIBREF5 , BIBREF9 . We will briefly describe the model from ground up as follows. Detailed descriptions can be found in BIBREF1 .
Lexicon Encoding Layer. We map the symbolic/surface feature of INLINEFORM0 and INLINEFORM1 into neural space via word embeddings , 16-dim part-of-speech (POS) tagging embeddings, 8-dim named-entity embeddings and 4-dim hard-rule features. Note that we use small embedding size of POS and NER to reduce model size and they mainly serve the role of coarse-grained word clusters. Additionally, we use question enhanced passages word embeddings which can viewwed as soft matching between questions and passages. At last, we use two separate two-layer position-wise Feed-Forward Networks (FFN) BIBREF11 , BIBREF1 to map both question and passage encodings into the same dimension. As results, we obtain the final lexicon embeddings for the tokens for INLINEFORM2 as a matrix INLINEFORM3 , and tokens in INLINEFORM4 as INLINEFORM5 .
Contextual Encoding Layer. A shared two-layers BiLSTM is used on the top to encode the contextual information of both passages and questions. To avoid overfitting, we concatenate a pre-trained 600-dimensional CoVe vectors BIBREF12 trained on German-English machine translation dataset, with the aforementioned lexicon embeddings as the final input of the contextual encoding layer, and also with the output of the first contextual encoding layer as the input of its second encoding layer. Thus, we obtain the final representation of the contextual encoding layer by a concatenation of the outputs of two BiLSTM: INLINEFORM0 for questions and INLINEFORM1 for passages.
Memory Generation Layer. In this layer, we generate a working memory by fusing information from both passages INLINEFORM0 and questions INLINEFORM1 . The attention function BIBREF11 is used to compute the similarity score between passages and questions as: INLINEFORM2
Note that INLINEFORM0 and INLINEFORM1 is transformed from INLINEFORM2 and INLINEFORM3 by one layer neural network INLINEFORM4 , respectively. A question-aware passage representation is computed as INLINEFORM5 . After that, we use the method of BIBREF13 to apply self attention to the passage: INLINEFORM6
where INLINEFORM0 means that we only drop diagonal elements on the similarity matrix (i.e., attention with itself). At last, INLINEFORM1 and INLINEFORM2 are concatenated and are passed through a BiLSTM to form the final memory: INLINEFORM3 .
Span detector. We adopt a multi-turn answer module for the span detector BIBREF1 . Formally, at time step INLINEFORM0 in the range of INLINEFORM1 , the state is defined by INLINEFORM2 . The initial state INLINEFORM3 is the summary of the INLINEFORM4 : INLINEFORM5 , where INLINEFORM6 . Here, INLINEFORM7 is computed from the previous state INLINEFORM8 and memory INLINEFORM9 : INLINEFORM10 and INLINEFORM11 . Finally, a bilinear function is used to find the begin and end point of answer spans at each reasoning step INLINEFORM12 : DISPLAYFORM0 DISPLAYFORM1
The final prediction is the average of each time step: INLINEFORM0 . We randomly apply dropout on the step level in each time step during training, as done in BIBREF1 .
Unanswerable classifier. We adopt a one-layer neural network as our unanswerable binary classifier: DISPLAYFORM0
, where INLINEFORM0 is the summary of the memory: INLINEFORM1 , where INLINEFORM2 . INLINEFORM3 denotes the probability of the question which is unanswerable.
Objective The objective function of the joint model has two parts: DISPLAYFORM0
Following BIBREF0 , the span loss function is defined: DISPLAYFORM0
The objective function of the binary classifier is defined: DISPLAYFORM0
where INLINEFORM0 is a binary variable: INLINEFORM1 indicates the question is unanswerable and INLINEFORM2 denotes the question is answerable.
Setup
We evaluate our system on SQuAD 2.0 dataset BIBREF14 , a new MRC dataset which is a combination of Stanford Question Answering Dataset (SQuAD) 1.0 BIBREF15 and additional unanswerable question-answer pairs. The answerable pairs are around 100K; while the unanswerable questions are around 53K. This dataset contains about 23K passages and they come from approximately 500 Wikipedia articles. All the questions and answers are obtained by crowd-sourcing. Two evaluation metrics are used: Exact Match (EM) and Macro-averaged F1 score (F1) BIBREF14 .
Implementation details
We utilize spaCy tool to tokenize the both passages and questions, and generate lemma, part-of-speech and named entity tags. The word embeddings are initialized with pre-trained 300-dimensional GloVe BIBREF10 . A 2-layer BiLSTM is used encoding the contextual information of both questions and passages. Regarding the hidden size of our model, we search greedily among INLINEFORM0 . During training, Adamax BIBREF16 is used as our optimizer. The min-batch size is set to 32. The learning rate is initialized to 0.002 and it is halved after every 10 epochs. The dropout rate is set to 0.1. To prevent overfitting, we also randomly set 0.5% words in both passages and questions as unknown words during the training. Here, we use a special token unk to indicate a word which doesn't appear in GloVe. INLINEFORM1 in Eq EQREF9 is set to 1.
Results
We would like to investigate effectiveness the proposed joint model. To do so, the same shared layer/architecture is employed in the following variants of the proposed model:
The results in terms of EM and F1 is summarized in Table TABREF20 . We observe that Joint SAN outperforms the SAN baseline with a large margin, e.g., 67.89 vs 69.27 (+1.38) and 70.68 vs 72.20 (+1.52) in terms of EM and F1 scores respectively, so it demonstrates the effectiveness of the joint optimization. By incorporating the output information of classifier into Joint SAN, it obtains a slight improvement, e.g., 72.2 vs 72.66 (+0.46) in terms of F1 score. By analyzing the results, we found that in most cases when our model extract an NULL string answer, the classifier also predicts it as an unanswerable question with a high probability.
Table TABREF21 reports comparison results in literature published . Our model achieves state-of-the-art on development dataset in setting without pre-trained large language model (ELMo). Comparing with the much complicated model R.M.-Reader + Verifier, which includes several components, our model still outperforms by 0.7 in terms of F1 score. Furthermore, we observe that ELMo gives a great boosting on the performance, e.g., 2.8 points in terms of F1 for DocQA. This encourages us to incorporate ELMo into our model in future.
Analysis. To better understand our model, we analyze the accuracy of the classifier in our joint model. We obtain 75.3 classification accuracy on the development with the threshold 0.5. By increasing value of INLINEFORM0 in Eq EQREF9 , the classification accuracy reached to 76.8 ( INLINEFORM1 ), however the final results of our model only have a small improvement (+0.2 in terms of F1 score). It shows that it is important to make balance between these two components: the span detector and unanswerable classifier.
Conclusion
To sum up, we proposed a simple yet efficient model based on SAN. It showed that the joint learning algorithm boosted the performance on SQuAD 2.0. We also would like to incorporate ELMo into our model in future.
Acknowledgments
We thank Yichong Xu, Shuohang Wang and Sheng Zhang for valuable discussions and comments. We also thank Robin Jia for the help on SQuAD evaluations.
| [
"Yes",
"Yes"
] | 1,687 | qasper | en | null | f3aba3579b9e3373ce708f10b33510d6a198c3bae58c5ad7 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Do they use attention?
Answer: | [
"Yes",
"Yes"
] | qasper | 128 | 92 | |
What datasets did they use for evaluation? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Bidirectional Encoder Representations from Transformers (BERT) is a novel Transformer BIBREF0 model, which recently achieved state-of-the-art performance in several language understanding tasks, such as question answering, natural language inference, semantic similarity, sentiment analysis, and others BIBREF1. While well-suited to dealing with relatively short sequences, Transformers suffer from a major issue that hinders their applicability in classification of long sequences, i.e. they are able to consume only a limited context of symbols as their input BIBREF2.
There are several natural language (NLP) processing tasks that involve such long sequences. Of particular interest are topic identification of spoken conversations BIBREF3, BIBREF4, BIBREF5 and call center customer satisfaction prediction BIBREF6, BIBREF7, BIBREF8, BIBREF9. Call center conversations, while usually quite short and to the point, often involve agents trying to solve very complex issues that the customers experience, resulting in some calls taking even an hour or more. For speech analytics purposes, these calls are typically transcribed using an automatic speech recognition (ASR) system, and processed in textual representations further down the NLP pipeline. These transcripts sometimes exceed the length of 5000 words. Furthermore, temporal information might play an important role in tasks like CSAT. For example, a customer may be angry at the beginning of the call, but after her issue is resolved, she would be very satisfied with the way it was handled. Therefore, simple bag of words models, or any model that does not include temporal dependencies between the inputs, may not be well-suited to handle this category of tasks. This motivates us to employ model such as BERT in this task.
In this paper, we propose a method that builds upon BERT's architecture. We split the input text sequence into shorter segments in order to obtain a representation for each of them using BERT. Then, we use either a recurrent LSTM BIBREF10 network, or another Transformer, to perform the actual classification. We call these techniques Recurrence over BERT (RoBERT) and Transformer over BERT (ToBERT). Given that these models introduce a hierarchy of representations (segment-wise and document-wise), we refer to them as Hierarchical Transformers. To the best of our knowledge, no attempt has been done before to use the Transformer architecture for classification of such long sequences.
Our novel contributions are:
Two extensions - RoBERT and ToBERT - to the BERT model, which enable its application in classification of long texts by performing segmentation and using another layer on top of the segment representations.
State-of-the-art results on the Fisher topic classification task.
Significant improvement on the CSAT prediction task over the MS-CNN model.
Related work
Several dimensionality reduction algorithms such as RBM, autoencoders, subspace multinomial models (SMM) are used to obtain a low dimensional representation of documents from a simple BOW representation and then classify it using a simple linear classifiers BIBREF11, BIBREF12, BIBREF13, BIBREF4. In BIBREF14 hierarchical attention networks are used for document classification. They evaluate their model on several datasets with average number of words around 150. Character-level CNN are explored in BIBREF15 but it is prohibitive for very long documents. In BIBREF16, dataset collected from arXiv papers is used for classification. For classification, they sample random blocks of words and use them together for classification instead of using full document which may work well as arXiv papers are usually coherent and well written on a well defined topic. Their method may not work well on spoken conversations as random block of words usually do not represent topic of full conversation.
Several researchers addressed the problem of predicting customer satisfaction BIBREF6, BIBREF7, BIBREF8, BIBREF9. In most of these works, logistic regression, SVM, CNN are applied on different kinds of representations.
In BIBREF17, authors use BERT for document classification but the average document length is less than BERT maximum length 512. TransformerXL BIBREF2 is an extension to the Transformer architecture that allows it to better deal with long inputs for the language modelling task. It relies on the auto-regressive property of the model, which is not the case in our tasks.
Method ::: BERT
Because our work builds heavily upon BERT, we provide a brief summary of its features. BERT is built upon the Transformer architecture BIBREF0, which uses self-attention, feed-forward layers, residual connections and layer normalization as the main building blocks. It has two pre-training objectives:
Masked language modelling - some of the words in a sentence are being masked and the model has to predict them based on the context (note the difference from the typical autoregressive language model training objective);
Next sentence prediction - given two input sequences, decide whether the second one is the next sentence or not.
BERT has been shown to beat the state-of-the-art performance on 11 tasks with no modifications to the model architecture, besides adding a task-specific output layer BIBREF1. We follow same procedure suggested in BIBREF1 for our tasks. Fig. FIGREF8 shows the BERT model for classification. We obtain two kinds of representation from BERT: pooled output from last transformer block, denoted by H, and posterior probabilities, denoted by P. There are two variants of BERT - BERT-Base and BERT-Large. In this work we are using BERT-Base for faster training and experimentation, however, our methods are applicable to BERT-Large as well. BERT-Base and BERT-Large are different in model parameters such as number of transformer blocks, number of self-attention heads. Total number of parameters in BERT-Base are 110M and 340M in BERT-Large.
BERT suffers from major limitations in terms of handling long sequences. Firstly, the self-attention layer has a quadratic complexity $O(n^2)$ in terms of the sequence length $n$ BIBREF0. Secondly, BERT uses a learned positional embeddings scheme BIBREF1, which means that it won't likely be able to generalize to positions beyond those seen in the training data.
To investigate the effect of fine-tuning BERT on task performance, we use either the pre-trained BERT weights, or the weights from a BERT fine-tuned on the task-specific dataset on a segment-level (i.e. we preserve the original label but fine-tune on each segment separately instead of on the whole text sequence). We compare these results to using the fine-tuned segment-level BERT predictions directly as inputs to the next layer.
Method ::: Recurrence over BERT
Given that BERT is limited to a particular input length, we split the input sequence into segments of a fixed size with overlap. For each of these segments, we obtain H or P from BERT model. We then stack these segment-level representations into a sequence, which serves as input to a small (100-dimensional) LSTM layer. Its output serves as a document embedding. Finally, we use two fully connected layers with ReLU (30-dimensional) and softmax (the same dimensionality as the number of classes) activations to obtain the final predictions.
With this approach, we overcome BERT's computational complexity, reducing it to $O(n/k * k^2) = O(nk)$ for RoBERT, with $k$ denoting the segment size (the LSTM component has negligible linear complexity $O(k)$). The positional embeddings are also no longer an issue.
Method ::: Transformer over BERT
Given that Transformers' edge over recurrent networks is their ability to effectively capture long distance relationships between words in a sequence BIBREF0, we experiment with replacing the LSTM recurrent layer in favor of a small Transformer model (2 layers of transformer building block containing self-attention, fully connected, etc.). To investigate if preserving the information about the input sequence order is important, we also build a variant of ToBERT which learns positional embeddings at the segment-level representations (but is limited to sequences of length seen during the training).
ToBERT's computational complexity $O(\frac{n^2}{k^2})$ is asymptotically inferior to RoBERT, as the top-level Transformer model again suffers from quadratic complexity in the number of segments. However, in practice this number is much smaller than the input sequence length (${\frac{n}{k}} << n$), so we haven't observed performance or memory issues with our datasets.
Experiments
We evaluated our models on 3 different datasets:
CSAT dataset for CSAT prediction, consisting of spoken transcripts (automatic via ASR).
20 newsgroups for topic identification task, consisting of written text;
Fisher Phase 1 corpus for topic identification task, consisting of spoken transcripts (manual);
Experiments ::: CSAT
CSAT dataset consists of US English telephone speech from call centers. For each call in this dataset, customers participated in that call gave a rating on his experience with agent. Originally, this dataset has labels rated on a scale 1-9 with 9 being extremely satisfied and 1 being extremely dissatisfied. Fig. FIGREF16 shows the histogram of ratings for our dataset. As the distribution is skewed towards extremes, we choose to do binary classification with ratings above 4.5 as satisfied and below 4.5 as dissatisfied. Quantization of ratings also helped us to create a balanced dataset. This dataset contains 4331 calls and we split them into 3 sets for our experiments: 2866 calls for training, 362 calls for validation and, finally, 1103 calls for testing.
We obtained the transcripts by employing an ASR system. The ASR system uses TDNN-LSTM acoustic model trained on Fisher and Switchboard datasets with lattice-free maximum mutual information criterion BIBREF18. The word error rates using four-gram language models were 9.2% and 17.3% respectively on Switchboard and CallHome portions of Eval2000 dataset.
Experiments ::: 20 newsgroups
20 newsgroups data set is one of the frequently used datasets in the text processing community for text classification and text clustering. This data set contains approximately 20,000 English documents from 20 topics to be identified, with 11314 documents for training and 7532 for testing. In this work, we used only 90% of documents for training and the remaining 10% for validation. For fair comparison with other publications, we used 53160 words vocabulary set available in the datasets website.
Experiments ::: Fisher
Fisher Phase 1 US English corpus is often used for automatic speech recognition in speech community. In this work, we used it for topic identification as in BIBREF3. The documents are 10-minute long telephone conversations between two people discussing a given topic. We used same training and test splits as BIBREF3 in which 1374 and 1372 documents are used for training and testing respectively. For validation of our model, we used 10% of training dataset and the remaining 90% was used for actual model training. The number of topics in this data set is 40.
Experiments ::: Dataset Statistics
Table TABREF22 shows statistics of our datasets. It can be observed that average length of Fisher is much higher than 20 newsgroups and CSAT. Cumulative distribution of document lengths for each dataset is shown in Fig. FIGREF21. It can be observed that almost all of the documents in Fisher dataset have length more than 1000 words. For CSAT, more than 50% of the documents have length greater than 500 and for 20newsgroups only 10% of the documents have length greater than 500. Note that, for CSAT and 20newsgroups, there are few documents with length more than 5000.
Experiments ::: Architecture and Training Details
In this work, we split document into segments of 200 tokens with a shift of 50 tokens to extract features from BERT model. For RoBERT, LSTM model is trained to minimize cross-entropy loss with Adam optimizer BIBREF19. The initial learning rate is set to $0.001$ and is reduced by a factor of $0.95$ if validation loss does not decrease for 3-epochs. For ToBERT, the Transformer is trained with the default BERT version of Adam optimizer BIBREF1 with an initial learning rate of $5e$-5. We report accuracy in all of our experiments. We chose a model with the best validation accuracy to calculate accuracy on the test set. To accomodate for non-determinism of some TensorFlow GPU operations, we report accuracy averaged over 5 runs.
Results
Table TABREF25 presents results using pre-trained BERT features. We extracted features from the pooled output of final transformer block as these were shown to be working well for most of the tasks BIBREF1. The features extracted from a pre-trained BERT model without any fine-tuning lead to a sub-par performance. However, We also notice that ToBERT model exploited the pre-trained BERT features better than RoBERT. It also converged faster than RoBERT. Table TABREF26 shows results using features extracted after fine-tuning BERT model with our datasets. Significant improvements can be observed compared to using pre-trained BERT features. Also, it can be noticed that ToBERT outperforms RoBERT on Fisher and 20newsgroups dataset by 13.63% and 0.81% respectively. On CSAT, ToBERT performs slightly worse than RoBERT but it is not statistically significant as this dataset is small.
Table TABREF27 presents results using fine-tuned BERT predictions instead of the pooled output from final transformer block. For each document, having obtained segment-wise predictions we can obtain final prediction for the whole document in three ways:
Compute the average of all segment-wise predictions and find the most probable class;
Find the most frequently predicted class;
Train a classification model.
It can be observed from Table TABREF27 that a simple averaging operation or taking most frequent predicted class works competitively for CSAT and 20newsgroups but not for the Fisher dataset. We believe the improvements from using RoBERT or ToBERT, compared to simple averaging or most frequent operations, are proportional to the fraction of long documents in the dataset. CSAT and 20newsgroups have (on average) significantly shorter documents than Fisher, as seen in Fig. FIGREF21. Also, significant improvements for Fisher could be because of less confident predictions from BERT model as this dataset has 40 classes. Fig. FIGREF31 presents the comparison of average voting and ToBERT for various document length ranges for Fisher dataset. We used fine-tuned BERT segment-level predictions (P) for this analysis. It can be observed that ToBERT outperforms average voting in every interval. To the best of our knowledge, this is a state-of-the-art result reported on the Fisher dataset.
Table TABREF32 presents the effect of position embeddings on the model performance. It can be observed that position embeddings did not significantly affect the model performance for Fisher and 20newsgroups, but they helped slightly in CSAT prediction (an absolute improvement of 0.64% F1-score). We think that this is explained by the fact that Fisher and 20newsgroups are topic identification tasks, and the topic does not change much throughout these documents. However, CSAT may vary during the call, and in some cases a naive assumption that the sequential nature of the transcripts is irrelevant may lead to wrong conclusions.
Table TABREF33 compares our results with previous works. It can be seen that our model ToBERT outperforms CNN based experiments by significant margin on CSAT and Fisher datasets. For CSAT dataset, we used multi-scale CNN (MS-CNN) as the baseline, given its strong results on Fisher and 20newsgroups. The setup was replicated from BIBREF5 for comparison. We also see that our result on 20 newsgroups is 0.6% worse than the state-of-the-art.
Conclusions
In this paper, we presented two methods for long documents using BERT model: RoBERT and ToBERT. We evaluated our experiments on two classification tasks - customer satisfaction prediction and topic identification - using 3 datasets: CSAT, 20newsgroups and Fisher. We observed that ToBERT outperforms RoBERT on pre-trained BERT features and fine-tuned BERT features for all our tasks. Also, we noticed that fine-tuned BERT performs better than pre-trained BERT. We have shown that both RoBERT and ToBERT improved the simple baselines of taking an average (or the most frequent) of segment-wise predictions for long documents to obtain final prediction. Position embeddings did not significantly affect our models performance, but slightly improved the accuracy on the CSAT task. We obtained the best results on Fisher dataset and good improvements for CSAT task compared to the CNN baseline. It is interesting to note that the longer the average input in a given task, the bigger improvement we observe w.r.t. the baseline for that task. Our results confirm that both RoBERT and ToBERT can be used for long sequences with competitive performance and quick fine-tuning procedure. For future work, we shall focus on training models on long documents directly (i.e. in an end-to-end manner).
| [
"CSAT dataset, 20 newsgroups, Fisher Phase 1 corpus",
"CSAT dataset , 20 newsgroups, Fisher Phase 1 corpus"
] | 2,652 | qasper | en | null | 466bd29bcab1cdfdef327777808236bd2677e2a54414a32a | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What datasets did they use for evaluation?
Answer: | [
"CSAT dataset, 20 newsgroups, Fisher Phase 1 corpus",
"CSAT dataset , 20 newsgroups, Fisher Phase 1 corpus"
] | qasper | 128 | 93 | |
What sentiment classification dataset is used? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Recurrent neural networks (RNNs), including gated variants such as the long short-term memory (LSTM) BIBREF0 have become the standard model architecture for deep learning approaches to sequence modeling tasks. RNNs repeatedly apply a function with trainable parameters to a hidden state. Recurrent layers can also be stacked, increasing network depth, representational power and often accuracy. RNN applications in the natural language domain range from sentence classification BIBREF1 to word- and character-level language modeling BIBREF2 . RNNs are also commonly the basic building block for more complex models for tasks such as machine translation BIBREF3 , BIBREF4 , BIBREF5 or question answering BIBREF6 , BIBREF7 . Unfortunately standard RNNs, including LSTMs, are limited in their capability to handle tasks involving very long sequences, such as document classification or character-level machine translation, as the computation of features or states for different parts of the document cannot occur in parallel.
Convolutional neural networks (CNNs) BIBREF8 , though more popular on tasks involving image data, have also been applied to sequence encoding tasks BIBREF9 . Such models apply time-invariant filter functions in parallel to windows along the input sequence. CNNs possess several advantages over recurrent models, including increased parallelism and better scaling to long sequences such as those often seen with character-level language data. Convolutional models for sequence processing have been more successful when combined with RNN layers in a hybrid architecture BIBREF10 , because traditional max- and average-pooling approaches to combining convolutional features across timesteps assume time invariance and hence cannot make full use of large-scale sequence order information.
We present quasi-recurrent neural networks for neural sequence modeling. QRNNs address both drawbacks of standard models: like CNNs, QRNNs allow for parallel computation across both timestep and minibatch dimensions, enabling high throughput and good scaling to long sequences. Like RNNs, QRNNs allow the output to depend on the overall order of elements in the sequence. We describe QRNN variants tailored to several natural language tasks, including document-level sentiment classification, language modeling, and character-level machine translation. These models outperform strong LSTM baselines on all three tasks while dramatically reducing computation time.
Model
Each layer of a quasi-recurrent neural network consists of two kinds of subcomponents, analogous to convolution and pooling layers in CNNs. The convolutional component, like convolutional layers in CNNs, allows fully parallel computation across both minibatches and spatial dimensions, in this case the sequence dimension. The pooling component, like pooling layers in CNNs, lacks trainable parameters and allows fully parallel computation across minibatch and feature dimensions.
Given an input sequence INLINEFORM0 of INLINEFORM1 INLINEFORM2 -dimensional vectors INLINEFORM3 , the convolutional subcomponent of a QRNN performs convolutions in the timestep dimension with a bank of INLINEFORM4 filters, producing a sequence INLINEFORM5 of INLINEFORM6 -dimensional candidate vectors INLINEFORM7 . In order to be useful for tasks that include prediction of the next token, the filters must not allow the computation for any given timestep to access information from future timesteps. That is, with filters of width INLINEFORM8 , each INLINEFORM9 depends only on INLINEFORM10 through INLINEFORM11 . This concept, known as a masked convolution BIBREF11 , is implemented by padding the input to the left by the convolution's filter size minus one.
We apply additional convolutions with separate filter banks to obtain sequences of vectors for the elementwise gates that are needed for the pooling function. While the candidate vectors are passed through a INLINEFORM0 nonlinearity, the gates use an elementwise sigmoid. If the pooling function requires a forget gate INLINEFORM1 and an output gate INLINEFORM2 at each timestep, the full set of computations in the convolutional component is then: DISPLAYFORM0
where INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 , each in INLINEFORM3 , are the convolutional filter banks and INLINEFORM4 denotes a masked convolution along the timestep dimension. Note that if the filter width is 2, these equations reduce to the LSTM-like DISPLAYFORM0
Convolution filters of larger width effectively compute higher INLINEFORM0 -gram features at each timestep; thus larger widths are especially important for character-level tasks.
Suitable functions for the pooling subcomponent can be constructed from the familiar elementwise gates of the traditional LSTM cell. We seek a function controlled by gates that can mix states across timesteps, but which acts independently on each channel of the state vector. The simplest option, which BIBREF12 term “dynamic average pooling”, uses only a forget gate: DISPLAYFORM0
We term these three options f-pooling, fo-pooling, and ifo-pooling respectively; in each case we initialize INLINEFORM0 or INLINEFORM1 to zero. Although the recurrent parts of these functions must be calculated for each timestep in sequence, their simplicity and parallelism along feature dimensions means that, in practice, evaluating them over even long sequences requires a negligible amount of computation time.
A single QRNN layer thus performs an input-dependent pooling, followed by a gated linear combination of convolutional features. As with convolutional neural networks, two or more QRNN layers should be stacked to create a model with the capacity to approximate more complex functions.
Variants
Motivated by several common natural language tasks, and the long history of work on related architectures, we introduce several extensions to the stacked QRNN described above. Notably, many extensions to both recurrent and convolutional models can be applied directly to the QRNN as it combines elements of both model types.
Regularization An important extension to the stacked QRNN is a robust regularization scheme inspired by recent work in regularizing LSTMs.
The need for an effective regularization method for LSTMs, and dropout's relative lack of efficacy when applied to recurrent connections, led to the development of recurrent dropout schemes, including variational inference–based dropout BIBREF13 and zoneout BIBREF14 . These schemes extend dropout to the recurrent setting by taking advantage of the repeating structure of recurrent networks, providing more powerful and less destructive regularization.
Variational inference–based dropout locks the dropout mask used for the recurrent connections across timesteps, so a single RNN pass uses a single stochastic subset of the recurrent weights. Zoneout stochastically chooses a new subset of channels to “zone out” at each timestep; for these channels the network copies states from one timestep to the next without modification.
As QRNNs lack recurrent weights, the variational inference approach does not apply. Thus we extended zoneout to the QRNN architecture by modifying the pooling function to keep the previous pooling state for a stochastic subset of channels. Conveniently, this is equivalent to stochastically setting a subset of the QRNN's INLINEFORM0 gate channels to 1, or applying dropout on INLINEFORM1 : DISPLAYFORM0
Thus the pooling function itself need not be modified at all. We note that when using an off-the-shelf dropout layer in this context, it is important to remove automatic rescaling functionality from the implementation if it is present. In many experiments, we also apply ordinary dropout between layers, including between word embeddings and the first QRNN layer.
Densely-Connected Layers We can also extend the QRNN architecture using techniques introduced for convolutional networks. For sequence classification tasks, we found it helpful to use skip-connections between every QRNN layer, a technique termed “dense convolution” by BIBREF15 . Where traditional feed-forward or convolutional networks have connections only between subsequent layers, a “DenseNet” with INLINEFORM0 layers has feed-forward or convolutional connections between every pair of layers, for a total of INLINEFORM1 . This can improve gradient flow and convergence properties, especially in deeper networks, although it requires a parameter count that is quadratic in the number of layers.
When applying this technique to the QRNN, we include connections between the input embeddings and every QRNN layer and between every pair of QRNN layers. This is equivalent to concatenating each QRNN layer's input to its output along the channel dimension before feeding the state into the next layer. The output of the last layer alone is then used as the overall encoding result.
Encoder–Decoder Models To demonstrate the generality of QRNNs, we extend the model architecture to sequence-to-sequence tasks, such as machine translation, by using a QRNN as encoder and a modified QRNN, enhanced with attention, as decoder. The motivation for modifying the decoder is that simply feeding the last encoder hidden state (the output of the encoder's pooling layer) into the decoder's recurrent pooling layer, analogously to conventional recurrent encoder–decoder architectures, would not allow the encoder state to affect the gate or update values that are provided to the decoder's pooling layer. This would substantially limit the representational power of the decoder.
Instead, the output of each decoder QRNN layer's convolution functions is supplemented at every timestep with the final encoder hidden state. This is accomplished by adding the result of the convolution for layer INLINEFORM0 (e.g., INLINEFORM1 , in INLINEFORM2 ) with broadcasting to a linearly projected copy of layer INLINEFORM3 's last encoder state (e.g., INLINEFORM4 , in INLINEFORM5 ): DISPLAYFORM0
where the tilde denotes that INLINEFORM0 is an encoder variable. Encoder–decoder models which operate on long sequences are made significantly more powerful with the addition of soft attention BIBREF3 , which removes the need for the entire input representation to fit into a fixed-length encoding vector. In our experiments, we computed an attentional sum of the encoder's last layer's hidden states. We used the dot products of these encoder hidden states with the decoder's last layer's un-gated hidden states, applying a INLINEFORM1 along the encoder timesteps, to weight the encoder states into an attentional sum INLINEFORM2 for each decoder timestep. This context, and the decoder state, are then fed into a linear layer followed by the output gate: DISPLAYFORM0
where INLINEFORM0 is the last layer.
While the first step of this attention procedure is quadratic in the sequence length, in practice it takes significantly less computation time than the model's linear and convolutional layers due to the simple and highly parallel dot-product scoring function.
Experiments
We evaluate the performance of the QRNN on three different natural language tasks: document-level sentiment classification, language modeling, and character-based neural machine translation. Our QRNN models outperform LSTM-based models of equal hidden size on all three tasks while dramatically improving computation speed. Experiments were implemented in Chainer BIBREF16 .
Sentiment Classification
We evaluate the QRNN architecture on a popular document-level sentiment classification benchmark, the IMDb movie review dataset BIBREF17 . The dataset consists of a balanced sample of 25,000 positive and 25,000 negative reviews, divided into equal-size train and test sets, with an average document length of 231 words BIBREF18 . We compare only to other results that do not make use of additional unlabeled data (thus excluding e.g., BIBREF19 ).
Our best performance on a held-out development set was achieved using a four-layer densely-connected QRNN with 256 units per layer and word vectors initialized using 300-dimensional cased GloVe embeddings BIBREF20 . Dropout of 0.3 was applied between layers, and we used INLINEFORM0 regularization of INLINEFORM1 . Optimization was performed on minibatches of 24 examples using RMSprop BIBREF21 with learning rate of INLINEFORM2 , INLINEFORM3 , and INLINEFORM4 .
Small batch sizes and long sequence lengths provide an ideal situation for demonstrating the QRNN's performance advantages over traditional recurrent architectures. We observed a speedup of 3.2x on IMDb train time per epoch compared to the optimized LSTM implementation provided in NVIDIA's cuDNN library. For specific batch sizes and sequence lengths, a 16x speed gain is possible. Figure FIGREF15 provides extensive speed comparisons.
In Figure FIGREF12 , we visualize the hidden state vectors INLINEFORM0 of the final QRNN layer on part of an example from the IMDb dataset. Even without any post-processing, changes in the hidden state are visible and interpretable in regards to the input. This is a consequence of the elementwise nature of the recurrent pooling function, which delays direct interaction between different channels of the hidden state until the computation of the next QRNN layer.
Language Modeling
We replicate the language modeling experiment of BIBREF2 and BIBREF13 to benchmark the QRNN architecture for natural language sequence prediction. The experiment uses a standard preprocessed version of the Penn Treebank (PTB) by BIBREF25 .
We implemented a gated QRNN model with medium hidden size: 2 layers with 640 units in each layer. Both QRNN layers use a convolutional filter width INLINEFORM0 of two timesteps. While the “medium” models used in other work BIBREF2 , BIBREF13 consist of 650 units in each layer, it was more computationally convenient to use a multiple of 32. As the Penn Treebank is a relatively small dataset, preventing overfitting is of considerable importance and a major focus of recent research. It is not obvious in advance which of the many RNN regularization schemes would perform well when applied to the QRNN. Our tests showed encouraging results from zoneout applied to the QRNN's recurrent pooling layer, implemented as described in Section SECREF5 .
The experimental settings largely followed the “medium” setup of BIBREF2 . Optimization was performed by stochastic gradient descent (SGD) without momentum. The learning rate was set at 1 for six epochs, then decayed by 0.95 for each subsequent epoch, for a total of 72 epochs. We additionally used INLINEFORM0 regularization of INLINEFORM1 and rescaled gradients with norm above 10. Zoneout was applied by performing dropout with ratio 0.1 on the forget gates of the QRNN, without rescaling the output of the dropout function. Batches consist of 20 examples, each 105 timesteps.
Comparing our results on the gated QRNN with zoneout to the results of LSTMs with both ordinary and variational dropout in Table TABREF14 , we see that the QRNN is highly competitive. The QRNN without zoneout strongly outperforms both our medium LSTM and the medium LSTM of BIBREF2 which do not use recurrent dropout and is even competitive with variational LSTMs. This may be due to the limited computational capacity that the QRNN's pooling layer has relative to the LSTM's recurrent weights, providing structural regularization over the recurrence.
Without zoneout, early stopping based upon validation loss was required as the QRNN would begin overfitting. By applying a small amount of zoneout ( INLINEFORM0 ), no early stopping is required and the QRNN achieves competitive levels of perplexity to the variational LSTM of BIBREF13 , which had variational inference based dropout of 0.2 applied recurrently. Their best performing variation also used Monte Carlo (MC) dropout averaging at test time of 1000 different masks, making it computationally more expensive to run.
When training on the PTB dataset with an NVIDIA K40 GPU, we found that the QRNN is substantially faster than a standard LSTM, even when comparing against the optimized cuDNN LSTM. In Figure FIGREF15 we provide a breakdown of the time taken for Chainer's default LSTM, the cuDNN LSTM, and QRNN to perform a full forward and backward pass on a single batch during training of the RNN LM on PTB. For both LSTM implementations, running time was dominated by the RNN computations, even with the highly optimized cuDNN implementation. For the QRNN implementation, however, the “RNN” layers are no longer the bottleneck. Indeed, there are diminishing returns from further optimization of the QRNN itself as the softmax and optimization overhead take equal or greater time. Note that the softmax, over a vocabulary size of only 10,000 words, is relatively small; for tasks with larger vocabularies, the softmax would likely dominate computation time.
It is also important to note that the cuDNN library's RNN primitives do not natively support any form of recurrent dropout. That is, running an LSTM that uses a state-of-the-art regularization scheme at cuDNN-like speeds would likely require an entirely custom kernel.
Character-level Neural Machine Translation
We evaluate the sequence-to-sequence QRNN architecture described in SECREF5 on a challenging neural machine translation task, IWSLT German–English spoken-domain translation, applying fully character-level segmentation. This dataset consists of 209,772 sentence pairs of parallel training data from transcribed TED and TEDx presentations, with a mean sentence length of 103 characters for German and 93 for English. We remove training sentences with more than 300 characters in English or German, and use a unified vocabulary of 187 Unicode code points.
Our best performance on a development set (TED.tst2013) was achieved using a four-layer encoder–decoder QRNN with 320 units per layer, no dropout or INLINEFORM0 regularization, and gradient rescaling to a maximum magnitude of 5. Inputs were supplied to the encoder reversed, while the encoder convolutions were not masked. The first encoder layer used convolutional filter width INLINEFORM1 , while the other encoder layers used INLINEFORM2 . Optimization was performed for 10 epochs on minibatches of 16 examples using Adam BIBREF28 with INLINEFORM3 , INLINEFORM4 , INLINEFORM5 , and INLINEFORM6 . Decoding was performed using beam search with beam width 8 and length normalization INLINEFORM7 . The modified log-probability ranking criterion is provided in the appendix.
Results using this architecture were compared to an equal-sized four-layer encoder–decoder LSTM with attention, applying dropout of 0.2. We again optimized using Adam; other hyperparameters were equal to their values for the QRNN and the same beam search procedure was applied. Table TABREF17 shows that the QRNN outperformed the character-level LSTM, almost matching the performance of a word-level attentional baseline.
Related Work
Exploring alternatives to traditional RNNs for sequence tasks is a major area of current research. Quasi-recurrent neural networks are related to several such recently described models, especially the strongly-typed recurrent neural networks (T-RNN) introduced by BIBREF12 . While the motivation and constraints described in that work are different, BIBREF12 's concepts of “learnware” and “firmware” parallel our discussion of convolution-like and pooling-like subcomponents. As the use of a fully connected layer for recurrent connections violates the constraint of “strong typing”, all strongly-typed RNN architectures (including the T-RNN, T-GRU, and T-LSTM) are also quasi-recurrent. However, some QRNN models (including those with attention or skip-connections) are not “strongly typed”. In particular, a T-RNN differs from a QRNN as described in this paper with filter size 1 and f-pooling only in the absence of an activation function on INLINEFORM0 . Similarly, T-GRUs and T-LSTMs differ from QRNNs with filter size 2 and fo- or ifo-pooling respectively in that they lack INLINEFORM1 on INLINEFORM2 and use INLINEFORM3 rather than sigmoid on INLINEFORM4 .
The QRNN is also related to work in hybrid convolutional–recurrent models. BIBREF31 apply CNNs at the word level to generate INLINEFORM0 -gram features used by an LSTM for text classification. BIBREF32 also tackle text classification by applying convolutions at the character level, with a stride to reduce sequence length, then feeding these features into a bidirectional LSTM. A similar approach was taken by BIBREF10 for character-level machine translation. Their model's encoder uses a convolutional layer followed by max-pooling to reduce sequence length, a four-layer highway network, and a bidirectional GRU. The parallelism of the convolutional, pooling, and highway layers allows training speed comparable to subword-level models without hard-coded text segmentation.
The QRNN encoder–decoder model shares the favorable parallelism and path-length properties exhibited by the ByteNet BIBREF33 , an architecture for character-level machine translation based on residual convolutions over binary trees. Their model was constructed to achieve three desired properties: parallelism, linear-time computational complexity, and short paths between any pair of words in order to better propagate gradient signals.
Conclusion
Intuitively, many aspects of the semantics of long sequences are context-invariant and can be computed in parallel (e.g., convolutionally), but some aspects require long-distance context and must be computed recurrently. Many existing neural network architectures either fail to take advantage of the contextual information or fail to take advantage of the parallelism. QRNNs exploit both parallelism and context, exhibiting advantages from both convolutional and recurrent neural networks. QRNNs have better predictive accuracy than LSTM-based models of equal hidden size, even though they use fewer parameters and run substantially faster. Our experiments show that the speed and accuracy advantages remain consistent across tasks and at both word and character levels.
Extensions to both CNNs and RNNs are often directly applicable to the QRNN, while the model's hidden states are more interpretable than those of other recurrent architectures as its channels maintain their independence across timesteps. We believe that QRNNs can serve as a building block for long-sequence tasks that were previously impractical with traditional RNNs.
Beam search ranking criterion
The modified log-probability ranking criterion we used in beam search for translation experiments is: DISPLAYFORM0
where INLINEFORM0 is a length normalization parameter BIBREF34 , INLINEFORM1 is the INLINEFORM2 th output character, and INLINEFORM3 is a “target length” equal to the source sentence length plus five characters. This reduces at INLINEFORM4 to ordinary beam search with probabilities: DISPLAYFORM0
and at INLINEFORM0 to beam search with probabilities normalized by length (up to the target length): DISPLAYFORM0
Conveniently, this ranking criterion can be computed at intermediate beam-search timesteps, obviating the need to apply a separate reranking on complete hypotheses.
| [
"the IMDb movie review dataset BIBREF17",
"IMDb movie review"
] | 3,432 | qasper | en | null | e7efd3969adf95459805233e580d6e0c7539a4de09b4441e | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What sentiment classification dataset is used?
Answer: | [
"the IMDb movie review dataset BIBREF17",
"IMDb movie review"
] | qasper | 128 | 94 | |
Were any of these tasks evaluated in any previous work? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
The recently introduced BERT model BIBREF0 exhibits strong performance on several language understanding benchmarks. To what extent does it capture syntax-sensitive structures?
Recent work examines the extent to which RNN-based models capture syntax-sensitive phenomena that are traditionally taken as evidence for the existence in hierarchical structure. In particular, in BIBREF1 we assess the ability of LSTMs to learn subject-verb agreement patterns in English, and evaluate on naturally occurring wikipedia sentences. BIBREF2 also consider subject-verb agreement, but in a “colorless green ideas” setting in which content words in naturally occurring sentences are replaced with random words with the same part-of-speech and inflection, thus ensuring a focus on syntax rather than on selectional-preferences based cues. BIBREF3 consider a wider range of syntactic phenomena (subject-verb agreement, reflexive anaphora, negative polarity items) using manually constructed stimuli, allowing for greater coverage and control than in the naturally occurring setting.
The BERT model is based on the “Transformer” architecture BIBREF4 , which—in contrast to RNNs—relies purely on attention mechanisms, and does not have an explicit notion of word order beyond marking each word with its absolute-position embedding. This reliance on attention may lead one to expect decreased performance on syntax-sensitive tasks compared to RNN (LSTM) models that do model word order directly, and explicitly track states across the sentence. Indeed, BIBREF5 finds that transformer-based models perform worse than LSTM models on the BIBREF1 agreement prediction dataset. In contrast, BIBREF6 find that self-attention performs on par with LSTM for syntax sensitive dependencies in the context of machine-translation, and performance on syntactic tasks is correlated with the number of attention heads in multi-head attention.
I adapt the evaluation protocol and stimuli of BIBREF1 , BIBREF2 and BIBREF3 to the bidirectional setting required by BERT, and evaluate the pre-trained BERT models (both the Large and the Base models). Surprisingly (at least to me), the out-of-the-box models (without any task-specific fine-tuning) perform very well on all the syntactic tasks.
Methodology
I use the stimuli provided by BIBREF1 , BIBREF2 , BIBREF3 , but change the experimental protocol to adapt it to the bidirectional nature of the BERT model. This requires discarding some of the stimuli, as described below. Thus, the numbers are not strictly comparable to those reported in previous work.
Previous setups
All three previous work use uni-directional language-model-like models.
BIBREF1 start with existing sentences from wikipedia that contain a present-tense verb. They feed each sentence word by word into an LSTM, stop right before the focus verb, and ask the model to predict a binary plural/singular decision (supervised setup) or compare the probability assigned by a pre-trained language model (LM) to the plural vs singular forms of the verb (LM setup). The evaluation is then performed on sentences with “agreement attractors” in which at there is at least one noun between the verb and its subject, and all of the nouns between the verb and subject are of the opposite number from the subject.
BIBREF2 also start with existing sentences. However, in order to control for the possibillity of the model learning to rely on “semantic” selectional-preferences cues rather than syntactic ones, they replace each content word with random words from the same part-of-speech and inflection. This results in “coloreless green ideas” nonce sentences. The evaluation is then performed similarly to the LM setup of BIBREF1 : the sentence is fed into a pre-traiend LSTM LM up to the focus verb, and the model is considered correct if the probability assigned to the correct inflection of the original verb form given the prefix is larger than that assigned to the incorrect inflection.
BIBREF3 focus on manually constructed and controlled stimuli, that also emphasizes linguistic structure over selectional preferences. They construct minimal pairs of grammatical and ungrammatical sentences, feed each one in its entirety into a pre-trained LSTM-LM, and compare the perplexity assigned by the model to the grammatical and ungrammatical sentences. The model is “correct” if it assigns the grammatical sentence a higher probability than to the ungrammatical one. Since the minimal pairs for most phenomena differ only in a single word (the focus verb), this scoring is very similar to the one used in the two previous works. However, it does consider the continuation of the sentence after the focus verb, and also allows for assessing phenomena that require change into two or more words (like negative polarity items).
Adaptation to the BERT model
In contrast to these works, the BERT model is bi-directional: it is trained to predict the identity of masked words based on both the prefix and suffix surrounding these words. I adapt the uni-directional setup by feeding into BERT the complete sentence, while masking out the single focus verb. I then ask BERT for its word predictions for the masked position, and compare the score assigned to the original correct verb to the score assigned to the incorrect one.
For example, for the sentence:
a 2002 systemic review of herbal products found that several herbs , including peppermint and caraway , have anti-dyspeptic effects for non-ulcer dyspepsia with “ encouraging safety profiles ” . (from BIBREF1 )
I feed into BERT:
[CLS] a 2002 systemic review of herbal products found that several herbs , including peppermint and caraway , [MASK] anti-dyspeptic effects for non-ulcer dyspepsia with “ encouraging safety profiles ” . and look for the score assigned to the words have and has at the masked position.
Similarly, for the pair
the game that the guard hates is bad .
the game that the guard hates are bad .
(from BIBREF3 ), I feed into BERT:
[CLS] the game that the guard hates [MASK] bad .
and compare the scores predicted for is and are.
This differs from BIBREF1 and BIBREF2 by considering the entire sentence (excluding the verb) and not just its prefix leading to the verb, and differs from BIBREF3 by conditioning the focus verb on bidirectional context.
I use the PyTorch implementation of BERT, with the pre-trained models supplied by Google. I experiment with the bert-large-uncased and bert-base-uncased models.
The bi-directional setup precludes using using the NPI stimuli of BIBREF3 , in which the minimal pair differs in two words position, which I discard from the evaluation. I also discard the agreement cases involving the verbs is or are in BIBREF1 and in BIBREF2 , because some of them are copular construction, in which strong agreement hints can be found also on the object following the verb. This is not an issue in the manually constructed BIBREF3 stimuli due to the patterns they chose.
Finally, I discard stimuli in which the focus verb or its plural/singular inflection does not appear as a single word in the BERT word-piece-based vocabulary (and hence cannot be predicted by the model). This include discarding BIBREF3 stimuli involving the words swims or admires, resulting in 23,368 discarded pairs (out of 152,300). I similarly discard 680 sentences from BIBREF1 where the focus verb or its inflection were one of 108 out-of-vocabulary tokens, and 28 sentence-pairs (8 tokens) from BIBREF2 .
The BERT results are not directly comparable to the numbers reported in previous work. Beyond the differences due to bidirectionality and the discarded stimuli, the BERT models are also trained on a different and larger corpus (covering both wikipedia and books).
Code is available at https://github.com/yoavg/bert-syntax.
Results
Tables 1 , 2 and 3 show the results. All cases exhibit high scores—in the vast majority of the cases substantially higher than reported in previous work. As discussed above, the results are not directly comparable to previous work: the BERT models are trained on different (and larger) data, are allowed to access the suffix of the sentence in addition to its prefix, and are evaluated on somewhat different data due to discarding OOV items. Still, taken together, the high performance numbers indicate that the purely attention-based BERT models are likely capable of capturing the same kind of syntactic regularities that LSTM-based models are capable of capturing, at least as well as the LSTM models and probably better.
Another noticeable and interesting trend is that larger is not necessarily better: the BERT-Base model outperforms the BERT-Large model on many of the syntactic conditions.
Discussion
The BERT models perform remarkably well on all the syntactic test cases. I expected the attention-based mechanism to fail on these (compared to the LSTM-based models), and am surprised by these results. The BIBREF2 and BIBREF3 conditions rule out the possibility of overly relying on selectional preference cues or memorizing the wikipedia training data, and suggest real syntactic generalization is taking place. Exploring the extent to which deep purely-attention-based architectures such as BERT are capable of capturing hierarchy-sensitive and syntactic dependencies—as well as the mechanisms by which this is achieved—is a fascinating area for future research.
| [
"Yes",
"Yes"
] | 1,464 | qasper | en | null | 6a78dbe4f8e30d35c13ea1f80e52df286f9e1f664f9c3d98 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Were any of these tasks evaluated in any previous work?
Answer: | [
"Yes",
"Yes"
] | qasper | 128 | 95 | |
Is datasets for sentiment analysis balanced? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
As social media, specially Twitter, takes on an influential role in presidential elections in the U.S., natural language processing of political tweets BIBREF0 has the potential to help with nowcasting and forecasting of election results as well as identifying the main issues with a candidate – tasks of much interest to journalists, political scientists, and campaign organizers BIBREF1. As a methodology to obtain training data for a machine learning system that analyzes political tweets, BIBREF2 devised a crowdsourcing scheme with variable crowdworker numbers based on the difficulty of the annotation task. They provided a dataset of tweets where the sentiments towards political candidates were labeled both by experts in political communication and by crowdworkers who were likely not domain experts. BIBREF2 revealed that crowdworkers can match expert performance relatively accurately and in a budget-efficient manner. Given this result, the authors envisioned future work in which groundtruth labels would be crowdsourced for a large number of tweets and then used to design an automated NLP tool for political tweet analysis.
The question we address here is: How accurate are existing NLP tools for political tweet analysis? These tools would provide a baseline performance that any new machine learning system for political tweet analysis would compete against. We here explore whether existing NLP systems can answer the questions "What sentiment?" and "Towards whom?" accurately for the dataset of political tweets provided by BIBREF2. In our analysis, we include NLP tools with publicly-available APIs, even if the tools were not specifically designed for short texts like tweets, and, in particular, political tweets.
Our experiments reveal that the task of entity-level sentiment analysis is difficult for existing tools to answer accurately while the recognition of the entity, here, which politician, was easier.
NLP Toolkits
NLP toolkits typically have the following capabilities: tokenization, part-of-speech (PoS) tagging, chunking, named entity recognition and sentiment analysis. In a study by BIBREF3, it is shown that the well-known NLP toolkits NLTK BIBREF4, Stanford CoreNLP BIBREF5, and TwitterNLP BIBREF6 have tokenization, PoS tagging and NER modules in their pipelines. There are two main approaches for NER: (1) rule-based and (2) statistical or machine learning based. The most ubiquitous algorithms for sequence tagging use Hidden Markov Models BIBREF7, Maximum Entropy Markov Models BIBREF7, BIBREF8, or Conditional Random Fields BIBREF9. Recent works BIBREF10, BIBREF11 have used recurrent neural networks with attention modules for NER.
Sentiment detection tools like SentiStrength BIBREF12 and TensiStrength BIBREF13 are rule-based tools, relying on various dictionaries of emoticons, slangs, idioms, and ironic phrases, and set of rules that can detect the sentiment of a sentence overall or a targeted sentiment. Given a list of keywords, TensiStrength (similar to SentiStrength) reports the sentiment towards selected entities in a sentence, based on five levels of relaxation and five levels of stress.
Among commercial NLP toolkits (e.g., BIBREF14, BIBREF15, BIBREF16), we selected BIBREF17 and BIBREF18 for our experiments, which, to the best of our knowledge, are the only publicly accessible commercial APIs for the task of entity-level sentiment analysis that is agnostic to the text domain. We also report results of TensiStrength BIBREF13, TwitterNLP BIBREF6, BIBREF19, CogComp-NLP BIBREF20, and Stanford NLP NER BIBREF21.
Dataset and Analysis Methodology
We used the 1,000-tweet dataset by BIBREF2 that contains the named-entities labels and entity-level sentiments for each of the four 2016 presidential primary candidates Bernie Sanders, Donald Trump, Hillary Clinton, and Ted Cruz, provided by crowdworkers, and by experts in political communication, whose labels are considered groundtruth. The crowdworkers were located in the US and hired on the BIBREF22 platform. For the task of entity-level sentiment analysis, a 3-scale rating of "negative," "neutral," and "positive" was used by the annotators.
BIBREF2 proposed a decision tree approach for computing the number of crowdworkers who should analyze a tweet based on the difficulty of the task. Tweets are labeled by 2, 3, 5, or 7 workers based on the difficulty of the task and the level of disagreement between the crowdworkers. The model computes the number of workers based on how long a tweet is, the presence of a link in a tweet, and the number of present sarcasm signals. Sarcasm is often used in political tweets and causes disagreement between the crowdworkers. The tweets that are deemed to be sarcastic by the decision tree model, are expected to be more difficult to annotate, and hence are allocated more crowdworkers to work on.
We conducted two sets of experiments. In the first set, we used BIBREF23, BIBREF17, and BIBREF18, for entity-level sentiment analysis; in the second set, BIBREF17, BIBREF19, BIBREF24, BIBREF25, and BIBREF26, BIBREF18 for named-entity recognition.
In the experiments that we conducted with TwitterNLP for named-entity recognition, we worked with the default values of the model. Furthermore, we selected the 3-class Stanford NER model, which uses the classes “person,” “organization,” and “location” because it resulted in higher accuracy compared to the 7-class model. For CogComp-NLP NER we used Ontonotes 5.0 NER model BIBREF27. For spaCy NER we used the `en_core_web_lg' model.
We report the experimental results for our two tasks in terms of the correct classification rate (CCR). For sentiment analysis, we have a three-class problem (positive, negative, and neutral), where the classes are mutually exclusive. The CCR, averaged for a set of tweets, is defined to be the number of correctly-predicted sentiments over the number of groundtruth sentiments in these tweets. For NER, we consider that each tweet may reference up to four candidates, i.e., targeted entities. The CCR, averaged for a set of tweets, is the number of correctly predicted entities (candidates) over the number of groundtruth entities (candidates) in this set.
Results and Discussion
The dataset of 1,000 randomly selected tweets contains more than twice as many tweets about Trump than about the other candidates. In the named-entity recognition experiment, the average CCR of crowdworkers was 98.6%, while the CCR of the automated systems ranged from 77.2% to 96.7%. For four of the automated systems, detecting the entity Trump was more difficult than the other entities (e.g., spaCy 72.7% for the entity Trump vs. above 91% for the other entities). An example of incorrect NER is shown in Figure FIGREF1 top. The difficulties the automated tools had in NER may be explained by the fact that the tools were not trained on tweets, except for TwitterNLP, which was not in active development when the data was created BIBREF1.
In the sentiment analysis experiments, we found that a tweet may contain multiple sentiments. The groundtruth labels contain 210 positive sentiments, 521 neutral sentiments, and 305 negative sentiments to the candidates. We measured the CCR, across all tweets, to be 31.7% for Rosette Text Analytics, 43.2% for Google Cloud, 44.2% for TensiStrength, and 74.7% for the crowdworkers. This means the difference between the performance of the tools and the crowdworkers is significant – more than 30 percent points.
Crowdworkers correctly identified 62% of the neutral, 85% of the positive, and 92% of the negative sentiments. Google Cloud correctly identified 88% of the neutral sentiments, but only 3% of the positive, and 19% of the negative sentiments. TensiStrength correctly identified 87.2% of the neutral sentiments, but 10.5% of the positive, and 8.1% of the negative sentiments. Rosette Text Analytics correctly identified 22.7% of neutral sentiments, 38.1% of negative sentiments and 40.9% of positive sentiments. The lowest and highest CCR pertains to tweets about Trump and Sanders for both Google Cloud and TensiStrength, Trump and Clinton for Rosette Text Analytics, and Clinton and Cruz for crowdworkers. An example of incorrect ELS analysis is shown in Figure FIGREF1 bottom.
Conclusions and Future Work
Our results show that existing NLP systems cannot accurately perform sentiment analysis of political tweets in the dataset we experimented with. Labeling by humans, even non-expert crowdworkers, yields accuracy results that are well above the results of existing automated NLP systems. In future work we will therefore use a crowdworker-labeled dataset to train a new machine-learning based NLP system for tweet analysis. We will ensure that the training data is balanced among classes. Our plan is to use state-of-the-art deep neural networks and compare their performance for entity-level sentiment analysis of political tweets.
Acknowledgments
Partial support of this work by the Hariri Institute for Computing and Computational Science & Engineering at Boston University (to L.G.) and a Google Faculty Research Award (to M.B. and L.G.) is gratefully acknowledged. Additionally, we would like to thank Daniel Khashabi for his help in running the CogComp-NLP Python API and Mike Thelwal for his help with TensiStrength. We are also grateful to the Stanford NLP group for clarifying some of the questions we had with regards to the Stanford NER tool.
| [
"No"
] | 1,441 | qasper | en | null | 188fe9331293312465b4564e11ab36dfbcb37191e62a969c | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: Is datasets for sentiment analysis balanced?
Answer: | [
"No"
] | qasper | 128 | 96 | |
What is the invertibility condition? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Data annotation is a major bottleneck for the application of supervised learning approaches to many problems. As a result, unsupervised methods that learn directly from unlabeled data are increasingly important. For tasks related to unsupervised syntactic analysis, discrete generative models have dominated in recent years – for example, for both part-of-speech (POS) induction BIBREF0 , BIBREF1 and unsupervised dependency parsing BIBREF2 , BIBREF3 , BIBREF4 . While similar models have had success on a range of unsupervised tasks, they have mostly ignored the apparent utility of continuous word representations evident from supervised NLP applications BIBREF5 , BIBREF6 . In this work, we focus on leveraging and explicitly representing continuous word embeddings within unsupervised models of syntactic structure.
Pre-trained word embeddings from massive unlabeled corpora offer a compact way of injecting a prior notion of word similarity into models that would otherwise treat words as discrete, isolated categories. However, the specific properties of language captured by any particular embedding scheme can be difficult to control, and, further, may not be ideally suited to the task at hand. For example, pre-trained skip-gram embeddings BIBREF7 with small context window size are found to capture the syntactic properties of language well BIBREF8 , BIBREF9 . However, if our goal is to separate syntactic categories, this embedding space is not ideal – POS categories correspond to overlapping interspersed regions in the embedding space, evident in Figure SECREF4 .
In our approach, we propose to learn a new latent embedding space as a projection of pre-trained embeddings (depicted in Figure SECREF5 ), while jointly learning latent syntactic structure – for example, POS categories or syntactic dependencies. To this end, we introduce a new generative model (shown in Figure FIGREF6 ) that first generates a latent syntactic representation (e.g. a dependency parse) from a discrete structured prior (which we also call the “syntax model”), then, conditioned on this representation, generates a sequence of latent embedding random variables corresponding to each word, and finally produces the observed (pre-trained) word embeddings by projecting these latent vectors through a parameterized non-linear function. The latent embeddings can be jointly learned with the structured syntax model in a completely unsupervised fashion.
By choosing an invertible neural network as our non-linear projector, and then parameterizing our model in terms of the projection's inverse, we are able to derive tractable exact inference and marginal likelihood computation procedures so long as inference is tractable in the underlying syntax model. In sec:learn-with-inv we show that this derivation corresponds to an alternate view of our approach whereby we jointly learn a mapping of observed word embeddings to a new embedding space that is more suitable for the syntax model, but include an additional Jacobian regularization term to prevent information loss.
Recent work has sought to take advantage of word embeddings in unsupervised generative models with alternate approaches BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 . BIBREF9 build an HMM with Gaussian emissions on observed word embeddings, but they do not attempt to learn new embeddings. BIBREF10 , BIBREF11 , and BIBREF12 extend HMM or dependency model with valence (DMV) BIBREF2 with multinomials that use word (or tag) embeddings in their parameterization. However, they do not represent the embeddings as latent variables.
In experiments, we instantiate our approach using both a Markov-structured syntax model and a tree-structured syntax model – specifically, the DMV. We evaluate on two tasks: part-of-speech (POS) induction and unsupervised dependency parsing without gold POS tags. Experimental results on the Penn Treebank BIBREF13 demonstrate that our approach improves the basic HMM and DMV by a large margin, leading to the state-of-the-art results on POS induction, and state-of-the-art results on unsupervised dependency parsing in the difficult training scenario where neither gold POS annotation nor punctuation-based constraints are available.
Model
As an illustrative example, we first present a baseline model for Markov syntactic structure (POS induction) that treats a sequence of pre-trained word embeddings as observations. Then, we propose our novel approach, again using Markov structure, that introduces latent word embedding variables and a neural projector. Lastly, we extend our approach to more general syntactic structures.
Example: Gaussian HMM
We start by describing the Gaussian hidden Markov model introduced by BIBREF9 , which is a locally normalized model with multinomial transitions and Gaussian emissions. Given a sentence of length INLINEFORM0 , we denote the latent POS tags as INLINEFORM1 , observed (pre-trained) word embeddings as INLINEFORM2 , transition parameters as INLINEFORM3 , and Gaussian emission parameters as INLINEFORM4 . The joint distribution of data and latent variables factors as:
DISPLAYFORM0
where INLINEFORM0 is the multinomial transition probability and INLINEFORM1 is the multivariate Gaussian emission probability.
While the observed word embeddings do inform this model with a notion of word similarity – lacking in the basic multinomial HMM – the Gaussian emissions may not be sufficiently flexible to separate some syntactic categories in the complex pre-trained embedding space – for example the skip-gram embedding space as visualized in Figure SECREF4 where different POS categories overlap. Next we introduce a new approach that adds flexibility to the emission distribution by incorporating new latent embedding variables.
Markov Structure with Neural Projector
To flexibly model observed embeddings and yield a new representation space that is more suitable for the syntax model, we propose to cascade a neural network as a projection function, deterministically transforming the simple space defined by the Gaussian HMM to the observed embedding space. We denote the latent embedding of the INLINEFORM0 word in a sentence as INLINEFORM1 , and the neural projection function as INLINEFORM2 , parameterized by INLINEFORM3 . In the case of sequential Markov structure, our new model corresponds to the following generative process:
For each time step INLINEFORM0 ,
[noitemsep, leftmargin=*]
Draw the latent state INLINEFORM0
Draw the latent embedding INLINEFORM0
Deterministically produce embedding
INLINEFORM0
The graphical model is depicted in Figure FIGREF6 . The deterministic projection can also be viewed as sampling each observation from a point mass at INLINEFORM0 . The joint distribution of our model is: DISPLAYFORM0
where INLINEFORM0 is a conditional Gaussian distribution, and INLINEFORM1 is the Dirac delta function centered at INLINEFORM2 : DISPLAYFORM0
General Structure with Neural Projector
Our approach can be applied to a broad family of structured syntax models. We denote latent embedding variables as INLINEFORM0 , discrete latent variables in the syntax model as INLINEFORM1 ( INLINEFORM2 ), where INLINEFORM3 are conditioned to generate INLINEFORM4 . The joint probability of our model factors as:
DISPLAYFORM0
where INLINEFORM0 represents the probability of the syntax model, and can encode any syntactic structure – though, its factorization structure will determine whether inference is tractable in our full model. As shown in Figure FIGREF6 , we focus on two syntax models for syntactic analysis in this paper. The first is Markov-structured, which we use for POS induction, and the second is DMV-structured, which we use to learn dependency parses without supervision.
The marginal data likelihood of our model is: DISPLAYFORM0
While the discrete variables INLINEFORM0 can be marginalized out with dynamic program in many cases, it is generally intractable to marginalize out the latent continuous variables, INLINEFORM1 , for an arbitrary projection INLINEFORM2 in Eq. ( EQREF17 ), which means inference and learning may be difficult. In sec:opt, we address this issue by constraining INLINEFORM3 to be invertible, and show that this constraint enables tractable exact inference and marginal likelihood computation.
Learning & Inference
In this section, we introduce an invertibility condition for our neural projector to tackle the optimization challenge. Specifically, we constrain our neural projector with two requirements: (1) INLINEFORM0 and (2) INLINEFORM1 exists. Invertible transformations have been explored before in independent components analysis BIBREF14 , gaussianization BIBREF15 , and deep density models BIBREF16 , BIBREF17 , BIBREF18 , for unstructured data. Here, we generalize this style of approach to structured learning, and augment it with discrete latent variables ( INLINEFORM2 ). Under the invertibility condition, we derive a learning algorithm and give another view of our approach revealed by the objective function. Then, we present the architecture of a neural projector we use in experiments: a volume-preserving invertible neural network proposed by BIBREF16 for independent components estimation.
Learning with Invertibility
For ease of exposition, we explain the learning algorithm in terms of Markov structure without loss of generality. As shown in Eq. ( EQREF17 ), the optimization challenge in our approach comes from the intractability of the marginalized emission factor INLINEFORM0 . If we can marginalize out INLINEFORM1 and compute INLINEFORM2 , then the posterior and marginal likelihood of our Markov-structured model can be computed with the forward-backward algorithm. We can apply Eq. ( EQREF14 ) and obtain : INLINEFORM3
By using the change of variable rule to the integration, which allows the integration variable INLINEFORM0 to be replaced by INLINEFORM1 , the marginal emission factor can be computed in closed-form when the invertibility condition is satisfied: DISPLAYFORM0
where INLINEFORM0 is a conditional Gaussian distribution, INLINEFORM1 is the Jacobian matrix of function INLINEFORM2 at INLINEFORM3 , and INLINEFORM4 represents the absolute value of its determinant. This Jacobian term is nonzero and differentiable if and only if INLINEFORM5 exists.
Eq. ( EQREF19 ) shows that we can directly calculate the marginal emission distribution INLINEFORM0 . Denote the marginal data likelihood of Gaussian HMM as INLINEFORM1 , then the log marginal data likelihood of our model can be directly written as: DISPLAYFORM0
where INLINEFORM0 represents the new sequence of embeddings after applying INLINEFORM1 to each INLINEFORM2 . Eq. ( EQREF20 ) shows that the training objective of our model is simply the Gaussian HMM log likelihood with an additional Jacobian regularization term. From this view, our approach can be seen as equivalent to reversely projecting the data through INLINEFORM3 to another manifold INLINEFORM4 that is directly modeled by the Gaussian HMM, with a regularization term. Intuitively, we optimize the reverse projection INLINEFORM5 to modify the INLINEFORM6 space, making it more appropriate for the syntax model. The Jacobian regularization term accounts for the volume expansion or contraction behavior of the projection. Maximizing it can be thought of as preventing information loss. In the extreme case, the Jacobian determinant is equal to zero, which means the projection is non-invertible and thus information is being lost through the projection. Such “information preserving” regularization is crucial during optimization, otherwise the trivial solution of always projecting data to the same single point to maximize likelihood is viable.
More generally, for an arbitrary syntax model the data likelihood of our approach is: DISPLAYFORM0
If the syntax model itself allows for tractable inference and marginal likelihood computation, the same dynamic program can be used to marginalize out INLINEFORM0 . Therefore, our joint model inherits the tractability of the underlying syntax model.
Invertible Volume-Preserving Neural Net
For the projection we can use an arbitrary invertible function, and given the representational power of neural networks they seem a natural choice. However, calculating the inverse and Jacobian of an arbitrary neural network can be difficult, as it requires that all component functions be invertible and also requires storage of large Jacobian matrices, which is memory intensive. To address this issue, several recent papers propose specially designed invertible networks that are easily trainable yet still powerful BIBREF16 , BIBREF17 , BIBREF19 . Inspired by these works, we use the invertible transformation proposed by BIBREF16 , which consists of a series of “coupling layers”. This architecture is specially designed to guarantee a unit Jacobian determinant (and thus the invertibility property).
From Eq. ( EQREF22 ) we know that only INLINEFORM0 is required for accomplishing learning and inference; we never need to explicitly construct INLINEFORM1 . Thus, we directly define the architecture of INLINEFORM2 . As shown in Figure FIGREF24 , the nonlinear transformation from the observed embedding INLINEFORM3 to INLINEFORM4 represents the first coupling layer. The input in this layer is partitioned into left and right halves of dimensions, INLINEFORM5 and INLINEFORM6 , respectively. A single coupling layer is defined as: DISPLAYFORM0
where INLINEFORM0 is the coupling function and can be any nonlinear form. This transformation satisfies INLINEFORM1 , and BIBREF16 show that its Jacobian matrix is triangular with all ones on the main diagonal. Thus the Jacobian determinant is always equal to one (i.e. volume-preserving) and the invertibility condition is naturally satisfied.
To be sufficiently expressive, we compose multiple coupling layers as suggested in BIBREF16 . Specifically, we exchange the role of left and right half vectors at each layer as shown in Figure FIGREF24 . For instance, from INLINEFORM0 to INLINEFORM1 the left subset INLINEFORM2 is unchanged, while from INLINEFORM3 to INLINEFORM4 the right subset INLINEFORM5 remains the same. Also note that composing multiple coupling layers does not change the volume-preserving and invertibility properties. Such a sequence of invertible transformations from the data space INLINEFORM6 to INLINEFORM7 is also called normalizing flow BIBREF20 .
Experiments
In this section, we first describe our datasets and experimental setup. We then instantiate our approach with Markov and DMV-structured syntax models, and report results on POS tagging and dependency grammar induction respectively. Lastly, we analyze the learned latent embeddings.
Data
For both POS tagging and dependency parsing, we run experiments on the Wall Street Journal (WSJ) portion of the Penn Treebank. To create the observed data embeddings, we train skip-gram word embeddings BIBREF7 that are found to capture syntactic properties well when trained with small context window BIBREF8 , BIBREF9 . Following BIBREF9 , the dimensionality INLINEFORM0 is set to 100, and the training context window size is set to 1 to encode more syntactic information. The skip-gram embeddings are trained on the one billion word language modeling benchmark dataset BIBREF21 in addition to the WSJ corpus.
General Experimental Setup
For the neural projector, we employ rectified networks as coupling function INLINEFORM0 following BIBREF16 . We use a rectified network with an input layer, one hidden layer, and linear output units, the number of hidden units is set to the same as the number of input units. The number of coupling layers are varied as 4, 8, 16 for both tasks. We optimize marginal data likelihood directly using Adam BIBREF22 . For both tasks in the fully unsupervised setting, we do not tune the hyper-parameters using supervised data.
Unsupervised POS tagging
For unsupervised POS tagging, we use a Markov-structured syntax model in our approach, which is a popular structure for unsupervised tagging tasks BIBREF9 , BIBREF10 .
Following existing literature, we train and test on the entire WSJ corpus (49208 sentences, 1M tokens). We use 45 tag clusters, the number of POS tags that appear in WSJ corpus. We train the discrete HMM and the Gaussian HMM BIBREF9 as baselines. For the Gaussian HMM, mean vectors of Gaussian emissions are initialized with the empirical mean of all word vectors with an additive noise. We assume diagonal covariance matrix for INLINEFORM0 and initialize it with the empirical variance of the word vectors. Following BIBREF9 , the covariance matrix is fixed during training. The multinomial probabilities are initialized as INLINEFORM1 , where INLINEFORM2 . For our approach, we initialize the syntax model and Gaussian parameters with the pre-trained Gaussian HMM. The weights of layers in the rectified network are initialized from a uniform distribution with mean zero and a standard deviation of INLINEFORM3 , where INLINEFORM4 is the input dimension. We evaluate the performance of POS tagging with both Many-to-One (M-1) accuracy BIBREF23 and V-Measure (VM) BIBREF24 . Given a model we found that the tagging performance is well-correlated with the training data likelihood, thus we use training data likelihood as a unsupervised criterion to select the trained model over 10 random restarts after training 50 epochs. We repeat this process 5 times and report the mean and standard deviation of performance.
We compare our approach with basic HMM, Gaussian HMM, and several state-of-the-art systems, including sophisticated HMM variants and clustering techniques with hand-engineered features. The results are presented in Table TABREF32 . Through the introduced latent embeddings and additional neural projection, our approach improves over the Gaussian HMM by 5.4 points in M-1 and 5.6 points in VM. Neural HMM (NHMM) BIBREF10 is a baseline that also learns word representation jointly. Both their basic model and extended Conv version does not outperform the Gaussian HMM. Their best model incorporates another LSTM to model long distance dependency and breaks the Markov assumption, yet our approach still achieves substantial improvement over it without considering more context information. Moreover, our method outperforms the best published result that benefits from hand-engineered features BIBREF27 by 2.0 points on VM.
We found that most tagging errors happen in noun subcategories. Therefore, we do the one-to-one mapping between gold POS tags and induced clusters and plot the normalized confusion matrix of noun subcategories in Figure FIGREF35 . The Gaussian HMM fails to identify “NN” and “NNS” correctly for most cases, and it often recognizes “NNPS” as “NNP”. In contrast, our approach corrects these errors well.
Unsupervised Dependency Parsing without gold POS tags
For the task of unsupervised dependency parse induction, we employ the Dependency Model with Valence (DMV) BIBREF2 as the syntax model in our approach. DMV is a generative model that defines a probability distribution over dependency parse trees and syntactic categories, generating tokens and dependencies in a head-outward fashion. While, traditionally, DMV is trained using gold POS tags as observed syntactic categories, in our approach, we treat each tag as a latent variable, as described in sec:general-neural.
Most existing approaches to this task are not fully unsupervised since they rely on gold POS tags following the original experimental setup for DMV. This is partially because automatically parsing from words is difficult even when using unsupervised syntactic categories BIBREF29 . However, inducing dependencies from words alone represents a more realistic experimental condition since gold POS tags are often unavailable in practice. Previous work that has trained from words alone often requires additional linguistic constraints (like sentence internal boundaries) BIBREF29 , BIBREF30 , BIBREF31 , BIBREF32 , acoustic cues BIBREF33 , additional training data BIBREF4 , or annotated data from related languages BIBREF34 . Our approach is naturally designed to train on word embeddings directly, thus we attempt to induce dependencies without using gold POS tags or other extra linguistic information.
Like previous work we use sections 02-21 of WSJ corpus as training data and evaluate on section 23, we remove punctuations and train the models on sentences of length INLINEFORM0 , “head-percolation” rules BIBREF39 are applied to obtain gold dependencies for evaluation. We train basic DMV, extended DMV (E-DMV) BIBREF35 and Gaussian DMV (which treats POS tag as unknown latent variables and generates observed word embeddings directly conditioned on them following Gaussian distribution) as baselines. Basic DMV and E-DMV are trained with Viterbi EM BIBREF40 on unsupervised POS tags induced from our Markov-structured model described in sec:pos. Multinomial parameters of the syntax model in both Gaussian DMV and our model are initialized with the pre-trained DMV baseline. Other parameters are initialized in the same way as in the POS tagging experiment. The directed dependency accuracy (DDA) is used for evaluation and we report accuracy on sentences of length INLINEFORM1 and all lengths. We train the parser until training data likelihood converges, and report the mean and standard deviation over 20 random restarts.
Our model directly observes word embeddings and does not require gold POS tags during training. Thus, results from related work trained on gold tags are not directly comparable. However, to measure how these systems might perform without gold tags, we run three recent state-of-the-art systems in our experimental setting: UR-A E-DMV BIBREF36 , Neural E-DMV BIBREF11 , and CRF Autoencoder (CRFAE) BIBREF37 . We use unsupervised POS tags (induced from our Markov-structured model) in place of gold tags. We also train basic DMV on gold tags and include several state-of-the-art results on gold tags as reference points.
As shown in Table TABREF39 , our approach is able to improve over the Gaussian DMV by 4.8 points on length INLINEFORM0 and 4.8 points on all lengths, which suggests the additional latent embedding layer and neural projector are helpful. The proposed approach yields, to the best of our knowledge, state-of-the-art performance without gold POS annotation and without sentence-internal boundary information. DMV, UR-A E-DMV, Neural E-DMV, and CRFAE suffer a large decrease in performance when trained on unsupervised tags – an effect also seen in previous work BIBREF29 , BIBREF34 . Since our approach induces latent POS tags jointly with dependency trees, it may be able to learn POS clusters that are more amenable to grammar induction than the unsupervised tags. We observe that CRFAE underperforms its gold-tag counterpart substantially. This may largely be a result of the model's reliance on prior linguistic rules that become unavailable when gold POS tag types are unknown. Many extensions to DMV can be considered orthogonal to our approach – they essentially focus on improving the syntax model. It is possible that incorporating these more sophisticated syntax models into our approach may lead to further improvements.
Sensitivity Analysis
In the above experiments we initialize the structured syntax components with the pre-trained Gaussian or discrete baseline, which is shown as a useful technique to help train our deep models. We further study the results with fully random initialization. In the POS tagging experiment, we report the results in Table TABREF48 . While the performance with 4 layers is comparable to the pre-trained Gaussian initialization, deeper projections (8 or 16 layers) result in a dramatic drop in performance. This suggests that the structured syntax model with very deep projections is difficult to train from scratch, and a simpler projection might be a good compromise in the random initialization setting.
Different from the Markov prior in POS tagging experiments, our parsing model seems to be quite sensitive to the initialization. For example, directed accuracy of our approach on sentences of length INLINEFORM0 is below 40.0 with random initialization. This is consistent with previous work that has noted the importance of careful initialization for DMV-based models such as the commonly used harmonic initializer BIBREF2 . However, it is not straightforward to apply the harmonic initializer for DMV directly in our model without using some kind of pre-training since we do not observe gold POS.
We investigate the effect of the choice of pre-trained embedding on performance while using our approach. To this end, we additionally include results using fastText embeddings BIBREF41 – which, in contrast with skip-gram embeddings, include character-level information. We set the context windows size to 1 and the dimension size to 100 as in the skip-gram training, while keeping other parameters set to their defaults. These results are summarized in Table TABREF50 and Table TABREF51 . While fastText embeddings lead to reduced performance with our model, our approach still yields an improvement over the Gaussian baseline with the new observed embeddings space.
Qualitative Analysis of Embeddings
We perform qualitative analysis to understand how the latent embeddings help induce syntactic structures. First we filter out low-frequency words and punctuations in WSJ, and visualize the rest words (10k) with t-SNE BIBREF42 under different embeddings. We assign each word with its most likely gold POS tags in WSJ and color them according to the gold POS tags.
For our Markov-structured model, we have displayed the embedding space in Figure SECREF5 , where the gold POS clusters are well-formed. Further, we present five example target words and their five nearest neighbors in terms of cosine similarity. As shown in Table TABREF53 , the skip-gram embedding captures both semantic and syntactic aspects to some degree, yet our embeddings are able to focus especially on the syntactic aspects of words, in an unsupervised fashion without using any extra morphological information.
In Figure FIGREF54 we depict the learned latent embeddings with the DMV-structured syntax model. Unlike the Markov structure, the DMV structure maps a large subset of singular and plural nouns to the same overlapping region. However, two clusters of singular and plural nouns are actually separated. We inspect the two clusters and the overlapping region in Figure FIGREF54 , it turns out that the nouns in the separated clusters are words that can appear as subjects and, therefore, for which verb agreement is important to model. In contrast, the nouns in the overlapping region are typically objects. This demonstrates that the latent embeddings are focusing on aspects of language that are specifically important for modeling dependency without ever having seen examples of dependency parses. Some previous work has deliberately created embeddings to capture different notions of similarity BIBREF43 , BIBREF44 , while they use extra morphology or dependency annotations to guide the embedding learning, our approach provides a potential alternative to create new embeddings that are guided by structured syntax model, only using unlabeled text corpora.
Related Work
Our approach is related to flow-based generative models, which are first described in NICE BIBREF16 and have recently received more attention BIBREF17 , BIBREF19 , BIBREF18 . This relevant work mostly adopts simple (e.g. Gaussian) and fixed priors and does not attempt to learn interpretable latent structures. Another related generative model class is variational auto-encoders (VAEs) BIBREF45 that optimize a lower bound on the marginal data likelihood, and can be extended to learn latent structures BIBREF46 , BIBREF47 . Against the flow-based models, VAEs remove the invertibility constraint but sacrifice the merits of exact inference and exact log likelihood computation, which potentially results in optimization challenges BIBREF48 . Our approach can also be viewed in connection with generative adversarial networks (GANs) BIBREF49 that is a likelihood-free framework to learn implicit generative models. However, it is non-trivial for a gradient-based method like GANs to propagate gradients through discrete structures.
Conclusion
In this work, we define a novel generative approach to leverage continuous word representations for unsupervised learning of syntactic structure. Experiments on both POS induction and unsupervised dependency parsing tasks demonstrate the effectiveness of our proposed approach. Future work might explore more sophisticated invertible projections, or recurrent projections that jointly transform the entire input sequence.
| [
"The neural projector must be invertible.",
"we constrain our neural projector with two requirements: (1) INLINEFORM0 and (2) INLINEFORM1 exists"
] | 4,323 | qasper | en | null | 0e83a6f7ee840931e1851402cc87bd34f52fe8bfa4dc1cab | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: What is the invertibility condition?
Answer: | [
"The neural projector must be invertible.",
"we constrain our neural projector with two requirements: (1) INLINEFORM0 and (2) INLINEFORM1 exists"
] | qasper | 128 | 97 | |
How does proposed qualitative annotation schema looks like? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
There is a recent spark of interest in the task of Question Answering (QA) over unstructured textual data, also referred to as Machine Reading Comprehension (MRC). This is mostly due to wide-spread success of advances in various facets of deep learning related research, such as novel architectures BIBREF0, BIBREF1 that allow for efficient optimisation of neural networks consisting of multiple layers, hardware designed for deep learning purposes and software frameworks BIBREF2, BIBREF3 that allow efficient development and testing of novel approaches. These factors enable researchers to produce models that are pre-trained on large scale corpora and provide contextualised word representations BIBREF4 that are shown to be a vital component towards solutions for a variety of natural language understanding tasks, including MRC BIBREF5. Another important factor that led to the recent success in MRC-related tasks is the widespread availability of various large datasets, e.g., SQuAD BIBREF6, that provide sufficient examples for optimising statistical models. The combination of these factors yields notable results, even surpassing human performance BIBREF7.
MRC is a generic task format that can be used to probe for various natural language understanding capabilities BIBREF8. Therefore it is crucially important to establish a rigorous evaluation methodology in order to be able to draw reliable conclusions from conducted experiments. While increasing effort is put into the evaluation of novel architectures, such as keeping the evaluation data from public access to prevent unintentional overfitting to test data, performing ablation and error studies and introducing novel metrics BIBREF9, surprisingly little is done to establish the quality of the data itself. Additionally, recent research arrived at worrisome findings: the data of those gold standards, which is usually gathered involving a crowd-sourcing step, suffers from flaws in design BIBREF10 or contains overly specific keywords BIBREF11. Furthermore, these gold standards contain “annotation artefacts”, cues that lead models into focusing on superficial aspects of text, such as lexical overlap and word order, instead of actual language understanding BIBREF12, BIBREF13. These weaknesses cast some doubt on whether the data can reliably evaluate the reading comprehension performance of the models they evaluate, i.e. if the models are indeed being assessed for their capability to read.
Figure FIGREF3 shows an example from HotpotQA BIBREF14, a dataset that exhibits the last kind of weakness mentioned above, i.e., the presence of unique keywords in both the question and the passage (in close proximity to the expected answer).
An evaluation methodology is vital to the fine-grained understanding of challenges associated with a single gold standard, in order to understand in greater detail which capabilities of MRC models it evaluates. More importantly, it allows to draw comparisons between multiple gold standards and between the results of respective state-of-the-art models that are evaluated on them.
In this work, we take a step back and propose a framework to systematically analyse MRC evaluation data, typically a set of questions and expected answers to be derived from accompanying passages. Concretely, we introduce a methodology to categorise the linguistic complexity of the textual data and the reasoning and potential external knowledge required to obtain the expected answer. Additionally we propose to take a closer look at the factual correctness of the expected answers, a quality dimension that appears under-explored in literature.
We demonstrate the usefulness of the proposed framework by applying it to precisely describe and compare six contemporary MRC datasets. Our findings reveal concerns about their factual correctness, the presence of lexical cues that simplify the task of reading comprehension and the lack of semantic altering grammatical modifiers. We release the raw data comprised of 300 paragraphs, questions and answers richly annotated under the proposed framework as a resource for researchers developing natural language understanding models and datasets to utilise further.
To the best of our knowledge this is the first attempt to introduce a common evaluation methodology for MRC gold standards and the first across-the-board qualitative evaluation of MRC datasets with respect to the proposed categories.
Framework for MRC Gold Standard Analysis ::: Problem definition
We define the task of machine reading comprehension, the target application of the proposed methodology as follows: Given a paragraph $P$ that consists of tokens (words) $p_1, \ldots , p_{n_P}$ and a question $Q$ that consists of tokens $q_1 \ldots q_{n_Q}$, the goal is to retrieve an answer $A$ with tokens $a_1 \ldots a_{n_A}$. $A$ is commonly constrained to be one of the following cases BIBREF15, illustrated in Figure FIGREF9:
Multiple choice, where the goal is to predict $A$ from a given set of choices $\mathcal {A}$.
Cloze-style, where $S$ is a sentence, and $A$ and $Q$ are obtained by removing a sequence of words such that $Q = S - A$. The task is to fill in the resulting gap in $Q$ with the expected answer $A$ to form $S$.
Span, where is a continuous subsequence of tokens from the paragraph ($A \subseteq P$). Flavours include multiple spans as the correct answer or $A \subseteq Q$.
Free form, where $A$ is an unconstrained natural language string.
A gold standard $G$ is composed of $m$ entries $(Q_i, A_i, P_i)_{i\in \lbrace 1,\ldots , m\rbrace }$.
The performance of an approach is established by comparing its answer predictions $A^*_{i}$ on the given input $(Q_i, T_i)$ (and $\mathcal {A}_i$ for the multiple choice setting) against the expected answer $A_i$ for all $i\in \lbrace 1,\ldots , m\rbrace $ under a performance metric. Typical performance metrics are exact match (EM) or accuracy, i.e. the percentage of exactly predicted answers, and the F1 score – the harmonic mean between the precision and the recall of the predicted tokens compared to expected answer tokens. The overall F1 score can either be computed by averaging the F1 scores for every instance or by first averaging the precision and recall and then computing the F1 score from those averages (macro F1). Free-text answers, meanwhile, are evaluated by means of text generation and summarisation metrics such as BLEU BIBREF16 or ROUGE-L BIBREF17.
Framework for MRC Gold Standard Analysis ::: Dimensions of Interest
In this section we describe a methodology to categorise gold standards according to linguistic complexity, required reasoning and background knowledge, and their factual correctness. Specifically, we use those dimensions as high-level categories of a qualitative annotation schema for annotating question, expected answer and the corresponding context. We further enrich the qualitative annotations by a metric based on lexical cues in order to approximate a lower bound for the complexity of the reading comprehension task. By sampling entries from each gold standard and annotating them, we obtain measurable results and thus are able to make observations about the challenges present in that gold standard data.
Framework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Problem setting
We are interested in different types of the expected answer. We differentiate between Span, where an answer is a continuous span taken from the passage, Paraphrasing, where the answer is a paraphrase of a text span, Unanswerable, where there is no answer present in the context, and Generated, if it does not fall into any of the other categories. It is not sufficient for an answer to restate the question or combine multiple Span or Paraphrasing answers to be annotated as Generated. It is worth mentioning that we focus our investigations on answerable questions. For a complementary qualitative analysis that categorises unanswerable questions, the reader is referred to Yatskar2019.
Furthermore, we mark a sentence as Supporting Fact if it contains evidence required to produce the expected answer, as they are used further in the complexity analysis.
Framework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Factual Correctness
An important factor for the quality of a benchmark is its factual correctness, because on the one hand, the presence of factually wrong or debatable examples introduces an upper bound for the achievable performance of models on those gold standards. On the other hand, it is hard to draw conclusions about the correctness of answers produced by a model that is evaluated on partially incorrect data.
One way by which developers of modern crowd-sourced gold standards ensure quality is by having the same entry annotated by multiple workers BIBREF18 and keeping only those with high agreement. We investigate whether this method is enough to establish a sound ground truth answer that is unambiguously correct. Concretely we annotate an answer as Debatable when the passage features multiple plausible answers, when multiple expected answers contradict each other, or an answer is not specific enough with respect to the question and a more specific answer is present. We annotate an answer as Wrong when it is factually wrong and a correct answer is present in the context.
Framework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Required Reasoning
It is important to understand what types of reasoning the benchmark evaluates, in order to be able to accredit various reasoning capabilities to the models it evaluates. Our proposed reasoning categories are inspired by those found in scientific question answering literature BIBREF19, BIBREF20, as research in this area focuses on understanding the required reasoning capabilities. We include reasoning about the Temporal succession of events, Spatial reasoning about directions and environment, and Causal reasoning about the cause-effect relationship between events. We further annotate (multiple-choice) answers that can only be answered By Exclusion of every other alternative.
We further extend the reasoning categories by operational logic, similar to those required in semantic parsing tasks BIBREF21, as solving those tasks typically requires “multi-hop” reasoning BIBREF14, BIBREF22. When an answer can only be obtained by combining information from different sentences joined by mentioning a common entity, concept, date, fact or event (from here on called entity), we annotate it as Bridge. We further annotate the cases, when the answer is a concrete entity that satisfies a Constraint specified in the question, when it is required to draw a Comparison of multiple entities' properties or when the expected answer is an Intersection of their properties (e.g. “What do Person A and Person B have in common?”)
We are interested in the linguistic reasoning capabilities probed by a gold standard, therefore we include the appropriate category used by Wang2019. Specifically, we annotate occurrences that require understanding of Negation, Quantifiers (such as “every”, “some”, or “all”), Conditional (“if ...then”) statements and the logical implications of Con-/Disjunction (i.e. “and” and “or”) in order to derive the expected answer.
Finally, we investigate whether arithmetic reasoning requirements emerge in MRC gold standards as this can probe for reasoning that is not evaluated by simple answer retrieval BIBREF23. To this end, we annotate the presence of of Addition and Subtraction, answers that require Ordering of numerical values, Counting and Other occurrences of simple mathematical operations.
An example can exhibit multiple forms of reasoning. Notably, we do not annotate any of the categories mentioned above if the expected answer is directly stated in the passage. For example, if the question asks “How many total points were scored in the game?” and the passage contains a sentence similar to “The total score of the game was 51 points”, it does not require any reasoning, in which case we annotate it as Retrieval.
Framework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Knowledge
Worthwhile knowing is whether the information presented in the context is sufficient to answer the question, as there is an increase of benchmarks deliberately designed to probe a model's reliance on some sort of background knowledge BIBREF24. We seek to categorise the type of knowledge required. Similar to Wang2019, on the one hand we annotate the reliance on factual knowledge, that is (Geo)political/Legal, Cultural/Historic, Technical/Scientific and Other Domain Specific knowledge about the world that can be expressed as a set of facts. On the other hand, we denote Intuitive knowledge requirements, which is challenging to express as a set of facts, such as the knowledge that a parenthetic numerical expression next to a person's name in a biography usually denotes his life span.
Framework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Linguistic Complexity
Another dimension of interest is the evaluation of various linguistic capabilities of MRC models BIBREF25, BIBREF26, BIBREF27. We aim to establish which linguistic phenomena are probed by gold standards and to which degree. To that end, we draw inspiration from the annotation schema used by Wang2019, and adapt it around lexical semantics and syntax.
More specifically, we annotate features that introduce variance between the supporting facts and the question. With regard to lexical semantics, we focus on the use of redundant words that do not alter the meaning of a sentence for the task of retrieving the expected answer (Redundancy), requirements on the understanding of words' semantic fields (Lexical Entailment) and the use of Synonyms and Paraphrases with respect to the question wording. Furthermore we annotate cases where supporting facts contain Abbreviations of concepts introduced in the question (and vice versa) and when a Dative case substitutes the use of a preposition (e.g. “I bought her a gift” vs “I bought a gift for her”). Regarding syntax, we annotate changes from passive to active Voice, the substitution of a Genitive case with a preposition (e.g. “of”) and changes from nominal to verbal style and vice versa (Nominalisation).
We recognise features that add ambiguity to the supporting facts, for example when information is only expressed implicitly by using an Ellipsis. As opposed to redundant words, we annotate Restrictivity and Factivity modifiers, words and phrases whose presence does change the meaning of a sentence with regard to the expected answer, and occurrences of intra- or inter-sentence Coreference in supporting facts (that is relevant to the question). Lastly, we mark ambiguous syntactic features, when their resolution is required in order to obtain the answer. Concretely, we mark argument collection with con- and disjunctions (Listing) and ambiguous Prepositions, Coordination Scope and Relative clauses/Adverbial phrases/Appositions.
Framework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Complexity
Finally, we want to approximate the presence of lexical cues that might simplify the reading required in order to arrive at the answer. Quantifying this allows for more reliable statements about and comparison of the complexity of gold standards, particularly regarding the evaluation of comprehension that goes beyond simple lexical matching. We propose the use of coarse metrics based on lexical overlap between question and context sentences. Intuitively, we aim to quantify how much supporting facts “stand out” from their surrounding passage context. This can be used as proxy for the capability to retrieve the answer BIBREF10. Specifically, we measure (i) the number of words jointly occurring in a question and a sentence, (ii) the length of the longest n-gram shared by question and sentence and (iii) whether a word or n-gram from the question uniquely appears in a sentence.
The resulting taxonomy of the framework is shown in Figure FIGREF10. The full catalogue of features, their description, detailed annotation guideline as well as illustrating examples can be found in Appendix .
Application of the Framework ::: Candidate Datasets
We select contemporary MRC benchmarks to represent all four commonly used problem definitions BIBREF15. In selecting relevant datasets, we do not consider those that are considered “solved”, i.e. where the state of the art performance surpasses human performance, as is the case with SQuAD BIBREF28, BIBREF7. Concretely, we selected gold standards that fit our problem definition and were published in the years 2016 to 2019, have at least $(2019 - publication\ year) \times 20$ citations, and bucket them according to the answer selection styles as described in Section SECREF4 We randomly draw one from each bucket and add two randomly drawn datasets from the candidate pool. This leaves us with the datasets described in Table TABREF19. For a more detailed description, we refer to Appendix .
Application of the Framework ::: Annotation Task
We randomly select 50 distinct question, answer and passage triples from the publicly available development sets of the described datasets. Training, development and the (hidden) test set are drawn from the same distribution defined by the data collection method of the respective dataset. For those collections that contain multiple questions over a single passage, we ensure that we are sampling unique paragraphs in order to increase the variety of investigated texts.
The samples were annotated by the first author of this paper, using the proposed schema. In order to validate our findings, we further take 20% of the annotated samples and present them to a second annotator (second author). Since at its core, the annotation is a multi-label task, we report the inter-annotator agreement by computing the (micro-averaged) F1 score, where we treat the first annotator's labels as gold. Table TABREF21 reports the agreement scores, the overall (micro) average F1 score of the annotations is 0.82, which means that on average, more than two thirds of the overall annotated labels were agreed on by both annotators. We deem this satisfactory, given the complexity of the annotation schema.
Application of the Framework ::: Qualitative Analysis
We present a concise view of the annotation results in Figure FIGREF23. The full annotation results can be found in Appendix . We centre our discussion around the following main points:
Application of the Framework ::: Qualitative Analysis ::: Linguistic Features
As observed in Figure FIGREF23 the gold standards feature a high degree of Redundancy, peaking at 76% of the annotated HotpotQA samples and synonyms and paraphrases (labelled Synonym), with ReCoRd samples containing 58% of them, likely to be attributed to the elaborating type of discourse of the dataset sources (encyclopedia and newswire). This is, however, not surprising, as it is fairly well understood in the literature that current state-of-the-art models perform well on distinguishing relevant words and phrases from redundant ones BIBREF32. Additionally, the representational capability of synonym relationships of word embeddings has been investigated and is well known BIBREF33. Finally, we observe the presence of syntactic features, such as ambiguous relative clauses, appositions and adverbial phrases, (RelAdvApp 40% in HotpotQA and ReCoRd) and those introducing variance, concretely switching between verbal and nominal styles (e.g. Nominalisation 10% in HotpotQA) and from passive to active voice (Voice, 8% in HotpotQA).
Syntactic features contributing to variety and ambiguity that we did not observe in our samples are the exploitation of verb symmetry, the use of dative and genitive cases or ambiguous prepositions and coordination scope (respectively Symmetry, Dative, Genitive, Prepositions, Scope). Therefore we cannot establish whether models are capable of dealing with those features by evaluating them on those gold standards.
Application of the Framework ::: Qualitative Analysis ::: Factual Correctness
We identify three common sources that surface in different problems regarding an answer's factual correctness, as reported in Figure FIGREF23 and illustrate their instantiations in Table TABREF31:
Design Constraints: Choosing the task design and the data collection method introduces some constraints that lead to factually debatable examples. For example, a span might have been arbitrarily selected from multiple spans that potentially answer a question, but only a single continuous answer span per question is allowed by design, as observed in the NewsQA and MsMarco samples (32% and 34% examples annotated as Debatable with 16% and 53% thereof exhibiting arbitrary selection, respectively). Sometimes, when additional passages are added after the annotation step, they can by chance contain passages that answer the question more precisely than the original span, as seen in HotpotQA (16% Debatable samples, 25% of them due to arbitrary selection). In the case of MultiRC it appears to be inconsistent, whether multiple correct answer choices are expected to be correct in isolation or in conjunction (28% Debatable with 29% of them exhibiting this problem). This might provide an explanation to its relatively weak human baseline performance of 84% F1 score BIBREF31.
Weak Quality assurance: When the (typically crowd-sourced) annotations are not appropriately validated, incorrect examples will find their way into the gold standards. This typically results in factually wrong expected answers (i.e. when a more correct answer is present in the context) or a question is expected to be Unanswerable, but is actually answerable from the provided context. The latter is observed in MsMarco (83% of examples annotated as Wrong) and NewsQA, where 60% of the examples annotated as Wrong are Unanswerable with an answer present.
Arbitrary Precision: There appears to be no clear guideline on how precise the answer is expected to be, when the passage expresses the answer in varying granularities. We annotated instances as Debatable when the expected answer was not the most precise given the context (44% and 29% of Debatable instances in NewsQA and MultiRC, respectively).
Application of the Framework ::: Qualitative Analysis ::: Semantics-altering grammatical modifiers
We took interest in whether any of the benchmarks contain what we call distracting lexical features (or distractors): grammatical modifiers that alter the semantics of a sentence for the final task of answering the given question while preserving a similar lexical form. An example of such features are cues for (double) Negation (e.g., “no”, “not”), which when introduced in a sentence, reverse its meaning. Other examples include modifiers denoting Restrictivity, Factivity and Reasoning (such as Monotonicity and Conditional cues). Examples of question-answer pairs containing a distractor are shown in Table FIGREF37.
We posit that the presence of such distractors would allow for evaluating reading comprehension beyond potential simple word matching. However, we observe no presence of such features in the benchmarks (beyond Negation in DROP, ReCoRd and HotpotQA, with 4%, 4% and 2% respectively). This results in gold standards that clearly express the evidence required to obtain the answer, lacking more challenging, i.e., distracting, sentences that can assess whether a model can truly understand meaning.
Application of the Framework ::: Qualitative Analysis ::: Other
In the Figure FIGREF23 we observe that Operational and Arithmetic reasoning moderately (6% to 8% combined) appears “in the wild”, i.e. when not enforced by the data design as is the case with HotpotQA (80% Operations combined) or DROP (68% Arithmetic combined). Causal reasoning is (exclusively) present in MultiRC (32%), whereas Temporal and Spatial reasoning requirements seem to not naturally emerge in gold standards. In ReCoRd, a fraction of 38% questions can only be answered By Exclusion of every other candidate, due to the design choice of allowing questions where the required information to answer them is not fully expressed in the accompanying paragraph.
Therefore, it is also a little surprising to observe that ReCoRd requires external resources with regard to knowledge, as seen in Figure FIGREF23. MultiRC requires technical or more precisely basic scientific knowledge (6% Technical/Scientific), as a portion of paragraphs is extracted from elementary school science textbooks BIBREF31. Other benchmarks moderately probe for factual knowledge (0% to 4% across all categories), while Intuitive knowledge is required to derive answers in each gold standard.
It is also worth pointing out, as done in Figure FIGREF23, that although MultiRC and MsMarco are not modelled as a span selection problem, their samples still contain 50% and 66% of answers that are directly taken from the context. DROP contains the biggest fraction of generated answers (60%), due to the requirement of arithmetic operations.
To conclude our analysis, we observe similar distributions of linguistic features and reasoning patterns, except where there are constraints enforced by dataset design, annotation guidelines or source text choice. Furthermore, careful consideration of design choices (such as single-span answers) is required, to avoid impairing the factual correctness of datasets, as pure crowd-worker agreement seems not sufficient in multiple cases.
Application of the Framework ::: Quantitative Results ::: Lexical overlap
We used the scores assigned by our proposed set of metrics (discussed in Section SECREF11 Dimensions of Interest: Complexity) to predict the supporting facts in the gold standard samples (that we included in our manual annotation). Concretely, we used the following five features capturing lexical overlap: (i) the number of words occurring in sentence and question, (ii) the length of the longest n-gram shared by sentence and question, whether a (iii) uni- and (iv) bigram from the question is unique to a sentence, and (v) the sentence index, as input to a logistic regression classifier. We optimised on each sample leaving one example for evaluation. We compute the average Precision, Recall and F1 score by means of leave-one-out validation with every sample entry. The averaged results after 5 runs are reported in Table TABREF41.
We observe that even by using only our five features based lexical overlap, the simple logistic regression baseline is able to separate out the supporting facts from the context to a varying degree. This is in line with the lack of semantics-altering grammatical modifiers discussed in the qualitative analysis section above. The classifier performs best on DROP (66% F1) and MultiRC (40% F1), which means that lexical cues can considerably facilitate the search for the answer in those gold standards. On MultiRC, yadav2019quick come to a similar conclusion, by using a more sophisticated approach based on overlap between question, sentence and answer choices.
Surprisingly, the classifier is able to pick up a signal from supporting facts even on data that has been pruned against lexical overlap heuristics by populating the context with additional documents that have high overlap scores with the question. This results in significantly higher scores than when guessing randomly (HotpotQA 26% F1, and MsMarco 11% F1). We observe similar results in the case the length of the question leaves few candidates to compute overlap with $6.3$ and $7.3$ tokens on average for MsMarco and NewsQA (26% F1), compared to $16.9$ tokens on average for the remaining four dataset samples.
Finally, it is worth mentioning that although the queries in ReCoRd are explicitly independent from the passage, the linear classifier is still capable of achieving 34% F1 score in predicting the supporting facts.
However, neural networks perform significantly better than our admittedly crude baseline (e.g. 66% F1 for supporting facts classification on HotpotQA BIBREF14), albeit utilising more training examples, and a richer sentence representation. This facts implies that those neural models are capable of solving more challenging problems than simple “text matching” as performed by the logistic regression baseline. However, they still circumvent actual reading comprehension as the respective gold standards are of limited suitability to evaluate this BIBREF34, BIBREF35. This suggests an exciting future research direction, that is categorising the scale between text matching and reading comprehension more precisely and respectively positioning state-of-the-art models thereon.
Related Work
Although not as prominent as the research on novel architecture, there has been steady progress in critically investigating the data and evaluation aspects of NLP and machine learning in general and MRC in particular.
Related Work ::: Adversarial Evaluation
The authors of the AddSent algorithm BIBREF11 show that MRC models trained and evaluated on the SQuAD dataset pay too little attention to details that might change the semantics of a sentence, and propose a crowd-sourcing based method to generate adversary examples to exploit that weakness. This method was further adapted to be fully automated BIBREF36 and applied to different gold standards BIBREF35. Our proposed approach differs in that we aim to provide qualitative justifications for those quantitatively measured issues.
Related Work ::: Sanity Baselines
Another line of research establishes sane baselines to provide more meaningful context to the raw performance scores of evaluated models. When removing integral parts of the task formulation such as question, the textual passage or parts thereof BIBREF37 or restricting model complexity by design in order to suppress some required form of reasoning BIBREF38, models are still able to perform comparably to the state-of-the-art. This raises concerns about the perceived benchmark complexity and is related to our work in a broader sense as one of our goals is to estimate the complexity of benchmarks.
Related Work ::: Benchmark evaluation in NLP
Beyond MRC, efforts similar to ours that pursue the goal of analysing the evaluation of established datasets exist in Natural Language Inference BIBREF13, BIBREF12. Their analyses reveal the existence of biases in training and evaluation data that can be approximated with simple majority-based heuristics. Because of these biases, trained models fail to extract the semantics that are required for the correct inference. Furthermore, a fair share of work was done to reveal gender bias in coreference resolution datasets and models BIBREF39, BIBREF40, BIBREF41.
Related Work ::: Annotation Taxonomies
Finally, related to our framework are works that introduce annotation categories for gold standards evaluation. Concretely, we build our annotation framework around linguistic features that were introduced in the GLUE suite BIBREF42 and the reasoning categories introduced in the WorldTree dataset BIBREF19. A qualitative analysis complementary to ours, with focus on the unanswerability patterns in datasets that feature unanswerable questions was done by Yatskar2019.
Conclusion
In this paper, we introduce a novel framework to characterise machine reading comprehension gold standards. This framework has potential applications when comparing different gold standards, considering the design choices for a new gold standard and performing qualitative error analyses for a proposed approach.
Furthermore we applied the framework to analyse popular state-of-the-art gold standards for machine reading comprehension: We reveal issues with their factual correctness, show the presence of lexical cues and we observe that semantics-altering grammatical modifiers are missing in all of the investigated gold standards. Studying how to introduce those modifiers into gold standards and observing whether state-of-the-art MRC models are capable of performing reading comprehension on text containing them, is a future research goal.
A future line of research is to extend the framework to be able to identify the different types of exploitable cues such as question or entity typing and concrete overlap patterns. This will allow the framework to serve as an interpretable estimate of reading comprehension complexity of gold standards. Finally, investigating gold standards under this framework where MRC models outperform the human baseline (e.g. SQuAD) will contribute to a deeper understanding of the seemingly superb performance of deep learning approaches on them.
| [
"The resulting taxonomy of the framework is shown in Figure FIGREF10",
"FIGREF10"
] | 4,958 | qasper | en | null | 894a0e08b526f2093c854d91c680190c898ae6acbc1ba131 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: How does proposed qualitative annotation schema looks like?
Answer: | [
"The resulting taxonomy of the framework is shown in Figure FIGREF10",
"FIGREF10"
] | qasper | 128 | 98 | |
what are the sizes of both datasets? | You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Article: Introduction
Text simplification aims to reduce the lexical and structural complexity of a text, while still retaining the semantic meaning, which can help children, non-native speakers, and people with cognitive disabilities, to understand text better. One of the methods of automatic text simplification can be generally divided into three categories: lexical simplification (LS) BIBREF0 , BIBREF1 , rule-based BIBREF2 , and machine translation (MT) BIBREF3 , BIBREF4 . LS is mainly used to simplify text by substituting infrequent and difficult words with frequent and easier words. However, there are several challenges for the LS approach: a great number of transformation rules are required for reasonable coverage and should be applied based on the specific context; third, the syntax and semantic meaning of the sentence is hard to retain. Rule-based approaches use hand-crafted rules for lexical and syntactic simplification, for example, substituting difficult words in a predefined vocabulary. However, such approaches need a lot of human-involvement to manually define these rules, and it is impossible to give all possible simplification rules. MT-based approach has attracted great attention in the last several years, which addresses text simplification as a monolingual machine translation problem translating from 'ordinary' and 'simplified' sentences.
In recent years, neural Machine Translation (NMT) is a newly-proposed deep learning approach and achieves very impressive results BIBREF5 , BIBREF6 , BIBREF7 . Unlike the traditional phrased-based machine translation system which operates on small components separately, NMT system is being trained end-to-end, without the need to have external decoders, language models or phrase tables. Therefore, the existing architectures in NMT are used for text simplification BIBREF8 , BIBREF4 . However, most recent work using NMT is limited to the training data that are scarce and expensive to build. Language models trained on simplified corpora have played a central role in statistical text simplification BIBREF9 , BIBREF10 . One main reason is the amount of available simplified corpora typically far exceeds the amount of parallel data. The performance of models can be typically improved when trained on more data. Therefore, we expect simplified corpora to be especially helpful for NMT models.
In contrast to previous work, which uses the existing NMT models, we explore strategy to include simplified training corpora in the training process without changing the neural network architecture. We first propose to pair simplified training sentences with synthetic ordinary sentences during training, and treat this synthetic data as additional training data. We obtain synthetic ordinary sentences through back-translation, i.e. an automatic translation of the simplified sentence into the ordinary sentence BIBREF11 . Then, we mix the synthetic data into the original (simplified-ordinary) data to train NMT model. Experimental results on two publicly available datasets show that we can improve the text simplification quality of NMT models by mixing simplified sentences into the training set over NMT model only using the original training data.
Related Work
Automatic TS is a complicated natural language processing (NLP) task, which consists of lexical and syntactic simplification levels BIBREF12 . It has attracted much attention recently as it could make texts more accessible to wider audiences, and used as a pre-processing step, improve performances of various NLP tasks and systems BIBREF13 , BIBREF14 , BIBREF15 . Usually, hand-crafted, supervised, and unsupervised methods based on resources like English Wikipedia and Simple English Wikipedia (EW-SEW) BIBREF10 are utilized for extracting simplification rules. It is very easy to mix up the automatic TS task and the automatic summarization task BIBREF3 , BIBREF16 , BIBREF6 . TS is different from text summarization as the focus of text summarization is to reduce the length and redundant content.
At the lexical level, lexical simplification systems often substitute difficult words using more common words, which only require a large corpus of regular text to obtain word embeddings to get words similar to the complex word BIBREF1 , BIBREF9 . Biran et al. BIBREF0 adopted an unsupervised method for learning pairs of complex and simpler synonyms from a corpus consisting of Wikipedia and Simple Wikipedia. At the sentence level, a sentence simplification model was proposed by tree transformation based on statistical machine translation (SMT) BIBREF3 . Woodsend and Lapata BIBREF17 presented a data-driven model based on a quasi-synchronous grammar, a formalism that can naturally capture structural mismatches and complex rewrite operations. Wubben et al. BIBREF18 proposed a phrase-based machine translation (PBMT) model that is trained on ordinary-simplified sentence pairs. Xu et al. BIBREF19 proposed a syntax-based machine translation model using simplification-specific objective functions and features to encourage simpler output.
Compared with SMT, neural machine translation (NMT) has shown to produce state-of-the-art results BIBREF5 , BIBREF7 . The central approach of NMT is an encoder-decoder architecture implemented by recurrent neural networks, which can represent the input sequence as a vector, and then decode that vector into an output sequence. Therefore, NMT models were used for text simplification task, and achieved good results BIBREF8 , BIBREF4 , BIBREF20 . The main limitation of the aforementioned NMT models for text simplification depended on the parallel ordinary-simplified sentence pairs. Because ordinary-simplified sentence pairs are expensive and time-consuming to build, the available largest data is EW-SEW that only have 296,402 sentence pairs. The dataset is insufficiency for NMT model if we want to NMT model can obtain the best parameters. Considering simplified data plays an important role in boosting fluency for phrase-based text simplification, and we investigate the use of simplified data for text simplification. We are the first to show that we can effectively adapt neural translation models for text simplifiation with simplified corpora.
Simplified Corpora
We collected a simplified dataset from Simple English Wikipedia that are freely available, which has been previously used for many text simplification methods BIBREF0 , BIBREF10 , BIBREF3 . The simple English Wikipedia is pretty easy to understand than normal English Wikipedia. We downloaded all articles from Simple English Wikipedia. For these articles, we removed stubs, navigation pages and any article that consisted of a single sentence. We then split them into sentences with the Stanford CorNLP BIBREF21 , and deleted these sentences whose number of words are smaller than 10 or large than 40. After removing repeated sentences, we chose 600K sentences as the simplified data with 11.6M words, and the size of vocabulary is 82K.
Text Simplification using Neural Machine Translation
Our work is built on attention-based NMT BIBREF5 as an encoder-decoder network with recurrent neural networks (RNN), which simultaneously conducts dynamic alignment and generation of the target simplified sentence.
The encoder uses a bidirectional RNN that consists of forward and backward RNN. Given a source sentence INLINEFORM0 , the forward RNN and backward RNN calculate forward hidden states INLINEFORM1 and backward hidden states INLINEFORM2 , respectively. The annotation vector INLINEFORM3 is obtained by concatenating INLINEFORM4 and INLINEFORM5 .
The decoder is a RNN that predicts a target simplificated sentence with Gated Recurrent Unit (GRU) BIBREF22 . Given the previously generated target (simplified) sentence INLINEFORM0 , the probability of next target word INLINEFORM1 is DISPLAYFORM0
where INLINEFORM0 is a non-linear function, INLINEFORM1 is the embedding of INLINEFORM2 , and INLINEFORM3 is a decoding state for time step INLINEFORM4 .
State INLINEFORM0 is calculated by DISPLAYFORM0
where INLINEFORM0 is the activation function GRU.
The INLINEFORM0 is the context vector computed as a weighted annotation INLINEFORM1 , computed by DISPLAYFORM0
where the weight INLINEFORM0 is computed by DISPLAYFORM0 DISPLAYFORM1
where INLINEFORM0 , INLINEFORM1 and INLINEFORM2 are weight matrices. The training objective is to maximize the likelihood of the training data. Beam search is employed for decoding.
Synthetic Simplified Sentences
We train an auxiliary system using NMT model from the simplified sentence to the ordinary sentence, which is first trained on the available parallel data. For leveraging simplified sentences to improve the quality of NMT model for text simplification, we propose to adapt the back-translation approach proposed by Sennrich et al. BIBREF11 to our scenario. More concretely, Given one sentence in simplified sentences, we use the simplified-ordinary system in translate mode with greedy decoding to translate it to the ordinary sentences, which is denoted as back-translation. This way, we obtain a synthetic parallel simplified-ordinary sentences. Both the synthetic sentences and the available parallel data are used as training data for the original NMT system.
Evaluation
We evaluate the performance of text simplification using neural machine translation on available parallel sentences and additional simplified sentences.
Dataset. We use two simplification datasets (WikiSmall and WikiLarge). WikiSmall consists of ordinary and simplified sentences from the ordinary and simple English Wikipedias, which has been used as benchmark for evaluating text simplification BIBREF17 , BIBREF18 , BIBREF8 . The training set has 89,042 sentence pairs, and the test set has 100 pairs. WikiLarge is also from Wikipedia corpus whose training set contains 296,402 sentence pairs BIBREF19 , BIBREF20 . WikiLarge includes 8 (reference) simplifications for 2,359 sentences split into 2,000 for development and 359 for testing.
Metrics. Three metrics in text simplification are chosen in this paper. BLEU BIBREF5 is one traditional machine translation metric to assess the degree to which translated simplifications differed from reference simplifications. FKGL measures the readability of the output BIBREF23 . A small FKGL represents simpler output. SARI is a recent text-simplification metric by comparing the output against the source and reference simplifications BIBREF20 .
We evaluate the output of all systems using human evaluation. The metric is denoted as Simplicity BIBREF8 . The three non-native fluent English speakers are shown reference sentences and output sentences. They are asked whether the output sentence is much simpler (+2), somewhat simpler (+1), equally (0), somewhat more difficult (-1), and much more difficult (-2) than the reference sentence.
Methods. We use OpenNMT BIBREF24 as the implementation of the NMT system for all experiments BIBREF5 . We generally follow the default settings and training procedure described by Klein et al.(2017). We replace out-of-vocabulary words with a special UNK symbol. At prediction time, we replace UNK words with the highest probability score from the attention layer. OpenNMT system used on parallel data is the baseline system. To obtain a synthetic parallel training set, we back-translate a random sample of 100K sentences from the collected simplified corpora. OpenNMT used on parallel data and synthetic data is our model. The benchmarks are run on a Intel(R) Core(TM) i7-5930K CPU@3.50GHz, 32GB Mem, trained on 1 GPU GeForce GTX 1080 (Pascal) with CUDA v. 8.0.
We choose three statistical text simplification systems. PBMT-R is a phrase-based method with a reranking post-processing step BIBREF18 . Hybrid performs sentence splitting and deletion operations based on discourse representation structures, and then simplifies sentences with PBMT-R BIBREF25 . SBMT-SARI BIBREF19 is syntax-based translation model using PPDB paraphrase database BIBREF26 and modifies tuning function (using SARI). We choose two neural text simplification systems. NMT is a basic attention-based encoder-decoder model which uses OpenNMT framework to train with two LSTM layers, hidden states of size 500 and 500 hidden units, SGD optimizer, and a dropout rate of 0.3 BIBREF8 . Dress is an encoder-decoder model coupled with a deep reinforcement learning framework, and the parameters are chosen according to the original paper BIBREF20 . For the experiments with synthetic parallel data, we back-translate a random sample of 60 000 sentences from the collected simplified sentences into ordinary sentences. Our model is trained on synthetic data and the available parallel data, denoted as NMT+synthetic.
Results. Table 1 shows the results of all models on WikiLarge dataset. We can see that our method (NMT+synthetic) can obtain higher BLEU, lower FKGL and high SARI compared with other models, except Dress on FKGL and SBMT-SARI on SARI. It verified that including synthetic data during training is very effective, and yields an improvement over our baseline NMF by 2.11 BLEU, 1.7 FKGL and 1.07 SARI. We also substantially outperform Dress, who previously reported SOTA result. The results of our human evaluation using Simplicity are also presented in Table 1. NMT on synthetic data is significantly better than PBMT-R, Dress, and SBMT-SARI on Simplicity. It indicates that our method with simplified data is effective at creating simpler output.
Results on WikiSmall dataset are shown in Table 2. We see substantial improvements (6.37 BLEU) than NMT from adding simplified training data with synthetic ordinary sentences. Compared with statistical machine translation models (PBMT-R, Hybrid, SBMT-SARI), our method (NMT+synthetic) still have better results, but slightly worse FKGL and SARI. Similar to the results in WikiLarge, the results of our human evaluation using Simplicity outperforms the other models. In conclusion, Our method produces better results comparing with the baselines, which demonstrates the effectiveness of adding simplified training data.
Conclusion
In this paper, we propose one simple method to use simplified corpora during training of NMT systems, with no changes to the network architecture. In the experiments on two datasets, we achieve substantial gains in all tasks, and new SOTA results, via back-translation of simplified sentences into the ordinary sentences, and treating this synthetic data as additional training data. Because we do not change the neural network architecture to integrate simplified corpora, our method can be easily applied to other Neural Text Simplification (NTS) systems. We expect that the effectiveness of our method not only varies with the quality of the NTS system used for back-translation, but also depends on the amount of available parallel and simplified corpora. In the paper, we have only utilized data from Wikipedia for simplified sentences. In the future, many other text sources are available and the impact of not only size, but also of domain should be investigated.
| [
"training set has 89,042 sentence pairs, and the test set has 100 pairs, training set contains 296,402, 2,000 for development and 359 for testing",
"WikiSmall 89 142 sentence pair and WikiLarge 298 761 sentence pairs. "
] | 2,266 | qasper | en | null | 0ce1ee7ab0f1557704a9d7f937e6f5182c665687a3e2b0d9 | Answer the question based on the above article as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable". Do not provide any explanation.
Question: what are the sizes of both datasets?
Answer: | [
"training set has 89,042 sentence pairs, and the test set has 100 pairs, training set contains 296,402, 2,000 for development and 359 for testing",
"WikiSmall 89 142 sentence pair and WikiLarge 298 761 sentence pairs. "
] | qasper | 128 | 99 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.