Datasets:
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, usi(...TRUNCATED) | ["GM$\\_$KL achieves better correlation than existing approaches for various metrics on SCWS dataset(...TRUNCATED) | 2,189 | qasper | en | null | 5f00d4f6e62f4b99484eb78491f803f8143cc1b13ad33816 | "Answer the question based on the above article as concisely as you can, using a single phrase or se(...TRUNCATED) | ["GM$\\_$KL achieves better correlation than existing approaches for various metrics on SCWS dataset(...TRUNCATED) | 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, usi(...TRUNCATED) | [
"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 se(...TRUNCATED) | [
"simply averaging the predictions from the constituent single models"
] | qasper | 128 | 9 |
End of preview. Expand in Data Studio
README.md exists but content is empty.
- Downloads last month
- 27