data/qasper.jsonl

{"input": "What is the improvement in performance for Estonian in the NER task?", "context": "Introduction\nWord embeddings are representations of words in numerical form, as vectors of typically several hundred dimensions. The vectors are used as an input to machine learning models; for complex language processing tasks these are typically deep neural networks. The embedding vectors are obtained from specialized learning tasks, based on neural networks, e.g., word2vec BIBREF0, GloVe BIBREF1, FastText BIBREF2, ELMo BIBREF3, and BERT BIBREF4. For training, the embeddings algorithms use large monolingual corpora that encode important information about word meaning as distances between vectors. In order to enable downstream machine learning on text understanding tasks, the embeddings shall preserve semantic relations between words, and this is true even across languages.\nProbably the best known word embeddings are produced by the word2vec method BIBREF5. The problem with word2vec embeddings is their failure to express polysemous words. During training of an embedding, all senses of a given word (e.g., paper as a material, as a newspaper, as a scientific work, and as an exam) contribute relevant information in proportion to their frequency in the training corpus. This causes the final vector to be placed somewhere in the weighted middle of all words' meanings. Consequently, rare meanings of words are poorly expressed with word2vec and the resulting vectors do not offer good semantic representations. For example, none of the 50 closest vectors of the word paper is related to science.\nThe idea of contextual embeddings is to generate a different vector for each context a word appears in and the context is typically defined sentence-wise. To a large extent, this solves the problems with word polysemy, i.e. the context of a sentence is typically enough to disambiguate different meanings of a word for humans and so it is for the learning algorithms. In this work, we describe high-quality models for contextual embeddings, called ELMo BIBREF3, precomputed for seven morphologically rich, less-resourced languages: Slovenian, Croatian, Finnish, Estonian, Latvian, Lithuanian, and Swedish. ELMo is one of the most successful approaches to contextual word embeddings. At time of its creation, ELMo has been shown to outperform previous word embeddings BIBREF3 like word2vec and GloVe on many NLP tasks, e.g., question answering, named entity extraction, sentiment analysis, textual entailment, semantic role labeling, and coreference resolution.\nThis report is split into further five sections. In section SECREF2, we describe the contextual embeddings ELMo. In Section SECREF3, we describe the datasets used and in Section SECREF4 we describe preprocessing and training of the embeddings. We describe the methodology for evaluation of created vectors and results in Section SECREF5. We present conclusion in Section SECREF6 where we also outline plans for further work.\nELMo\nTypical word embeddings models or representations, such as word2vec BIBREF0, GloVe BIBREF1, or FastText BIBREF2, are fast to train and have been pre-trained for a number of different languages. They do not capture the context, though, so each word is always given the same vector, regardless of its context or meaning. This is especially problematic for polysemous words. ELMo (Embeddings from Language Models) embedding BIBREF3 is one of the state-of-the-art pretrained transfer learning models, that remedies the problem and introduces a contextual component.\nELMo model`s architecture consists of three neural network layers. The output of the model after each layer gives one set of embeddings, altogether three sets. The first layer is a CNN layer, which operates on a character level. It is context independent, so each word always gets the same embedding, regardless of its context. It is followed by two biLM layers. A biLM layer consists of two concatenated LSTMs. In the first LSTM, we try to predict the following word, based on the given past words, where each word is represented by the embeddings from the CNN layer. In the second LSTM, we try to predict the preceding word, based on the given following words. It is equivalent to the first LSTM, just reading the text in reverse.\nIn NLP tasks, any set of these embeddings may be used; however, a weighted average is usually used. The weights of the average are learned during the training of the model for the specific task. Additionally, an entire ELMo model can be fine-tuned on a specific end task.\nAlthough ELMo is trained on character level and is able to handle out-of-vocabulary words, a vocabulary file containing most common tokens is used for efficiency during training and embedding generation. The original ELMo model was trained on a one billion word large English corpus, with a given vocabulary file of about 800,000 words. Later, ELMo models for other languages were trained as well, but limited to larger languages with many resources, like German and Japanese.\nELMo ::: ELMoForManyLangs\nRecently, ELMoForManyLangs BIBREF6 project released pre-trained ELMo models for a number of different languages BIBREF7. These models, however, were trained on a significantly smaller datasets. They used 20-million-words data randomly sampled from the raw text released by the CoNLL 2017 Shared Task - Automatically Annotated Raw Texts and Word Embeddings BIBREF8, which is a combination of Wikipedia dump and common crawl. The quality of these models is questionable. For example, we compared the Latvian model by ELMoForManyLangs with a model we trained on a complete (wikidump + common crawl) Latvian corpus, which has about 280 million tokens. The difference of each model on the word analogy task is shown in Figure FIGREF16 in Section SECREF5. As the results of the ELMoForManyLangs embeddings are significantly worse than using the full corpus, we can conclude that these embeddings are not of sufficient quality. For that reason, we computed ELMo embeddings for seven languages on much larger corpora. As this effort requires access to large amount of textual data and considerable computational resources, we made the precomputed models publicly available by depositing them to Clarin repository.\nTraining Data\nWe trained ELMo models for seven languages: Slovenian, Croatian, Finnish, Estonian, Latvian, Lithuanian and Swedish. To obtain high-quality embeddings, we used large monolingual corpora from various sources for each language. Some corpora are available online under permissive licences, others are available only for research purposes or have limited availability. The corpora used in training datasets are a mix of news articles and general web crawl, which we preprocessed and deduplicated. Below we shortly describe the used corpora in alphabetical order of the involved languages. Their names and sizes are summarized in Table TABREF3.\nCroatian dataset include hrWaC 2.1 corpus BIBREF9, Riznica BIBREF10, and articles of Croatian branch of Styria media house, made available to us through partnership in a joint project. hrWaC was built by crawling the .hr internet domain in 2011 and 2014. Riznica is composed of Croatian fiction and non-fiction prose, poetry, drama, textbooks, manuals, etc. The Styria dataset consists of 570,219 news articles published on the Croatian 24sata news portal and niche portals related to 24sata.\nEstonian dataset contains texts from two sources, CoNLL 2017 Shared Task - Automatically Annotated Raw Texts and Word Embeddings BIBREF8, and news articles made available to us by Ekspress Meedia due to partnership in the project. Ekspress Meedia dataset is composed of Estonian news articles between years 2009 and 2019. The CoNLL 2017 corpus is composed of Estonian Wikipedia and webcrawl.\nFinnish dataset contains articles by Finnish news agency STT, Finnish part of the CoNLL 2017 dataset, and Ylilauta downloadable version BIBREF11. STT news articles were published between years 1992 and 2018. Ylilauta is a Finnish online discussion board; the corpus contains parts of the discussions from 2012 to 2014.\nLatvian dataset consists only of the Latvian portion of the ConLL 2017 corpus.\nLithuanian dataset is composed of Lithuanian Wikipedia articles from 2018, DGT-UD corpus, and LtTenTen. DGT-UD is a parallel corpus of 23 official languages of the EU, composed of JRC DGT translation memory of European law, automatically annotated with UD-Pipe 1.2. LtTenTen is Lithuanian web corpus made up of texts collected from the internet in April 2014 BIBREF12.\nSlovene dataset is formed from the Gigafida 2.0 corpus BIBREF13. It is a general language corpus composed of various sources, mostly newspapers, internet pages, and magazines, but also fiction and non-fiction prose, textbooks, etc.\nSwedish dataset is composed of STT Swedish articles and Swedish part of CoNLL 2017. The Finnish news agency STT publishes some of its articles in Swedish language. They were made available to us through partnership in a joint project. The corpus contains those articles from 1992 to 2017.\nPreprocessing and Training\nPrior to training the ELMo models, we sentence and word tokenized all the datasets. The text was formatted in such a way that each sentence was in its own line with tokens separated by white spaces. CoNLL 2017, DGT-UD and LtTenTen14 corpora were already pre-tokenized. We tokenized the others using the NLTK library and its tokenizers for each of the languages. There is no tokenizer for Croatian in NLTK library, so we used Slovene tokenizer instead. After tokenization, we deduplicated the datasets for each language separately, using the Onion (ONe Instance ONly) tool for text deduplication. We applied the tool on paragraph level for corpora that did not have sentences shuffled and on sentence level for the rest. We considered 9-grams with duplicate content threshold of 0.9.\nFor each language we prepared a vocabulary file, containing roughly one million most common tokens, i.e. tokens that appear at least $n$ times in the corpus, where $n$ is between 15 and 25, depending on the dataset size. We included the punctuation marks among the tokens. We trained each ELMo model using default values used to train the original English ELMo (large) model.\nEvaluation\nWe evaluated the produced ELMo models for all languages using two evaluation tasks: a word analogy task and named entity recognition (NER) task. Below, we first shortly describe each task, followed by the evaluation results.\nEvaluation ::: Word Analogy Task\nThe word analogy task was popularized by mikolov2013distributed. The goal is to find a term $y$ for a given term $x$ so that the relationship between $x$ and $y$ best resembles the given relationship $a : b$. There are two main groups of categories: 5 semantic and 10 syntactic. To illustrate a semantic relationship, consider for example that the word pair $a : b$ is given as “Finland : Helsinki”. The task is to find the term $y$ corresponding to the relationship “Sweden : $y$”, with the expected answer being $y=$ Stockholm. In syntactic categories, the two words in a pair have a common stem (in some cases even same lemma), with all the pairs in a given category having the same morphological relationship. For example, given the word pair “long : longer”, we see that we have an adjective in its base form and the same adjective in a comparative form. That task is then to find the term $y$ corresponding to the relationship “dark : $y$”, with the expected answer being $y=$ darker, that is a comparative form of the adjective dark.\nIn the vector space, the analogy task is transformed into vector arithmetic and search for nearest neighbours, i.e. we compute the distance between vectors: d(vec(Finland), vec(Helsinki)) and search for word $y$ which would give the closest result in distance d(vec(Sweden), vec($y$)). In the analogy dataset the analogies are already pre-specified, so we are measuring how close are the given pairs. In the evaluation below, we use analogy datasets for all tested languages based on the English dataset by BIBREF14 . Due to English-centered bias of this dataset, we used a modified dataset which was first written in Slovene language and then translated into other languages BIBREF15.\nAs each instance of analogy contains only four words, without any context, the contextual models (such as ELMo) do not have enough context to generate sensible embeddings. We therefore used some additional text to form simple sentences using the four analogy words, while taking care that their noun case stays the same. For example, for the words \"Rome\", \"Italy\", \"Paris\" and \"France\" (forming the analogy Rome is to Italy as Paris is to $x$, where the correct answer is $x=$France), we formed the sentence \"If the word Rome corresponds to the word Italy, then the word Paris corresponds to the word France\". We generated embeddings for those four words in the constructed sentence, substituted the last word with each word in our vocabulary and generated the embeddings again. As typical for non-contextual analogy task, we measure the cosine distance ($d$) between the last word ($w_4$) and the combination of the first three words ($w_2-w_1+w_3$). We use the CSLS metric BIBREF16 to find the closest candidate word ($w_4$). If we find the correct word among the five closest words, we consider that entry as successfully identified. The proportion of correctly identified words forms a statistic called accuracy@5, which we report as the result.\nWe first compare existing Latvian ELMo embeddings from ELMoForManyLangs project with our Latvian embeddings, followed by the detailed analysis of our ELMo embeddings. We trained Latvian ELMo using only CoNLL 2017 corpora. Since this is the only language, where we trained the embedding model on exactly the same corpora as ELMoForManyLangs models, we chose it for comparison between our ELMo model with ELMoForManyLangs. In other languages, additional or other corpora were used, so a direct comparison would also reflect the quality of the corpora used for training. In Latvian, however, only the size of the training dataset is different. ELMoForManyLangs uses only 20 million tokens and we use the whole corpus of 270 million tokens.\nThe Latvian ELMo model from ELMoForManyLangs project performs significantly worse than EMBEDDIA ELMo Latvian model on all categories of word analogy task (Figure FIGREF16). We also include the comparison with our Estonian ELMo embeddings in the same figure. This comparison shows that while differences between our Latvian and Estonian embeddings can be significant for certain categories, the accuracy score of ELMoForManyLangs is always worse than either of our models. The comparison of Estonian and Latvian models leads us to believe that a few hundred million tokens is a sufficiently large corpus to train ELMo models (at least for word analogy task), but 20-million token corpora used in ELMoForManyLangs are too small.\nThe results for all languages and all ELMo layers, averaged over semantic and syntactic categories, are shown in Table TABREF17. The embeddings after the first LSTM layer perform best in semantic categories. In syntactic categories, the non-contextual CNN layer performs the best. Syntactic categories are less context dependent and much more morphology and syntax based, so it is not surprising that the non-contextual layer performs well. The second LSTM layer embeddings perform the worst in syntactic categories, though still outperforming CNN layer embeddings in semantic categories. Latvian ELMo performs worse compared to other languages we trained, especially in semantic categories, presumably due to smaller training data size. Surprisingly, the original English ELMo performs very poorly in syntactic categories and only outperforms Latvian in semantic categories. The low score can be partially explained by English model scoring $0.00$ in one syntactic category “opposite adjective”, which we have not been able to explain.\nEvaluation ::: Named Entity Recognition\nFor evaluation of ELMo models on a relevant downstream task, we used named entity recognition (NER) task. NER is an information extraction task that seeks to locate and classify named entity mentions in unstructured text into pre-defined categories such as the person names, organizations, locations, medical codes, time expressions, quantities, monetary values, percentages, etc. To allow comparison of results between languages, we used an adapted version of this task, which uses a reduced set of labels, available in NER datasets for all processed languages. The labels in the used NER datasets are simplified to a common label set of three labels (person - PER, location - LOC, organization - ORG). Each word in the NER dataset is labeled with one of the three mentioned labels or a label 'O' (other, i.e. not a named entity) if it does not fit any of the other three labels. The number of words having each label is shown in Table TABREF19.\nTo measure the performance of ELMo embeddings on the NER task we proceeded as follows. We embedded the text in the datasets sentence by sentence, producing three vectors (one from each ELMo layer) for each token in a sentence. We calculated the average of the three vectors and used it as the input of our recognition model. The input layer was followed by a single LSTM layer with 128 LSTM cells and a dropout layer, randomly dropping 10% of the neurons on both the output and the recurrent branch. The final layer of our model was a time distributed softmax layer with 4 neurons.\nWe used ADAM optimiser BIBREF17 with the learning rate 0.01 and $10^{-5}$ learning rate decay. We used categorical cross-entropy as a loss function and trained the model for 3 epochs. We present the results using the Macro $F_1$ score, that is the average of $F_1$-scores for each of the three NE classes (the class Other is excluded).\nSince the differences between the tested languages depend more on the properties of the NER datasets than on the quality of embeddings, we can not directly compare ELMo models. For this reason, we take the non-contextual fastText embeddings as a baseline and predict named entities using them. The architecture of the model using fastText embeddings is the same as the one using ELMo embeddings, except that the input uses 300 dimensional fastText embedding vectors, and the model was trained for 5 epochs (instead of 3 as for ELMo). In both cases (ELMo and fastText) we trained and evaluated the model five times, because there is some random component involved in initialization of the neural network model. By training and evaluating multiple times, we minimise this random component.\nThe results are presented in Table TABREF21. We included the evaluation of the original ELMo English model in the same table. NER models have little difficulty distinguishing between types of named entities, but recognizing whether a word is a named entity or not is more difficult. For languages with the smallest NER datasets, Croatian and Lithuanian, ELMo embeddings show the largest improvement over fastText embeddings. However, we can observe significant improvements with ELMo also on English and Finnish, which are among the largest datasets (English being by far the largest). Only on Slovenian dataset did ELMo perform slightly worse than fastText, on all other EMBEDDIA languages, the ELMo embeddings improve the results.\nConclusion\nWe prepared precomputed ELMo contextual embeddings for seven languages: Croatian, Estonian, Finnish, Latvian, Lithuanian, Slovenian, and Swedish. We present the necessary background on embeddings and contextual embeddings, the details of training the embedding models, and their evaluation. We show that the size of used training sets importantly affects the quality of produced embeddings, and therefore the existing publicly available ELMo embeddings for the processed languages are inadequate. We trained new ELMo embeddings on larger training sets and analysed their properties on the analogy task and on the NER task. The results show that the newly produced contextual embeddings produce substantially better results compared to the non-contextual fastText baseline. In future work, we plan to use the produced contextual embeddings on the problems of news media industry. The pretrained ELMo models will be deposited to the CLARIN repository by the time of the final version of this paper.\nAcknowledgments\nThe work was partially supported by the Slovenian Research Agency (ARRS) core research programme P6-0411. This paper is supported by European Union's Horizon 2020 research and innovation programme under grant agreement No 825153, project EMBEDDIA (Cross-Lingual Embeddings for Less-Represented Languages in European News Media). The results of this publication reflects only the authors' view and the EU Commission is not responsible for any use that may be made of the information it contains.", "answers": ["5 percent points.", "0.05 F1"], "length": 3290, "dataset": "qasper", "language": "en", "all_classes": null, "_id": "bfc5d4d72997fdcc0107cd5cab9b9718777c43b86be45eb2"}
{"input": "what previous RNN models do they compare with?", "context": "Introduction\nLong short term memory (LSTM) units BIBREF1 are popular for many sequence modeling tasks and are used extensively in language modeling. A key to their success is their articulated gating structure, which allows for more control over the information passed along the recurrence. However, despite the sophistication of the gating mechanisms employed in LSTMs and similar recurrent units, the input and context vectors are treated with simple linear transformations prior to gating. Non-linear transformations such as convolutions BIBREF2 have been used, but these have not achieved the performance of well regularized LSTMs for language modeling BIBREF3 .\nA natural way to improve the expressiveness of linear transformations is to increase the number of dimensions of the input and context vectors, but this comes with a significant increase in the number of parameters which may limit generalizability. An example is shown in Figure FIGREF1 , where LSTMs performance decreases with the increase in dimensions of the input and context vectors. Moreover, the semantics of the input and context vectors are different, suggesting that each may benefit from specialized treatment.\nGuided by these insights, we introduce a new recurrent unit, the Pyramidal Recurrent Unit (PRU), which is based on the LSTM gating structure. Figure FIGREF2 provides an overview of the PRU. At the heart of the PRU is the pyramidal transformation (PT), which uses subsampling to effect multiple views of the input vector. The subsampled representations are combined in a pyramidal fusion structure, resulting in richer interactions between the individual dimensions of the input vector than is possible with a linear transformation. Context vectors, which have already undergone this transformation in the previous cell, are modified with a grouped linear transformation (GLT) which allows the network to learn latent representations in high dimensional space with fewer parameters and better generalizability (see Figure FIGREF1 ).\nWe show that PRUs can better model contextual information and demonstrate performance gains on the task of language modeling. The PRU improves the perplexity of the current state-of-the-art language model BIBREF0 by up to 1.3 points, reaching perplexities of 56.56 and 64.53 on the Penn Treebank and WikiText2 datasets while learning 15-20% fewer parameters. Replacing an LSTM with a PRU results in improvements in perplexity across a variety of experimental settings. We provide detailed ablations which motivate the design of the PRU architecture, as well as detailed analysis of the effect of the PRU on other components of the language model.\nRelated work\nMultiple methods, including a variety of gating structures and transformations, have been proposed to improve the performance of recurrent neural networks (RNNs). We first describe these approaches and then provide an overview of recent work in language modeling.\nPyramidal Recurrent Units\nWe introduce Pyramidal Recurrent Units (PRUs), a new RNN architecture which improves modeling of context by allowing for higher dimensional vector representations while learning fewer parameters. Figure FIGREF2 provides an overview of PRU. We first elaborate on the details of the pyramidal transformation and the grouped linear transformation. We then describe our recurrent unit, PRU.\nPyramidal transformation for input\nThe basic transformation in many recurrent units is a linear transformation INLINEFORM0 defined as: DISPLAYFORM0\nwhere INLINEFORM0 are learned weights that linearly map INLINEFORM1 to INLINEFORM2 . To simplify notation, we omit the biases.\nMotivated by successful applications of sub-sampling in computer vision (e.g., BIBREF22 , BIBREF23 , BIBREF9 , BIBREF24 ), we subsample input vector INLINEFORM0 into INLINEFORM1 pyramidal levels to achieve representation of the input vector at multiple scales. This sub-sampling operation produces INLINEFORM2 vectors, represented as INLINEFORM3 , where INLINEFORM4 is the sampling rate and INLINEFORM5 . We learn scale-specific transformations INLINEFORM6 for each INLINEFORM7 . The transformed subsamples are concatenated to produce the pyramidal analog to INLINEFORM8 , here denoted as INLINEFORM9 : DISPLAYFORM0\nwhere INLINEFORM0 indicates concatenation. We note that pyramidal transformation with INLINEFORM1 is the same as the linear transformation.\nTo improve gradient flow inside the recurrent unit, we combine the input and output using an element-wise sum (when dimension matches) to produce residual analog of pyramidal transformation, as shown in Figure FIGREF2 BIBREF25 .\nWe sub-sample the input vector INLINEFORM0 into INLINEFORM1 pyramidal levels using the kernel-based approach BIBREF8 , BIBREF9 . Let us assume that we have a kernel INLINEFORM2 with INLINEFORM3 elements. Then, the input vector INLINEFORM4 can be sub-sampled as: DISPLAYFORM0\nwhere INLINEFORM0 represents the stride and INLINEFORM1 .\nThe number of parameters learned by the linear transformation and the pyramidal transformation with INLINEFORM0 pyramidal levels to map INLINEFORM1 to INLINEFORM2 are INLINEFORM3 and INLINEFORM4 respectively. Thus, pyramidal transformation reduces the parameters of a linear transformation by a factor of INLINEFORM5 . For example, the pyramidal transformation (with INLINEFORM6 and INLINEFORM7 ) learns INLINEFORM8 fewer parameters than the linear transformation.\nGrouped linear transformation for context\nMany RNN architectures apply linear transformations to both the input and context vector. However, this may not be ideal due to the differing semantics of each vector. In many NLP applications including language modeling, the input vector is a dense word embedding which is shared across all contexts for a given word in a dataset. In contrast, the context vector is highly contextualized by the current sequence. The differences between the input and context vector motivate their separate treatment in the PRU architecture.\nThe weights learned using the linear transformation (Eq. EQREF9 ) are reused over multiple time steps, which makes them prone to over-fitting BIBREF26 . To combat over-fitting, various methods, such as variational dropout BIBREF26 and weight dropout BIBREF0 , have been proposed to regularize these recurrent connections. To further improve generalization abilities while simultaneously enabling the recurrent unit to learn representations at very high dimensional space, we propose to use grouped linear transformation (GLT) instead of standard linear transformation for recurrent connections BIBREF27 . While pyramidal and linear transformations can be applied to transform context vectors, our experimental results in Section SECREF39 suggests that GLTs are more effective.\nThe linear transformation INLINEFORM0 maps INLINEFORM1 linearly to INLINEFORM2 . Grouped linear transformations break the linear interactions by factoring the linear transformation into two steps. First, a GLT splits the input vector INLINEFORM3 into INLINEFORM4 smaller groups such that INLINEFORM5 . Second, a linear transformation INLINEFORM6 is applied to map INLINEFORM7 linearly to INLINEFORM8 , for each INLINEFORM9 . The INLINEFORM10 resultant output vectors INLINEFORM11 are concatenated to produce the final output vector INLINEFORM12 . DISPLAYFORM0\nGLTs learn representations at low dimensionality. Therefore, a GLT requires INLINEFORM0 fewer parameters than the linear transformation. We note that GLTs are subset of linear transformations. In a linear transformation, each neuron receives an input from each element in the input vector while in a GLT, each neuron receives an input from a subset of the input vector. Therefore, GLT is the same as a linear transformation when INLINEFORM1 .\nPyramidal Recurrent Unit\nWe extend the basic gating architecture of LSTM with the pyramidal and grouped linear transformations outlined above to produce the Pyramidal Recurrent Unit (PRU), whose improved sequence modeling capacity is evidenced in Section SECREF4 .\nAt time INLINEFORM0 , the PRU combines the input vector INLINEFORM1 and the previous context vector (or previous hidden state vector) INLINEFORM2 using the following transformation function as: DISPLAYFORM0\nwhere INLINEFORM0 indexes the various gates in the LSTM model, and INLINEFORM1 and INLINEFORM2 represents the pyramidal and grouped linear transformations defined in Eqns. EQREF10 and EQREF15 , respectively.\nWe will now incorporate INLINEFORM0 into LSTM gating architecture to produce PRU. At time INLINEFORM1 , a PRU cell takes INLINEFORM2 , INLINEFORM3 , and INLINEFORM4 as inputs to produce forget INLINEFORM5 , input INLINEFORM6 , output INLINEFORM7 , and content INLINEFORM8 gate signals. The inputs are combined with these gate signals to produce context vector INLINEFORM9 and cell state INLINEFORM10 . Mathematically, the PRU with the LSTM gating architecture can be defined as: DISPLAYFORM0\nwhere INLINEFORM0 represents the element-wise multiplication operation, and INLINEFORM1 and INLINEFORM2 are the sigmoid and hyperbolic tangent activation functions. We note that LSTM is a special case of PRU when INLINEFORM3 = INLINEFORM4 =1.\nExperiments\nTo showcase the effectiveness of the PRU, we evaluate the performance on two standard datasets for word-level language modeling and compare with state-of-the-art methods. Additionally, we provide a detailed examination of the PRU and its behavior on the language modeling tasks.\nSet-up\nFollowing recent works, we compare on two widely used datasets, the Penn Treebank (PTB) BIBREF28 as prepared by BIBREF29 and WikiText2 (WT-2) BIBREF20 . For both datasets, we follow the same training, validation, and test splits as in BIBREF0 .\nWe extend the language model, AWD-LSTM BIBREF0 , by replacing LSTM layers with PRU. Our model uses 3-layers of PRU with an embedding size of 400. The number of parameters learned by state-of-the-art methods vary from 18M to 66M with majority of the methods learning about 22M to 24M parameters on the PTB dataset. For a fair comparison with state-of-the-art methods, we fix the model size to 19M and vary the value of INLINEFORM0 and hidden layer sizes so that total number of learned parameters is similar across different configurations. We use 1000, 1200, and 1400 as hidden layer sizes for values of INLINEFORM1 =1,2, and 4, respectively. We use the same settings for the WT-2 dataset. We set the number of pyramidal levels INLINEFORM2 to two in our experiments and use average pooling for sub-sampling. These values are selected based on our ablation experiments on the validation set (Section SECREF39 ). We measure the performance of our models in terms of word-level perplexity. We follow the same training strategy as in BIBREF0 .\nTo understand the effect of regularization methods on the performance of PRUs, we perform experiments under two different settings: (1) Standard dropout: We use a standard dropout BIBREF12 with probability of 0.5 after embedding layer, the output between LSTM layers, and the output of final LSTM layer. (2) Advanced dropout: We use the same dropout techniques with the same dropout values as in BIBREF0 . We call this model as AWD-PRU.\nResults\nTable TABREF23 compares the performance of the PRU with state-of-the-art methods. We can see that the PRU achieves the best performance with fewer parameters.\nPRUs achieve either the same or better performance than LSTMs. In particular, the performance of PRUs improves with the increasing value of INLINEFORM0 . At INLINEFORM1 , PRUs outperform LSTMs by about 4 points on the PTB dataset and by about 3 points on the WT-2 dataset. This is explained in part by the regularization effect of the grouped linear transformation (Figure FIGREF1 ). With grouped linear and pyramidal transformations, PRUs learn rich representations at very high dimensional space while learning fewer parameters. On the other hand, LSTMs overfit to the training data at such high dimensions and learn INLINEFORM2 to INLINEFORM3 more parameters than PRUs.\nWith the advanced dropouts, the performance of PRUs improves by about 4 points on the PTB dataset and 7 points on the WT-2 dataset. This further improves with finetuning on the PTB (about 2 points) and WT-2 (about 1 point) datasets.\nFor similar number of parameters, the PRU with standard dropout outperforms most of the state-of-the-art methods by large margin on the PTB dataset (e.g. RAN BIBREF7 by 16 points with 4M less parameters, QRNN BIBREF33 by 16 points with 1M more parameters, and NAS BIBREF31 by 1.58 points with 6M less parameters). With advanced dropouts, the PRU delivers the best performance. On both datasets, the PRU improves the perplexity by about 1 point while learning 15-20% fewer parameters.\nPRU is a drop-in replacement for LSTM, therefore, it can improve language models with modern inference techniques such as dynamic evaluation BIBREF21 . When we evaluate PRU-based language models (only with standard dropout) with dynamic evaluation on the PTB test set, the perplexity of PRU ( INLINEFORM0 ) improves from 62.42 to 55.23 while the perplexity of an LSTM ( INLINEFORM1 ) with similar settings improves from 66.29 to 58.79; suggesting that modern inference techniques are equally applicable to PRU-based language models.\nAnalysis\nIt is shown above that the PRU can learn representations at higher dimensionality with more generalization power, resulting in performance gains for language modeling. A closer analysis of the impact of the PRU in a language modeling system reveals several factors that help explain how the PRU achieves these gains.\nAs exemplified in Table TABREF34 , the PRU tends toward more confident decisions, placing more of the probability mass on the top next-word prediction than the LSTM. To quantify this effect, we calculate the entropy of the next-token distribution for both the PRU and the LSTM using 3687 contexts from the PTB validation set. Figure FIGREF32 shows a histogram of the entropies of the distribution, where bins of size 0.23 are used to effect categories. We see that the PRU more often produces lower entropy distributions corresponding to higher confidences for next-token choices. This is evidenced by the mass of the red PRU curve lying in the lower entropy ranges compared to the blue LSTM's curve. The PRU can produce confident decisions in part because more information is encoded in the higher dimensional context vectors.\nThe PRU has the ability to model individual words at different resolutions through the pyramidal transform; which provides multiple paths for the gradient to the embedding layer (similar to multi-task learning) and improves the flow of information. When considering the embeddings by part of speech, we find that the pyramid level 1 embeddings exhibit higher variance than the LSTM across all POS categories (Figure FIGREF33 ), and that pyramid level 2 embeddings show extremely low variance. We hypothesize that the LSTM must encode both coarse group similarities and individual word differences into the same vector space, reducing the space between individual words of the same category. The PRU can rely on the subsampled embeddings to account for coarse-grained group similarities, allowing for finer individual word distinctions in the embedding layer. This hypothesis is strengthened by the entropy results described above: a model which can make finer distinctions between individual words can more confidently assign probability mass. A model that cannot make these distinctions, such as the LSTM, must spread its probability mass across a larger class of similar words.\nSaliency analysis using gradients help identify relevant words in a test sequence that contribute to the prediction BIBREF34 , BIBREF35 , BIBREF36 . These approaches compute the relevance as the squared norm of the gradients obtained through back-propagation. Table TABREF34 visualizes the heatmaps for different sequences. PRUs, in general, give more relevance to contextual words than LSTMs, such as southeast (sample 1), cost (sample 2), face (sample 4), and introduced (sample 5), which help in making more confident decisions. Furthermore, when gradients during back-propagation are visualized BIBREF37 (Table TABREF34 ), we find that PRUs have better gradient coverage than LSTMs, suggesting PRUs use more features than LSTMs that contributes to the decision. This also suggests that PRUs update more parameters at each iteration which results in faster training. Language model in BIBREF0 takes 500 and 750 epochs to converge with PRU and LSTM as a recurrent unit, respectively.\nAblation studies\nIn this section, we provide a systematic analysis of our design choices. Our training methodology is the same as described in Section SECREF19 with the standard dropouts. For a thorough understanding of our design choices, we use a language model with a single layer of PRU and fix the size of embedding and hidden layers to 600. The word-level perplexities are reported on the validation sets of the PTB and the WT-2 datasets.\nThe two hyper-parameters that control the trade-off between performance and number of parameters in PRUs are the number of pyramidal levels INLINEFORM0 and groups INLINEFORM1 . Figure FIGREF35 provides a trade-off between perplexity and recurrent unit (RU) parameters.\nVariable INLINEFORM0 and fixed INLINEFORM1 : When we increase the number of pyramidal levels INLINEFORM2 at a fixed value of INLINEFORM3 , the performance of the PRU drops by about 1 to 4 points while reducing the total number of recurrent unit parameters by up to 15%. We note that the PRU with INLINEFORM4 at INLINEFORM5 delivers similar performance as the LSTM while learning about 15% fewer recurrent unit parameters.\nFixed INLINEFORM0 and variable INLINEFORM1 : When we vary the value of INLINEFORM2 at fixed number of pyramidal levels INLINEFORM3 , the total number of recurrent unit parameters decreases significantly with a minimal impact on the perplexity. For example, PRUs with INLINEFORM4 and INLINEFORM5 learns 77% fewer recurrent unit parameters while its perplexity (lower is better) increases by about 12% in comparison to LSTMs. Moreover, the decrease in number of parameters at higher value of INLINEFORM6 enables PRUs to learn the representations in high dimensional space with better generalizability (Table TABREF23 ).\nTable TABREF43 shows the impact of different transformations of the input vector INLINEFORM0 and the context vector INLINEFORM1 . We make following observations: (1) Using the pyramidal transformation for the input vectors improves the perplexity by about 1 point on both the PTB and WT-2 datasets while reducing the number of recurrent unit parameters by about 14% (see R1 and R4). We note that the performance of the PRU drops by up to 1 point when residual connections are not used (R4 and R6). (2) Using the grouped linear transformation for context vectors reduces the total number of recurrent unit parameters by about 75% while the performance drops by about 11% (see R3 and R4). When we use the pyramidal transformation instead of the linear transformation, the performance drops by up to 2% while there is no significant drop in the number of parameters (R4 and R5).\nWe set sub-sampling kernel INLINEFORM0 (Eq. EQREF12 ) with stride INLINEFORM1 and size of 3 ( INLINEFORM2 ) in four different ways: (1) Skip: We skip every other element in the input vector. (2) Convolution: We initialize the elements of INLINEFORM3 randomly from normal distribution and learn them during training the model. We limit the output values between -1 and 1 using INLINEFORM4 activation function to make training stable. (3) Avg. pool: We initialize the elements of INLINEFORM5 to INLINEFORM6 . (4) Max pool: We select the maximum value in the kernel window INLINEFORM7 .\nTable TABREF45 compares the performance of the PRU with different sampling methods. Average pooling performs the best while skipping give comparable performance. Both of these methods enable the network to learn richer word representations while representing the input vector in different forms, thus delivering higher performance. Surprisingly, a convolution-based sub-sampling method does not perform as well as the averaging method. The INLINEFORM0 function used after convolution limits the range of output values which are further limited by the LSTM gating structure, thereby impeding in the flow of information inside the cell. Max pooling forces the network to learn representations from high magnitude elements, thus distinguishing features between elements vanishes, resulting in poor performance.\nConclusion\nWe introduce the Pyramidal Recurrent Unit, which better model contextual information by admitting higher dimensional representations with good generalizability. When applied to the task of language modeling, PRUs improve perplexity across several settings, including recent state-of-the-art systems. Our analysis shows that the PRU improves the flow of gradient and expand the word embedding subspace, resulting in more confident decisions. Here we have shown improvements for language modeling. In future, we plan to study the performance of PRUs on different tasks, including machine translation and question answering. In addition, we will study the performance of the PRU on language modeling with more recent inference techniques, such as dynamic evaluation and mixture of softmax.\nAcknowledgments\nThis research was supported by NSF (IIS 1616112, III 1703166), Allen Distinguished Investigator Award, and gifts from Allen Institute for AI, Google, Amazon, and Bloomberg. We are grateful to Aaron Jaech, Hannah Rashkin, Mandar Joshi, Aniruddha Kembhavi, and anonymous reviewers for their helpful comments.", "answers": ["Variational LSTM, CharCNN, Pointer Sentinel-LSTM, RHN, NAS Cell, SRU, QRNN, RAN, 4-layer skip-connection LSTM, AWD-LSTM, Quantized LSTM"], "length": 3319, "dataset": "qasper", "language": "en", "all_classes": null, "_id": "eeab2b9167f294f68d3058752acc01e1145eb7d89437d5c4"}
{"input": "How many different types of entities exist in the dataset?", "context": "Introduction\nNamed Entity Recognition (NER) is a foremost NLP task to label each atomic elements of a sentence into specific categories like \"PERSON\", \"LOCATION\", \"ORGANIZATION\" and othersBIBREF0. There has been an extensive NER research on English, German, Dutch and Spanish language BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5, and notable research on low resource South Asian languages like HindiBIBREF6, IndonesianBIBREF7 and other Indian languages (Kannada, Malayalam, Tamil and Telugu)BIBREF8. However, there has been no study on developing neural NER for Nepali language. In this paper, we propose a neural based Nepali NER using latest state-of-the-art architecture based on grapheme-level which doesn't require any hand-crafted features and no data pre-processing.\nRecent neural architecture like BIBREF1 is used to relax the need to hand-craft the features and need to use part-of-speech tag to determine the category of the entity. However, this architecture have been studied for languages like English, and German and not been applied to languages like Nepali which is a low resource language i.e limited data set to train the model. Traditional methods like Hidden Markov Model (HMM) with rule based approachesBIBREF9,BIBREF10, and Support Vector Machine (SVM) with manual feature-engineeringBIBREF11 have been applied but they perform poor compared to neural. However, there has been no research in Nepali NER using neural network. Therefore, we created the named entity annotated dataset partly with the help of Dataturk to train a neural model. The texts used for this dataset are collected from various daily news sources from Nepal around the year 2015-2016.\nFollowing are our contributions:\nWe present a novel Named Entity Recognizer (NER) for Nepali language. To best of our knowledge we are the first to propose neural based Nepali NER.\nAs there are not good quality dataset to train NER we release a dataset to support future research\nWe perform empirical evaluation of our model with state-of-the-art models with relative improvement of upto 10%\nIn this paper, we present works similar to ours in Section SECREF2. We describe our approach and dataset statistics in Section SECREF3 and SECREF4, followed by our experiments, evaluation and discussion in Section SECREF5, SECREF6, and SECREF7. We conclude with our observations in Section SECREF8.\nTo facilitate further research our code and dataset will be made available at github.com/link-yet-to-be-updated\nRelated Work\nThere has been a handful of research on Nepali NER task based on approaches like Support Vector Machine and gazetteer listBIBREF11 and Hidden Markov Model and gazetteer listBIBREF9,BIBREF10.\nBIBREF11 uses SVM along with features like first word, word length, digit features and gazetteer (person, organization, location, middle name, verb, designation and others). It uses one vs rest classification model to classify each word into different entity classes. However, it does not the take context word into account while training the model. Similarly, BIBREF9 and BIBREF10 uses Hidden Markov Model with n-gram technique for extracting POS-tags. POS-tags with common noun, proper noun or combination of both are combined together, then uses gazetteer list as look-up table to identify the named entities.\nResearchers have shown that the neural networks like CNNBIBREF12, RNNBIBREF13, LSTMBIBREF14, GRUBIBREF15 can capture the semantic knowledge of language better with the help of pre-trained embbeddings like word2vecBIBREF16, gloveBIBREF17 or fasttextBIBREF18.\nSimilar approaches has been applied to many South Asian languages like HindiBIBREF6, IndonesianBIBREF7, BengaliBIBREF19 and In this paper, we present the neural network architecture for NER task in Nepali language, which doesn't require any manual feature engineering nor any data pre-processing during training. First we are comparing BiLSTMBIBREF14, BiLSTM+CNNBIBREF20, BiLSTM+CRFBIBREF1, BiLSTM+CNN+CRFBIBREF2 models with CNN modelBIBREF0 and Stanford CRF modelBIBREF21. Secondly, we show the comparison between models trained on general word embeddings, word embedding + character-level embedding, word embedding + part-of-speech(POS) one-hot encoding and word embedding + grapheme clustered or sub-word embeddingBIBREF22. The experiments were performed on the dataset that we created and on the dataset received from ILPRL lab. Our extensive study shows that augmenting word embedding with character or grapheme-level representation and POS one-hot encoding vector yields better results compared to using general word embedding alone.\nApproach\nIn this section, we describe our approach in building our model. This model is partly inspired from multiple models BIBREF20,BIBREF1, andBIBREF2\nApproach ::: Bidirectional LSTM\nWe used Bi-directional LSTM to capture the word representation in forward as well as reverse direction of a sentence. Generally, LSTMs take inputs from left (past) of the sentence and computes the hidden state. However, it is proven beneficialBIBREF23 to use bi-directional LSTM, where, hidden states are computed based from right (future) of sentence and both of these hidden states are concatenated to produce the final output as $h_t$=[$\\overrightarrow{h_t}$;$\\overleftarrow{h_t}$], where $\\overrightarrow{h_t}$, $\\overleftarrow{h_t}$ = hidden state computed in forward and backward direction respectively.\nApproach ::: Features ::: Word embeddings\nWe have used Word2Vec BIBREF16, GloVe BIBREF17 and FastText BIBREF18 word vectors of 300 dimensions. These vectors were trained on the corpus obtained from Nepali National Corpus. This pre-lemmatized corpus consists of 14 million words from books, web-texts and news papers. This corpus was mixed with the texts from the dataset before training CBOW and skip-gram version of word2vec using gensim libraryBIBREF24. This trained model consists of vectors for 72782 unique words.\nLight pre-processing was performed on the corpus before training it. For example, invalid characters or characters other than Devanagari were removed but punctuation and numbers were not removed. We set the window context at 10 and the rare words whose count is below 5 are dropped. These word embeddings were not frozen during the training session because fine-tuning word embedding help achieve better performance compared to frozen oneBIBREF20.\nWe have used fasttext embeddings in particular because of its sub-word representation ability, which is very useful in highly inflectional language as shown in Table TABREF25. We have trained the word embedding in such a way that the sub-word size remains between 1 and 4. We particularly chose this size because in Nepali language a single letter can also be a word, for example e, t, C, r, l, n, u and a single character (grapheme) or sub-word can be formed after mixture of dependent vowel signs with consonant letters for example, C + O + = CO, here three different consonant letters form a single sub-word.\nThe two-dimensional visualization of an example word npAl is shown in FIGREF14. Principal Component Analysis (PCA) technique was used to generate this visualization which helps use to analyze the nearest neighbor words of a given sample word. 84 and 104 nearest neighbors were observed using word2vec and fasttext embedding respectively on the same corpus.\nApproach ::: Features ::: Character-level embeddings\nBIBREF20 and BIBREF2 successfully presented that the character-level embeddings, extracted using CNN, when combined with word embeddings enhances the NER model performance significantly, as it is able to capture morphological features of a word. Figure FIGREF7 shows the grapheme-level CNN used in our model, where inputs to CNN are graphemes. Character-level CNN is also built in similar fashion, except the inputs are characters. Grapheme or Character -level embeddings are randomly initialized from [0,1] with real values with uniform distribution of dimension 30.\nApproach ::: Features ::: Grapheme-level embeddings\nGrapheme is atomic meaningful unit in writing system of any languages. Since, Nepali language is highly morphologically inflectional, we compared grapheme-level representation with character-level representation to evaluate its effect. For example, in character-level embedding, each character of a word npAl results into n + + p + A + l has its own embedding. However, in grapheme level, a word npAl is clustered into graphemes, resulting into n + pA + l. Here, each grapheme has its own embedding. This grapheme-level embedding results good scores on par with character-level embedding in highly inflectional languages like Nepali, because graphemes also capture syntactic information similar to characters. We created grapheme clusters using uniseg package which is helpful in unicode text segmentations.\nApproach ::: Features ::: Part-of-speech (POS) one hot encoding\nWe created one-hot encoded vector of POS tags and then concatenated with pre-trained word embeddings before passing it to BiLSTM network. A sample of data is shown in figure FIGREF13.\nDataset Statistics ::: OurNepali dataset\nSince, we there was no publicly available standard Nepali NER dataset and did not receive any dataset from the previous researchers, we had to create our own dataset. This dataset contains the sentences collected from daily newspaper of the year 2015-2016. This dataset has three major classes Person (PER), Location (LOC) and Organization (ORG). Pre-processing was performed on the text before creation of the dataset, for example all punctuations and numbers besides ',', '-', '|' and '.' were removed. Currently, the dataset is in standard CoNLL-2003 IO formatBIBREF25.\nSince, this dataset is not lemmatized originally, we lemmatized only the post-positions like Ek, kO, l, mA, m, my, jF, sg, aEG which are just the few examples among 299 post positions in Nepali language. We obtained these post-positions from sanjaalcorps and added few more to match our dataset. We will be releasing this list in our github repository. We found out that lemmatizing the post-positions boosted the F1 score by almost 10%.\nIn order to label our dataset with POS-tags, we first created POS annotated dataset of 6946 sentences and 16225 unique words extracted from POS-tagged Nepali National Corpus and trained a BiLSTM model with 95.14% accuracy which was used to create POS-tags for our dataset.\nThe dataset released in our github repository contains each word in newline with space separated POS-tags and Entity-tags. The sentences are separated by empty newline. A sample sentence from the dataset is presented in table FIGREF13.\nDataset Statistics ::: ILPRL dataset\nAfter much time, we received the dataset from Bal Krishna Bal, ILPRL, KU. This dataset follows standard CoNLL-2003 IOB formatBIBREF25 with POS tags. This dataset is prepared by ILPRL Lab, KU and KEIV Technologies. Few corrections like correcting the NER tags had to be made on the dataset. The statistics of both the dataset is presented in table TABREF23.\nTable TABREF24 presents the total entities (PER, LOC, ORG and MISC) from both of the dataset used in our experiments. The dataset is divided into three parts with 64%, 16% and 20% of the total dataset into training set, development set and test set respectively.\nExperiments\nIn this section, we present the details about training our neural network. The neural network architecture are implemented using PyTorch framework BIBREF26. The training is performed on a single Nvidia Tesla P100 SXM2. We first run our experiment on BiLSTM, BiLSTM-CNN, BiLSTM-CRF BiLSTM-CNN-CRF using the hyper-parameters mentioned in Table TABREF30. The training and evaluation was done on sentence-level. The RNN variants are initialized randomly from $(-\\sqrt{k},\\sqrt{k})$ where $k=\\frac{1}{hidden\\_size}$.\nFirst we loaded our dataset and built vocabulary using torchtext library. This eased our process of data loading using its SequenceTaggingDataset class. We trained our model with shuffled training set using Adam optimizer with hyper-parameters mentioned in table TABREF30. All our models were trained on single layer of LSTM network. We found out Adam was giving better performance and faster convergence compared to Stochastic Gradient Descent (SGD). We chose those hyper-parameters after many ablation studies. The dropout of 0.5 is applied after LSTM layer.\nFor CNN, we used 30 different filters of sizes 3, 4 and 5. The embeddings of each character or grapheme involved in a given word, were passed through the pipeline of Convolution, Rectified Linear Unit and Max-Pooling. The resulting vectors were concatenated and applied dropout of 0.5 before passing into linear layer to obtain the embedding size of 30 for the given word. This resulting embedding is concatenated with word embeddings, which is again concatenated with one-hot POS vector.\nExperiments ::: Tagging Scheme\nCurrently, for our experiments we trained our model on IO (Inside, Outside) format for both the dataset, hence the dataset does not contain any B-type annotation unlike in BIO (Beginning, Inside, Outside) scheme.\nExperiments ::: Early Stopping\nWe used simple early stopping technique where if the validation loss does not decrease after 10 epochs, the training was stopped, else the training will run upto 100 epochs. In our experience, training usually stops around 30-50 epochs.\nExperiments ::: Hyper-parameters Tuning\nWe ran our experiment looking for the best hyper-parameters by changing learning rate from (0,1, 0.01, 0.001, 0.0001), weight decay from [$10^{-1}$, $10^{-2}$, $10^{-3}$, $10^{-4}$, $10^{-5}$, $10^{-6}$, $10^{-7}$], batch size from [1, 2, 4, 8, 16, 32, 64, 128], hidden size from [8, 16, 32, 64, 128, 256, 512 1024]. Table TABREF30 shows all other hyper-parameter used in our experiment for both of the dataset.\nExperiments ::: Effect of Dropout\nFigure FIGREF31 shows how we end up choosing 0.5 as dropout rate. When the dropout layer was not used, the F1 score are at the lowest. As, we slowly increase the dropout rate, the F1 score also gradually increases, however after dropout rate = 0.5, the F1 score starts falling down. Therefore, we have chosen 0.5 as dropout rate for all other experiments performed.\nEvaluation\nIn this section, we present the details regarding evaluation and comparison of our models with other baselines.\nTable TABREF25 shows the study of various embeddings and comparison among each other in OurNepali dataset. Here, raw dataset represents such dataset where post-positions are not lemmatized. We can observe that pre-trained embeddings significantly improves the score compared to randomly initialized embedding. We can deduce that Skip Gram models perform better compared CBOW models for word2vec and fasttext. Here, fastText_Pretrained represents the embedding readily available in fastText website, while other embeddings are trained on the Nepali National Corpus as mentioned in sub-section SECREF11. From this table TABREF25, we can clearly observe that model using fastText_Skip Gram embeddings outperforms all other models.\nTable TABREF35 shows the model architecture comparison between all the models experimented. The features used for Stanford CRF classifier are words, letter n-grams of upto length 6, previous word and next word. This model is trained till the current function value is less than $1\\mathrm {e}{-2}$. The hyper-parameters of neural network experiments are set as shown in table TABREF30. Since, word embedding of character-level and grapheme-level is random, their scores are near.\nAll models are evaluated using CoNLL-2003 evaluation scriptBIBREF25 to calculate entity-wise precision, recall and f1 score.\nDiscussion\nIn this paper we present that we can exploit the power of neural network to train the model to perform downstream NLP tasks like Named Entity Recognition even in Nepali language. We showed that the word vectors learned through fasttext skip gram model performs better than other word embedding because of its capability to represent sub-word and this is particularly important to capture morphological structure of words and sentences in highly inflectional like Nepali. This concept can come handy in other Devanagari languages as well because the written scripts have similar syntactical structure.\nWe also found out that stemming post-positions can help a lot in improving model performance because of inflectional characteristics of Nepali language. So when we separate out its inflections or morphemes, we can minimize the variations of same word which gives its root word a stronger word vector representations compared to its inflected versions.\nWe can clearly imply from tables TABREF23, TABREF24, and TABREF35 that we need more data to get better results because OurNepali dataset volume is almost ten times bigger compared to ILPRL dataset in terms of entities.\nConclusion and Future work\nIn this paper, we proposed a novel NER for Nepali language and achieved relative improvement of upto 10% and studies different factors effecting the performance of the NER for Nepali language.\nWe also present a neural architecture BiLSTM+CNN(grapheme-level) which turns out to be performing on par with BiLSTM+CNN(character-level) under the same configuration. We believe this will not only help Nepali language but also other languages falling under the umbrellas of Devanagari languages. Our model BiLSTM+CNN(grapheme-level) and BiLSTM+CNN(G)+POS outperforms all other model experimented in OurNepali and ILPRL dataset respectively.\nSince this is the first named entity recognition research in Nepal language using neural network, there are many rooms for improvement. We believe initializing the grapheme-level embedding with fasttext embeddings might help boosting the performance, rather than randomly initializing it. In future, we plan to apply other latest techniques like BERT, ELMo and FLAIR to study its effect on low-resource language like Nepali. We also plan to improve the model using cross-lingual or multi-lingual parameter sharing techniques by jointly training with other Devanagari languages like Hindi and Bengali.\nFinally, we would like to contribute our dataset to Nepali NLP community to move forward the research going on in language understanding domain. We believe there should be special committee to create and maintain such dataset for Nepali NLP and organize various competitions which would elevate the NLP research in Nepal.\nSome of the future works are listed below:\nProper initialization of grapheme level embedding from fasttext embeddings.\nApply robust POS-tagger for Nepali dataset\nLemmatize the OurNepali dataset with robust and efficient lemmatizer\nImprove Nepali language score with cross-lingual learning techniques\nCreate more dataset using Wikipedia/Wikidata framework\nAcknowledgments\nThe authors of this paper would like to express sincere thanks to Bal Krishna Bal, Kathmandu University Professor for providing us the POS-tagged Nepali NER data.", "answers": ["OurNepali contains 3 different types of entities, ILPRL contains 4 different types of entities", "three"], "length": 2851, "dataset": "qasper", "language": "en", "all_classes": null, "_id": "ac6538e9c173ba7b453d6bf62480d56eb3761b5f6f73c328"}
{"input": "What datasets did they use for evaluation?", "context": "Introduction\nBidirectional Encoder Representations from Transformers (BERT) is a novel Transformer BIBREF0 model, which recently achieved state-of-the-art performance in several language understanding tasks, such as question answering, natural language inference, semantic similarity, sentiment analysis, and others BIBREF1. While well-suited to dealing with relatively short sequences, Transformers suffer from a major issue that hinders their applicability in classification of long sequences, i.e. they are able to consume only a limited context of symbols as their input BIBREF2.\nThere are several natural language (NLP) processing tasks that involve such long sequences. Of particular interest are topic identification of spoken conversations BIBREF3, BIBREF4, BIBREF5 and call center customer satisfaction prediction BIBREF6, BIBREF7, BIBREF8, BIBREF9. Call center conversations, while usually quite short and to the point, often involve agents trying to solve very complex issues that the customers experience, resulting in some calls taking even an hour or more. For speech analytics purposes, these calls are typically transcribed using an automatic speech recognition (ASR) system, and processed in textual representations further down the NLP pipeline. These transcripts sometimes exceed the length of 5000 words. Furthermore, temporal information might play an important role in tasks like CSAT. For example, a customer may be angry at the beginning of the call, but after her issue is resolved, she would be very satisfied with the way it was handled. Therefore, simple bag of words models, or any model that does not include temporal dependencies between the inputs, may not be well-suited to handle this category of tasks. This motivates us to employ model such as BERT in this task.\nIn this paper, we propose a method that builds upon BERT's architecture. We split the input text sequence into shorter segments in order to obtain a representation for each of them using BERT. Then, we use either a recurrent LSTM BIBREF10 network, or another Transformer, to perform the actual classification. We call these techniques Recurrence over BERT (RoBERT) and Transformer over BERT (ToBERT). Given that these models introduce a hierarchy of representations (segment-wise and document-wise), we refer to them as Hierarchical Transformers. To the best of our knowledge, no attempt has been done before to use the Transformer architecture for classification of such long sequences.\nOur novel contributions are:\nTwo extensions - RoBERT and ToBERT - to the BERT model, which enable its application in classification of long texts by performing segmentation and using another layer on top of the segment representations.\nState-of-the-art results on the Fisher topic classification task.\nSignificant improvement on the CSAT prediction task over the MS-CNN model.\nRelated work\nSeveral dimensionality reduction algorithms such as RBM, autoencoders, subspace multinomial models (SMM) are used to obtain a low dimensional representation of documents from a simple BOW representation and then classify it using a simple linear classifiers BIBREF11, BIBREF12, BIBREF13, BIBREF4. In BIBREF14 hierarchical attention networks are used for document classification. They evaluate their model on several datasets with average number of words around 150. Character-level CNN are explored in BIBREF15 but it is prohibitive for very long documents. In BIBREF16, dataset collected from arXiv papers is used for classification. For classification, they sample random blocks of words and use them together for classification instead of using full document which may work well as arXiv papers are usually coherent and well written on a well defined topic. Their method may not work well on spoken conversations as random block of words usually do not represent topic of full conversation.\nSeveral researchers addressed the problem of predicting customer satisfaction BIBREF6, BIBREF7, BIBREF8, BIBREF9. In most of these works, logistic regression, SVM, CNN are applied on different kinds of representations.\nIn BIBREF17, authors use BERT for document classification but the average document length is less than BERT maximum length 512. TransformerXL BIBREF2 is an extension to the Transformer architecture that allows it to better deal with long inputs for the language modelling task. It relies on the auto-regressive property of the model, which is not the case in our tasks.\nMethod ::: BERT\nBecause our work builds heavily upon BERT, we provide a brief summary of its features. BERT is built upon the Transformer architecture BIBREF0, which uses self-attention, feed-forward layers, residual connections and layer normalization as the main building blocks. It has two pre-training objectives:\nMasked language modelling - some of the words in a sentence are being masked and the model has to predict them based on the context (note the difference from the typical autoregressive language model training objective);\nNext sentence prediction - given two input sequences, decide whether the second one is the next sentence or not.\nBERT has been shown to beat the state-of-the-art performance on 11 tasks with no modifications to the model architecture, besides adding a task-specific output layer BIBREF1. We follow same procedure suggested in BIBREF1 for our tasks. Fig. FIGREF8 shows the BERT model for classification. We obtain two kinds of representation from BERT: pooled output from last transformer block, denoted by H, and posterior probabilities, denoted by P. There are two variants of BERT - BERT-Base and BERT-Large. In this work we are using BERT-Base for faster training and experimentation, however, our methods are applicable to BERT-Large as well. BERT-Base and BERT-Large are different in model parameters such as number of transformer blocks, number of self-attention heads. Total number of parameters in BERT-Base are 110M and 340M in BERT-Large.\nBERT suffers from major limitations in terms of handling long sequences. Firstly, the self-attention layer has a quadratic complexity $O(n^2)$ in terms of the sequence length $n$ BIBREF0. Secondly, BERT uses a learned positional embeddings scheme BIBREF1, which means that it won't likely be able to generalize to positions beyond those seen in the training data.\nTo investigate the effect of fine-tuning BERT on task performance, we use either the pre-trained BERT weights, or the weights from a BERT fine-tuned on the task-specific dataset on a segment-level (i.e. we preserve the original label but fine-tune on each segment separately instead of on the whole text sequence). We compare these results to using the fine-tuned segment-level BERT predictions directly as inputs to the next layer.\nMethod ::: Recurrence over BERT\nGiven that BERT is limited to a particular input length, we split the input sequence into segments of a fixed size with overlap. For each of these segments, we obtain H or P from BERT model. We then stack these segment-level representations into a sequence, which serves as input to a small (100-dimensional) LSTM layer. Its output serves as a document embedding. Finally, we use two fully connected layers with ReLU (30-dimensional) and softmax (the same dimensionality as the number of classes) activations to obtain the final predictions.\nWith this approach, we overcome BERT's computational complexity, reducing it to $O(n/k * k^2) = O(nk)$ for RoBERT, with $k$ denoting the segment size (the LSTM component has negligible linear complexity $O(k)$). The positional embeddings are also no longer an issue.\nMethod ::: Transformer over BERT\nGiven that Transformers' edge over recurrent networks is their ability to effectively capture long distance relationships between words in a sequence BIBREF0, we experiment with replacing the LSTM recurrent layer in favor of a small Transformer model (2 layers of transformer building block containing self-attention, fully connected, etc.). To investigate if preserving the information about the input sequence order is important, we also build a variant of ToBERT which learns positional embeddings at the segment-level representations (but is limited to sequences of length seen during the training).\nToBERT's computational complexity $O(\\frac{n^2}{k^2})$ is asymptotically inferior to RoBERT, as the top-level Transformer model again suffers from quadratic complexity in the number of segments. However, in practice this number is much smaller than the input sequence length (${\\frac{n}{k}} << n$), so we haven't observed performance or memory issues with our datasets.\nExperiments\nWe evaluated our models on 3 different datasets:\nCSAT dataset for CSAT prediction, consisting of spoken transcripts (automatic via ASR).\n20 newsgroups for topic identification task, consisting of written text;\nFisher Phase 1 corpus for topic identification task, consisting of spoken transcripts (manual);\nExperiments ::: CSAT\nCSAT dataset consists of US English telephone speech from call centers. For each call in this dataset, customers participated in that call gave a rating on his experience with agent. Originally, this dataset has labels rated on a scale 1-9 with 9 being extremely satisfied and 1 being extremely dissatisfied. Fig. FIGREF16 shows the histogram of ratings for our dataset. As the distribution is skewed towards extremes, we choose to do binary classification with ratings above 4.5 as satisfied and below 4.5 as dissatisfied. Quantization of ratings also helped us to create a balanced dataset. This dataset contains 4331 calls and we split them into 3 sets for our experiments: 2866 calls for training, 362 calls for validation and, finally, 1103 calls for testing.\nWe obtained the transcripts by employing an ASR system. The ASR system uses TDNN-LSTM acoustic model trained on Fisher and Switchboard datasets with lattice-free maximum mutual information criterion BIBREF18. The word error rates using four-gram language models were 9.2% and 17.3% respectively on Switchboard and CallHome portions of Eval2000 dataset.\nExperiments ::: 20 newsgroups\n20 newsgroups data set is one of the frequently used datasets in the text processing community for text classification and text clustering. This data set contains approximately 20,000 English documents from 20 topics to be identified, with 11314 documents for training and 7532 for testing. In this work, we used only 90% of documents for training and the remaining 10% for validation. For fair comparison with other publications, we used 53160 words vocabulary set available in the datasets website.\nExperiments ::: Fisher\nFisher Phase 1 US English corpus is often used for automatic speech recognition in speech community. In this work, we used it for topic identification as in BIBREF3. The documents are 10-minute long telephone conversations between two people discussing a given topic. We used same training and test splits as BIBREF3 in which 1374 and 1372 documents are used for training and testing respectively. For validation of our model, we used 10% of training dataset and the remaining 90% was used for actual model training. The number of topics in this data set is 40.\nExperiments ::: Dataset Statistics\nTable TABREF22 shows statistics of our datasets. It can be observed that average length of Fisher is much higher than 20 newsgroups and CSAT. Cumulative distribution of document lengths for each dataset is shown in Fig. FIGREF21. It can be observed that almost all of the documents in Fisher dataset have length more than 1000 words. For CSAT, more than 50% of the documents have length greater than 500 and for 20newsgroups only 10% of the documents have length greater than 500. Note that, for CSAT and 20newsgroups, there are few documents with length more than 5000.\nExperiments ::: Architecture and Training Details\nIn this work, we split document into segments of 200 tokens with a shift of 50 tokens to extract features from BERT model. For RoBERT, LSTM model is trained to minimize cross-entropy loss with Adam optimizer BIBREF19. The initial learning rate is set to $0.001$ and is reduced by a factor of $0.95$ if validation loss does not decrease for 3-epochs. For ToBERT, the Transformer is trained with the default BERT version of Adam optimizer BIBREF1 with an initial learning rate of $5e$-5. We report accuracy in all of our experiments. We chose a model with the best validation accuracy to calculate accuracy on the test set. To accomodate for non-determinism of some TensorFlow GPU operations, we report accuracy averaged over 5 runs.\nResults\nTable TABREF25 presents results using pre-trained BERT features. We extracted features from the pooled output of final transformer block as these were shown to be working well for most of the tasks BIBREF1. The features extracted from a pre-trained BERT model without any fine-tuning lead to a sub-par performance. However, We also notice that ToBERT model exploited the pre-trained BERT features better than RoBERT. It also converged faster than RoBERT. Table TABREF26 shows results using features extracted after fine-tuning BERT model with our datasets. Significant improvements can be observed compared to using pre-trained BERT features. Also, it can be noticed that ToBERT outperforms RoBERT on Fisher and 20newsgroups dataset by 13.63% and 0.81% respectively. On CSAT, ToBERT performs slightly worse than RoBERT but it is not statistically significant as this dataset is small.\nTable TABREF27 presents results using fine-tuned BERT predictions instead of the pooled output from final transformer block. For each document, having obtained segment-wise predictions we can obtain final prediction for the whole document in three ways:\nCompute the average of all segment-wise predictions and find the most probable class;\nFind the most frequently predicted class;\nTrain a classification model.\nIt can be observed from Table TABREF27 that a simple averaging operation or taking most frequent predicted class works competitively for CSAT and 20newsgroups but not for the Fisher dataset. We believe the improvements from using RoBERT or ToBERT, compared to simple averaging or most frequent operations, are proportional to the fraction of long documents in the dataset. CSAT and 20newsgroups have (on average) significantly shorter documents than Fisher, as seen in Fig. FIGREF21. Also, significant improvements for Fisher could be because of less confident predictions from BERT model as this dataset has 40 classes. Fig. FIGREF31 presents the comparison of average voting and ToBERT for various document length ranges for Fisher dataset. We used fine-tuned BERT segment-level predictions (P) for this analysis. It can be observed that ToBERT outperforms average voting in every interval. To the best of our knowledge, this is a state-of-the-art result reported on the Fisher dataset.\nTable TABREF32 presents the effect of position embeddings on the model performance. It can be observed that position embeddings did not significantly affect the model performance for Fisher and 20newsgroups, but they helped slightly in CSAT prediction (an absolute improvement of 0.64% F1-score). We think that this is explained by the fact that Fisher and 20newsgroups are topic identification tasks, and the topic does not change much throughout these documents. However, CSAT may vary during the call, and in some cases a naive assumption that the sequential nature of the transcripts is irrelevant may lead to wrong conclusions.\nTable TABREF33 compares our results with previous works. It can be seen that our model ToBERT outperforms CNN based experiments by significant margin on CSAT and Fisher datasets. For CSAT dataset, we used multi-scale CNN (MS-CNN) as the baseline, given its strong results on Fisher and 20newsgroups. The setup was replicated from BIBREF5 for comparison. We also see that our result on 20 newsgroups is 0.6% worse than the state-of-the-art.\nConclusions\nIn this paper, we presented two methods for long documents using BERT model: RoBERT and ToBERT. We evaluated our experiments on two classification tasks - customer satisfaction prediction and topic identification - using 3 datasets: CSAT, 20newsgroups and Fisher. We observed that ToBERT outperforms RoBERT on pre-trained BERT features and fine-tuned BERT features for all our tasks. Also, we noticed that fine-tuned BERT performs better than pre-trained BERT. We have shown that both RoBERT and ToBERT improved the simple baselines of taking an average (or the most frequent) of segment-wise predictions for long documents to obtain final prediction. Position embeddings did not significantly affect our models performance, but slightly improved the accuracy on the CSAT task. We obtained the best results on Fisher dataset and good improvements for CSAT task compared to the CNN baseline. It is interesting to note that the longer the average input in a given task, the bigger improvement we observe w.r.t. the baseline for that task. Our results confirm that both RoBERT and ToBERT can be used for long sequences with competitive performance and quick fine-tuning procedure. For future work, we shall focus on training models on long documents directly (i.e. in an end-to-end manner).", "answers": ["CSAT dataset, 20 newsgroups, Fisher Phase 1 corpus", "CSAT dataset , 20 newsgroups, Fisher Phase 1 corpus"], "length": 2652, "dataset": "qasper", "language": "en", "all_classes": null, "_id": "466bd29bcab1cdfdef327777808236bd2677e2a54414a32a"}
{"input": "Were any of these tasks evaluated in any previous work?", "context": "Introduction\nThe recently introduced BERT model BIBREF0 exhibits strong performance on several language understanding benchmarks. To what extent does it capture syntax-sensitive structures?\nRecent work examines the extent to which RNN-based models capture syntax-sensitive phenomena that are traditionally taken as evidence for the existence in hierarchical structure. In particular, in BIBREF1 we assess the ability of LSTMs to learn subject-verb agreement patterns in English, and evaluate on naturally occurring wikipedia sentences. BIBREF2 also consider subject-verb agreement, but in a “colorless green ideas” setting in which content words in naturally occurring sentences are replaced with random words with the same part-of-speech and inflection, thus ensuring a focus on syntax rather than on selectional-preferences based cues. BIBREF3 consider a wider range of syntactic phenomena (subject-verb agreement, reflexive anaphora, negative polarity items) using manually constructed stimuli, allowing for greater coverage and control than in the naturally occurring setting.\nThe BERT model is based on the “Transformer” architecture BIBREF4 , which—in contrast to RNNs—relies purely on attention mechanisms, and does not have an explicit notion of word order beyond marking each word with its absolute-position embedding. This reliance on attention may lead one to expect decreased performance on syntax-sensitive tasks compared to RNN (LSTM) models that do model word order directly, and explicitly track states across the sentence. Indeed, BIBREF5 finds that transformer-based models perform worse than LSTM models on the BIBREF1 agreement prediction dataset. In contrast, BIBREF6 find that self-attention performs on par with LSTM for syntax sensitive dependencies in the context of machine-translation, and performance on syntactic tasks is correlated with the number of attention heads in multi-head attention.\nI adapt the evaluation protocol and stimuli of BIBREF1 , BIBREF2 and BIBREF3 to the bidirectional setting required by BERT, and evaluate the pre-trained BERT models (both the Large and the Base models). Surprisingly (at least to me), the out-of-the-box models (without any task-specific fine-tuning) perform very well on all the syntactic tasks.\nMethodology\nI use the stimuli provided by BIBREF1 , BIBREF2 , BIBREF3 , but change the experimental protocol to adapt it to the bidirectional nature of the BERT model. This requires discarding some of the stimuli, as described below. Thus, the numbers are not strictly comparable to those reported in previous work.\nPrevious setups\nAll three previous work use uni-directional language-model-like models.\nBIBREF1 start with existing sentences from wikipedia that contain a present-tense verb. They feed each sentence word by word into an LSTM, stop right before the focus verb, and ask the model to predict a binary plural/singular decision (supervised setup) or compare the probability assigned by a pre-trained language model (LM) to the plural vs singular forms of the verb (LM setup). The evaluation is then performed on sentences with “agreement attractors” in which at there is at least one noun between the verb and its subject, and all of the nouns between the verb and subject are of the opposite number from the subject.\nBIBREF2 also start with existing sentences. However, in order to control for the possibillity of the model learning to rely on “semantic” selectional-preferences cues rather than syntactic ones, they replace each content word with random words from the same part-of-speech and inflection. This results in “coloreless green ideas” nonce sentences. The evaluation is then performed similarly to the LM setup of BIBREF1 : the sentence is fed into a pre-traiend LSTM LM up to the focus verb, and the model is considered correct if the probability assigned to the correct inflection of the original verb form given the prefix is larger than that assigned to the incorrect inflection.\nBIBREF3 focus on manually constructed and controlled stimuli, that also emphasizes linguistic structure over selectional preferences. They construct minimal pairs of grammatical and ungrammatical sentences, feed each one in its entirety into a pre-trained LSTM-LM, and compare the perplexity assigned by the model to the grammatical and ungrammatical sentences. The model is “correct” if it assigns the grammatical sentence a higher probability than to the ungrammatical one. Since the minimal pairs for most phenomena differ only in a single word (the focus verb), this scoring is very similar to the one used in the two previous works. However, it does consider the continuation of the sentence after the focus verb, and also allows for assessing phenomena that require change into two or more words (like negative polarity items).\nAdaptation to the BERT model\nIn contrast to these works, the BERT model is bi-directional: it is trained to predict the identity of masked words based on both the prefix and suffix surrounding these words. I adapt the uni-directional setup by feeding into BERT the complete sentence, while masking out the single focus verb. I then ask BERT for its word predictions for the masked position, and compare the score assigned to the original correct verb to the score assigned to the incorrect one.\nFor example, for the sentence:\na 2002 systemic review of herbal products found that several herbs , including peppermint and caraway , have anti-dyspeptic effects for non-ulcer dyspepsia with “ encouraging safety profiles ” . (from BIBREF1 )\nI feed into BERT:\n[CLS] a 2002 systemic review of herbal products found that several herbs , including peppermint and caraway , [MASK] anti-dyspeptic effects for non-ulcer dyspepsia with “ encouraging safety profiles ” . and look for the score assigned to the words have and has at the masked position.\nSimilarly, for the pair\nthe game that the guard hates is bad .\nthe game that the guard hates are bad .\n(from BIBREF3 ), I feed into BERT:\n[CLS] the game that the guard hates [MASK] bad .\nand compare the scores predicted for is and are.\nThis differs from BIBREF1 and BIBREF2 by considering the entire sentence (excluding the verb) and not just its prefix leading to the verb, and differs from BIBREF3 by conditioning the focus verb on bidirectional context.\nI use the PyTorch implementation of BERT, with the pre-trained models supplied by Google. I experiment with the bert-large-uncased and bert-base-uncased models.\nThe bi-directional setup precludes using using the NPI stimuli of BIBREF3 , in which the minimal pair differs in two words position, which I discard from the evaluation. I also discard the agreement cases involving the verbs is or are in BIBREF1 and in BIBREF2 , because some of them are copular construction, in which strong agreement hints can be found also on the object following the verb. This is not an issue in the manually constructed BIBREF3 stimuli due to the patterns they chose.\nFinally, I discard stimuli in which the focus verb or its plural/singular inflection does not appear as a single word in the BERT word-piece-based vocabulary (and hence cannot be predicted by the model). This include discarding BIBREF3 stimuli involving the words swims or admires, resulting in 23,368 discarded pairs (out of 152,300). I similarly discard 680 sentences from BIBREF1 where the focus verb or its inflection were one of 108 out-of-vocabulary tokens, and 28 sentence-pairs (8 tokens) from BIBREF2 .\nThe BERT results are not directly comparable to the numbers reported in previous work. Beyond the differences due to bidirectionality and the discarded stimuli, the BERT models are also trained on a different and larger corpus (covering both wikipedia and books).\nCode is available at https://github.com/yoavg/bert-syntax.\nResults\nTables 1 , 2 and 3 show the results. All cases exhibit high scores—in the vast majority of the cases substantially higher than reported in previous work. As discussed above, the results are not directly comparable to previous work: the BERT models are trained on different (and larger) data, are allowed to access the suffix of the sentence in addition to its prefix, and are evaluated on somewhat different data due to discarding OOV items. Still, taken together, the high performance numbers indicate that the purely attention-based BERT models are likely capable of capturing the same kind of syntactic regularities that LSTM-based models are capable of capturing, at least as well as the LSTM models and probably better.\nAnother noticeable and interesting trend is that larger is not necessarily better: the BERT-Base model outperforms the BERT-Large model on many of the syntactic conditions.\nDiscussion\nThe BERT models perform remarkably well on all the syntactic test cases. I expected the attention-based mechanism to fail on these (compared to the LSTM-based models), and am surprised by these results. The BIBREF2 and BIBREF3 conditions rule out the possibility of overly relying on selectional preference cues or memorizing the wikipedia training data, and suggest real syntactic generalization is taking place. Exploring the extent to which deep purely-attention-based architectures such as BERT are capable of capturing hierarchy-sensitive and syntactic dependencies—as well as the mechanisms by which this is achieved—is a fascinating area for future research.", "answers": ["Yes", "Yes"], "length": 1464, "dataset": "qasper", "language": "en", "all_classes": null, "_id": "6a78dbe4f8e30d35c13ea1f80e52df286f9e1f664f9c3d98"}
{"input": "How is the dataset of hashtags sourced?", "context": "Introduction\nA 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 .\nHashtags 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.\nThe goal of our study is to build efficient methods for automatically splitting a hashtag into a meaningful word sequence. Our contributions are:\nOur 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.\nBackground and Preliminaries\nCurrent 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.\nGenerating 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\nwhere 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.\nHowever, 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.\nMulti-task Pairwise Neural Ranking\nWe 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 ).\nSegmentation as Pairwise Ranking\nThe 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 ).\nWe 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\nWe 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.\nPairwise Neural Ranking Model\nFor 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\nwhere INLINEFORM0 is the number of training examples.\nTo 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.\nMargin Ranking (MR) Loss\nAs 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\nwhere INLINEFORM0 is the number of training samples. The architecture of this model is presented in Figure FIGREF11 (b).\nAdaptive Multi-task Learning\nBoth 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.\nTo 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\nwhere 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.\nMore 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\nWe 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\nFeatures\nWe 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 ).\nSimilarly, 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.\nImplementation Details\nWe 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.\nFor 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 .\nHashtag Segmentation Data\nWe 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.\nExperiments\nIn 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.\nExisting Methods\nWe compare our pairwise neural ranker with the following baseline and state-of-the-art approaches:\nThe original hashtag as a single token;\nA rule-based segmenter, which employs a set of word-shape rules with an English dictionary BIBREF13 ;\nA Viterbi model which uses word frequencies from a book corpus BIBREF0 ;\nThe specially developed GATE Hashtag Tokenizer from the open source toolkit, which combines dictionaries and gazetteers in a Viterbi-like algorithm BIBREF11 ;\nA 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 ;\nOur reimplementation of the Word Breaker algorithm which uses beam search and a Twitter ngram language model BIBREF16 ;\nA 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.\nEvaluation Metrics\nWe 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.\nResults\nTables 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).\nThe 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.\nAnalysis and Discussion\nTo 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.\nAs 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 ).\nSince 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.\nLanguage 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.\nExtrinsic Evaluation: Twitter Sentiment Analysis\nWe 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.\nExperimental Setup\nWe 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.\nWe 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.\nResults and Analysis\nIn 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%.\nOther Related Work\nAutomatic 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 .\nConclusion\nWe 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.\nAcknowledgments\nWe 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.\nWord-shape rules\nOur 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.", "answers": ["1,268 randomly selected tweets in the Stanford Sentiment Analysis Dataset BIBREF36, all 12,594 unique English hashtags and their associated tweets from the same Stanford dataset", "Stanford Sentiment Analysis Dataset BIBREF36"], "length": 3756, "dataset": "qasper", "language": "en", "all_classes": null, "_id": "3278efeb1fb7f569dfd6ef8627cdefb5955b93b612fe2902"}
{"input": "What is the tagging scheme employed?", "context": "Introduction\nThere exists a class of language construction known as pun in natural language texts and utterances, where a certain word or other lexical items are used to exploit two or more separate meanings. It has been shown that understanding of puns is an important research question with various real-world applications, such as human-computer interaction BIBREF0 , BIBREF1 and machine translation BIBREF2 . Recently, many researchers show their interests in studying puns, like detecting pun sentences BIBREF3 , locating puns in the text BIBREF4 , interpreting pun sentences BIBREF5 and generating sentences containing puns BIBREF6 , BIBREF7 , BIBREF8 . A pun is a wordplay in which a certain word suggests two or more meanings by exploiting polysemy, homonymy, or phonological similarity to another sign, for an intended humorous or rhetorical effect. Puns can be generally categorized into two groups, namely heterographic puns (where the pun and its latent target are phonologically similar) and homographic puns (where the two meanings of the pun reflect its two distinct senses) BIBREF9 . Consider the following two examples:\nThe first punning joke exploits the sound similarity between the word “propane\" and the latent target “profane\", which can be categorized into the group of heterographic puns. Another categorization of English puns is homographic pun, exemplified by the second instance leveraging distinct senses of the word “gut\".\nPun detection is the task of detecting whether there is a pun residing in the given text. The goal of pun location is to find the exact word appearing in the text that implies more than one meanings. Most previous work addresses such two tasks separately and develop separate systems BIBREF10 , BIBREF5 . Typically, a system for pun detection is built to make a binary prediction on whether a sentence contains a pun or not, where all instances (with or without puns) are taken into account during training. For the task of pun location, a separate system is used to make a single prediction as to which word in the given sentence in the text that trigger more than one semantic interpretations of the text, where the training data involves only sentences that contain a pun. Therefore, if one is interested in solving both problems at the same time, a pipeline approach that performs pun detection followed by pun location can be used.\nCompared to the pipeline methods, joint learning has been shown effective BIBREF11 , BIBREF12 since it is able to reduce error propagation and allows information exchange between tasks which is potentially beneficial to all the tasks. In this work, we demonstrate that the detection and location of puns can be jointly addressed by a single model. The pun detection and location tasks can be combined as a sequence labeling problem, which allows us to jointly detect and locate a pun in a sentence by assigning each word a tag. Since each context contains a maximum of one pun BIBREF9 , we design a novel tagging scheme to capture this structural constraint. Statistics on the corpora also show that a pun tends to appear in the second half of a context. To capture such a structural property, we also incorporate word position knowledge into our structured prediction model. Experiments on the benchmark datasets show that detection and location tasks can reinforce each other, leading to new state-of-the-art performance on these two tasks. To the best of our knowledge, this is the first work that performs joint detection and location of English puns by using a sequence labeling approach.\nProblem Definition\nWe first design a simple tagging scheme consisting of two tags { INLINEFORM0 }:\nINLINEFORM0 tag means the current word is not a pun.\nINLINEFORM0 tag means the current word is a pun.\nIf the tag sequence of a sentence contains a INLINEFORM0 tag, then the text contains a pun and the word corresponding to INLINEFORM1 is the pun.\nThe contexts have the characteristic that each context contains a maximum of one pun BIBREF9 . In other words, there exists only one pun if the given sentence is detected as the one containing a pun. Otherwise, there is no pun residing in the text. To capture this interesting property, we propose a new tagging scheme consisting of three tags, namely { INLINEFORM0 }.\nINLINEFORM0 tag indicates that the current word appears before the pun in the given context.\nINLINEFORM0 tag highlights the current word is a pun.\nINLINEFORM0 tag indicates that the current word appears after the pun.\nWe empirically show that the INLINEFORM0 scheme can guarantee the context property that there exists a maximum of one pun residing in the text.\nGiven a context from the training set, we will be able to generate its corresponding gold tag sequence using a deterministic procedure. Under the two schemes, if a sentence does not contain any puns, all words will be tagged with INLINEFORM0 or INLINEFORM1 , respectively. Exemplified by the second sentence “Some diets cause a gut reaction,\" the pun is given as “gut.\" Thus, under the INLINEFORM2 scheme, it should be tagged with INLINEFORM3 , while the words before it are assigned with the tag INLINEFORM4 and words after it are with INLINEFORM5 , as illustrated in Figure FIGREF8 . Likewise, the INLINEFORM6 scheme tags the word “gut\" with INLINEFORM7 , while other words are tagged with INLINEFORM8 . Therefore, we can combine the pun detection and location tasks into one problem which can be solved by the sequence labeling approach.\nModel\nNeural models have shown their effectiveness on sequence labeling tasks BIBREF13 , BIBREF14 , BIBREF15 . In this work, we adopt the bidirectional Long Short Term Memory (BiLSTM) BIBREF16 networks on top of the Conditional Random Fields BIBREF17 (CRF) architecture to make labeling decisions, which is one of the classical models for sequence labeling. Our model architecture is illustrated in Figure FIGREF8 with a running example. Given a context/sentence INLINEFORM0 where INLINEFORM1 is the length of the context, we generate the corresponding tag sequence INLINEFORM2 based on our designed tagging schemes and the original annotations for pun detection and location provided by the corpora. Our model is then trained on pairs of INLINEFORM3 .\nInput. The contexts in the pun corpus hold the property that each pun contains exactly one content word, which can be either a noun, a verb, an adjective, or an adverb. To capture this characteristic, we consider lexical features at the character level. Similar to the work of BIBREF15 , the character embeddings are trained by the character-level LSTM networks on the unannotated input sequences. Nonlinear transformations are then applied to the character embeddings by highway networks BIBREF18 , which map the character-level features into different semantic spaces.\nWe also observe that a pun tends to appear at the end of a sentence. Specifically, based on the statistics, we found that sentences with a pun that locate at the second half of the text account for around 88% and 92% in homographic and heterographic datasets, respectively. We thus introduce a binary feature that indicates if a word is located at the first or the second half of an input sentence to capture such positional information. A binary indicator can be mapped to a vector representation using a randomly initialized embedding table BIBREF19 , BIBREF20 . In this work, we directly adopt the value of the binary indicator as part of the input.\nThe concatenation of the transformed character embeddings, the pre-trained word embeddings BIBREF21 , and the position indicators are taken as input of our model.\nTagging. The input is then fed into a BiLSTM network, which will be able to capture contextual information. For a training instance INLINEFORM0 , we suppose the output by the word-level BiLSTM is INLINEFORM1 . The CRF layer is adopted to capture label dependencies and make final tagging decisions at each position, which has been included in many state-of-the-art sequence labeling models BIBREF14 , BIBREF15 . The conditional probability is defined as:\nwhere INLINEFORM0 is a set of all possible label sequences consisting of tags from INLINEFORM1 (or INLINEFORM2 ), INLINEFORM3 and INLINEFORM4 are weight and bias parameters corresponding to the label pair INLINEFORM5 . During training, we minimize the negative log-likelihood summed over all training instances:\nwhere INLINEFORM0 refers to the INLINEFORM1 -th instance in the training set. During testing, we aim to find the optimal label sequence for a new input INLINEFORM2 :\nThis search process can be done efficiently using the Viterbi algorithm.\nDatasets and Settings\nWe evaluate our model on two benchmark datasets BIBREF9 . The homographic dataset contains 2,250 contexts, 1,607 of which contain a pun. The heterographic dataset consists of 1,780 contexts with 1,271 containing a pun. We notice there is no standard splitting information provided for both datasets. Thus we apply 10-fold cross validation. To make direct comparisons with prior studies, following BIBREF4 , we accumulated the predictions for all ten folds and calculate the scores in the end.\nFor each fold, we randomly select 10% of the instances from the training set for development. Word embeddings are initialized with the 100-dimensional Glove BIBREF21 . The dimension of character embeddings is 30 and they are randomly initialized, which can be fine tuned during training. The pre-trained word embeddings are not updated during training. The dimensions of hidden vectors for both char-level and word-level LSTM units are set to 300. We adopt stochastic gradient descent (SGD) BIBREF26 with a learning rate of 0.015.\nFor the pun detection task, if the predicted tag sequence contains at least one INLINEFORM0 tag, we regard the output (i.e., the prediction of our pun detection model) for this task as true, otherwise false. For the pun location task, a predicted pun is regarded as correct if and only if it is labeled as the gold pun in the dataset. As to pun location, to make fair comparisons with prior studies, we only consider the instances that are labeled as the ones containing a pun. We report precision, recall and INLINEFORM1 score in Table TABREF11 . A list of prior works that did not employ joint learning are also shown in the first block of Table TABREF11 .\nResults\nWe also implemented a baseline model based on conditional random fields (CRF), where features like POS tags produced by the Stanford POS tagger BIBREF27 , n-grams, label transitions, word suffixes and relative position to the end of the text are considered. We can see that our model with the INLINEFORM0 tagging scheme yields new state-of-the-art INLINEFORM1 scores on pun detection and competitive results on pun location, compared to baselines that do not adopt joint learning in the first block. For location on heterographic puns, our model's performance is slightly lower than the system of BIBREF25 , which is a rule-based locator. Compared to CRF, we can see that our model, either with the INLINEFORM2 or the INLINEFORM3 scheme, yields significantly higher recall on both detection and location tasks, while the precisions are relatively close. This demonstrates the effectiveness of BiLSTM, which learns the contextual features of given texts – such information appears to be helpful in recalling more puns.\nCompared to the INLINEFORM0 scheme, the INLINEFORM1 tagging scheme is able to yield better performance on these two tasks. After studying outputs from these two approaches, we found that one leading source of error for the INLINEFORM2 approach is that there exist more than one words in a single instance that are assigned with the INLINEFORM3 tag. However, according to the description of pun in BIBREF9 , each context contains a maximum of one pun. Thus, such a useful structural constraint is not well captured by the simple approach based on the INLINEFORM4 tagging scheme. On the other hand, by applying the INLINEFORM5 tagging scheme, such a constraint is properly captured in the model. As a result, the results for such a approach are significantly better than the approach based on the INLINEFORM6 tagging scheme, as we can observe from the table. Under the same experimental setup, we also attempted to exclude word position features. Results are given by INLINEFORM7 - INLINEFORM8 . It is expected that the performance of pun location drops, since such position features are able to capture the interesting property that a pun tends to appear in the second half of a sentence. While such knowledge is helpful for the location task, interestingly, a model without position knowledge yields improved performance on the pun detection task. One possible reason is that detecting whether a sentence contains a pun is not concerned with such word position information.\nAdditionally, we conduct experiments over sentences containing a pun only, namely 1,607 and 1,271 instances from homographic and heterographic pun corpora separately. It can be regarded as a “pipeline” method where the classifier for pun detection is regarded as perfect. Following the prior work of BIBREF4 , we apply 10-fold cross validation. Since we are given that all input sentences contain a pun, we only report accumulated results on pun location, denoted as Pipeline in Table TABREF11 . Compared with our approaches, the performance of such an approach drops significantly. On the other hand, such a fact demonstrates that the two task, detection and location of puns, can reinforce each other. These figures demonstrate the effectiveness of our sequence labeling method to detect and locate English puns in a joint manner.\nError Analysis\nWe studied the outputs from our system and make some error analysis. We found the errors can be broadly categorized into several types, and we elaborate them here. 1) Low word coverage: since the corpora are relatively small, there exist many unseen words in the test set. Learning the representations of such unseen words is challenging, which affects the model's performance. Such errors contribute around 40% of the total errors made by our system. 2) Detection errors: we found many errors are due to the model's inability to make correct pun detection. Such inability harms both pun detection and pun location. Although our approach based on the INLINEFORM0 tagging scheme yields relatively higher scores on the detection task, we still found that 40% of the incorrectly predicted instances fall into this group. 3) Short sentences: we found it was challenging for our model to make correct predictions when the given text is short. Consider the example “Superglue! Tom rejoined,\" here the word rejoined is the corresponding pun. However, it would be challenging to figure out the pun with such limited contextual information.\nRelated Work\nMost existing systems address pun detection and location separately. BIBREF22 applied word sense knowledge to conduct pun detection. BIBREF24 trained a bidirectional RNN classifier for detecting homographic puns. Next, a knowledge-based approach is adopted to find the exact pun. Such a system is not applicable to heterographic puns. BIBREF28 applied Google n-gram and word2vec to make decisions. The phonetic distance via the CMU Pronouncing Dictionary is computed to detect heterographic puns. BIBREF10 used the hidden Markov model and a cyclic dependency network with rich features to detect and locate puns. BIBREF23 used a supervised approach to pun detection and a weakly supervised approach to pun location based on the position within the context and part of speech features. BIBREF25 proposed a rule-based system for pun location that scores candidate words according to eleven simple heuristics. Two systems are developed to conduct detection and location separately in the system known as UWAV BIBREF3 . The pun detector combines predictions from three classifiers. The pun locator considers word2vec similarity between every pair of words in the context and position to pinpoint the pun. The state-of-the-art system for homographic pun location is a neural method BIBREF4 , where the word senses are incorporated into a bidirectional LSTM model. This method only supports the pun location task on homographic puns. Another line of research efforts related to this work is sequence labeling, such as POS tagging, chunking, word segmentation and NER. The neural methods have shown their effectiveness in this task, such as BiLSTM-CNN BIBREF13 , GRNN BIBREF29 , LSTM-CRF BIBREF30 , LSTM-CNN-CRF BIBREF14 , LM-LSTM-CRF BIBREF15 .\nIn this work, we combine pun detection and location tasks as a single sequence labeling problem. Inspired by the work of BIBREF15 , we also adopt a LSTM-CRF with character embeddings to make labeling decisions.\nConclusion\nIn this paper, we propose to perform pun detection and location tasks in a joint manner from a sequence labeling perspective. We observe that each text in our corpora contains a maximum of one pun. Hence, we design a novel tagging scheme to incorporate such a constraint. Such a scheme guarantees that there is a maximum of one word that will be tagged as a pun during the testing phase. We also found the interesting structural property such as the fact that most puns tend to appear at the second half of the sentences can be helpful for such a task, but was not explored in previous works. Furthermore, unlike many previous approaches, our approach, though simple, is generally applicable to both heterographic and homographic puns. Empirical results on the benchmark datasets prove the effectiveness of the proposed approach that the two tasks of pun detection and location can be addressed by a single model from a sequence labeling perspective.\nFuture research includes the investigations on how to make use of richer semantic and linguistic information for detection and location of puns. Research on puns for other languages such as Chinese is still under-explored, which could also be an interesting direction for our future studies.\nAcknowledgments\nWe would like to thank the three anonymous reviewers for their thoughtful and constructive comments. This work is supported by Singapore Ministry of Education Academic Research Fund (AcRF) Tier 2 Project MOE2017-T2-1-156, and is partially supported by SUTD project PIE-SGP-AI-2018-01.", "answers": ["A new tagging scheme that tags the words before and after the pun as well as the pun words.", "a new tagging scheme consisting of three tags, namely { INLINEFORM0 }"], "length": 2974, "dataset": "qasper", "language": "en", "all_classes": null, "_id": "c276162e5903240b254238163922fb0ef59df8b8c706790f"}
{"input": "Were other baselines tested to compare with the neural baseline?", "context": "Introduction\nPrivacy policies are the documents which disclose the ways in which a company gathers, uses, shares and manages a user's data. As legal documents, they function using the principle of notice and choice BIBREF0, where companies post their policies, and theoretically, users read the policies and decide to use a company's products or services only if they find the conditions outlined in its privacy policy acceptable. Many legal jurisdictions around the world accept this framework, including the United States and the European Union BIBREF1, BIBREF2. However, the legitimacy of this framework depends upon users actually reading and understanding privacy policies to determine whether company practices are acceptable to them BIBREF3. In practice this is seldom the case BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10. This is further complicated by the highly individual and nuanced compromises that users are willing to make with their data BIBREF11, discouraging a `one-size-fits-all' approach to notice of data practices in privacy documents.\nWith devices constantly monitoring our environment, including our personal space and our bodies, lack of awareness of how our data is being used easily leads to problematic situations where users are outraged by information misuse, but companies insist that users have consented. The discovery of increasingly egregious uses of data by companies, such as the scandals involving Facebook and Cambridge Analytica BIBREF12, have further brought public attention to the privacy concerns of the internet and ubiquitous computing. This makes privacy a well-motivated application domain for NLP researchers, where advances in enabling users to quickly identify the privacy issues most salient to them can potentially have large real-world impact.\n[1]https://play.google.com/store/apps/details?id=com.gotokeep.keep.intl [2]https://play.google.com/store/apps/details?id=com.viber.voip [3]A question might not have any supporting evidence for an answer within the privacy policy.\nMotivated by this need, we contribute PrivacyQA, a corpus consisting of 1750 questions about the contents of privacy policies, paired with over 3500 expert annotations. The goal of this effort is to kickstart the development of question-answering methods for this domain, to address the (unrealistic) expectation that a large population should be reading many policies per day. In doing so, we identify several understudied challenges to our ability to answer these questions, with broad implications for systems seeking to serve users' information-seeking intent. By releasing this resource, we hope to provide an impetus to develop systems capable of language understanding in this increasingly important domain.\nRelated Work\nPrior work has aimed to make privacy policies easier to understand. Prescriptive approaches towards communicating privacy information BIBREF21, BIBREF22, BIBREF23 have not been widely adopted by industry. Recently, there have been significant research effort devoted to understanding privacy policies by leveraging NLP techniques BIBREF24, BIBREF25, BIBREF26, BIBREF27, BIBREF28, especially by identifying specific data practices within a privacy policy. We adopt a personalized approach to understanding privacy policies, that allows users to query a document and selectively explore content salient to them. Most similar is the PolisisQA corpus BIBREF29, which examines questions users ask corporations on Twitter. Our approach differs in several ways: 1) The PrivacyQA dataset is larger, containing 10x as many questions and answers. 2) Answers are formulated by domain experts with legal training. 3) PrivacyQA includes diverse question types, including unanswerable and subjective questions.\nOur work is also related to reading comprehension in the open domain, which is frequently based upon Wikipedia passages BIBREF16, BIBREF17, BIBREF15, BIBREF30 and news articles BIBREF20, BIBREF31, BIBREF32. Table.TABREF4 presents the desirable attributes our dataset shares with past approaches. This work is also tied into research in applying NLP approaches to legal documents BIBREF33, BIBREF34, BIBREF35, BIBREF36, BIBREF37, BIBREF38, BIBREF39. While privacy policies have legal implications, their intended audience consists of the general public rather than individuals with legal expertise. This arrangement is problematic because the entities that write privacy policies often have different goals than the audience. feng2015applying, tan-EtAl:2016:P16-1 examine question answering in the insurance domain, another specialized domain similar to privacy, where the intended audience is the general public.\nData Collection\nWe describe the data collection methodology used to construct PrivacyQA. With the goal of achieving broad coverage across application types, we collect privacy policies from 35 mobile applications representing a number of different categories in the Google Play Store. One of our goals is to include both policies from well-known applications, which are likely to have carefully-constructed privacy policies, and lesser-known applications with smaller install bases, whose policies might be considerably less sophisticated. Thus, setting 5 million installs as a threshold, we ensure each category includes applications with installs on both sides of this threshold. All policies included in the corpus are in English, and were collected before April 1, 2018, predating many companies' GDPR-focused BIBREF41 updates. We leave it to future studies BIBREF42 to look at the impact of the GDPR (e.g., to what extent GDPR requirements contribute to making it possible to provide users with more informative answers, and to what extent their disclosures continue to omit issues that matter to users).\nData Collection ::: Crowdsourced Question Elicitation\nThe intended audience for privacy policies consists of the general public. This informs the decision to elicit questions from crowdworkers on the contents of privacy policies. We choose not to show the contents of privacy policies to crowdworkers, a procedure motivated by a desire to avoid inadvertent biases BIBREF43, BIBREF44, BIBREF45, BIBREF46, BIBREF47, and encourage crowdworkers to ask a variety of questions beyond only asking questions based on practices described in the document.\nInstead, crowdworkers are presented with public information about a mobile application available on the Google Play Store including its name, description and navigable screenshots. Figure FIGREF9 shows an example of our user interface. Crowdworkers are asked to imagine they have access to a trusted third-party privacy assistant, to whom they can ask any privacy question about a given mobile application. We use the Amazon Mechanical Turk platform and recruit crowdworkers who have been conferred “master” status and are located within the United States of America. Turkers are asked to provide five questions per mobile application, and are paid $2 per assignment, taking ~eight minutes to complete the task.\nData Collection ::: Answer Selection\nTo identify legally sound answers, we recruit seven experts with legal training to construct answers to Turker questions. Experts identify relevant evidence within the privacy policy, as well as provide meta-annotation on the question's relevance, subjectivity, OPP-115 category BIBREF49, and how likely any privacy policy is to contain the answer to the question asked.\nData Collection ::: Analysis\nTable.TABREF17 presents aggregate statistics of the PrivacyQA dataset. 1750 questions are posed to our imaginary privacy assistant over 35 mobile applications and their associated privacy documents. As an initial step, we formulate the problem of answering user questions as an extractive sentence selection task, ignoring for now background knowledge, statistical data and legal expertise that could otherwise be brought to bear. The dataset is partitioned into a training set featuring 27 mobile applications and 1350 questions, and a test set consisting of 400 questions over 8 policy documents. This ensures that documents in training and test splits are mutually exclusive. Every question is answered by at least one expert. In addition, in order to estimate annotation reliability and provide for better evaluation, every question in the test set is answered by at least two additional experts.\nTable TABREF14 describes the distribution over first words of questions posed by crowdworkers. We also observe low redundancy in the questions posed by crowdworkers over each policy, with each policy receiving ~49.94 unique questions despite crowdworkers independently posing questions. Questions are on average 8.4 words long. As declining to answer a question can be a legally sound response but is seldom practically useful, answers to questions where a minority of experts abstain to answer are filtered from the dataset. Privacy policies are ~3000 words long on average. The answers to the question asked by the users typically have ~100 words of evidence in the privacy policy document.\nData Collection ::: Analysis ::: Categories of Questions\nQuestions are organized under nine categories from the OPP-115 Corpus annotation scheme BIBREF49:\nFirst Party Collection/Use: What, why and how information is collected by the service provider\nThird Party Sharing/Collection: What, why and how information shared with or collected by third parties\nData Security: Protection measures for user information\nData Retention: How long user information will be stored\nUser Choice/Control: Control options available to users\nUser Access, Edit and Deletion: If/how users can access, edit or delete information\nPolicy Change: Informing users if policy information has been changed\nInternational and Specific Audiences: Practices pertaining to a specific group of users\nOther: General text, contact information or practices not covered by other categories.\nFor each question, domain experts indicate one or more relevant OPP-115 categories. We mark a category as relevant to a question if it is identified as such by at least two annotators. If no such category exists, the category is marked as `Other' if atleast one annotator has identified the `Other' category to be relevant. If neither of these conditions is satisfied, we label the question as having no agreement. The distribution of questions in the corpus across OPP-115 categories is as shown in Table.TABREF16. First party and third party related questions are the largest categories, forming nearly 66.4% of all questions asked to the privacy assistant.\nData Collection ::: Analysis ::: Answer Validation\nWhen do experts disagree? We would like to analyze the reasons for potential disagreement on the annotation task, to ensure disagreements arise due to valid differences in opinion rather than lack of adequate specification in annotation guidelines. It is important to note that the annotators are experts rather than crowdworkers. Accordingly, their judgements can be considered valid, legally-informed opinions even when their perspectives differ. For the sake of this question we randomly sample 100 instances in the test data and analyze them for likely reasons for disagreements. We consider a disagreement to have occurred when more than one expert does not agree with the majority consensus. By disagreement we mean there is no overlap between the text identified as relevant by one expert and another.\nWe find that the annotators agree on the answer for 74% of the questions, even if the supporting evidence they identify is not identical i.e full overlap. They disagree on the remaining 26%. Sources of apparent disagreement correspond to situations when different experts: have differing interpretations of question intent (11%) (for example, when a user asks 'who can contact me through the app', the questions admits multiple interpretations, including seeking information about the features of the app, asking about first party collection/use of data or asking about third party collection/use of data), identify different sources of evidence for questions that ask if a practice is performed or not (4%), have differing interpretations of policy content (3%), identify a partial answer to a question in the privacy policy (2%) (for example, when the user asks `who is allowed to use the app' a majority of our annotators decline to answer, but the remaining annotators highlight partial evidence in the privacy policy which states that children under the age of 13 are not allowed to use the app), and other legitimate sources of disagreement (6%) which include personal subjective views of the annotators (for example, when the user asks `is my DNA information used in any way other than what is specified', some experts consider the boilerplate text of the privacy policy which states that it abides to practices described in the policy document as sufficient evidence to answer this question, whereas others do not).\nExperimental Setup\nWe evaluate the ability of machine learning methods to identify relevant evidence for questions in the privacy domain. We establish baselines for the subtask of deciding on the answerability (§SECREF33) of a question, as well as the overall task of identifying evidence for questions from policies (§SECREF37). We describe aspects of the question that can render it unanswerable within the privacy domain (§SECREF41).\nExperimental Setup ::: Answerability Identification Baselines\nWe define answerability identification as a binary classification task, evaluating model ability to predict if a question can be answered, given a question in isolation. This can serve as a prior for downstream question-answering. We describe three baselines on the answerability task, and find they considerably improve performance over a majority-class baseline.\nSVM: We define 3 sets of features to characterize each question. The first is a simple bag-of-words set of features over the question (SVM-BOW), the second is bag-of-words features of the question as well as length of the question in words (SVM-BOW + LEN), and lastly we extract bag-of-words features, length of the question in words as well as part-of-speech tags for the question (SVM-BOW + LEN + POS). This results in vectors of 200, 201 and 228 dimensions respectively, which are provided to an SVM with a linear kernel.\nCNN: We utilize a CNN neural encoder for answerability prediction. We use GloVe word embeddings BIBREF50, and a filter size of 5 with 64 filters to encode questions.\nBERT: BERT BIBREF51 is a bidirectional transformer-based language-model BIBREF52. We fine-tune BERT-base on our binary answerability identification task with a learning rate of 2e-5 for 3 epochs, with a maximum sequence length of 128.\nExperimental Setup ::: Privacy Question Answering\nOur goal is to identify evidence within a privacy policy for questions asked by a user. This is framed as an answer sentence selection task, where models identify a set of evidence sentences from all candidate sentences in each policy.\nExperimental Setup ::: Privacy Question Answering ::: Evaluation Metric\nOur evaluation metric for answer-sentence selection is sentence-level F1, implemented similar to BIBREF30, BIBREF16. Precision and recall are implemented by measuring the overlap between predicted sentences and sets of gold-reference sentences. We report the average of the maximum F1 from each n$-$1 subset, in relation to the heldout reference.\nExperimental Setup ::: Privacy Question Answering ::: Baselines\nWe describe baselines on this task, including a human performance baseline.\nNo-Answer Baseline (NA) : Most of the questions we receive are difficult to answer in a legally-sound way on the basis of information present in the privacy policy. We establish a simple baseline to quantify the effect of identifying every question as unanswerable.\nWord Count Baseline : To quantify the effect of using simple lexical matching to answer the questions, we retrieve the top candidate policy sentences for each question using a word count baseline BIBREF53, which counts the number of question words that also appear in a sentence. We include the top 2, 3 and 5 candidates as baselines.\nBERT: We implement two BERT-based baselines BIBREF51 for evidence identification. First, we train BERT on each query-policy sentence pair as a binary classification task to identify if the sentence is evidence for the question or not (Bert). We also experiment with a two-stage classifier, where we separately train the model on questions only to predict answerability. At inference time, if the answerable classifier predicts the question is answerable, the evidence identification classifier produces a set of candidate sentences (Bert + Unanswerable).\nHuman Performance: We pick each reference answer provided by an annotator, and compute the F1 with respect to the remaining references, as described in section 4.2.1. Each reference answer is treated as the prediction, and the remaining n-1 answers are treated as the gold reference. The average of the maximum F1 across all reference answers is computed as the human baseline.\nResults and Discussion\nThe results of the answerability baselines are presented in Table TABREF31, and on answer sentence selection in Table TABREF32. We observe that bert exhibits the best performance on a binary answerability identification task. However, most baselines considerably exceed the performance of a majority-class baseline. This suggests considerable information in the question, indicating it's possible answerability within this domain.\nTable.TABREF32 describes the performance of our baselines on the answer sentence selection task. The No-answer (NA) baseline performs at 28 F1, providing a lower bound on performance at this task. We observe that our best-performing baseline, Bert + Unanswerable achieves an F1 of 39.8. This suggest that bert is capable of making some progress towards answering questions in this difficult domain, while still leaving considerable headroom for improvement to reach human performance. Bert + Unanswerable performance suggests that incorporating information about answerability can help in this difficult domain. We examine this challenging phenomena of unanswerability further in Section .\nResults and Discussion ::: Error Analysis\nDisagreements are analyzed based on the OPP-115 categories of each question (Table.TABREF34). We compare our best performing BERT variant against the NA model and human performance. We observe significant room for improvement across all categories of questions but especially for first party, third party and data retention categories.\nWe analyze the performance of our strongest BERT variant, to identify classes of errors and directions for future improvement (Table.8). We observe that a majority of answerability mistakes made by the BERT model are questions which are in fact answerable, but are identified as unanswerable by BERT. We observe that BERT makes 124 such mistakes on the test set. We collect expert judgments on relevance, subjectivity , silence and information about how likely the question is to be answered from the privacy policy from our experts. We find that most of these mistakes are relevant questions. However many of them were identified as subjective by the annotators, and at least one annotator marked 19 of these questions as having no answer within the privacy policy. However, only 6 of these questions were unexpected or do not usually have an answer in privacy policies. These findings suggest that a more nuanced understanding of answerability might help improve model performance in his challenging domain.\nResults and Discussion ::: What makes Questions Unanswerable?\nWe further ask legal experts to identify potential causes of unanswerability of questions. This analysis has considerable implications. While past work BIBREF17 has treated unanswerable questions as homogeneous, a question answering system might wish to have different treatments for different categories of `unanswerable' questions. The following factors were identified to play a role in unanswerability:\nIncomprehensibility: If a question is incomprehensible to the extent that its meaning is not intelligible.\nRelevance: Is this question in the scope of what could be answered by reading the privacy policy.\nIll-formedness: Is this question ambiguous or vague. An ambiguous statement will typically contain expressions that can refer to multiple potential explanations, whereas a vague statement carries a concept with an unclear or soft definition.\nSilence: Other policies answer this type of question but this one does not.\nAtypicality: The question is of a nature such that it is unlikely for any policy policy to have an answer to the question.\nOur experts attempt to identify the different `unanswerable' factors for all 573 such questions in the corpus. 4.18% of the questions were identified as being incomprehensible (for example, `any difficulties to occupy the privacy assistant'). Amongst the comprehendable questions, 50% were identified as likely to have an answer within the privacy policy, 33.1% were identified as being privacy-related questions but not within the scope of a privacy policy (e.g., 'has Viber had any privacy breaches in the past?') and 16.9% of questions were identified as completely out-of-scope (e.g., `'will the app consume much space?'). In the questions identified as relevant, 32% were ill-formed questions that were phrased by the user in a manner considered vague or ambiguous. Of the questions that were both relevant as well as `well-formed', 95.7% of the questions were not answered by the policy in question but it was reasonable to expect that a privacy policy would contain an answer. The remaining 4.3% were described as reasonable questions, but of a nature generally not discussed in privacy policies. This suggests that the answerability of questions over privacy policies is a complex issue, and future systems should consider each of these factors when serving user's information seeking intent.\nWe examine a large-scale dataset of “natural” unanswerable questions BIBREF54 based on real user search engine queries to identify if similar unanswerability factors exist. It is important to note that these questions have previously been filtered, according to a criteria for bad questions defined as “(questions that are) ambiguous, incomprehensible, dependent on clear false presuppositions, opinion-seeking, or not clearly a request for factual information.” Annotators made the decision based on the content of the question without viewing the equivalent Wikipedia page. We randomly sample 100 questions from the development set which were identified as unanswerable, and find that 20% of the questions are not questions (e.g., “all I want for christmas is you mariah carey tour”). 12% of questions are unlikely to ever contain an answer on Wikipedia, corresponding closely to our atypicality category. 3% of questions are unlikely to have an answer anywhere (e.g., `what guides Santa home after he has delivered presents?'). 7% of questions are incomplete or open-ended (e.g., `the south west wind blows across nigeria between'). 3% of questions have an unresolvable coreference (e.g., `how do i get to Warsaw Missouri from here'). 4% of questions are vague, and a further 7% have unknown sources of error. 2% still contain false presuppositions (e.g., `what is the only fruit that does not have seeds?') and the remaining 42% do not have an answer within the document. This reinforces our belief that though they have been understudied in past work, any question answering system interacting with real users should expect to receive such unanticipated and unanswerable questions.\nConclusion\nWe present PrivacyQA, the first significant corpus of privacy policy questions and more than 3500 expert annotations of relevant answers. The goal of this work is to promote question-answering research in the specialized privacy domain, where it can have large real-world impact. Strong neural baselines on PrivacyQA achieve a performance of only 39.8 F1 on this corpus, indicating considerable room for future research. Further, we shed light on several important considerations that affect the answerability of questions. We hope this contribution leads to multidisciplinary efforts to precisely understand user intent and reconcile it with information in policy documents, from both the privacy and NLP communities.\nAcknowledgements\nThis research was supported in part by grants from the National Science Foundation Secure and Trustworthy Computing program (CNS-1330596, CNS-1330214, CNS-15-13957, CNS-1801316, CNS-1914486, CNS-1914444) and a DARPA Brandeis grant on Personalized Privacy Assistants (FA8750-15-2-0277). The US Government is authorized to reproduce and distribute reprints for Governmental purposes not withstanding any copyright notation. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the NSF, DARPA, or the US Government. The authors would like to extend their gratitude to Elias Wright, Gian Mascioli, Kiara Pillay, Harrison Kay, Eliel Talo, Alexander Fagella and N. Cameron Russell for providing their valuable expertise and insight to this effort. The authors are also grateful to Eduard Hovy, Lorrie Cranor, Florian Schaub, Joel Reidenberg, Aditya Potukuchi and Igor Shalyminov for helpful discussions related to this work, and to the three anonymous reviewers of this draft for their constructive feedback. Finally, the authors would like to thank all crowdworkers who consented to participate in this study.", "answers": ["SVM, No-Answer Baseline (NA) , Word Count Baseline, Human Performance", "No-Answer Baseline (NA), Word Count Baseline, Human Performance"], "length": 3855, "dataset": "qasper", "language": "en", "all_classes": null, "_id": "e48405471fa8e3af3ddf67e7e9deb109523dfab05dc9d659"}
{"input": "In what language are the tweets?", "context": "Introduction\nOffensive content has become pervasive in social media and a reason of concern for government organizations, online communities, and social media platforms. One of the most common strategies to tackle the problem is to train systems capable of recognizing offensive content, which then can be deleted or set aside for human moderation. In the last few years, there have been several studies published on the application of computational methods to deal with this problem. Most prior work focuses on a different aspect of offensive language such as abusive language BIBREF0 , BIBREF1 , (cyber-)aggression BIBREF2 , (cyber-)bullying BIBREF3 , BIBREF4 , toxic comments INLINEFORM0 , hate speech BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , and offensive language BIBREF11 . Prior work has focused on these aspects of offensive language in Twitter BIBREF3 , BIBREF7 , BIBREF8 , BIBREF11 , Wikipedia comments, and Facebook posts BIBREF2 .\nRecently, Waseem et. al. ( BIBREF12 ) acknowledged the similarities among prior work and discussed the need for a typology that differentiates between whether the (abusive) language is directed towards a specific individual or entity or towards a generalized group and whether the abusive content is explicit or implicit. Wiegand et al. ( BIBREF11 ) followed this trend as well on German tweets. In their evaluation, they have a task to detect offensive vs not offensive tweets and a second task for distinguishing between the offensive tweets as profanity, insult, or abuse. However, no prior work has explored the target of the offensive language, which is important in many scenarios, e.g., when studying hate speech with respect to a specific target.\nTherefore, we expand on these ideas by proposing a a hierarchical three-level annotation model that encompasses:\nUsing this annotation model, we create a new large publicly available dataset of English tweets. The key contributions of this paper are as follows:\nRelated Work\nDifferent abusive and offense language identification sub-tasks have been explored in the past few years including aggression identification, bullying detection, hate speech, toxic comments, and offensive language.\nAggression identification: The TRAC shared task on Aggression Identification BIBREF2 provided participants with a dataset containing 15,000 annotated Facebook posts and comments in English and Hindi for training and validation. For testing, two different sets, one from Facebook and one from Twitter were provided. Systems were trained to discriminate between three classes: non-aggressive, covertly aggressive, and overtly aggressive.\nBullying detection: Several studies have been published on bullying detection. One of them is the one by xu2012learning which apply sentiment analysis to detect bullying in tweets. xu2012learning use topic models to to identify relevant topics in bullying. Another related study is the one by dadvar2013improving which use user-related features such as the frequency of profanity in previous messages to improve bullying detection.\nHate speech identification: It is perhaps the most widespread abusive language detection sub-task. There have been several studies published on this sub-task such as kwok2013locate and djuric2015hate who build a binary classifier to distinguish between `clean' comments and comments containing hate speech and profanity. More recently, Davidson et al. davidson2017automated presented the hate speech detection dataset containing over 24,000 English tweets labeled as non offensive, hate speech, and profanity.\nOffensive language: The GermEval BIBREF11 shared task focused on Offensive language identification in German tweets. A dataset of over 8,500 annotated tweets was provided for a course-grained binary classification task in which systems were trained to discriminate between offensive and non-offensive tweets and a second task where the organizers broke down the offensive class into three classes: profanity, insult, and abuse.\nToxic comments: The Toxic Comment Classification Challenge was an open competition at Kaggle which provided participants with comments from Wikipedia labeled in six classes: toxic, severe toxic, obscene, threat, insult, identity hate.\nWhile each of these sub-tasks tackle a particular type of abuse or offense, they share similar properties and the hierarchical annotation model proposed in this paper aims to capture this. Considering that, for example, an insult targeted at an individual is commonly known as cyberbulling and that insults targeted at a group are known as hate speech, we pose that OLID's hierarchical annotation model makes it a useful resource for various offensive language identification sub-tasks.\nHierarchically Modelling Offensive Content\nIn the OLID dataset, we use a hierarchical annotation model split into three levels to distinguish between whether language is offensive or not (A), and type (B) and target (C) of the offensive language. Each level is described in more detail in the following subsections and examples are shown in Table TABREF10 .\nLevel A: Offensive language Detection\nLevel A discriminates between offensive (OFF) and non-offensive (NOT) tweets.\nNot Offensive (NOT): Posts that do not contain offense or profanity;\nOffensive (OFF): We label a post as offensive if it contains any form of non-acceptable language (profanity) or a targeted offense, which can be veiled or direct. This category includes insults, threats, and posts containing profane language or swear words.\nLevel B: Categorization of Offensive Language\nLevel B categorizes the type of offense and two labels are used: targeted (TIN) and untargeted (INT) insults and threats.\nTargeted Insult (TIN): Posts which contain an insult/threat to an individual, group, or others (see next layer);\nUntargeted (UNT): Posts containing non-targeted profanity and swearing. Posts with general profanity are not targeted, but they contain non-acceptable language.\nLevel C: Offensive Language Target Identification\nLevel C categorizes the targets of insults and threats as individual (IND), group (GRP), and other (OTH).\nIndividual (IND): Posts targeting an individual. It can be a a famous person, a named individual or an unnamed participant in the conversation. Insults and threats targeted at individuals are often defined as cyberbulling.\nGroup (GRP): The target of these offensive posts is a group of people considered as a unity due to the same ethnicity, gender or sexual orientation, political affiliation, religious belief, or other common characteristic. Many of the insults and threats targeted at a group correspond to what is commonly understood as hate speech.\nOther (OTH): The target of these offensive posts does not belong to any of the previous two categories (e.g. an organization, a situation, an event, or an issue).\nData Collection\nThe data included in OLID has been collected from Twitter. We retrieved the data using the Twitter API by searching for keywords and constructions that are often included in offensive messages, such as `she is' or `to:BreitBartNews'. We carried out a first round of trial annotation of 300 instances with six experts. The goal of the trial annotation was to 1) evaluate the proposed tagset; 2) evaluate the data retrieval method; and 3) create a gold standard with instances that could be used as test questions in the training and test setting annotation which was carried out using crowdsourcing. The breakdown of keywords and their offensive content in the trial data of 300 tweets is shown in Table TABREF14 . We included a left (@NewYorker) and far-right (@BreitBartNews) news accounts because there tends to be political offense in the comments. One of the best offensive keywords was tweets that were flagged as not being safe by the Twitter `safe' filter (the `-' indicates `not safe'). The vast majority of content on Twitter is not offensive so we tried different strategies to keep a reasonable number of tweets in the offensive class amounting to around 30% of the dataset including excluding some keywords that were not high in offensive content such as `they are` and `to:NewYorker`. Although `he is' is lower in offensive content we kept it as a keyword to avoid gender bias. In addition to the keywords in the trial set, we searched for more political keywords which tend to be higher in offensive content, and sampled our dataset such that 50% of the the tweets come from political keywords and 50% come from non-political keywords. In addition to the keywords `gun control', and `to:BreitbartNews', political keywords used to collect these tweets are `MAGA', `antifa', `conservative' and `liberal'. We computed Fliess' INLINEFORM0 on the trial set for the five annotators on 21 of the tweets. INLINEFORM1 is .83 for Layer A (OFF vs NOT) indicating high agreement. As to normalization and anonymization, no user metadata or Twitter IDs have been stored, and URLs and Twitter mentions have been substituted to placeholders. We follow prior work in related areas (burnap2015cyber,davidson2017automated) and annotate our data using crowdsourcing using the platform Figure Eight. We ensure data quality by: 1) we only received annotations from individuals who were experienced in the platform; and 2) we used test questions to discard annotations of individuals who did not reach a certain threshold. Each instance in the dataset was annotated by multiple annotators and inter-annotator agreement has been calculated. We first acquired two annotations for each instance. In case of 100% agreement, we considered these as acceptable annotations, and in case of disagreement, we requested more annotations until the agreement was above 66%. After the crowdsourcing annotation, we used expert adjudication to guarantee the quality of the annotation. The breakdown of the data into training and testing for the labels from each level is shown in Table TABREF15 .\nExperiments and Evaluation\nWe assess our dataset using traditional and deep learning methods. Our simplest model is a linear SVM trained on word unigrams. SVMs have produced state-of-the-art results for many text classification tasks BIBREF13 . We also train a bidirectional Long Short-Term-Memory (BiLSTM) model, which we adapted from the sentiment analysis system of sentimentSystem,rasooli2018cross and altered to predict offensive labels instead. It consists of (1) an input embedding layer, (2) a bidirectional LSTM layer, (3) an average pooling layer of input features. The concatenation of the LSTM's and average pool layer is passed through a dense layer and the output is passed through a softmax function. We set two input channels for the input embedding layers: pre-trained FastText embeddings BIBREF14 , as well as updatable embeddings learned by the model during training. Finally, we also apply a Convolutional Neural Network (CNN) model based on the architecture of BIBREF15 , using the same multi-channel inputs as the above BiLSTM.\nOur models are trained on the training data, and evaluated by predicting the labels for the held-out test set. The distribution is described in Table TABREF15 . We evaluate and compare the models using the macro-averaged F1-score as the label distribution is highly imbalanced. Per-class Precision (P), Recall (R), and F1-score (F1), also with other averaged metrics are also reported. The models are compared against baselines of predicting all labels as the majority or minority classes.\nOffensive Language Detection\nThe performance on discriminating between offensive (OFF) and non-offensive (NOT) posts is reported in Table TABREF18 . We can see that all systems perform significantly better than chance, with the neural models being substantially better than the SVM. The CNN outperforms the RNN model, achieving a macro-F1 score of 0.80.\nCategorization of Offensive Language\nIn this experiment, the two systems were trained to discriminate between insults and threats (TIN) and untargeted (UNT) offenses, which generally refer to profanity. The results are shown in Table TABREF19 .\nThe CNN system achieved higher performance in this experiment compared to the BiLSTM, with a macro-F1 score of 0.69. All systems performed better at identifying target and threats (TIN) than untargeted offenses (UNT).\nOffensive Language Target Identification\nThe results of the offensive target identification experiment are reported in Table TABREF20 . Here the systems were trained to distinguish between three targets: a group (GRP), an individual (IND), or others (OTH). All three models achieved similar results far surpassing the random baselines, with a slight performance edge for the neural models.\nThe performance of all systems for the OTH class is 0. This poor performances can be explained by two main factors. First, unlike the two other classes, OTH is a heterogeneous collection of targets. It includes offensive tweets targeted at organizations, situations, events, etc. making it more challenging for systems to learn discriminative properties of this class. Second, this class contains fewer training instances than the other two. There are only 395 instances in OTH, and 1,075 in GRP, and 2,407 in IND.\nConclusion and Future Work\nThis paper presents OLID, a new dataset with annotation of type and target of offensive language. OLID is the official dataset of the shared task SemEval 2019 Task 6: Identifying and Categorizing Offensive Language in Social Media (OffensEval) BIBREF16 . In OffensEval, each annotation level in OLID is an independent sub-task. The dataset contains 14,100 tweets and is released freely to the research community. To the best of our knowledge, this is the first dataset to contain annotation of type and target of offenses in social media and it opens several new avenues for research in this area. We present baseline experiments using SVMs and neural networks to identify the offensive tweets, discriminate between insults, threats, and profanity, and finally to identify the target of the offensive messages. The results show that this is a challenging task. A CNN-based sentence classifier achieved the best results in all three sub-tasks.\nIn future work, we would like to make a cross-corpus comparison of OLID and datasets annotated for similar tasks such as aggression identification BIBREF2 and hate speech detection BIBREF8 . This comparison is, however, far from trivial as the annotation of OLID is different.\nAcknowledgments\nThe research presented in this paper was partially supported by an ERAS fellowship awarded by the University of Wolverhampton.", "answers": ["English", "English ", "English"], "length": 2240, "dataset": "qasper", "language": "en", "all_classes": null, "_id": "85a7083bae278e83df0fea37003937a9294638f488a18a83"}