BERT-based architectures currently give state-of-the-art performance on many NLP tasks, but little is known about the exact mechanisms that contribute to its success. In the current work, we focus on the interpretation of self-attention, which is one of the fundamental underlying components of BERT. Using a subset of GLUE tasks and a set of handcrafted features-of-interest, we propose the methodology and carry out a qualitative and quantitative analysis of the information encoded by the individual BERT’s heads. Our findings suggest that there is a limited set of attention patterns that are repeated across different heads, indicating the overall model overparametrization. While different heads consistently use the same attention patterns, they have varying impact on performance across different tasks. We show that manually disabling attention in certain heads leads to a performance improvement over the regular fine-tuned BERT models.

In this paper, we present a method for adversarial decomposition of text representation. This method can be used to decompose a representation of an input sentence into several independent vectors, each of them responsible for a specific aspect of the input sentence. We evaluate the proposed method on two case studies: the conversion between different social registers and diachronic language change. We show that the proposed method is capable of fine-grained controlled change of these aspects of the input sentence. It is also learning a continuous (rather than categorical) representation of the style of the sentence, which is more linguistically realistic. The model uses adversarial-motivational training and includes a special motivational loss, which acts opposite to the discriminator and encourages a better decomposition. Furthermore, we evaluate the obtained meaning embeddings on a downstream task of paraphrase detection and show that they significantly outperform the embeddings of a regular autoencoder.

There is a growing body of work that proposes methods for mitigating bias in machine learning systems. These methods typically rely on access to protected attributes such as race, gender, or age. However, this raises two significant challenges: (1) protected attributes may not be available or it may not be legal to use them, and (2) it is often desirable to simultaneously consider multiple protected attributes, as well as their intersections. In the context of mitigating bias in occupation classification, we propose a method for discouraging correlation between the predicted probability of an individual’s true occupation and a word embedding of their name. This method leverages the societal biases that are encoded in word embeddings, eliminating the need for access to protected attributes. Crucially, it only requires access to individuals’ names at training time and not at deployment time. We evaluate two variations of our proposed method using a large-scale dataset of online biographies. We find that both variations simultaneously reduce race and gender biases, with almost no reduction in the classifier’s overall true positive rate.

State of the art models using deep neural networks have become very good in learning an accurate mapping from inputs to outputs. However, they still lack generalization capabilities in conditions that differ from the ones encountered during training. This is even more challenging in specialized, and knowledge intensive domains, where training data is limited. To address this gap, we introduce MedNLI - a dataset annotated by doctors, performing a natural language inference task (NLI), grounded in the medical history of patients. We present strategies to: 1) leverage transfer learning using datasets from the open domain, (e.g. SNLI) and 2) incorporate domain knowledge from external data and lexical sources (e.g. medical terminologies). Our results demonstrate performance gains using both strategies.

This paper addresses the problem of representation learning. Using an autoencoder framework, we propose and evaluate several loss functions that can be used as an alternative to the commonly used cross-entropy reconstruction loss. The proposed loss functions use similarities between words in the embedding space, and can be used to train any neural model for text generation. We show that the introduced loss functions amplify semantic diversity of reconstructed sentences, while preserving the original meaning of the input. We test the derived autoencoder-generated representations on paraphrase detection and language inference tasks and demonstrate performance improvement compared to the traditional cross-entropy loss.

This paper presents RuSentiment, a new dataset for sentiment analysis of social media posts in Russian, and a new set of comprehensive annotation guidelines that are extensible to other languages. RuSentiment is currently the largest in its class for Russian, with 31,185 posts annotated with Fleiss’ kappa of 0.58 (3 annotations per post). To diversify the dataset, 6,950 posts were pre-selected with an active learning-style strategy. We report baseline classification results, and we also release the best-performing embeddings trained on 3.2B tokens of Russian VKontakte posts.

Language generation tasks that seek to mimic human ability to use language creatively are difficult to evaluate, since one must consider creativity, style, and other non-trivial aspects of the generated text. The goal of this paper is to develop evaluations methods for one such task, ghostwriting of rap lyrics, and to provide an explicit, quantifiable foundation for the goals and future directions for this task. Ghostwriting must produce text that is similar in style to the emulated artist, yet distinct in content. We develop a novel evaluation methodology that addresses several complementary aspects of this task, and illustrate how such evaluation can be used to meaning fully analyze system performance. We provide a corpus of lyrics for 13 rap artists, annotated for stylistic similarity, which allows us to assess the feasibility of manual evaluation for generated verse.

This paper describes a new shared task for humor understanding that attempts to eschew the ubiquitous binary approach to humor detection and focus on comparative humor ranking instead. The task is based on a new dataset of funny tweets posted in response to shared hashtags, collected from the ‘Hashtag Wars’ segment of the TV show @midnight. The results are evaluated in two subtasks that require the participants to generate either the correct pairwise comparisons of tweets (subtask A), or the correct ranking of the tweets (subtask B) in terms of how funny they are. 7 teams participated in subtask A, and 5 teams participated in subtask B. The best accuracy in subtask A was 0.675. The best (lowest) rank edit distance for subtask B was 0.872.

This paper describes the winning system for SemEval-2017 Task 6: #HashtagWars: Learning a Sense of Humor. Humor detection has up until now been predominantly addressed using feature-based approaches. Our system utilizes recurrent deep learning methods with dense embeddings to predict humorous tweets from the @midnight show #HashtagWars. In order to include both meaning and sound in the analysis, GloVe embeddings are combined with a novel phonetic representation to serve as input to an LSTM component. The output is combined with a character-based CNN model, and an XGBoost component in an ensemble model which achieves 0.675 accuracy on the evaluation data.

In this paper, we propose to use a set of simple, uniform in architecture LSTM-based models to recover different kinds of temporal relations from text. Using the shortest dependency path between entities as input, the same architecture is used to extract intra-sentence, cross-sentence, and document creation time relations. A “double-checking” technique reverses entity pairs in classification, boosting the recall of positive cases and reducing misclassifications between opposite classes. An efficient pruning algorithm resolves conflicts globally. Evaluated on QA-TempEval (SemEval2015 Task 5), our proposed technique outperforms state-of-the-art methods by a large margin. We also conduct intrinsic evaluation and post state-of-the-art results on Timebank-Dense.

In order to determine argument structure in text, one must understand how individual components of the overall argument are linked. This work presents the first neural network-based approach to link extraction in argument mining. Specifically, we propose a novel architecture that applies Pointer Network sequence-to-sequence attention modeling to structural prediction in discourse parsing tasks. We then develop a joint model that extends this architecture to simultaneously address the link extraction task and the classification of argument components. The proposed joint model achieves state-of-the-art results on two separate evaluation corpora, showing far superior performance than the previously proposed corpus-specific and heavily feature-engineered models. Furthermore, our results demonstrate that jointly optimizing for both tasks is crucial for high performance.

This paper addresses the task of identifying the bias in news articles published during a political or social conflict. We create a silver-standard corpus based on the actions of users in social media. Specifically, we reconceptualize bias in terms of how likely a given article is to be shared or liked by each of the opposing sides. We apply our methodology to a dataset of links collected in relation to the Russia-Ukraine Maidan crisis from 2013-2014. We show that on the task of predicting which side is likely to prefer a given article, a Naive Bayes classifier can record 90.3% accuracy looking only at domain names of the news sources. The best accuracy of 93.5% is achieved by a feed forward neural network. We also apply our methodology to gold-labeled set of articles annotated for bias, where the aforementioned Naive Bayes classifier records 82.6% accuracy and a feed-forward neural networks records 85.6% accuracy.