We propose the task of emotion style transfer, which is particularly challenging, as emotions (here: anger, disgust, fear, joy, sadness, surprise) are on the fence between content and style. To understand the particular difficulties of this task, we design a transparent emotion style transfer pipeline based on three steps: (1) select the words that are promising to be substituted to change the emotion (with a brute-force approach and selection based on the attention mechanism of an emotion classifier), (2) find sets of words as candidates for substituting the words (based on lexical and distributional semantics), and (3) select the most promising combination of substitutions with an objective function which consists of components for content (based on BERT sentence embeddings), emotion (based on an emotion classifier), and fluency (based on a neural language model). This comparably straight-forward setup enables us to explore the task and understand in what cases lexical substitution can vary the emotional load of texts, how changes in content and style interact and if they are at odds. We further evaluate our pipeline quantitatively in an automated and an annotation study based on Tweets and find, indeed, that simultaneous adjustments of content and emotion are conflicting objectives: as we show in a qualitative analysis motivated by Scherer's emotion component model, this is particularly the case for implicit emotion expressions based on cognitive appraisal or descriptions of bodily reactions.
Few works in the literature of event extraction have gone beyond individual sentences to make extraction decisions. This is problematic when the information needed to recognize an event argument is spread across multiple sentences. We argue that document-level event extraction is a difficult task since it requires a view of a larger context to determine which spans of text correspond to event role fillers. We first investigate how end-to-end neural sequence models (with pre-trained language model representations) perform on document-level role filler extraction, as well as how the length of context captured affects the models' performance. To dynamically aggregate information captured by neural representations learned at different levels of granularity (e.g., the sentence- and paragraph-level), we propose a novel multi-granularity reader. We evaluate our models on the MUC-4 event extraction dataset, and show that our best system performs substantially better than prior work. We also report findings on the relationship between context length and neural model performance on the task.
Hate speech detection is a critical problem in social media platforms, being often accused for enabling the spread of hatred and igniting physical violence. Hate speech detection requires overwhelming resources including high-performance computing for online posts and tweets monitoring as well as thousands of human experts for daily screening of suspected posts or tweets. Recently, Deep Learning (DL)-based solutions have been proposed for automatic detection of hate speech, using modest-sized training datasets of few thousands of hate speech sequences. While these methods perform well on the specific datasets, their ability to detect new hate speech sequences is limited and has not been investigated. Being a data-driven approach, it is well known that DL surpasses other methods whenever a scale-up in train dataset size and diversity is achieved. Therefore, we first present a dataset of 1 million realistic hate and non-hate sequences, produced by a deep generative language model. We further utilize the generated dataset to train a well-studied DL-based hate speech detector, and demonstrate consistent and significant performance improvements across five public hate speech datasets. Therefore, the proposed solution enables high sensitivity detection of a very large variety of hate speech sequences, paving the way to a fully automatic solution.
We present a simple approach for text infilling, the task of predicting missing spans of text at any position in a document. While infilling could enable rich functionality especially for writing assistance tools, more attention has been devoted to language modeling---a special case of infilling where text is predicted at the end of a document. In this paper, we aim to extend the capabilities of language models (LMs) to the more general task of infilling. To this end, we train (or fine-tune) off-the-shelf LMs on sequences containing the concatenation of artificially-masked text and the text which was masked. We show that this approach, which we call infilling by language modeling, can enable LMs to infill entire sentences effectively on three different domains: short stories, scientific abstracts, and lyrics. Furthermore, we show that humans have difficulty identifying sentences infilled by our approach as machine-generated in the domain of short stories.
It is commonly believed that knowledge of syntactic structure should improve language modeling. However, effectively and computationally efficiently incorporating syntactic structure into neural language models has been a challenging topic. In this paper, we make use of a multi-task objective, i.e., the models simultaneously predict words as well as ground truth parse trees in a form called "syntactic distances", where information between these two separate objectives shares the same intermediate representation. Experimental results on the Penn Treebank and Chinese Treebank datasets show that when ground truth parse trees are provided as additional training signals, the model is able to achieve lower perplexity and induce trees with better quality.
Neural language models are becoming the prevailing methodology for the tasks of query answering, text classification, disambiguation, completion and translation. Commonly comprised of hundreds of millions of parameters, these neural network models offer state-of-the-art performance at the cost of interpretability; humans are no longer capable of tracing and understanding how decisions are being made. The attention mechanism, introduced initially for the task of translation, has been successfully adopted for other language-related tasks. We propose AttViz, an online toolkit for exploration of self-attention---real values associated with individual text tokens. We show how existing deep learning pipelines can produce outputs suitable for AttViz, offering novel visualizations of the attention heads and their aggregations with minimal effort, online. We show on examples of news segments how the proposed system can be used to inspect and potentially better understand what a model has learned (or emphasized).
Recent work has found evidence that Multilingual BERT (mBERT), a transformer-based multilingual masked language model, is capable of zero-shot cross-lingual transfer, suggesting that some aspects of its representations are shared cross-lingually. To better understand this overlap, we extend recent work on finding syntactic trees in neural networks' internal representations to the multilingual setting. We show that subspaces of mBERT representations recover syntactic tree distances in languages other than English, and that these subspaces are approximately shared across languages. Motivated by these results, we present an unsupervised analysis method that provides evidence mBERT learns representations of syntactic dependency labels, in the form of clusters which largely agree with the Universal Dependencies taxonomy. This evidence suggests that even without explicit supervision, multilingual masked language models learn certain linguistic universals.
We propose a sentence-level language model which selects the next sentence in a story from a finite set of fluent alternatives. Since it does not need to model fluency, the sentence-level language model can focus on longer range dependencies, which are crucial for multi-sentence coherence. Rather than dealing with individual words, our method treats the story so far as a list of pre-trained sentence embeddings and predicts an embedding for the next sentence, which is more efficient than predicting word embeddings. Notably this allows us to consider a large number of candidates for the next sentence during training. We demonstrate the effectiveness of our approach with state-of-the-art accuracy on the unsupervised Story Cloze task and with promising results on larger-scale next sentence prediction tasks.
While state-of-the-art neural network models continue to achieve lower perplexity scores on language modeling benchmarks, it remains unknown whether optimizing for broad-coverage predictive performance leads to human-like syntactic knowledge. Furthermore, existing work has not provided a clear picture about the model properties required to produce proper syntactic generalizations. We present a systematic evaluation of the syntactic knowledge of neural language models, testing 20 combinations of model types and data sizes on a set of 34 English-language syntactic test suites. We find substantial differences in syntactic generalization performance by model architecture, with sequential models underperforming other architectures. Factorially manipulating model architecture and training dataset size (1M--40M words), we find that variability in syntactic generalization performance is substantially greater by architecture than by dataset size for the corpora tested in our experiments. Our results also reveal a dissociation between perplexity and syntactic generalization performance.
Automatic sentence summarization produces a shorter version of a sentence, while preserving its most important information. A good summary is characterized by language fluency and high information overlap with the source sentence. We model these two aspects in an unsupervised objective function, consisting of language modeling and semantic similarity metrics. We search for a high-scoring summary by discrete optimization. Our proposed method achieves a new state-of-the art for unsupervised sentence summarization according to ROUGE scores. Additionally, we demonstrate that the commonly reported ROUGE F1 metric is sensitive to summary length. Since this is unwillingly exploited in recent work, we emphasize that future evaluation should explicitly group summarization systems by output length brackets.
Generating coherent, grammatically correct, and meaningful text is very challenging, however, it is crucial to many modern NLP systems. So far, research has mostly focused on English language, for other languages both standardized datasets, as well as experiments with state-of-the-art models, are rare. In this work, we i) provide a novel reference dataset for Russian language modeling, ii) experiment with popular modern methods for text generation, namely variational autoencoders, and generative adversarial networks, which we trained on the new dataset. We evaluate the generated text regarding metrics such as perplexity, grammatical correctness and lexical diversity.
Functional Distributional Semantics provides a linguistically interpretable framework for distributional semantics, by representing the meaning of a word as a function (a binary classifier), instead of a vector. However, the large number of latent variables means that inference is computationally expensive, and training a model is therefore slow to converge. In this paper, I introduce the Pixie Autoencoder, which augments the generative model of Functional Distributional Semantics with a graph-convolutional neural network to perform amortised variational inference. This allows the model to be trained more effectively, achieving better results on two tasks (semantic similarity in context and semantic composition), and outperforming BERT, a large pre-trained language model.
MaskGAN opens the query for the conditional language model by filling in the blanks between the given tokens. In this paper, we focus on addressing the limitations caused by having to specify blanks to be filled. We decompose conditional text generation problem into two tasks, make-a-blank and fill-in-the-blank, and extend the former to handle more complex manipulations on the given tokens. We cast these tasks as a hierarchical multi agent RL problem and introduce a conditional adversarial learning that allows the agents to reach a goal, producing realistic texts, in cooperative setting. We show that the proposed model not only addresses the limitations but also provides good results without compromising the performance in terms of quality and diversity.
Large-scale pre-trained language models, such as BERT and GPT-2, have achieved excellent performance in language representation learning and free-form text generation. However, these models cannot be directly employed to generate text under specified lexical constraints. To address this challenge, we present POINTER (PrOgressive INsertion-based TransformER), a simple yet novel insertion-based approach for hard-constrained text generation. The proposed method operates by progressively inserting new tokens between existing tokens in a parallel manner. This procedure is recursively applied until a sequence is completed. The resulting coarse-to-fine hierarchy makes the generation process intuitive and interpretable. We pre-train our model with the proposed progressive insertion-based objective on a 12GB Wikipedia dataset, and fine-tune it on downstream hard-constrained generation tasks. Non-autoregressive decoding yields an empirically logarithmic time complexity during inference time. Experimental results on both News and Yelp datasets demonstrate that POINTER achieves state-of-the-art performance on constrained text generation. We released the pre-trained models and the source code to facilitate future research (https://github.com/dreasysnail/POINTER).
Task-oriented dialogue is often decomposed into three tasks: understanding user input, deciding actions, and generating a response. While such decomposition might suggest a dedicated model for each sub-task, we find a simple, unified approach leads to state-of-the-art performance on the MultiWOZ dataset. SimpleTOD is a simple approach to task-oriented dialogue that uses a single, causal language model trained on all sub-tasks recast as a single sequence prediction problem. This allows SimpleTOD to fully leverage transfer learning from pre-trained, open domain, causal language models such as GPT-2. SimpleTOD improves over the prior state-of-the-art in joint goal accuracy for dialogue state tracking, and our analysis reveals robustness to noisy annotations in this setting. SimpleTOD also improves the main metrics used to evaluate action decisions and response generation in an end-to-end setting: inform rate by 8.1 points, success rate by 9.7 points, and combined score by 7.2 points.
Humans acquire language continually with much more limited access to data samples at a time, as compared to contemporary NLP systems. To study this human-like language acquisition ability, we present VisCOLL, a visually grounded language learning task, which simulates the continual acquisition of compositional phrases from streaming visual scenes. In the task, models are trained on a paired image-caption stream which has shifting object distribution; while being constantly evaluated by a visually-grounded masked language prediction task on held-out test sets. VisCOLL compounds the challenges of continual learning (i.e., learning from continuously shifting data distribution) and compositional generalization (i.e., generalizing to novel compositions). To facilitate research on VisCOLL, we construct two datasets, COCO-shift and Flickr-shift, and benchmark them using different continual learning methods. Results reveal that SoTA continual learning approaches provide little to no improvements on VisCOLL, since storing examples of all possible compositions is infeasible. We conduct further ablations and analysis to guide future work.
Recent advances in NLP demonstrate the effectiveness of training large-scale language models and transferring them to downstream tasks. Can fine-tuning these models on tasks other than language modeling further improve performance? In this paper, we conduct an extensive study of the transferability between 33 NLP tasks across three broad classes of problems (text classification, question answering, and sequence labeling). Our results show that transfer learning is more beneficial than previously thought, especially when target task data is scarce, and can improve performance even when the source task is small or differs substantially from the target task (e.g., part-of-speech tagging transfers well to the DROP QA dataset). We also develop task embeddings that can be used to predict the most transferable source tasks for a given target task, and we validate their effectiveness in experiments controlled for source and target data size. Overall, our experiments reveal that factors such as source data size, task and domain similarity, and task complexity all play a role in determining transferability.
Khandelwal et al. (2020) use a k-nearest-neighbor (kNN) component to improve language model performance. We show that this idea is beneficial for open-domain question answering (QA). To improve the recall of facts encountered during training, we combine BERT (Devlin et al., 2019) with a traditional information retrieval step (IR) and a kNN search over a large datastore of an embedded text collection. Our contributions are as follows: i) BERT-kNN outperforms BERT on cloze-style QA by large margins without any further training. ii) We show that BERT often identifies the correct response category (e.g., US city), but only kNN recovers the factually correct answer (e.g., "Miami"). iii) Compared to BERT, BERT-kNN excels for rare facts. iv) BERT-kNN can easily handle facts not covered by BERT's training set, e.g., recent events.
Question answering (QA) tasks have been posed using a variety of formats, such as extractive span selection, multiple choice, etc. This has led to format-specialized models, and even to an implicit division in the QA community. We argue that such boundaries are artificial and perhaps unnecessary, given the reasoning abilities we seek to teach are not governed by the format. As evidence, we use the latest advances in language modeling to build a single pre-trained QA model, UnifiedQA, that performs surprisingly well across 17 QA datasets spanning 4 diverse formats. UnifiedQA performs on par with 9 different models that were trained on individual datasets themselves. Even when faced with 12 unseen datasets of observed formats, UnifiedQA performs surprisingly well, showing strong generalization from its out-of-format training data. Finally, simply fine-tuning this pre-trained QA model into specialized models results in a new state of the art on 6 datasets, establishing UnifiedQA as a strong starting point for building QA systems.
Commonsense question answering (QA) requires background knowledge which is not explicitly stated in a given context. Prior works use commonsense knowledge graphs (KGs) to obtain this knowledge for reasoning. However, relying entirely on these KGs may not suffice, considering their limited coverage and the contextual dependence of their knowledge. In this paper, we augment a general commonsense QA framework with a knowledgeable path generator. By extrapolating over existing paths in a KG with a state-of-the-art language model, our generator learns to connect a pair of entities in text with a dynamic, and potentially novel, multi-hop relational path. Such paths can provide structured evidence for solving commonsense questions without fine-tuning the path generator. Experiments on two datasets show the superiority of our method over previous works which fully rely on knowledge from KGs (with up to 6% improvement in accuracy), across various amounts of training data. Further evaluation suggests that the generated paths are typically interpretable, novel, and relevant to the task.