EMNLP 2019 Papers

Selective rationalization has become a common mechanism to ensure that predictive models reveal how they use any available features. The selection may be soft or hard, and identifies a subset of input features relevant for prediction. The setup can be viewed as a co-operate game between the selector (aka rationale generator) and the predictor making use of only the selected features. The co-operative setting may, however, be compromised for two reasons. First, the generator typically has no direct access to the outcome it aims to justify, resulting in poor performance. Second, there’s typically no control exerted on the information left outside the selection. We revise the overall co-operative framework to address these challenges. We introduce an introspective model which explicitly predicts and incorporates the outcome into the selection process. Moreover, we explicitly control the rationale complement via an adversary so as not to leave any useful information out of the selection. We show that the two complementary mechanisms maintain both high predictive accuracy and lead to comprehensive rationales.

Neural networks are part of many contemporary NLP systems, yet their empirical successes come at the price of vulnerability to adversarial attacks. Previous work has used adversarial training and data augmentation to partially mitigate such brittleness, but these are unlikely to find worst-case adversaries due to the complexity of the search space arising from discrete text perturbations. In this work, we approach the problem from the opposite direction: to formally verify a system’s robustness against a predefined class of adversarial attacks. We study text classification under synonym replacements or character flip perturbations. We propose modeling these input perturbations as a simplex and then using Interval Bound Propagation – a formal model verification method. We modify the conventional log-likelihood training objective to train models that can be efficiently verified, which would otherwise come with exponential search complexity. The resulting models show only little difference in terms of nominal accuracy, but have much improved verified accuracy under perturbations and come with an efficiently computable formal guarantee on worst case adversaries.

State-of-the-art models often make use of superficial patterns in the data that do not generalize well to out-of-domain or adversarial settings. For example, textual entailment models often learn that particular key words imply entailment, irrespective of context, and visual question answering models learn to predict prototypical answers, without considering evidence in the image. In this paper, we show that if we have prior knowledge of such biases, we can train a model to be more robust to domain shift. Our method has two stages: we (1) train a naive model that makes predictions exclusively based on dataset biases, and (2) train a robust model as part of an ensemble with the naive one in order to encourage it to focus on other patterns in the data that are more likely to generalize. Experiments on five datasets with out-of-domain test sets show significantly improved robustness in all settings, including a 12 point gain on a changing priors visual question answering dataset and a 9 point gain on an adversarial question answering test set.

The exploding cost and time needed for data labeling and model training are bottlenecks for training DNN models on large datasets. Identifying smaller representative data samples with strategies like active learning can help mitigate such bottlenecks. Previous works on active learning in NLP identify the problem of sampling bias in the samples acquired by uncertainty-based querying and develop costly approaches to address it. Using a large empirical study, we demonstrate that active set selection using the posterior entropy of deep models like FastText.zip (FTZ) is robust to sampling biases and to various algorithmic choices (query size and strategies) unlike that suggested by traditional literature. We also show that FTZ based query strategy produces sample sets similar to those from more sophisticated approaches (e.g ensemble networks). Finally, we show the effectiveness of the selected samples by creating tiny high-quality datasets, and utilizing them for fast and cheap training of large models. Based on the above, we propose a simple baseline for deep active text classification that outperforms the state of the art. We expect the presented work to be useful and informative for dataset compression and for problems involving active, semi-supervised or online learning scenarios. Code and models are available at: https://github.com/drimpossible/Sampling-Bias-Active-Learning.

While broadly applicable to many natural language processing (NLP) tasks, variational autoencoders (VAEs) are hard to train due to the posterior collapse issue where the latent variable fails to encode the input data effectively. Various approaches have been proposed to alleviate this problem to improve the capability of the VAE. In this paper, we propose to introduce a mutual information (MI) term between the input and its latent variable to regularize the objective of the VAE. Since estimating the MI in the high-dimensional space is intractable, we employ neural networks for the estimation of the MI and provide a training algorithm based on the convex duality approach. Our experimental results on three benchmark datasets demonstrate that the proposed model, compared to the state-of-the-art baselines, exhibits less posterior collapse and has comparable or better performance in language modeling and text generation. We also qualitatively evaluate the inferred latent space and show that the proposed model can generate more reasonable and diverse sentences via linear interpolation in the latent space.

While NLP systems become more pervasive, their accountability gains value as a focal point of effort. Epistemological opaqueness of nonlinear learning methods, such as deep learning models, can be a major drawback for their adoptions. In this paper, we discuss the application of Layerwise Relevance Propagation over a linguistically motivated neural architecture, the Kernel-based Deep Architecture, in order to trace back connections between linguistic properties of input instances and system decisions. Such connections then guide the construction of argumentations on network’s inferences, i.e., explanations based on real examples, semantically related to the input. We propose here a methodology to evaluate the transparency and coherence of analogy-based explanations modeling an audit stage for the system. Quantitative analysis on two semantic tasks, i.e., question classification and semantic role labeling, show that the explanatory capabilities (native in KDAs) are effective and they pave the way to more complex argumentation methods.

Recently, kernelized locality sensitive hashcodes have been successfully employed as representations of natural language text, especially showing high relevance to biomedical relation extraction tasks. In this paper, we propose to optimize the hashcode representations in a nearly unsupervised manner, in which we only use data points, but not their class labels, for learning. The optimized hashcode representations are then fed to a supervised classifi er following the prior work. This nearly unsupervised approach allows fine-grained optimization of each hash function, which is particularly suitable for building hashcode representations generalizing from a training set to a test set. We empirically evaluate the proposed approach for biomedical relation extraction tasks, obtaining significant accuracy improvements w.r.t. state-of-the-art supervised and semi-supervised approaches.

We introduce the dialog intent induction task and present a novel deep multi-view clustering approach to tackle the problem. Dialog intent induction aims at discovering user intents from user query utterances in human-human conversations such as dialogs between customer support agents and customers. Motivated by the intuition that a dialog intent is not only expressed in the user query utterance but also captured in the rest of the dialog, we split a conversation into two independent views and exploit multi-view clustering techniques for inducing the dialog intent. In par- ticular, we propose alternating-view k-means (AV-KMEANS) for joint multi-view represen- tation learning and clustering analysis. The key innovation is that the instance-view representations are updated iteratively by predicting the cluster assignment obtained from the alternative view, so that the multi-view representations of the instances lead to similar cluster assignments. Experiments on two public datasets show that AV-KMEANS can induce better dialog intent clusters than state-of-the-art unsupervised representation learning methods and standard multi-view clustering approaches.

This paper presents a new sequence-to-sequence (seq2seq) pre-training method PoDA (Pre-training of Denoising Autoencoders), which learns representations suitable for text generation tasks. Unlike encoder-only (e.g., BERT) or decoder-only (e.g., OpenAI GPT) pre-training approaches, PoDA jointly pre-trains both the encoder and decoder by denoising the noise-corrupted text, and it also has the advantage of keeping the network architecture unchanged in the subsequent fine-tuning stage. Meanwhile, we design a hybrid model of Transformer and pointer-generator networks as the backbone architecture for PoDA. We conduct experiments on two text generation tasks: abstractive summarization, and grammatical error correction. Results on four datasets show that PoDA can improve model performance over strong baselines without using any task-specific techniques and significantly speed up convergence.

We consider a document classification problem where document labels are absent but only relevant keywords of a target class and unlabeled documents are given. Although heuristic methods based on pseudo-labeling have been considered, theoretical understanding of this problem has still been limited. Moreover, previous methods cannot easily incorporate well-developed techniques in supervised text classification. In this paper, we propose a theoretically guaranteed learning framework that is simple to implement and has flexible choices of models, e.g., linear models or neural networks. We demonstrate how to optimize the area under the receiver operating characteristic curve (AUC) effectively and also discuss how to adjust it to optimize other well-known evaluation metrics such as the accuracy and F1-measure. Finally, we show the effectiveness of our framework using benchmark datasets.

BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019) has set a new state-of-the-art performance on sentence-pair regression tasks like semantic textual similarity (STS). However, it requires that both sentences are fed into the network, which causes a massive computational overhead: Finding the most similar pair in a collection of 10,000 sentences requires about 50 million inference computations (~65 hours) with BERT. The construction of BERT makes it unsuitable for semantic similarity search as well as for unsupervised tasks like clustering. In this publication, we present Sentence-BERT (SBERT), a modification of the pretrained BERT network that use siamese and triplet network structures to derive semantically meaningful sentence embeddings that can be compared using cosine-similarity. This reduces the effort for finding the most similar pair from 65 hours with BERT / RoBERTa to about 5 seconds with SBERT, while maintaining the accuracy from BERT. We evaluate SBERT and SRoBERTa on common STS tasks and transfer learning tasks, where it outperforms other state-of-the-art sentence embeddings methods.

We conduct a large-scale, systematic study to evaluate the existing evaluation methods for natural language generation in the context of generating online product reviews. We compare human-based evaluators with a variety of automated evaluation procedures, including discriminative evaluators that measure how well machine-generated text can be distinguished from human-written text, as well as word overlap metrics that assess how similar the generated text compares to human-written references. We determine to what extent these different evaluators agree on the ranking of a dozen of state-of-the-art generators for online product reviews. We find that human evaluators do not correlate well with discriminative evaluators, leaving a bigger question of whether adversarial accuracy is the correct objective for natural language generation. In general, distinguishing machine-generated text is challenging even for human evaluators, and human decisions correlate better with lexical overlaps. We find lexical diversity an intriguing metric that is indicative of the assessments of different evaluators. A post-experiment survey of participants provides insights into how to evaluate and improve the quality of natural language generation systems.

Text emotion distribution learning (EDL) aims to develop models that can predict the intensity values of a sentence across a set of emotion categories. Existing methods based on supervised learning require a large amount of well-labelled training data, which is difficult to obtain due to inconsistent perception of fine-grained emotion intensity. In this paper, we propose a meta-learning approach to learn text emotion distributions from a small sample. Specifically, we propose to learn low-rank sentence embeddings by tensor decomposition to capture their contextual semantic similarity, and use K-nearest neighbors (KNNs) of each sentence in the embedding space to generate sample clusters. We then train a meta-learner that can adapt to new data with only a few training samples on the clusters, and further fit the meta-learner on KNNs of a testing sample for EDL. In this way, we effectively augment the learning ability of a model on the small sample. To demonstrate the performance, we compare the proposed approach with state-of-the-art EDL methods on a widely used EDL dataset: SemEval 2007 Task 14 (Strapparava and Mihalcea, 2007). Results show the superiority of our method on small-sample emotion distribution learning.

Deep latent variable models (LVM) such as variational auto-encoder (VAE) have recently played an important role in text generation. One key factor is the exploitation of smooth latent structures to guide the generation. However, the representation power of VAEs is limited due to two reasons: (1) the Gaussian assumption is often made on the variational posteriors; and meanwhile (2) a notorious “posterior collapse” issue occurs. In this paper, we advocate sample-based representations of variational distributions for natural language, leading to implicit latent features, which can provide flexible representation power compared with Gaussian-based posteriors. We further develop an LVM to directly match the aggregated posterior to the prior. It can be viewed as a natural extension of VAEs with a regularization of maximizing mutual information, mitigating the “posterior collapse” issue. We demonstrate the effectiveness and versatility of our models in various text generation scenarios, including language modeling, unaligned style transfer, and dialog response generation. The source code to reproduce our experimental results is available on GitHub.

This paper shows that standard assessment methodology for style transfer has several significant problems. First, the standard metrics for style accuracy and semantics preservation vary significantly on different re-runs. Therefore one has to report error margins for the obtained results. Second, starting with certain values of bilingual evaluation understudy (BLEU) between input and output and accuracy of the sentiment transfer the optimization of these two standard metrics diverge from the intuitive goal of the style transfer task. Finally, due to the nature of the task itself, there is a specific dependence between these two metrics that could be easily manipulated. Under these circumstances, we suggest taking BLEU between input and human-written reformulations into consideration for benchmarks. We also propose three new architectures that outperform state of the art in terms of this metric.

While neural models show remarkable accuracy on individual predictions, their internal beliefs can be inconsistent across examples. In this paper, we formalize such inconsistency as a generalization of prediction error. We propose a learning framework for constraining models using logic rules to regularize them away from inconsistency. Our framework can leverage both labeled and unlabeled examples and is directly compatible with off-the-shelf learning schemes without model redesign. We instantiate our framework on natural language inference, where experiments show that enforcing invariants stated in logic can help make the predictions of neural models both accurate and consistent.

Zero-shot text classification (0Shot-TC) is a challenging NLU problem to which little attention has been paid by the research community. 0Shot-TC aims to associate an appropriate label with a piece of text, irrespective of the text domain and the aspect (e.g., topic, emotion, event, etc.) described by the label. And there are only a few articles studying 0Shot-TC, all focusing only on topical categorization which, we argue, is just the tip of the iceberg in 0Shot-TC. In addition, the chaotic experiments in literature make no uniform comparison, which blurs the progress. This work benchmarks the 0Shot-TC problem by providing unified datasets, standardized evaluations, and state-of-the-art baselines. Our contributions include: i) The datasets we provide facilitate studying 0Shot-TC relative to conceptually different and diverse aspects: the “topic” aspect includes “sports” and “politics” as labels; the “emotion” aspect includes “joy” and “anger”; the “situation” aspect includes “medical assistance” and “water shortage”. ii) We extend the existing evaluation setup (label-partially-unseen) – given a dataset, train on some labels, test on all labels – to include a more challenging yet realistic evaluation label-fully-unseen 0Shot-TC (Chang et al., 2008), aiming at classifying text snippets without seeing task specific training data at all. iii) We unify the 0Shot-TC of diverse aspects within a textual entailment formulation and study it this way.

Text classification tends to struggle when data is deficient or when it needs to adapt to unseen classes. In such challenging scenarios, recent studies have used meta-learning to simulate the few-shot task, in which new queries are compared to a small support set at the sample-wise level. However, this sample-wise comparison may be severely disturbed by the various expressions in the same class. Therefore, we should be able to learn a general representation of each class in the support set and then compare it to new queries. In this paper, we propose a novel Induction Network to learn such a generalized class-wise representation, by innovatively leveraging the dynamic routing algorithm in meta-learning. In this way, we find the model is able to induce and generalize better. We evaluate the proposed model on a well-studied sentiment classification dataset (English) and a real-world dialogue intent classification dataset (Chinese). Experiment results show that on both datasets, the proposed model significantly outperforms the existing state-of-the-art approaches, proving the effectiveness of class-wise generalization in few-shot text classification.

We propose a novel on-device sequence model for text classification using recurrent projections. Our model ProSeqo uses dynamic recurrent projections without the need to store or look up any pre-trained embeddings. This results in fast and compact neural networks that can perform on-device inference for complex short and long text classification tasks. We conducted exhaustive evaluation on multiple text classification tasks. Results show that ProSeqo outperformed state-of-the-art neural and on-device approaches for short text classification tasks such as dialog act and intent prediction. To the best of our knowledge, ProSeqo is the first on-device long text classification neural model. It achieved comparable results to previous neural approaches for news article, answers and product categorization, while preserving small memory footprint and maintaining high accuracy.