Large-scale pre-trained models (PTMs) show great zero-shot capabilities.In this paper, we study how to leverage them for zero-shot visual question answering (VQA).Our approach is motivated by a few observations.First, VQA questions often require multiple steps of reasoning, which is still a capability that most PTMs lack.Second, different steps in VQA reasoning chains require different skills such as object detection and relational reasoning, but a single PTM may not possess all these skills.Third, recent work on zero-shot VQA does not explicitly consider multi-step reasoning chains, which makes them less interpretable compared with a decomposition-based approach.We propose a modularized zero-shot network that explicitly decomposes questions into sub reasoning steps and is highly interpretable. We convert sub reasoning tasks to acceptable objectives of PTMs and assign tasks to proper PTMs without any adaptation.Our experiments on two VQA benchmarks under the zero-shot setting demonstrate the effectiveness of our method and better interpretability compared with several baselines.

Recently, a new training oaxe loss has proven effective to ameliorate the effect of multimodality for non-autoregressive translation (NAT), which removes the penalty of word order errors in the standard cross-entropy loss. Starting from the intuition that reordering generally occurs between phrases, we extend oaxe by only allowing reordering between ngram phrases and still requiring a strict match of word order within the phrases. Extensive experiments on NAT benchmarks across language pairs and data scales demonstrate the effectiveness and universality of our approach. Further analyses show that ngram noaxe indeed improves the translation of ngram phrases, and produces more fluent translation with a better modeling of sentence structure.

In this paper, we revisit math word problems (MWPs) from the cross-lingual and multilingual perspective. We construct our MWP solvers over pretrained multilingual language models using the sequence-to-sequence model with copy mechanism. We compare how the MWP solvers perform in cross-lingual and multilingual scenarios. To facilitate the comparison of cross-lingual performance, we first adapt the large-scale English dataset MathQA as a counterpart of the Chinese dataset Math23K. Then we extend several English datasets to bilingual datasets through machine translation plus human annotation. Our experiments show that the MWP solvers may not be transferred to a different language even if the target expressions share the same numerical constants and operator set. However, it can be better generalized if problem types exist on both source language and target language.

Distant supervision uses triple facts in knowledge graphs to label a corpus for relation extraction, leading to wrong labeling and long-tail problems. Some works use the hierarchy of relations for knowledge transfer to long-tail relations. However, a coarse-grained relation often implies only an attribute (e.g., domain or topic) of the distant fact, making it hard to discriminate relations based solely on sentence semantics. One solution is resorting to entity types, but open questions remain about how to fully leverage the information of entity types and how to align multi-granular entity types with sentences. In this work, we propose a novel model to enrich distantly-supervised sentences with entity types. It consists of (1) a pairwise type-enriched sentence encoding module injecting both context-free and -related backgrounds to alleviate sentence-level wrong labeling, and (2) a hierarchical type-sentence alignment module enriching a sentence with the triple fact’s basic attributes to support long-tail relations. Our model achieves new state-of-the-art results in overall and long-tail performance on benchmarks.

While GPT has become the de-facto method for text generation tasks, its application to pinyin input method remains unexplored.In this work, we make the first exploration to leverage Chinese GPT for pinyin input method.We find that a frozen GPT achieves state-of-the-art performance on perfect pinyin.However, the performance drops dramatically when the input includes abbreviated pinyin.A reason is that an abbreviated pinyin can be mapped to many perfect pinyin, which links to even larger number of Chinese characters.We mitigate this issue with two strategies,including enriching the context with pinyin and optimizing the training process to help distinguish homophones. To further facilitate the evaluation of pinyin input method, we create a dataset consisting of 270K instances from fifteen domains.Results show that our approach improves the performance on abbreviated pinyin across all domains.Model analysis demonstrates that both strategiescontribute to the performance boost.

A recent study by Feldman (2020) proposed a long-tail theory to explain the memorization behavior of deep learning models. However, memorization has not been empirically verified in the context of NLP, a gap addressed by this work. In this paper, we use three different NLP tasks to check if the long-tail theory holds. Our experiments demonstrate that top-ranked memorized training instances are likely atypical, and removing the top-memorized training instances leads to a more serious drop in test accuracy compared with removing training instances randomly. Furthermore, we develop an attribution method to better understand why a training instance is memorized. We empirically show that our memorization attribution method is faithful, and share our interesting finding that the top-memorized parts of a training instance tend to be features negatively correlated with the class label.

Translate-train is a general training approach to multilingual tasks. The key idea is to use the translator of the target language to generate training data to mitigate the gap between the source and target languages. However, its performance is often hampered by the artifacts in the translated texts (translationese). We discover that such artifacts have common patterns in different languages and can be modeled by deep learning, and subsequently propose an approach to conduct translate-train using Translationese Embracing the effect of Artifacts (TEA). TEA learns to mitigate such effect on the training data of a source language (whose original and translationese are both available), and applies the learned module to facilitate the inference on the target language. Extensive experiments on the multilingual QA dataset TyDiQA demonstrate that TEA outperforms strong baselines.

Hateful meme classification is a challenging multimodal task that requires complex reasoning and contextual background knowledge. Ideally, we could leverage an explicit external knowledge base to supplement contextual and cultural information in hateful memes. However, there is no known explicit external knowledge base that could provide such hate speech contextual information. To address this gap, we propose PromptHate, a simple yet effective prompt-based model that prompts pre-trained language models (PLMs) for hateful meme classification. Specifically, we construct simple prompts and provide a few in-context examples to exploit the implicit knowledge in the pre-trained RoBERTa language model for hateful meme classification. We conduct extensive experiments on two publicly available hateful and offensive meme datasets. Our experiment results show that PromptHate is able to achieve a high AUC of 90.96, outperforming state-of-the-art baselines on the hateful meme classification task. We also perform fine-grain analyses and case studies on various prompt settings and demonstrate the effectiveness of the prompts on hateful meme classification.

Out-of-distribution (OOD) settings are used to measure a model’s performance when the distribution of the test data is different from that of the training data. NLU models are known to suffer in OOD. We study this issue from the perspective of causality, which sees confounding bias as the reason for models to learn spurious correlations. While a common solution is to perform intervention, existing methods handle only known and single confounder, but in many NLU tasks the confounders can be both unknown and multifactorial. In this paper, we propose a novel interventional training method called Bottom-up Automatic Intervention (BAI) that performs multi-granular intervention with identified multifactorial confounders. Our experiments on three NLU tasks, namely, natural language inference, fact verification and paraphrase identification, show the effectiveness of BAI for tackling OOD settings.

In this paper we study how to measure stereotypical bias in pre-trained vision-language models. We leverage a recently released text-only dataset, StereoSet, which covers a wide range of stereotypical bias, and extend it into a vision-language probing dataset called VLStereoSet to measure stereotypical bias in vision-language models. We analyze the differences between text and image and propose a probing task that detects bias by evaluating a model’s tendency to pick stereotypical statements as captions for anti-stereotypical images. We further define several metrics to measure both a vision-language model’s overall stereotypical bias and its intra-modal and inter-modal bias. Experiments on six representative pre-trained vision-language models demonstrate that stereotypical biases clearly exist in most of these models and across all four bias categories, with gender bias slightly more evident. Further analysis using gender bias data and two vision-language models also suggest that both intra-modal and inter-modal bias exist.

There has been much interest in rumor detection using deep learning models in recent years. A well-known limitation of deep learning models is that they tend to learn superficial patterns, which restricts their generalization ability. We find that this is also true for cross-topic rumor detection. In this paper, we propose a method inspired by the “mixture of experts” paradigm. We assume that the prediction of the rumor class label given an instance is dependent on the topic distribution of the instance. After deriving a vector representation for each topic, given an instance, we derive a “topic mixture” vector for the instance based on its topic distribution. This topic mixture is combined with the vector representation of the instance itself to make rumor predictions. Our experiments show that our proposed method can outperform two baseline debiasing methods in a cross-topic setting. In a synthetic setting when we removed topic-specific words, our method also works better than the baselines, showing that our method does not rely on superficial features.

While diverse question answering (QA) datasets have been proposed and contributed significantly to the development of deep learning models for QA tasks, the existing datasets fall short in two aspects. First, we lack QA datasets covering complex questions that involve answers as well as the reasoning processes to get them. As a result, the state-of-the-art QA research on numerical reasoning still focuses on simple calculations and does not provide the mathematical expressions or evidence justifying the answers. Second, the QA community has contributed a lot of effort to improve the interpretability of QA models. However, they fail to explicitly show the reasoning process, such as the evidence order for reasoning and the interactions between different pieces of evidence. To address the above shortcoming, we introduce NOAHQA, a conversational and bilingual QA dataset with questions requiring numerical reasoning with compound mathematical expressions. With NOAHQA, we develop an interpretable reasoning graph as well as the appropriate evaluation metric to measure the answer quality. We evaluate the state-of-the-art QA models trained using existing QA datasets on NOAHQA and show that the best among them can only achieve 55.5 exact match scores, while the human performance is 89.7. We also present a new QA model for generating a reasoning graph where the reasoning graph metric still has a large gap compared with that of humans, eg, 28 scores.

Pre-trained multilingual language models, e.g., multilingual-BERT, are widely used in cross-lingual tasks, yielding the state-of-the-art performance. However, such models suffer from a large performance gap between source and target languages, especially in the zero-shot setting, where the models are fine-tuned only on English but tested on other languages for the same task. We tackle this issue by incorporating language-agnostic information, specifically, universal syntax such as dependency relations and POS tags, into language models, based on the observation that universal syntax is transferable across different languages. Our approach, called COunterfactual SYntax (COSY), includes the design of SYntax-aware networks as well as a COunterfactual training method to implicitly force the networks to learn not only the semantics but also the syntax. To evaluate COSY, we conduct cross-lingual experiments on natural language inference and question answering using mBERT and XLM-R as network backbones. Our results show that COSY achieves the state-of-the-art performance for both tasks, without using auxiliary training data.

Conversational KBQA is about answering a sequence of questions related to a KB. Follow-up questions in conversational KBQA often have missing information referring to entities from the conversation history. In this paper, we propose to model these implied entities, which we refer to as the focal entities of the conversation. We propose a novel graph-based model to capture the transitions of focal entities and apply a graph neural network to derive a probability distribution of focal entities for each question, which is then combined with a standard KBQA module to perform answer ranking. Our experiments on two datasets demonstrate the effectiveness of our proposed method.

We study the task of learning and evaluating Chinese idiom embeddings. We first construct a new evaluation dataset that contains idiom synonyms and antonyms. Observing that existing Chinese word embedding methods may not be suitable for learning idiom embeddings, we further present a BERT-based method that directly learns embedding vectors for individual idioms. We empirically compare representative existing methods and our method. We find that our method substantially outperforms existing methods on the evaluation dataset we have constructed.

Understanding idioms is important in NLP. In this paper, we study to what extent pre-trained BERT model can encode the meaning of a potentially idiomatic expression (PIE) in a certain context. We make use of a few existing datasets and perform two probing tasks: PIE usage classification and idiom paraphrase identification. Our experiment results suggest that BERT indeed can separate the literal and idiomatic usages of a PIE with high accuracy. It is also able to encode the idiomatic meaning of a PIE to some extent.

This paper investigates a new co-attention mechanism in neural transduction models for machine translation tasks. We propose a paradigm, termed Two-Headed Monster (THM), which consists of two symmetric encoder modules and one decoder module connected with co-attention. As a specific and concrete implementation of THM, Crossed Co-Attention Networks (CCNs) are designed based on the Transformer model. We test CCNs on WMT 2014 EN-DE and WMT 2016 EN-FI translation tasks and show both advantages and disadvantages of the proposed method. Our model outperforms the strong Transformer baseline by 0.51 (big) and 0.74 (base) BLEU points on EN-DE and by 0.17 (big) and 0.47 (base) BLEU points on EN-FI but the epoch time increases by circa 75%.

Many graph embedding approaches have been proposed for knowledge graph completion via link prediction. Among those, translating embedding approaches enjoy the advantages of light-weight structure, high efficiency and great interpretability. Especially when extended to complex vector space, they show the capability in handling various relation patterns including symmetry, antisymmetry, inversion and composition. However, previous translating embedding approaches defined in complex vector space suffer from two main issues: 1) representing and modeling capacities of the model are limited by the translation function with rigorous multiplication of two complex numbers; and 2) embedding ambiguity caused by one-to-many relations is not explicitly alleviated. In this paper, we propose a relation-adaptive translation function built upon a novel weighted product in complex space, where the weights are learnable, relation-specific and independent to embedding size. The translation function only requires eight more scalar parameters each relation, but improves expressive power and alleviates embedding ambiguity problem. Based on the function, we then present our Relation-adaptive translating Embedding (RatE) approach to score each graph triple. Moreover, a novel negative sampling method is proposed to utilize both prior knowledge and self-adversarial learning for effective optimization. Experiments verify RatE achieves state-of-the-art performance on four link prediction benchmarks.

Chinese idioms are special fixed phrases usually derived from ancient stories, whose meanings are oftentimes highly idiomatic and non-compositional. The Chinese idiom prediction task is to select the correct idiom from a set of candidate idioms given a context with a blank. We propose a BERT-based dual embedding model to encode the contextual words as well as to learn dual embeddings of the idioms. Specifically, we first match the embedding of each candidate idiom with the hidden representation corresponding to the blank in the context. We then match the embedding of each candidate idiom with the hidden representations of all the tokens in the context thorough context pooling. We further propose to use two separate idiom embeddings for the two kinds of matching. Experiments on a recently released Chinese idiom cloze test dataset show that our proposed method performs better than the existing state of the art. Ablation experiments also show that both context pooling and dual embedding contribute to the improvement of performance.

Wrong labeling problem and long-tail relations are two main challenges caused by distant supervision in relation extraction. Recent works alleviate the wrong labeling by selective attention via multi-instance learning, but cannot well handle long-tail relations even if hierarchies of the relations are introduced to share knowledge. In this work, we propose a novel neural network, Collaborating Relation-augmented Attention (CoRA), to handle both the wrong labeling and long-tail relations. Particularly, we first propose relation-augmented attention network as base model. It operates on sentence bag with a sentence-to-relation attention to minimize the effect of wrong labeling. Then, facilitated by the proposed base model, we introduce collaborating relation features shared among relations in the hierarchies to promote the relation-augmenting process and balance the training data for long-tail relations. Besides the main training objective to predict the relation of a sentence bag, an auxiliary objective is utilized to guide the relation-augmenting process for a more accurate bag-level representation. In the experiments on the popular benchmark dataset NYT, the proposed CoRA improves the prior state-of-the-art performance by a large margin in terms of Precision@N, AUC and Hits@K. Further analyses verify its superior capability in handling long-tail relations in contrast to the competitors.

Previous work on answering complex questions from knowledge bases usually separately addresses two types of complexity: questions with constraints and questions with multiple hops of relations. In this paper, we handle both types of complexity at the same time. Motivated by the observation that early incorporation of constraints into query graphs can more effectively prune the search space, we propose a modified staged query graph generation method with more flexible ways to generate query graphs. Our experiments clearly show that our method achieves the state of the art on three benchmark KBQA datasets.

In this paper, we study Multimodal Named Entity Recognition (MNER) for social media posts. Existing approaches for MNER mainly suffer from two drawbacks: (1) despite generating word-aware visual representations, their word representations are insensitive to the visual context; (2) most of them ignore the bias brought by the visual context. To tackle the first issue, we propose a multimodal interaction module to obtain both image-aware word representations and word-aware visual representations. To alleviate the visual bias, we further propose to leverage purely text-based entity span detection as an auxiliary module, and design a Unified Multimodal Transformer to guide the final predictions with the entity span predictions. Experiments show that our unified approach achieves the new state-of-the-art performance on two benchmark datasets.

In this paper, we propose Cross-Thought, a novel approach to pre-training sequence encoder, which is instrumental in building reusable sequence embeddings for large-scale NLP tasks such as question answering. Instead of using the original signals of full sentences, we train a Transformer-based sequence encoder over a large set of short sequences, which allows the model to automatically select the most useful information for predicting masked words. Experiments on question answering and textual entailment tasks demonstrate that our pre-trained encoder can outperform state-of-the-art encoders trained with continuous sentence signals as well as traditional masked language modeling baselines. Our proposed approach also achieves new state of the art on HotpotQA (full-wiki setting) by improving intermediate information retrieval performance.

The prevalent use of social media enables rapid spread of rumors on a massive scale, which leads to the emerging need of automatic rumor verification (RV). A number of previous studies focus on leveraging stance classification to enhance RV with multi-task learning (MTL) methods. However, most of these methods failed to employ pre-trained contextualized embeddings such as BERT, and did not exploit inter-task dependencies by using predicted stance labels to improve the RV task. Therefore, in this paper, to extend BERT to obtain thread representations, we first propose a Hierarchical Transformer, which divides each long thread into shorter subthreads, and employs BERT to separately represent each subthread, followed by a global Transformer layer to encode all the subthreads. We further propose a Coupled Transformer Module to capture the inter-task interactions and a Post-Level Attention layer to use the predicted stance labels for RV, respectively. Experiments on two benchmark datasets show the superiority of our Coupled Hierarchical Transformer model over existing MTL approaches.

Commonsense reasoning is fundamental to natural language understanding. While traditional methods rely heavily on human-crafted features and knowledge bases, we explore learning commonsense knowledge from a large amount of raw text via unsupervised learning. We propose two neural network models based on the Deep Structured Semantic Models (DSSM) framework to tackle two classic commonsense reasoning tasks, Winograd Schema challenges (WSC) and Pronoun Disambiguation (PDP). Evaluation shows that the proposed models effectively capture contextual information in the sentence and co-reference information between pronouns and nouns, and achieve significant improvement over previous state-of-the-art approaches.

Neural networks equipped with self-attention have parallelizable computation, light-weight structure, and the ability to capture both long-range and local dependencies. Further, their expressive power and performance can be boosted by using a vector to measure pairwise dependency, but this requires to expand the alignment matrix to a tensor, which results in memory and computation bottlenecks. In this paper, we propose a novel attention mechanism called “Multi-mask Tensorized Self-Attention” (MTSA), which is as fast and as memory-efficient as a CNN, but significantly outperforms previous CNN-/RNN-/attention-based models. MTSA 1) captures both pairwise (token2token) and global (source2token) dependencies by a novel compatibility function composed of dot-product and additive attentions, 2) uses a tensor to represent the feature-wise alignment scores for better expressive power but only requires parallelizable matrix multiplications, and 3) combines multi-head with multi-dimensional attentions, and applies a distinct positional mask to each head (subspace), so the memory and computation can be distributed to multiple heads, each with sequential information encoded independently. The experiments show that a CNN/RNN-free model based on MTSA achieves state-of-the-art or competitive performance on nine NLP benchmarks with compelling memory- and time-efficiency.

Multi-choice reading comprehension is a challenging task, which involves the matching between a passage and a question-answer pair. This paper proposes a new co-matching approach to this problem, which jointly models whether a passage can match both a question and a candidate answer. Experimental results on the RACE dataset demonstrate that our approach achieves state-of-the-art performance.

In this paper, we study how we can improve a deep learning approach to textual entailment by incorporating lexical entailment relations from WordNet. Our idea is to embed the lexical entailment knowledge contained in WordNet in specially-learned word vectors, which we call “entailment vectors.” We present a standard neural network model and a novel set-theoretic model to learn these entailment vectors from word pairs with known lexical entailment relations derived from WordNet. We further incorporate these entailment vectors into a decomposable attention model for textual entailment and evaluate the model on the SICK and the SNLI dataset. We find that using these special entailment word vectors, we can significantly improve the performance of textual entailment compared with a baseline that uses only standard word2vec vectors. The final performance of our model is close to or above the state of the art, but our method does not rely on any manually-crafted rules or extensive syntactic features.

In this paper, we target at improving the performance of multi-label emotion classification with the help of sentiment classification. Specifically, we propose a new transfer learning architecture to divide the sentence representation into two different feature spaces, which are expected to respectively capture the general sentiment words and the other important emotion-specific words via a dual attention mechanism. Experimental results on two benchmark datasets demonstrate the effectiveness of our proposed method.

In this paper, we study how to improve the domain adaptability of a deletion-based Long Short-Term Memory (LSTM) neural network model for sentence compression. We hypothesize that syntactic information helps in making such models more robust across domains. We propose two major changes to the model: using explicit syntactic features and introducing syntactic constraints through Integer Linear Programming (ILP). Our evaluation shows that the proposed model works better than the original model as well as a traditional non-neural-network-based model in a cross-domain setting.

In this paper, we study domain adaptation with a state-of-the-art hierarchical neural network for document-level sentiment classification. We first design a new auxiliary task based on sentiment scores of domain-independent words. We then propose two neural network architectures to respectively induce document embeddings and sentence embeddings that work well for different domains. When these document and sentence embeddings are used for sentiment classification, we find that with both pseudo and external sentiment lexicons, our proposed methods can perform similarly to or better than several highly competitive domain adaptation methods on a benchmark dataset of product reviews.

Relation classification is the task of classifying the semantic relations between entity pairs in text. Observing that existing work has not fully explored using different representations for relation instances, especially in order to better handle the asymmetry of relation types, in this paper, we propose a neural network based method for relation classification that combines the raw sequence and the shortest dependency path representations of relation instances and uses mirror instances to perform pairwise relation classification. We evaluate our proposed models on the SemEval-2010 Task 8 dataset. The empirical results show that with two additional features, our model achieves the state-of-the-art result of F1 score of 85.7.

Microblogging services allow users to create hashtags to categorize their posts. In recent years, the task of recommending hashtags for microblogs has been given increasing attention. However, most of existing methods depend on hand-crafted features. Motivated by the successful use of long short-term memory (LSTM) for many natural language processing tasks, in this paper, we adopt LSTM to learn the representation of a microblog post. Observing that hashtags indicate the primary topics of microblog posts, we propose a novel attention-based LSTM model which incorporates topic modeling into the LSTM architecture through an attention mechanism. We evaluate our model using a large real-world dataset. Experimental results show that our model significantly outperforms various competitive baseline methods. Furthermore, the incorporation of topical attention mechanism gives more than 7.4% improvement in F1 score compared with standard LSTM method.