EMNLP 2022 Papers

Stance detection aims to identify people’s standpoints expressed in the text towards a target, which can provide powerful information for various downstream tasks.Recent studies have proposed multi-task learning models that introduce sentiment information to boost stance detection.However, they neglect to explore capturing the fine-grained task-specific interaction between stance detection and sentiment tasks, thus degrading performance.To address this issue, this paper proposes a novel multi-task interaction network (MTIN) for improving the performance of stance detection and sentiment analysis tasks simultaneously.Specifically, we construct heterogeneous task-related graphs to automatically identify and adapt the roles that a word plays with respect to a specific task. Also, a multi-task interaction module is designed to capture the word-level interaction between tasks, so as to obtain richer task representations.Extensive experiments on two real-world datasets show that our proposed approach outperforms state-of-the-art methods in both stance detection and sentiment analysis tasks.

In this paper, we show that the combination of Phrase Pair Injection and Corpus Filtering boosts the performance of Neural Machine Translation (NMT) systems. We extract parallel phrases and sentences from the pseudo-parallel corpus and augment it with the parallel corpus to train the NMT models. With the proposed approach, we observe an improvement in the Machine Translation (MT) system for 3 low-resource language pairs, Hindi-Marathi, English-Marathi, and English-Pashto, and 6 translation directions by up to 2.7 BLEU points, on the FLORES test data. These BLEU score improvements are over the models trained using the whole pseudo-parallel corpus augmented with the parallel corpus.

Pre-trained multilingual language models are the foundation of many NLP approaches, including cross-lingual transfer solutions. However, languages with small available monolingual corpora are often not well-supported by these models leading to poor performance. We propose an unsupervised approach to improve the cross-lingual representations of low-resource languages by bootstrapping word translation pairs from monolingual corpora and using them to improve language alignment in pre-trained language models. We perform experiments on nine languages, using contextual word retrieval and zero-shot named entity recognition to measure both intrinsic cross-lingual word representation quality and downstream task performance, showing improvements on both tasks. Our results show that it is possible to improve pre-trained multilingual language models by relying only on non-parallel resources.

This paper addresses the quality issues in existing Twitter-based paraphrase datasets, and discusses the necessity of using two separate definitions of paraphrase for identification and generation tasks. We present a new Multi-Topic Paraphrase in Twitter (MultiPIT) corpus that consists of a total of 130k sentence pairs with crowdsoursing (MultiPIT_crowd) and expert (MultiPIT_expert) annotations using two different paraphrase definitions for paraphrase identification, in addition to a multi-reference test set (MultiPIT_NMR) and a large automatically constructed training set (MultiPIT_Auto) for paraphrase generation. With improved data annotation quality and task-specific paraphrase definition, the best pre-trained language model fine-tuned on our dataset achieves the state-of-the-art performance of 84.2 F1 for automatic paraphrase identification. Furthermore, our empirical results also demonstrate that the paraphrase generation models trained on MultiPIT_Auto generate more diverse and high-quality paraphrases compared to their counterparts fine-tuned on other corpora such as Quora, MSCOCO, and ParaNMT.

This paper presents an approach to identify samples from live traffic where the customer implicitly communicated satisfaction with Alexa’s responses, by leveraging interpretations of model behavior. Such customer signals are noisy and adding a large number of samples from live traffic to training set makes re-training infeasible. Our work addresses these challenges by identifying a small number of samples that grow training set by ~0.05% while producing statistically significant improvements in both offline and online tests.

Iterative text revision improves text quality by fixing grammatical errors, rephrasing for better readability or contextual appropriateness, or reorganizing sentence structures throughout a document.Most recent research has focused on understanding and classifying different types of edits in the iterative revision process from human-written text instead of building accurate and robust systems for iterative text revision.In this work, we aim to build an end-to-end text revision system that can iteratively generate helpful edits by explicitly detecting editable spans (where-to-edit) with their corresponding edit intents and then instructing a revision model to revise the detected edit spans.Leveraging datasets from other related text editing NLP tasks, combined with the specification of editable spans, leads our system to more accurately model the process of iterative text refinement, as evidenced by empirical results and human evaluations.Our system significantly outperforms previous baselines on our text revision tasks and other standard text revision tasks, including grammatical error correction, text simplification, sentence fusion, and style transfer.Through extensive qualitative and quantitative analysis, we make vital connections between edit intentions and writing quality, and better computational modeling of iterative text revisions.

Word sense disambiguation (WSD) is the task of identifying the intended sense of a word in context. While prior work on unsupervised WSD has leveraged lexical knowledge bases, such as WordNet and BabelNet, these resources have proven to be less effective for Chinese. Instead, the most widely used lexical knowledge base for Chinese is HowNet. Previous HowNet-based WSD methods have not exploited contextual translation information. In this paper, we present the first HowNet-based WSD system which combines monolingual contextual information from a pretrained neural language model with bilingual information obtained via machine translation and sense translation information from HowNet. The results of our evaluation experiment on a test set from prior work demonstrate that our new method achieves a new state of the art for unsupervised Chinese WSD.

Knowledge Distillation (KD) is a commonly used technique for improving the generalization of compact Pre-trained Language Models (PLMs) on downstream tasks. However, such methods impose the additional burden of training a separate teacher model for every new dataset.Alternatively, one may directly work on the improvement of the optimization procedure of the compact model towards better generalization. Recent works observe that the flatness of the local minimum correlates well with better generalization.In this work, we adapt Stochastic Weight Averaging (SWA), a method encouraging convergence to a flatter minimum, to fine-tuning PLMs. We conduct extensive experiments on various NLP tasks (text classification, question answering, and generation) and different model architectures and demonstrate that our adaptation improves the generalization without extra computation cost. Moreover, we observe that this simple optimization technique is able to outperform the state-of-the-art KD methods for compact models.

Named entity disambiguation (NED) is a critical subtask of entity linking, which seeks to connect knowledge base entities with textual mentions of those entities. Naturally, the performance of a model depends on the domain it was trained on; thus, reducing the amount of data required to train models is advantageous. In this work, we leverage recent research on pattern exploitation for NED and explore whether it can reduce the amount of data required for domain adaptation by reformulating the disambiguation task as a masked language modeling problem. Using ADAPET (Tam et al., 2021), which implements a new approach for few-shot learning using fine-tuned transformer-based language models, we produce an NED model which yields, without any sacrifice of in-domain accuracy, a 7% improvement in zero-shot cross-domain performance as evaluated on NEDMed, a new NED dataset of mental health news which we release with this work.

Current abstractive summarization systems tend to hallucinate content that is unfaithful to the source document, posing a risk of misinformation. To mitigate hallucination, we must teach the model to distinguish hallucinated summaries from faithful ones. However, the commonly used maximum likelihood training does not disentangle factual errors from other model errors. To address this issue,we propose a back-translation-style approach to augment negative samples that mimic factual errors made by the model. Specifically, we train an elaboration model that generates hallucinated documents given the reference summaries, and then generates negative summaries from the fake documents. We incorporate the negative samples into training through a controlled generator, which produces faithful/unfaithful summaries conditioned on the control codes. Additionally, we find that adding textual entailment data through multitasking further boosts the performance. Experiments on three datasets (XSum, Gigaword, and WikiHow) show that our method consistently improves faithfulness without sacrificing informativeness according to both human and automatic evaluation

State-of-the-art summarization models still struggle to be factually consistent with the input text. A model-agnostic way to address this problem is post-editing the generated summaries. However, existing approaches typically fail to remove entity errors if a suitable input entity replacement is not available or may insert erroneous content. In our work, we focus on removing extrinsic entity errors, or entities not in the source, to improve consistency while retaining the summary’s essential information and form. We propose to use sentence-compression data to train the post-editing model to take a summary with extrinsic entity errors marked with special tokens and output a compressed, well-formed summary with those errors removed. We show that this model improves factual consistency while maintaining ROUGE, improving entity precision by up to 30% on XSum, and that this model can be applied on top of another post-editor, improving entity precision by up to a total of 38%. We perform an extensive comparison of post-editing approaches that demonstrate trade-offs between factual consistency, informativeness, and grammaticality, and we analyze settings where post-editors show the largest improvements.

Event coreference resolution (ECR) aims to cluster event mentions that refer to the same real-world events. Deep learning methods have achieved SOTA results on the ECR task. However, due to the encoding length limitation, previous methods either adopt classical pairwise models based on sentence-level context or split each document into multiple chunks and encode them separately. They failed to capture the interactions and contextual cues among those long-distance event mentions. Besides, high-level information, such as event topics, is rarely considered to enhance representation learning for ECR. To address the above two issues, we first apply a Longformer-based encoder to obtain the document-level embeddings and an encoder with a trigger-mask mechanism to learn sentence-level embeddings based on local context. In addition, we propose an event topic generator to infer the latent topic-level representations. Finally, using the above event embeddings, we employ a multiple tensor matching method to capture their interactions at the document, sentence, and topic levels. Experimental results on the KBP 2017 dataset show that our model outperforms the SOTA baselines.

Transliteration is an important task in natural language processing (NLP) which aims to convert a name in the source language to the target language without changing its pronunciation. Particularly, transliteration from English to Arabic is highly needed in many applications, especially in countries (e.g., United Arab Emirates (UAE)) whose most citizens are foreigners but the official language is Arabic. In such a task-oriented scenario, namely transliterating the English names to the corresponding Arabic ones, the performance of the transliteration model is highly important. However, most existing neural approaches mainly apply a universal transliteration model with advanced encoders and decoders to the task, where limited attention is paid to leveraging the phonemic association between English and Arabic to further improve model performance. In this paper, we focus on transliteration of people’s names from English to Arabic for the general public. In doing so, we collect a corpus named EANames by extracting high quality name pairs from online resources which better represent the names in the general public than linked Wikipedia entries that are always names of famous people). We propose a model for English-Arabic transliteration, where a memory module modeling the phonemic association between English and Arabic is used to guide the transliteration process. We run experiments on the collected data and the results demonstrate the effectiveness of our approach for English-Arabic transliteration.

We propose an approach for comparing curricula of study programs in higher education. Pre-trained word embeddings are fine-tuned in a study program classification task, where each curriculum is represented by the names and content of its courses. By combining metric learning with a novel course-guided attention mechanism, our method obtains more accurate curriculum representations than strong baselines. Experiments on a new dataset with curricula of computing programs demonstrate the intuitive power of our approach via attention weights, topic modeling, and embeddings visualizations. We also present a use case comparing computing curricula from USA and Latin America to showcase the capabilities of our improved embeddings representations.

Complex knowledge base question answering can be achieved by converting questions into sequences of predefined actions. However, there is a significant semantic and structural gap between natural language and action sequences, which makes this conversion difficult. In this paper, we introduce an alignment-enhanced complex question answering framework, called ALCQA, which mitigates this gap through question-to-action alignment and question-to-question alignment. We train a question rewriting model to align the question and each action, and utilize a pretrained language model to implicitly align the question and KG artifacts. Moreover, considering that similar questions correspond to similar action sequences, we retrieve top-k similar question-answer pairs at the inference stage through question-to-question alignment and propose a novel reward-guided action sequence selection strategy to select from candidate action sequences. We conduct experiments on CQA and WQSP datasets, and the results show that our approach outperforms state-of-the-art methods and obtains a 9.88% improvements in the F1 metric on CQA dataset. Our source code is available at https://github.com/TTTTTTTTy/ALCQA.

Chinese spelling check (CSC) is a fundamental NLP task that detects and corrects spelling errors in Chinese texts. As most of these spelling errors are caused by phonetic similarity, effectively modeling the pronunciation of Chinese characters is a key factor for CSC. In this paper, we consider introducing an auxiliary task of Chinese pronunciation prediction (CPP) to improve CSC, and, for the first time, systematically discuss the adaptivity and granularity of this auxiliary task. We propose SCOPE which builds upon a shared encoder two parallel decoders, one for the primary CSC task and the other for a fine-grained auxiliary CPP task, with a novel adaptive weighting scheme to balance the two tasks. In addition, we design a delicate iterative correction strategy for further improvements during inference. Empirical evaluation shows that SCOPE achieves new state-of-the-art on three CSC benchmarks, demonstrating the effectiveness and superiority of the auxiliary CPP task. Comprehensive ablation studies further verify the positive effects of adaptivity and granularity of the task.

Bilingual lexicon induction (BLI) with limited bilingual supervision is a crucial yet challenging task in multilingual NLP. Current state-of-the-art BLI methods rely on the induction of cross-lingual word embeddings (CLWEs) to capture cross-lingual word similarities; such CLWEs are obtained <b>1)</b> via traditional static models (e.g., VecMap), or <b>2)</b> by extracting type-level CLWEs from multilingual pretrained language models (mPLMs), or <b>3)</b> through combining the former two options. In this work, we propose a novel semi-supervised <i>post-hoc</i> reranking method termed <b>BLICEr</b> (<b>BLI</b> with <b>C</b>ross-<b>E</b>ncoder <b>R</b>eranking), applicable to any precalculated CLWE space, which improves their BLI capability. The key idea is to ‘extract’ cross-lingual lexical knowledge from mPLMs, and then combine it with the original CLWEs. This crucial step is done via <b>1)</b> creating a word similarity dataset, comprising positive word pairs (i.e., true translations) and hard negative pairs induced from the original CLWE space, and then <b>2)</b> fine-tuning an mPLM (e.g., mBERT or XLM-R) in a cross-encoder manner to predict the similarity scores. At inference, we <b>3)</b> combine the similarity score from the original CLWE space with the score from the BLI-tuned cross-encoder. BLICEr establishes new state-of-the-art results on two standard BLI benchmarks spanning a wide spectrum of diverse languages: it substantially outperforms a series of strong baselines across the board. We also validate the robustness of BLICEr with different CLWEs.

Recently, aspect sentiment quad prediction (ASQP) has become a popular task in the field of aspect-level sentiment analysis. Previous work utilizes a predefined template to paraphrase the original sentence into a structure target sequence, which can be easily decoded as quadruplets of the form (aspect category, aspect term, opinion term, sentiment polarity). The template involves the four elements in a fixed order. However, we observe that this solution contradicts with the order-free property of the ASQP task, since there is no need to fix the template order as long as the quadruplet is extracted correctly. Inspired by the observation, we study the effects of template orders and find that some orders help the generative model achieve better performance. It is hypothesized that different orders provide various views of the quadruplet. Therefore, we propose a simple but effective method to identify the most proper orders, and further combine multiple proper templates as data augmentation to improve the ASQP task. Specifically, we use the pre-trained language model to select the orders with minimal entropy. By fine-tuning the pre-trained language model with these template orders, our approach improves the performance of quad prediction, and outperforms state-of-the-art methods significantly in low-resource settings.

We offer a two-stage reranking method for grammatical error correction: the first model serves as edit generator, while the second classifies the proposed edits as correct or false. We show how to use both encoder-decoder and sequence labeling models for the first step of our pipeline. We achieve state-of-the-art quality on BEA 2019 English dataset even using weak BERT-GEC edit generator. Combining our roberta-base scorer with state-of-the-art GECToR edit generator, we surpass GECToR by 2-3%. With a larger model we establish a new SOTA on BEA development and test sets. Our model also sets a new SOTA on Russian, despite using smaller models and less data than the previous approaches.