EMNLP 2022 论文列表

In this paper, we focus on the robustness evaluation of Chinese Question Matching (QM) models. Most of the previous work on analyzing robustness issues focus on just one or a few types of artificial adversarial examples. Instead, we argue that a comprehensive evaluation should be conducted on natural texts, which takes into account the fine-grained linguistic capabilities of QM models. For this purpose, we create a Chinese dataset namely DuQM which contains natural questions with linguistic perturbations to evaluate the robustness of QM models. DuQM contains 3 categories and 13 subcategories with 32 linguistic perturbations. The extensive experiments demonstrate that DuQM has a better ability to distinguish different models. Importantly, the detailed breakdown of evaluation by the linguistic phenomena in DuQM helps us easily diagnose the strength and weakness of different models. Additionally, our experiment results show that the effect of artificial adversarial examples does not work on natural texts. Our baseline codes and a leaderboard are now publicly available.

Overfitting is a notorious problem when there is insufficient data to train deep neural networks in machine learning tasks. Data augmentation regularization methods such as Dropout, Mixup, and their enhanced variants are effective and prevalent, and achieve promising performance to overcome overfitting. However, in text learning, most of the existing regularization approaches merely adopt ideas from computer vision without considering the importance of dimensionality in natural language processing. In this paper, we argue that the property is essential to overcome overfitting in text learning. Accordingly, we present a saliency map informed textual data augmentation and regularization framework, which combines Dropout and Mixup, namely DropMix, to mitigate the overfitting problem in text learning. In addition, we design a procedure that drops and patches fine grained shapes of the saliency map under the DropMix framework to enhance regularization. Empirical studies confirm the effectiveness of the proposed approach on 12 text classification tasks.

A large body of recent work highlights the fallacies of zero-shot cross-lingual transfer (ZS-XLT) with large multilingual language models. Namely, their performance varies substantially for different target languages and is the weakest where needed the most: for low-resource languages distant to the source language. One remedy is few-shot transfer (FS-XLT), where leveraging only a few task-annotated instances in the target language(s) may yield sizable performance gains. However, FS-XLT also succumbs to large variation, as models easily overfit to the small datasets. In this work, we present a systematic study focused on a spectrum of FS-XLT fine-tuning regimes, analyzing key properties such as effectiveness, (in)stability, and modularity. We conduct extensive experiments on both higher-level (NLI, paraphrasing) and lower-level tasks (NER, POS), presenting new FS-XLT strategies that yield both improved and more stable FS-XLT across the board. Our findings challenge established FS-XLT methods: e.g., we propose to replace sequential fine-tuning with joint fine-tuning on source and target language instances, offering consistent gains with different number of shots (including resource-rich scenarios). We also show that further gains can be achieved with multi-stage FS-XLT training in which joint multilingual fine-tuning precedes the bilingual source-target specialization.

Masked language models like BERT can perform text classification in a zero-shot fashion by reformulating downstream tasks as text infilling. However, this approach is highly sensitive to the template used to prompt the model, yet practitioners are blind when designing them in strict zero-shot settings. In this paper, we propose an alternative mining-based approach for zero-shot learning. Instead of prompting language models, we use regular expressions to mine labeled examples from unlabeled corpora, which can optionally be filtered through prompting, and used to finetune a pretrained model. Our method is more flexible and interpretable than prompting, and outperforms it on a wide range of tasks when using comparable templates. Our results suggest that the success of prompting can partly be explained by the model being exposed to similar examples during pretraining, which can be directly retrieved through regular expressions.

Transformer-based pre-trained models are known to encode societal biases not only in their contextual representations, but also in downstream predictions when fine-tuned on task-specific data.We present D-Bias, an approach that selectively eliminates stereotypical associations (e.g, co-occurrence statistics) at fine-tuning, such that the model doesn’t learn to excessively rely on those signals.D-Bias attenuates biases from both identity words and frequently co-occurring proxies, which we select using pointwise mutual information.We apply D-Bias to a) occupation classification, and b) toxicity classification and find that our approach substantially reduces downstream biases (e.g. by > 60% in toxicity classification, for identities that are most frequently flagged as toxic on online platforms).In addition, we show that D-Bias dramatically improves upon scrubbing, i.e., removing only the identity words in question.We also demonstrate that D-Bias easily extends to multiple identities, and achieves competitive performance with two recently proposed debiasing approaches: R-LACE and INLP.

Embodied dialogue instruction following requires an agent to complete a complex sequence of tasks from a natural language exchange. The recent introduction of benchmarks raises the question of how best to train and evaluate models for this multi-turn, multi-agent, long-horizon task. This paper contributes to that conversation, by arguing that imitation learning (IL) and related low-level metrics are actually misleading and do not align with the goals of embodied dialogue research and may hinder progress.We provide empirical comparisons of metrics, analysis of three models, and make suggestions for how the field might best progress. First, we observe that models trained with IL take spurious actions during evaluation. Second, we find that existing models fail to ground query utterances, which are essential for task completion. Third, we argue evaluation should focus on higher-level semantic goals. We will release code to additionally filter the data and benchmark models for improved evaluation.

Although a machine translation model trained with a large in-domain parallel corpus achieves remarkable results, it still works poorly when no in-domain data are available. This situation restricts the applicability of machine translation when the target domain’s data are limited. However, there is great demand for high-quality domain-specific machine translation models for many domains. We propose a framework that efficiently and effectively collects parallel sentences in a target domain from the web with the help of crowdworkers.With the collected parallel data, we can quickly adapt a machine translation model to the target domain. Our experiments show that the proposed method can collect target-domain parallel data over a few days at a reasonable cost. We tested it with five domains, and the domain-adapted model improved the BLEU scores to +19.7 by an average of +7.8 points compared to a general-purpose translation model.

Entity typing aims at predicting one or more words that describe the type(s) of a specific mention in a sentence. Due to shortcuts from surface patterns to annotated entity labels and biased training, existing entity typing models are subject to the problem of spurious correlations. To comprehensively investigate the faithfulness and reliability of entity typing methods, we first systematically define distinct kinds of model biases that are reflected mainly from spurious correlations. Particularly, we identify six types of existing model biases, including mention-context bias, lexical overlapping bias, named entity bias, pronoun bias, dependency bias, and overgeneralization bias. To mitigate model biases, we then introduce a counterfactual data augmentation method. By augmenting the original training set with their debiasedcounterparts, models are forced to fully comprehend sentences and discover the fundamental cues for entity typing, rather than relying on spurious correlations for shortcuts. Experimental results on the UFET dataset show our counterfactual data augmentation approach helps improve generalization of different entity typing models with consistently better performance on both the original and debiased test sets.

In simultaneous speech translation (SimulST), finding the best trade-off between high output quality and low latency is a challenging task. To meet the latency constraints posed by different application scenarios, multiple dedicated SimulST models are usually trained and maintained, generating high computational costs. In this paper, also motivated by the increased sensitivity towards sustainable AI, we investigate whether a single model trained offline can serve both offline and simultaneous applications under different latency regimes without additional training or adaptation. Experiments on en->de, es show that, aside from facilitating the adoption of well-established offline architectures and training strategies without affecting latency, offline training achieves similar or better quality compared to the standard SimulST training protocol, also being competitive with the state-of-the-art system.

Rationalization is fundamental to human reasoning and learning. NLP models trained to produce rationales along with predictions, called self-rationalization models, have been investigated for their interpretability and utility to end-users. However, the extent to which training with human-written rationales facilitates learning remains an under-explored question. We ask whether training models to self-rationalize can aid in their learning to solve tasks for the right reasons. Specifically, we evaluate how training self-rationalization models with free-text rationales affects robustness to spurious correlations in fine-tuned encoder-decoder and decoder-only models of six different sizes. We evaluate robustness to spurious correlations by measuring performance on 1) manually annotated challenge datasets and 2) subsets of original test sets where reliance on spurious correlations would fail to produce correct answers. We find that while self-rationalization can improve robustness to spurious correlations in low-resource settings, it tends to hurt robustness in higher-resource settings. Furthermore, these effects depend on model family and size, as well as on rationale content. Together, our results suggest that explainability can come at the cost of robustness; thus, appropriate care should be taken when training self-rationalizing models with the goal of creating more trustworthy models.

Currently, in speech translation, the straightforward approach - cascading a recognition system with a translation system - delivers state-of-the-art results.However, fundamental challenges such as error propagation from the automatic speech recognition system still remain.To mitigate these problems, recently, people turn their attention to direct data and propose various joint training methods.In this work, we seek to answer the question of whether joint training really helps cascaded speech translation.We review recent papers on the topic and also investigate a joint training criterion by marginalizing the transcription posterior probabilities.Our findings show that a strong cascaded baseline can diminish any improvements obtained using joint training, and we suggest alternatives to joint training.We hope this work can serve as a refresher of the current speech translation landscape, and motivate research in finding more efficient and creative ways to utilize the direct data for speech translation.

The vast majority of non-English corpora are derived from automatically filtered versions of CommonCrawl. While prior work has identified major issues on the quality of these datasets (Kreutzer et al., 2021), it is not clear how this impacts downstream performance. Taking representation learning in Basque as a case study, we explore tailored crawling (manually identifying and scraping websites with high-quality content) as an alternative to filtering CommonCrawl. Our new corpus, called EusCrawl, is similar in size to the Basque portion of popular multilingual corpora like CC100 and mC4, yet it has a much higher quality according to native annotators. For instance, 66% of documents are rated as high-quality for EusCrawl, in contrast with <33% for both mC4 and CC100. Nevertheless, we obtain similar results on downstream NLU tasks regardless of the corpus used for pre-training. Our work suggests that NLU performance in low-resource languages is not primarily constrained by the quality of the data, and other factors like corpus size and domain coverage can play a more important role.

We present DocInfer - a novel, end-to-end Document-level Natural Language Inference model that builds a hierarchical document graph enriched through inter-sentence relations (topical, entity-based, concept-based), performs paragraph pruning using the novel SubGraph Pooling layer, followed by optimal evidence selection based on REINFORCE algorithm to identify the most important context sentences for a given hypothesis. Our evidence selection mechanism allows it to transcend the input length limitation of modern BERT-like Transformer models while presenting the entire evidence together for inferential reasoning. We show this is an important property needed to reason on large documents where the evidence may be fragmented and located arbitrarily far from each other. Extensive experiments on popular corpora - DocNLI, ContractNLI, and ConTRoL datasets, and our new proposed dataset called CaseHoldNLI on the task of legal judicial reasoning, demonstrate significant performance gains of 8-12% over SOTA methods. Our ablation studies validate the impact of our model. Performance improvement of 3-6% on annotation-scarce downstream tasks of fact verification, multiple-choice QA, and contract clause retrieval demonstrates the usefulness of DocInfer beyond primary NLI tasks.

Financial prediction is complex due to the stochastic nature of the stock market. Semi-structured financial documents present comprehensive financial data in tabular formats, such as earnings, profit-loss statements, and balance sheets, and can often contain rich technical analysis along with a textual discussion of corporate history, and management analysis, compliance, and risks. Existing research focuses on the textual and audio modalities of financial disclosures from company conference calls to forecast stock volatility and price movement, but ignores the rich tabular data available in financial reports. Moreover, the economic realm is still plagued with a severe under-representation of various communities spanning diverse demographics, gender, and native speakers. In this work, we show that combining tabular data from financial semi-structured documents with text transcripts and audio recordings not only improves stock volatility and price movement prediction by 5-12% but also reduces gender bias caused due to audio-based neural networks by over 30%.

This paper introduces Doc2Bot, a novel dataset for building machines that help users seek information via conversations. This is of particular interest for companies and organizations that own a large number of manuals or instruction books. Despite its potential, the nature of our task poses several challenges: (1) documents contain various structures that hinder the ability of machines to comprehend, and (2) user information needs are often underspecified. Compared to prior datasets that either focus on a single structural type or overlook the role of questioning to uncover user needs, the Doc2Bot dataset is developed to target such challenges systematically. Our dataset contains over 100,000 turns based on Chinese documents from five domains, larger than any prior document-grounded dialog dataset for information seeking. We propose three tasks in Doc2Bot: (1) dialog state tracking to track user intentions, (2) dialog policy learning to plan system actions and contents, and (3) response generation which generates responses based on the outputs of the dialog policy. Baseline methods based on the latest deep learning models are presented, indicating that our proposed tasks are challenging and worthy of further research.

Recent advances in vision-and-language modeling have seen the development of Transformer architectures that achieve remarkable performance on multimodal reasoning tasks.Yet, the exact capabilities of these black-box models are still poorly understood. While much of previous work has focused on studying their ability to learn meaning at the word-level, their ability to track syntactic dependencies between words has received less attention.We take a first step in closing this gap by creating a new multimodal task targeted at evaluating understanding of predicate-noun dependencies in a controlled setup.We evaluate a range of state-of-the-art models and find that their performance on the task varies considerably, with some models performing relatively well and others at chance level. In an effort to explain this variability, our analyses indicate that the quality (and not only sheer quantity) of pretraining data is essential. Additionally, the best performing models leverage fine-grained multimodal pretraining objectives in addition to the standard image-text matching objectives.This study highlights that targeted and controlled evaluations are a crucial step for a precise and rigorous test of the multimodal knowledge of vision-and-language models.

Traditional multi-task learning architectures learn a single model across multiple tasks through a shared encoder followed by task-specific decoders. Learning these models often requires specialized training algorithms that address task-conflict in the shared parameter updates, which otherwise can lead to negative transfer. A new type of multi-task learning within NLP homogenizes multi-task architectures as a shared encoder and language model decoder, which does surprisingly well across a range of diverse tasks. Does this new architecture suffer from task-conflicts that require specialized training algorithms? We study how certain factors in the shift towards text-to-text models affects multi-task conflict and negative transfer, finding that both directional conflict and transfer are surprisingly constant across architectures.

Recent success of pre-trained language models (PLMs) has stimulated interest in their ability to understand and work with numbers. Yet, the numerical reasoning over measurements has not been formally studied despite their importance. In this study, we show that PLMs lack the capability required for reasoning over measurements. Furthermore, we find that a language model trained on a measurement-rich corpus shows better performance on understanding measurements. We propose a simple embedding strategy to better distinguish between numbers and units, which leads to a significant improvement in the probing tasks.

Compiling comprehensive repositories of commonsense knowledge is a long-standing problem in AI. Many concerns revolve around the issue of reporting bias, i.e., that frequency in text sources is not a good proxy for relevance or truth. This paper explores whether children’s texts hold the key to commonsense knowledge compilation, based on the hypothesis that such content makes fewer assumptions on the reader’s knowledge, and therefore spells out commonsense more explicitly. An analysis with several corpora shows that children’s texts indeed contain much more, and more typical commonsense assertions. Moreover, experiments show that this advantage can be leveraged in popular language-model-based commonsense knowledge extraction settings, where task-unspecific fine-tuning on small amounts of children texts (childBERT) already yields significant improvements. This provides a refreshing perspective different from the common trend of deriving progress from ever larger models and corpora.