EMNLP 2022 论文列表

Large-scale language models coupled with prompts have shown remarkable performance on few-shot learning. However, through systematic experiments, we find that the few-shot performance of small language models is poor, and using prompts on them brings fewer improvements than on larger ones. In this paper, we propose SMASH, an approach to improve SMAll language models’ few-SHot ability by training on intermediate tasks before prompt-based fine-tuning on downstream tasks. We design intermediate tasks for sentence-pair tasks and sentiment classification tasks by creating training examples with prompt templates similar to downstream tasks using sentences sampled from a large-scale unsupervised corpus, and apply knowledge distillation to distill from outputs of larger pre-trained models as the training objective. We conduct extensive experiments and show that SMASH can make a 6-layer DistilRoBRETa-base achieve comparable performance on few-shot datasets with a 12-layer RoBERTa-base at a low cost.

Automatic product attribute value extraction refers to the task of identifying values of an attribute from the product information. Product attributes are essential in improving online shopping experience for customers. Most existing methods focus on extracting attribute values from product title and description.However, in many real-world applications, a product is usually represented by multiple modalities beyond title and description, such as product specifications, text and visual information from the product image, etc. In this paper, we propose SMARTAVE, a Structure Mltimodal trAnsformeR for producT Attribute Value Extraction, which jointly encodes the structured product information from multiple modalities. Specifically, in SMARTAVE encoder, we introduce hyper-tokens to represent the modality-level information, and local-tokens to represent the original text and visual inputs. Structured attention patterns are designed among the hyper-tokens and local-tokens for learning effective product representation. The attribute values are then extracted based on the learned embeddings. We conduct extensive experiments on two multimodal product datasets. Experimental results demonstrate the superior performance of the proposed approach over several state-of-the-art methods. Ablation studies validate the effectiveness of the structured attentions in modeling the multimodal product information.

To understand what kinds of linguistic knowledge are encoded by pretrained Chinese language models (LMs), we introduce the benchmark of Sino LINGuistics (SLING), which consists of 38K minimal sentence pairs in Mandarin Chinese grouped into 9 high-level linguistic phenomena. Each pair demonstrates the acceptability contrast of a specific syntactic or semantic phenomenon (e.g., The keys are lost vs. The keys is lost), and an LM should assign lower perplexity to the acceptable sentence. In contrast to the CLiMP dataset (Xiang et al., 2021), which also contains Chinese minimal pairs and was created by translating the vocabulary of the English BLiMP dataset, the minimal pairs in SLING are derived primarily by applying syntactic and lexical transformations to naturally-occurring, linguist-annotated sentences from the Chinese Treebank 9.0, thus addressing severe issues in CLiMP’s data generation process. We test 18 publicly available pretrained monolingual (e.g., BERT-base-zh, CPM) and multi-lingual (e.g., mT5, XLM) language models on SLING. Our experiments show that the average accuracy for LMs is far below human performance (69.7% vs. 97.1%), while BERT-base-zh achieves the highest accuracy (84.8%) of all tested LMs, even much larger ones. Additionally, we find that most LMs have a strong gender and number (singular/plural) bias, and they perform better on local phenomena than hierarchical ones.

Large multilingual language models generally demonstrate impressive results in zero-shot cross-lingual transfer, yet often fail to successfully transfer to low-resource languages, even for token-level prediction tasks like named entity recognition (NER). In this work, we introduce a simple yet highly effective approach for improving zero-shot transfer for NER to low-resource languages. We observe that NER fine-tuning in the source language decontextualizes token representations, i.e., tokens increasingly attend to themselves. This increased reliance on token information itself, we hypothesize, triggers a type of overfitting to properties that NE tokens within the source languages share, but are generally not present in NE mentions of target languages. As a remedy, we propose a simple yet very effective sliced fine-tuning for NER (SLICER) that forces stronger token contextualization in the Transformer: we divide the transformed token representations and classifier into disjoint slices that are then independently classified during training. We evaluate SLICER on two standard benchmarks for NER that involve low-resource languages, WikiANN and MasakhaNER, and show that it (i) indeed reduces decontextualization (i.e., extent to which NE tokens attend to themselves), consequently (ii) yielding consistent transfer gains, especially prominent for low-resource target languages distant from the source language.

We present SLATE, a sequence labeling approach for extracting tasks from free-form content such as digitally handwritten (or “inked”) notes on a virtual whiteboard. Our approach allows us to create a single, low-latency model to simultaneously perform sentence segmentation and classification of these sentences into task/non-task sentences. SLATE greatly outperforms a baseline two-model (sentence segmentation followed by classification model) approach, achieving a task F1 score of 84.4%, a sentence segmentation (boundary similarity) score of 88.4% and three times lower latency compared to the baseline. Furthermore, we provide insights into tackling challenges of performing NLP on the inking domain. We release both our code and dataset for this novel task.

The large population of home cooks with dietary restrictions is under-served by existing cooking resources and recipe generation models. To help them, we propose the task of controllable recipe editing: adapt a base recipe to satisfy a user-specified dietary constraint. This task is challenging, and cannot be adequately solved with human-written ingredient substitution rules or existing end-to-end recipe generation models. We tackle this problem with SHARE: a System for Hierarchical Assistive Recipe Editing, which performs simultaneous ingredient substitution before generating natural-language steps using the edited ingredients. By decoupling ingredient and step editing, our step generator can explicitly integrate the available ingredients. Experiments on the novel RecipePairs dataset—83K pairs of similar recipes where each recipe satisfies one of seven dietary constraints—demonstrate that SHARE produces convincing, coherent recipes that are appropriate for a target dietary constraint. We further show through human evaluations and real-world cooking trials that recipes edited by SHARE can be easily followed by home cooks to create appealing dishes.

This paper investigates the sentiment analysis task from a novel perspective by incorporating sentiment knowledge and eye movement into a graph architecture, aiming to draw the eye movement-based sentiment relationships for learning the sentiment expression of the context. To be specific, we first explore a linguistic probing eye movement paradigm to extract eye movement features based on the close relationship between linguistic features and the early and late processes of human reading behavior. Furthermore, to derive eye movement features with sentiment concepts, we devise a novel weighting strategy to integrate sentiment scores extracted from affective commonsense knowledge into eye movement features, called sentiment-eye movement weights. Then, the sentiment-eye movement weights are exploited to build the sentiment-eye movement guided graph (SEMGraph) model, so as to model the intricate sentiment relationships in the context. Experimental results on two sentiment analysis datasets with eye movement signals and three sentiment analysis datasets without eye movement signals show that the proposed SEMGraph achieves state-of-the-art performance, and can also be directly generalized to those sentiment analysis datasets without eye movement signals.

Recent work has introduced an important yet relatively under-explored NLP task called Semantic Overlap Summarization (SOS) that entails generating a summary from multiple alternative narratives which conveys the common information provided by those narratives. Previous work also published a benchmark dataset for this task by collecting 2,925 alternative narrative pairs from the web and manually annotating 411 different reference summaries by engaging human annotators. In this paper, we exclusively focus on the automated evaluation of the SOS task using the benchmark dataset. More specifically, we first use the popular ROUGE metric from text-summarization literature and conduct a systematic study to evaluate the SOS task. Our experiments discover that ROUGE is not suitable for this novel task and therefore, we propose a new sentence-level precision-recall style automated evaluation metric, called SEM-F1 (Semantic F1). It is inspired by the benefits of the sentence-wise annotation technique using overlap labels reported by the previous work. Our experiments show that the proposed SEM-F1 metric yields a higher correlation with human judgment and higher inter-rater agreement compared to the ROUGE metric.

When recalling life experiences, people often forget or confuse life events, which necessitates information recall services. Previous work on information recall focuses on providing such assistance reactively, i.e., by retrieving the life event of a given query. Proactively detecting the need for information recall services is rarely discussed. In this paper, we use a human-annotated life experience retelling dataset to detect the right time to trigger the information recall service. We propose a pilot model—structured event enhancement network (SEEN) that detects life event inconsistency, additional information in life events, and forgotten events. A fusing mechanism is also proposed to incorporate event graphs of stories and enhance the textual representations. To explain the need detection results, SEEN simultaneously provides support evidence by selecting the related nodes from the event graph. Experimental results show that SEEN achieves promising performance in detecting information needs. In addition, the extracted evidence can be served as complementary information to remind users what events they may want to recall.

NLP benchmarks have largely focused on short texts, such as sentences and paragraphs, even though long texts comprise a considerable amount of natural language in the wild. We introduce SCROLLS, a suite of tasks that require reasoning over long texts. We examine existing long-text datasets, and handpick ones where the text is naturally long, while prioritizing tasks that involve synthesizing information across the input. SCROLLS contains summarization, question answering, and natural language inference tasks, covering multiple domains, including literature, science, business, and entertainment. Initial baselines, including Longformer Encoder-Decoder, indicate that there is ample room for improvement on SCROLLS. We make all datasets available in a unified text-to-text format and host a live leaderboard to facilitate research on model architecture and pretraining methods.

Self-training methods have been explored in recent years and have exhibited great performance in improving semi-supervised learning. This work presents a simple instance-adaptive self-training method (SAT) for semi-supervised text classification. SAT first generates two augmented views for each unlabeled data, and then trains a meta learner to automatically identify the relative strength of augmentations based on the similarity between the original view and the augmented views. The weakly-augmented view is fed to the model to produce a pseudo-label and the strongly-augmented view is used to train the model to predict the same pseudo-label. We conducted extensive experiments and analyses on three text classification datasets and found that with varying sizes of labeled training data, SAT consistently shows competitive performance compared to existing semi-supervised learning methods.

Traditional machine translation (MT) metrics provide an average measure of translation quality that is insensitive to the long tail of behavioral problems. Examples include translation of numbers, physical units, dropped content and hallucinations. These errors, which occur rarely and unpredictably in Neural Machine Translation (NMT), greatly undermine the reliability of state-of-the-art MT systems. Consequently, it is important to have visibility into these problems during model development.Towards this end, we introduce SALTED, a specifications-based framework for behavioral testing of NMT models. At the core of our approach is the use of high-precision detectors that flag errors (or alternatively, verify output correctness) between a source sentence and a system output. These detectors provide fine-grained measurements of long-tail errors, providing a trustworthy view of problems that were previously invisible. We demonstrate that such detectors could be used not just to identify salient long-tail errors in MT systems, but also for higher-recall filtering of the training data, fixing targeted errors with model fine-tuning in NMT and generating novel data for metamorphic testing to elicit further bugs in models.

Linguistic acceptability (LA) attracts the attention of the research community due to its many uses, such as testing the grammatical knowledge of language models and filtering implausible texts with acceptability classifiers.However, the application scope of LA in languages other than English is limited due to the lack of high-quality resources.To this end, we introduce the Russian Corpus of Linguistic Acceptability (RuCoLA), built from the ground up under the well-established binary LA approach. RuCoLA consists of 9.8k in-domain sentences from linguistic publications and 3.6k out-of-domain sentences produced by generative models. The out-of-domain set is created to facilitate the practical use of acceptability for improving language generation.Our paper describes the data collection protocol and presents a fine-grained analysis of acceptability classification experiments with a range of baseline approaches.In particular, we demonstrate that the most widely used language models still fall behind humans by a large margin, especially when detecting morphological and semantic errors. We release RuCoLA, the code of experiments, and a public leaderboard to assess the linguistic competence of language models for Russian.

As multimodal learning finds applications in a wide variety of high-stakes societal tasks, investigating their robustness becomes important. Existing work has focused on understanding the robustness of vision-and-language models to imperceptible variations on benchmark tasks. In this work, we investigate the robustness of multimodal classifiers to cross-modal dilutions – a plausible variation. We develop a model that, given a multimodal (image + text) input, generates additional dilution text that (a) maintains relevance and topical coherence with the image and existing text, and (b) when added to the original text, leads to misclassification of the multimodal input. Via experiments on Crisis Humanitarianism and Sentiment Detection tasks, we find that the performance of task-specific fusion-based multimodal classifiers drops by 23.3% and 22.5%, respectively, in the presence of dilutions generated by our model. Metric-based comparisons with several baselines and human evaluations indicate that our dilutions show higher relevance and topical coherence, while simultaneously being more effective at demonstrating the brittleness of the multimodal classifiers. Our work aims to highlight and encourage further research on the robustness of deep multimodal models to realistic variations, especially in human-facing societal applications.

Demonstration-based learning has shown great potential in stimulating pretrained language models’ ability under limited data scenario. Simply augmenting the input with some demonstrations can significantly improve performance on few-shot NER. However, why such demonstrations are beneficial for the learning process remains unclear since there is no explicit alignment between the demonstrations and the predictions. In this paper, we design pathological demonstrations by gradually removing intuitively useful information from the standard ones to take a deep dive of the robustness of demonstration-based sequence labeling and show that (1) demonstrations composed of random tokens still make the model a better few-shot learner; (2) the length of random demonstrations and the relevance of random tokens are the main factors affecting the performance; (3) demonstrations increase the confidence of model predictions on captured superficial patterns. We have publicly released our code at https://github.com/SALT-NLP/RobustDemo.

For text classification tasks, finetuned language models perform remarkably well. Yet, they tend to rely on spurious patterns in training data, thus limiting their performance on out-of-distribution (OOD) test data. Among recent models aiming to avoid this spurious pattern problem, adding extra counterfactual samples to the training data has proven to be very effective. Yet, counterfactual data generation is costly since it relies on human annotation. Thus, we propose a novel solution that only requires annotation of a small fraction (e.g., 1%) of the original training data, and uses automatic generation of extra counterfactuals in an encoding vector space. We demonstrate the effectiveness of our approach in sentiment classification, using IMDb data for training and other sets for OOD tests (i.e., Amazon, SemEval and Yelp). We achieve noticeable accuracy improvements by adding only 1% manual counterfactuals: +3% compared to adding +100% in-distribution training samples, +1.3% compared to alternate counterfactual approaches.

Transformers have been shown to be able to perform deductive reasoning on inputs containing rules and statements written in the English natural language. However, it is unclear if these models indeed follow rigorous logical reasoning to arrive at the prediction or rely on spurious correlation patterns in making decisions. A strong deductive reasoning model should consistently understand the semantics of different logical operators. To this end, we present RobustLR, a diagnostic benchmark that evaluates the robustness of language models to minimal logical edits in the inputs and different logical equivalence conditions. In our experiments with RoBERTa, T5, and GPT3 we show that the models trained on deductive reasoning datasets do not perform consistently on the RobustLR test set, thus showing that the models are not robust to our proposed logical perturbations. Further, we observe that the models find it especially hard to learn logical negation operators. Our results demonstrate the shortcomings of current language models in logical reasoning and call for the development of better inductive biases to teach the logical semantics to language models. All the datasets and code base have been made publicly available.

Data artifacts incentivize machine learning models to learn non-transferable generalizations by taking advantage of shortcuts in the data, andthere is growing evidence that data artifacts play a role for the strong results that deep learning models achieve in recent natural language processing benchmarks.In this paper, we focus on task-oriented dialogue and investigate whether popular datasets such as MultiWOZ contain such data artifacts.We found that by only keeping frequent phrases in the trainingexamples, state-of-the-art models perform similarly compared to the variant trained with full data, suggesting they exploit these spurious correlationsto solve the task. Motivated by this, we propose a contrastive learning based framework to encourage the model to ignore these cues and focus on learning generalisable patterns. We also experiment with adversarial filtering to remove easy training instances so that the model would focus on learning from the harder instances. We conduct a number of generalization experiments — e.g., cross-domain/dataset and adversarial tests — to assess the robustness of our approach and found that it works exceptionally well.

A deployed question answering (QA) model can easily fail when the test data has a distribution shift compared to the training data. Robustness tuning (RT) methods have been widely studied to enhance model robustness against distribution shifts before model deployment. However, can we improve a model after deployment? To answer this question, we evaluate test-time adaptation (TTA) to improve a model after deployment. We first introduce ColdQA, a unified evaluation benchmark for robust QA against text corruption and changes in language and domain. We then evaluate previous TTA methods on ColdQA and compare them to RT methods. We also propose a novel TTA method called online imitation learning (OIL). Through extensive experiments, we find that TTA is comparable to RT methods, and applying TTA after RT can significantly boost the performance on ColdQA. Our proposed OIL improves TTA to be more robust to variation in hyper-parameters and test distributions over time.