NAACL 2021 Papers

We annotate 17,000 SNS posts with both the writer’s subjective emotional intensity and the reader’s objective one to construct a Japanese emotion analysis dataset. In this study, we explore the difference between the emotional intensity of the writer and that of the readers with this dataset. We found that the reader cannot fully detect the emotions of the writer, especially anger and trust. In addition, experimental results in estimating the emotional intensity show that it is more difficult to estimate the writer’s subjective labels than the readers’. The large gap between the subjective and objective emotions imply the complexity of the mapping from a post to the subjective emotion intensities, which also leads to a lower performance with machine learning models.

Cross-document event coreference resolution is a foundational task for NLP applications involving multi-text processing. However, existing corpora for this task are scarce and relatively small, while annotating only modest-size clusters of documents belonging to the same topic. To complement these resources and enhance future research, we present Wikipedia Event Coreference (WEC), an efficient methodology for gathering a large-scale dataset for cross-document event coreference from Wikipedia, where coreference links are not restricted within predefined topics. We apply this methodology to the English Wikipedia and extract our large-scale WEC-Eng dataset. Notably, our dataset creation method is generic and can be applied with relatively little effort to other Wikipedia languages. To set baseline results, we develop an algorithm that adapts components of state-of-the-art models for within-document coreference resolution to the cross-document setting. Our model is suitably efficient and outperforms previously published state-of-the-art results for the task.

We investigate video-aided grammar induction, which learns a constituency parser from both unlabeled text and its corresponding video. Existing methods of multi-modal grammar induction focus on grammar induction from text-image pairs, with promising results showing that the information from static images is useful in induction. However, videos provide even richer information, including not only static objects but also actions and state changes useful for inducing verb phrases. In this paper, we explore rich features (e.g. action, object, scene, audio, face, OCR and speech) from videos, taking the recent Compound PCFG model as the baseline. We further propose a Multi-Modal Compound PCFG model (MMC-PCFG) to effectively aggregate these rich features from different modalities. Our proposed MMC-PCFG is trained end-to-end and outperforms each individual modality and previous state-of-the-art systems on three benchmarks, i.e. DiDeMo, YouCook2 and MSRVTT, confirming the effectiveness of leveraging video information for unsupervised grammar induction.

Video Question Answering (VidQA) evaluation metrics have been limited to a single-word answer or selecting a phrase from a fixed set of phrases. These metrics limit the VidQA models’ application scenario. In this work, we leverage semantic roles derived from video descriptions to mask out certain phrases, to introduce VidQAP which poses VidQA as a fill-in-the-phrase task. To enable evaluation of answer phrases, we compute the relative improvement of the predicted answer compared to an empty string. To reduce the influence of language bias in VidQA datasets, we retrieve a video having a different answer for the same question. To facilitate research, we construct ActivityNet-SRL-QA and Charades-SRL-QA and benchmark them by extending three vision-language models. We perform extensive analysis and ablative studies to guide future work. Code and data are public.

First-order meta-learning algorithms have been widely used in practice to learn initial model parameters that can be quickly adapted to new tasks due to their efficiency and effectiveness. However, existing studies find that meta-learner can overfit to some specific adaptation when we have heterogeneous tasks, leading to significantly degraded performance. In Natural Language Processing (NLP) applications, datasets are often diverse and each task has its unique characteristics. Therefore, to address the overfitting issue when applying first-order meta-learning to NLP applications, we propose to reduce the variance of the gradient estimator used in task adaptation. To this end, we develop a variance-reduced first-order meta-learning algorithm. The core of our algorithm is to introduce a novel variance reduction term to the gradient estimation when performing the task adaptation. Experiments on two NLP applications: few-shot text classification and multi-domain dialog state tracking demonstrate the superior performance of our proposed method.

Morphological analysis (MA) and lexical normalization (LN) are both important tasks for Japanese user-generated text (UGT). To evaluate and compare different MA/LN systems, we have constructed a publicly available Japanese UGT corpus. Our corpus comprises 929 sentences annotated with morphological and normalization information, along with category information we classified for frequent UGT-specific phenomena. Experiments on the corpus demonstrated the low performance of existing MA/LN methods for non-general words and non-standard forms, indicating that the corpus would be a challenging benchmark for further research on UGT.

Pre-trained contextual vision-and-language (V&L) models have achieved impressive performance on various benchmarks. However, existing models require a large amount of parallel image-caption data for pre-training. Such data are costly to collect and require cumbersome curation. Inspired by unsupervised machine translation, we investigate if a strong V&L representation model can be learned through unsupervised pre-training without image-caption corpora. In particular, we propose to conduct “mask-and-predict” pre-training on text-only and image-only corpora and introduce the object tags detected by an object recognition model as anchor points to bridge two modalities. We find that such a simple approach achieves performance close to a model pre-trained with aligned data, on four English V&L benchmarks. Our work challenges the widely held notion that aligned data is necessary for V&L pre-training, while significantly reducing the amount of supervision needed for V&L models.

Obtaining training data for multi-hop question answering (QA) is time-consuming and resource-intensive. We explore the possibility to train a well-performed multi-hop QA model without referencing any human-labeled multi-hop question-answer pairs, i.e., unsupervised multi-hop QA. We propose MQA-QG, an unsupervised framework that can generate human-like multi-hop training data from both homogeneous and heterogeneous data sources. MQA-QG generates questions by first selecting/generating relevant information from each data source and then integrating the multiple information to form a multi-hop question. Using only generated training data, we can train a competent multi-hop QA which achieves 61% and 83% of the supervised learning performance for the HybridQA and the HotpotQA dataset, respectively. We also show that pretraining the QA system with the generated data would greatly reduce the demand for human-annotated training data. Our codes are publicly available at https://github.com/teacherpeterpan/Unsupervised-Multi-hop-QA.

Measuring document similarity plays an important role in natural language processing tasks. Most existing document similarity approaches suffer from the information gap caused by context and vocabulary mismatches when comparing varying-length texts. In this paper, we propose an unsupervised concept representation learning approach to address the above issues. Specifically, we propose a novel Concept Generation Network (CGNet) to learn concept representations from the perspective of the entire text corpus. Moreover, a concept-based document matching method is proposed to leverage advances in the recognition of local phrase features and corpus-level concept features. Extensive experiments on real-world data sets demonstrate that new method can achieve a considerable improvement in comparing length-varying texts. In particular, our model achieved 6.5% better F1 Score compared to the best of the baseline models for a concept-project benchmark dataset.

Universal Semantic Tagging aims to provide lightweight unified analysis for all languages at the word level. Though the proposed annotation scheme is conceptually promising, the feasibility is only examined in four Indo–European languages. This paper is concerned with extending the annotation scheme to handle Mandarin Chinese and empirically study the plausibility of unifying meaning representations for multiple languages. We discuss a set of language-specific semantic phenomena, propose new annotation specifications and build a richly annotated corpus. The corpus consists of 1100 English–Chinese parallel sentences, where compositional semantic analysis is available for English, and another 1000 Chinese sentences which has enriched syntactic analysis. By means of the new annotations, we also evaluate a series of neural tagging models to gauge how successful semantic tagging can be: accuracies of 92.7% and 94.6% are obtained for Chinese and English respectively. The English tagging performance is remarkably better than the state-of-the-art by 7.7%.

Recent work has demonstrated the vulnerability of modern text classifiers to universal adversarial attacks, which are input-agnostic sequences of words added to text processed by classifiers. Despite being successful, the word sequences produced in such attacks are often ungrammatical and can be easily distinguished from natural text. We develop adversarial attacks that appear closer to natural English phrases and yet confuse classification systems when added to benign inputs. We leverage an adversarially regularized autoencoder (ARAE) to generate triggers and propose a gradient-based search that aims to maximize the downstream classifier’s prediction loss. Our attacks effectively reduce model accuracy on classification tasks while being less identifiable than prior models as per automatic detection metrics and human-subject studies. Our aim is to demonstrate that adversarial attacks can be made harder to detect than previously thought and to enable the development of appropriate defenses.

While cross-lingual techniques are finding increasing success in a wide range of Natural Language Processing tasks, their application to Semantic Role Labeling (SRL) has been strongly limited by the fact that each language adopts its own linguistic formalism, from PropBank for English to AnCora for Spanish and PDT-Vallex for Czech, inter alia. In this work, we address this issue and present a unified model to perform cross-lingual SRL over heterogeneous linguistic resources. Our model implicitly learns a high-quality mapping for different formalisms across diverse languages without resorting to word alignment and/or translation techniques. We find that, not only is our cross-lingual system competitive with the current state of the art but that it is also robust to low-data scenarios. Most interestingly, our unified model is able to annotate a sentence in a single forward pass with all the inventories it was trained with, providing a tool for the analysis and comparison of linguistic theories across different languages. We release our code and model at https://github.com/SapienzaNLP/unify-srl.

Code summarization and generation empower conversion between programming language (PL) and natural language (NL), while code translation avails the migration of legacy code from one PL to another. This paper introduces PLBART, a sequence-to-sequence model capable of performing a broad spectrum of program and language understanding and generation tasks. PLBART is pre-trained on an extensive collection of Java and Python functions and associated NL text via denoising autoencoding. Experiments on code summarization in the English language, code generation, and code translation in seven programming languages show that PLBART outperforms or rivals state-of-the-art models. Moreover, experiments on discriminative tasks, e.g., program repair, clone detection, and vulnerable code detection, demonstrate PLBART’s effectiveness in program understanding. Furthermore, analysis reveals that PLBART learns program syntax, style (e.g., identifier naming convention), logical flow (e.g., “if“ block inside an “else“ block is equivalent to “else if“ block) that are crucial to program semantics and thus excels even with limited annotations.

Transformer architecture achieves great success in abundant natural language processing tasks. The over-parameterization of the Transformer model has motivated plenty of works to alleviate its overfitting for superior performances. With some explorations, we find simple techniques such as dropout, can greatly boost model performance with a careful design. Therefore, in this paper, we integrate different dropout techniques into the training of Transformer models. Specifically, we propose an approach named UniDrop to unites three different dropout techniques from fine-grain to coarse-grain, i.e., feature dropout, structure dropout, and data dropout. Theoretically, we demonstrate that these three dropouts play different roles from regularization perspectives. Empirically, we conduct experiments on both neural machine translation and text classification benchmark datasets. Extensive results indicate that Transformer with UniDrop can achieve around 1.5 BLEU improvement on IWSLT14 translation tasks, and better accuracy for the classification even using strong pre-trained RoBERTa as backbone.

Negation is a core construction in natural language. Despite being very successful on many tasks, state-of-the-art pre-trained language models often handle negation incorrectly. To improve language models in this regard, we propose to augment the language modeling objective with an unlikelihood objective that is based on negated generic sentences from a raw text corpus. By training BERT with the resulting combined objective we reduce the mean top 1 error rate to 4% on the negated LAMA dataset. We also see some improvements on the negated NLI benchmarks.

The choice of negative examples is important in noise contrastive estimation. Recent works find that hard negatives—highest-scoring incorrect examples under the model—are effective in practice, but they are used without a formal justification. We develop analytical tools to understand the role of hard negatives. Specifically, we view the contrastive loss as a biased estimator of the gradient of the cross-entropy loss, and show both theoretically and empirically that setting the negative distribution to be the model distribution results in bias reduction. We also derive a general form of the score function that unifies various architectures used in text retrieval. By combining hard negatives with appropriate score functions, we obtain strong results on the challenging task of zero-shot entity linking.

Modern summarization models generate highly fluent but often factually unreliable outputs. This motivated a surge of metrics attempting to measure the factuality of automatically generated summaries. Due to the lack of common benchmarks, these metrics cannot be compared. Moreover, all these methods treat factuality as a binary concept and fail to provide deeper insights on the kinds of inconsistencies made by different systems. To address these limitations, we devise a typology of factual errors and use it to collect human annotations of generated summaries from state-of-the-art summarization systems for the CNN/DM and XSum datasets. Through these annotations we identify the proportion of different categories of factual errors and benchmark factuality metrics, showing their correlation with human judgement as well as their specific strengths and weaknesses.

Contextual word embedding models, such as BioBERT and Bio_ClinicalBERT, have achieved state-of-the-art results in biomedical natural language processing tasks by focusing their pre-training process on domain-specific corpora. However, such models do not take into consideration structured expert domain knowledge from a knowledge base. We introduce UmlsBERT, a contextual embedding model that integrates domain knowledge during the pre-training process via a novel knowledge augmentation strategy. More specifically, the augmentation on UmlsBERT with the Unified Medical Language System (UMLS) Metathesaurus is performed in two ways: i) connecting words that have the same underlying ‘concept’ in UMLS and ii) leveraging semantic type knowledge in UMLS to create clinically meaningful input embeddings. By applying these two strategies, UmlsBERT can encode clinical domain knowledge into word embeddings and outperform existing domain-specific models on common named-entity recognition (NER) and clinical natural language inference tasks.

In this work we explore Unsupervised Domain Adaptation (UDA) of pretrained language models for downstream tasks. We introduce UDALM, a fine-tuning procedure, using a mixed classification and Masked Language Model loss, that can adapt to the target domain distribution in a robust and sample efficient manner. Our experiments show that performance of models trained with the mixed loss scales with the amount of available target data and the mixed loss can be effectively used as a stopping criterion during UDA training. Furthermore, we discuss the relationship between A-distance and the target error and explore some limitations of the Domain Adversarial Training approach. Our method is evaluated on twelve domain pairs of the Amazon Reviews Sentiment dataset, yielding 91.74% accuracy, which is an 1.11% absolute improvement over the state-of-the-art.