NAACL 2022 论文列表

We present a procedure for learning to ground symbols from a sequence of stimuli consisting of an arbitrarily complex noun phrase (e.g. “all but one green square above both red circles.”) and its designation in the visual scene. Our distinctive approach combines: a) lazy few-shot learning to relate open-class words like green and above to their visual percepts; and b) symbolic reasoning with closed-class word categories like quantifiers and negation. We use this combination to estimate new training examples for grounding symbols that occur within a noun phrase but aren’t designated by that noun phase (e.g, red in the above example), thereby potentially gaining data efficiency. We evaluate the approach in a visual reference resolution task, in which the learner starts out unaware of concepts that are part of the domain model and how they relate to visual percepts.

Interactive summarization is a task that facilitates user-guided exploration of information within a document set. While one would like to employ state of the art neural models to improve the quality of interactive summarization, many such technologies cannot ingest the full document set or cannot operate at sufficient speed for interactivity. To that end, we propose two novel deep reinforcement learning models for the task that address, respectively, the subtask of summarizing salient information that adheres to user queries, and the subtask of listing suggested queries to assist users throughout their exploration. In particular, our models allow encoding the interactive session state and history to refrain from redundancy. Together, these models compose a state of the art solution that addresses all of the task requirements. We compare our solution to a recent interactive summarization system, and show through an experimental study involving real users that our models are able to improve informativeness while preserving positive user experience.

A key challenge in the creation and refinement of virtual assistants is the ability to mine unlabeled utterance data to discover common intents. We develop an approach to this problem that combines large-scale pre-training and multi-task learning to derive a semantic embedding that can be leveraged to identify clusters of utterances that correspond to unhandled intents. An utterance encoder is first trained with a language modeling objective and subsequently adapted to predict intent labels from a large collection of cross-domain enterprise virtual assistant data using a multi-task cosine softmax loss. Experimental evaluation shows significant advantages for this multi-step pre-training approach, with large gains in downstream clustering accuracy on new applications compared to standard sentence embedding approaches. The approach has been incorporated into an interactive discovery tool that enables visualization and exploration of intents by system analysts and builders.

Intent Detection is a crucial component of Dialogue Systems wherein the objective is to classify a user utterance into one of multiple pre-defined intents. A pre-requisite for developing an effective intent identifier is a training dataset labeled with all possible user intents. However, even skilled domain experts are often unable to foresee all possible user intents at design time and for practical applications, novel intents may have to be inferred incrementally on-the-fly from user utterances. Therefore, for any real-world dialogue system, the number of intents increases over time and new intents have to be discovered by analyzing the utterances outside the existing set of intents. In this paper, our objective is to i) detect known intent utterances from a large number of unlabeled utterance samples given a few labeled samples and ii) discover new unknown intents from the remaining unlabeled samples. Existing SOTA approaches address this problem via alternate representation learning and clustering wherein pseudo labels are used for updating the representations and clustering is used for generating the pseudo labels. Unlike existing approaches that rely on epoch wise cluster alignment, we propose an end-to-end deep contrastive clustering algorithm that jointly updates model parameters and cluster centers via supervised and self-supervised learning and optimally utilizes both labeled and unlabeled data. Our proposed approach outperforms competitive baselines on five public datasets for both settings: (i) where the number of undiscovered intents are known in advance, and (ii) where the number of intents are estimated by an algorithm. We also propose a human-in-the-loop variant of our approach for practical deployment which does not require an estimate of new intents and outperforms the end-to-end approach.

Understanding human language often necessitates understanding entities and their place in a taxonomy of knowledge—their types.Previous methods to learn entity types rely on training classifiers on datasets with coarse, noisy, and incomplete labels. We introduce a method to instill fine-grained type knowledge in language models with text-to-text pre-training on type-centric questions leveraging knowledge base documents and knowledge graphs.We create the WikiWiki dataset: entities and passages from 10M Wikipedia articles linked to the Wikidata knowledge graph with 41K types.Models trained on WikiWiki achieve state-of-the-art performance in zero-shot dialog state tracking benchmarks, accurately infer entity types in Wikipedia articles, and can discover new types deemed useful by human judges.

Spurious correlations are a threat to the trustworthiness of natural language processing systems, motivating research into methods for identifying and eliminating them. However, addressing the problem of spurious correlations requires more clarity on what they are and how they arise in language data. Gardner et al (2021) argue that the compositional nature of language implies that all correlations between labels and individual “input features” are spurious. This paper analyzes this proposal in the context of a toy example, demonstrating three distinct conditions that can give rise to feature-label correlations in a simple PCFG. Linking the toy example to a structured causal model shows that (1) feature-label correlations can arise even when the label is invariant to interventions on the feature, and (2) feature-label correlations may be absent even when the label is sensitive to interventions on the feature. Because input features will be individually correlated with labels in all but very rare circumstances, domain knowledge must be applied to identify spurious correlations that pose genuine robustness threats.

Transition-based parsers for Abstract Meaning Representation (AMR) rely on node-to-word alignments. These alignments are learned separately from parser training and require a complex pipeline of rule-based components, pre-processing, and post-processing to satisfy domain-specific constraints. Parsers also train on a point-estimate of the alignment pipeline, neglecting the uncertainty due to the inherent ambiguity of alignment. In this work we explore two avenues for overcoming these limitations. First, we propose a neural aligner for AMR that learns node-to-word alignments without relying on complex pipelines. We subsequently explore a tighter integration of aligner and parser training by considering a distribution over oracle action sequences arising from aligner uncertainty. Empirical results show this approach leads to more accurate alignments and generalization better from the AMR2.0 to AMR3.0 corpora. We attain a new state-of-the art for gold-only trained models, matching silver-trained performance without the need for beam search on AMR3.0.

In recent years, transformer-based coreference resolution systems have achieved remarkable improvements on the CoNLL dataset. However, how coreference resolvers can benefit from discourse coherence is still an open question. In this paper, we propose to incorporate centering transitions derived from centering theory in the form of a graph into a neural coreference model. Our method improves the performance over the SOTA baselines, especially on pronoun resolution in long documents, formal well-structured text, and clusters with scattered mentions.

Single-task models have proven pivotal in solving specific tasks; however, they have limitations in real-world applications where multi-tasking is necessary and domain shifts are exhibited. Recently, instructional prompts have shown significant improvement towards multi-task generalization; however, the effect of instructional prompts and Multi-Task Learning (MTL) has not been systematically studied in the biomedical domain. Motivated by this, this paper explores the impact of instructional prompts for biomedical MTL. We introduce the BoX, a collection of 32 instruction tasks for Biomedical NLP across (X) various categories. Using this meta-dataset, we propose a unified model termed as In-BoXBART, that can jointly learn all tasks of the BoX without any task-specific modules. To the best of our knowledge, this is the first attempt to propose a unified model in the biomedical domain and use instructions to achieve generalization across several biomedical tasks. Experimental results indicate that the proposed model: 1) outperforms single-task baseline by ~3% and multi-task (without instruction) baseline by ~18% on an average, and 2) shows ~23% improvement compared to single-task baseline in few-shot learning (i.e., 32 instances per task) on an average. Our analysis indicates that there is significant room for improvement across tasks in the BoX, implying the scope for future research direction.

Abstractive summarization systems leveraging pre-training language models have achieved superior results on benchmark datasets. However, such models have been shown to be more prone to hallucinate facts that are unfaithful to the input context. In this paper, we propose a method to remedy entity-level extrinsic hallucinations with Entity Coverage Control (ECC). We first compute entity coverage precision and prepend the corresponding control code for each training example, which implicitly guides the model to recognize faithfulness contents in the training phase. We further extend our method via intermediate fine-tuning on large but noisy data extracted from Wikipedia to unlock zero-shot summarization. We show that the proposed method leads to more faithful and salient abstractive summarization in supervised fine-tuning and zero-shot settings according to our experimental results on three benchmark datasets XSum, Pubmed, and SAMSum of very different domains and styles.

Negation is a common linguistic feature that is crucial in many language understanding tasks, yet it remains a hard problem due to diversity in its expression in different types of text. Recent works show that state-of-the-art NLP models underperform on samples containing negation in various tasks, and that negation detection models do not transfer well across domains. We propose a new negation-focused pre-training strategy, involving targeted data augmentation and negation masking, to better incorporate negation information into language models. Extensive experiments on common benchmarks show that our proposed approach improves negation detection performance and generalizability over the strong baseline NegBERT (Khandelwal and Sawant, 2020).

Many clinical informatics tasks that are based on electronic health records (EHR) need relevant patient cohorts to be selected based on findings, symptoms and diseases. Frequently, these conditions are described in radiology reports which can be retrieved using information retrieval (IR) methods. The latest of these techniques utilize neural IR models such as BERT trained on clinical text. However, these methods still lack semantic understanding of the underlying clinical conditions as well as ruled out findings, resulting in poor precision during retrieval. In this paper we combine clinical finding detection with supervised query match learning. Specifically, we use lexicon-driven concept detection to detect relevant findings in sentences. These findings are used as queries to train a Sentence-BERT (SBERT) model using triplet loss on matched and unmatched query-sentence pairs. We show that the proposed supervised training task remarkably improves the retrieval performance of SBERT. The trained model generalizes well to unseen queries and reports from different collections.

A notable challenge in Multi-Document Summarization (MDS) is the extremely-long length of the input. In this paper, we present an extract-then-abstract Transformer framework to overcome the problem. Specifically, we leverage pre-trained language models to construct a hierarchical extractor for salient sentence selection across documents and an abstractor for rewriting the selected contents as summaries. However, learning such a framework is challenging since the optimal contents for the abstractor are generally unknown. Previous works typically create pseudo extraction oracle to enable the supervised learning for both the extractor and the abstractor. Nevertheless, we argue that the performance of such methods could be restricted due to the insufficient information for prediction and inconsistent objectives between training and testing. To this end, we propose a loss weighting mechanism that makes the model aware of the unequal importance for the sentences not in the pseudo extraction oracle, and leverage the fine-tuned abstractor to generate summary references as auxiliary signals for learning the extractor. Moreover, we propose a reinforcement learning method that can efficiently apply to the extractor for harmonizing the optimization between training and testing. Experiment results show that our framework substantially outperforms strong baselines with comparable model sizes and achieves the best results on the Multi-News, Multi-XScience, and WikiCatSum corpora.

Self-supervised pretraining has made few-shot learning possible for many NLP tasks. But the pretraining objectives are not typically adapted specifically for in-context few-shot learning. In this paper, we propose to use self-supervision in an intermediate training stage between pretraining and downstream few-shot usage with the goal to teach the model to perform in-context few shot learning. We propose and evaluate four self-supervised objectives on two benchmarks. We find that the intermediate self-supervision stage produces models that outperform strong baselines. Ablation study shows that several factors affect the downstream performance, such as the amount of training data and the diversity of the self-supervised objectives. Human-annotated cross-task supervision and self-supervision are complementary. Qualitative analysis suggests that the self-supervised-trained models are better at following task requirements.

Few-shot relation classification is difficult because the few instances available may not represent well the relation patterns. Some existing approaches explored extra information such as relation definition, in addition to the instances, to learn a better relation representation. However, the encoding of the extra information has been performed independently from the labeled instances. In this paper, we propose to learn a prototype encoder from relation definition in a way that is useful for relation instance classification. To this end, we use a joint training approach to train both a prototype encoder from definition and an instance encoder. Extensive experiments on several datasets demonstrate the effectiveness and usefulness of our prototype encoder from definition text, enabling us to outperform state-of-the-art approaches.

Humans can obtain the knowledge of novel visual concepts from language descriptions, and we thus use the few-shot image classification task to investigate whether a machine learning model can have this capability. Our proposed model, LIDE (Learning from Image and DEscription), has a text decoder to generate the descriptions and a text encoder to obtain the text representations of machine- or user-generated descriptions. We confirmed that LIDE with machine-generated descriptions outperformed baseline models. Moreover, the performance was improved further with high-quality user-generated descriptions. The generated descriptions can be viewed as the explanations of the model’s predictions, and we observed that such explanations were consistent with prediction results. We also investigated why the language description improves the few-shot image classification performance by comparing the image representations and the text representations in the feature spaces.

Recent work in entity disambiguation (ED) has typically neglected structured knowledge base (KB) facts, and instead relied on a limited subset of KB information, such as entity descriptions or types. This limits the range of contexts in which entities can be disambiguated. To allow the use of all KB facts, as well as descriptions and types, we introduce an ED model which links entities by reasoning over a symbolic knowledge base in a fully differentiable fashion. Our model surpasses state-of-the-art baselines on six well-established ED datasets by 1.3 F1 on average. By allowing access to all KB information, our model is less reliant on popularity-based entity priors, and improves performance on the challenging ShadowLink dataset (which emphasises infrequent and ambiguous entities) by 12.7 F1.

A key challenge of Conversational Recommendation Systems (CRS) is to integrate the recommendation function and the dialog generation function smoothly. Previous works employ graph neural networks with external knowledge graphs (KG) to model individual recommendation items and integrate KGs with language models through attention mechanism for response generation. Although previous approaches prove effective, there is still room for improvement. For example, KG-based approaches only rely on entity relations and bag-of-words to recommend items and neglect the information in the conversational context. We propose to improve the usage of dialog context for both recommendation and response generation using an encoding architecture along with the self-attention mechanism of transformers. In this paper, we propose a simple yet effective architecture comprising a pre-trained language model (PLM) and an item metadata encoder to integrate the recommendation and the dialog generation better. The proposed item encoder learns to map item metadata to embeddings reflecting the rich information of the item, which can be matched with dialog context. The PLM then consumes the context-aware item embeddings and dialog context to generate high-quality recommendations and responses. Experimental results on the benchmark dataset ReDial show that our model obtains state-of-the-art results on both recommendation and response generation tasks.

Though achieving impressive results on many NLP tasks, the BERT-like masked language models (MLM) encounter the discrepancy between pre-training and inference. In light of this gap, we investigate the contextual representation of pre-training and inference from the perspective of word probability distribution. We discover that BERT risks neglecting the contextual word similarity in pre-training. To tackle this issue, we propose an auxiliary gloss regularizer module to BERT pre-training (GR-BERT), to enhance word semantic similarity. By predicting masked words and aligning contextual embeddings to corresponding glosses simultaneously, the word similarity can be explicitly modeled. We design two architectures for GR-BERT and evaluate our model in downstream tasks. Experimental results show that the gloss regularizer benefits BERT in word-level and sentence-level semantic representation. The GR-BERT achieves new state-of-the-art in lexical substitution task and greatly promotes BERT sentence representation in both unsupervised and supervised STS tasks.