NAACL 2022 Papers

We propose a generative model for text generation, which exhibits disentangled latent representations of syntax and semantics. Contrary to previous work, this model does not need syntactic information such as constituency parses, or semantic information such as paraphrase pairs. Our model relies solely on the inductive bias found in attention-based architectures such as Transformers. In the attention of Transformers, keys handle information selection while values specify what information is conveyed. Our model, dubbed QKVAE, uses Attention in its decoder to read latent variables where one latent variable infers keys while another infers values. We run experiments on latent representations and experiments on syntax/semantics transfer which show that QKVAE displays clear signs of disentangled syntax and semantics. We also show that our model displays competitive syntax transfer capabilities when compared to supervised models and that comparable supervised models need a fairly large amount of data (more than 50K samples) to outperform it on both syntactic and semantic transfer. The code for our experiments is publicly available.

Pretrained Transformer based models finetuned on domain specific corpora have changed the landscape of NLP. However, training or fine-tuning these models for individual tasks can be time consuming and resource intensive. Thus, a lot of current research is focused on using transformers for multi-task learning (Raffel et al., 2020) and how to group the tasks to help a multi-task model to learn effective representations that can be shared across tasks (Standley et al., 2020; Fifty et al., 2021) . In this work, we show that a single multi-tasking model can match the performance of task specific model when the task specific models show similar representations across all of their hidden layers and their gradients are aligned, i.e. their gradients follow the same direction. We hypothesize that the above observations explain the effectiveness of multi-task learning. We validate our observations on our internal radiologist-annotated datasets on the cervical and lumbar spine. Our method is simple and intuitive, and can be used in a wide range of NLP problems.

In the context of data labeling, NLP researchers are increasingly interested in having humans select rationales, a subset of input tokens relevant to the chosen label. We conducted a 332-participant online user study to understand how humans select rationales, especially how different instructions and user interface affordances impact the rationales chosen. Participants labeled ten movie reviews as positive or negative, selecting words and phrases supporting their label as rationales. We varied the instructions given, the rationale-selection task, and the user interface. Participants often selected about 12% of input tokens as rationales, but selected fewer if unable to drag over multiple tokens at once. Whereas participants were near unanimous in their data labels, they were far less consistent in their rationales. The user interface affordances and task greatly impacted the types of rationales chosen. We also observed large variance across participants.

Warning: This paper contains content that is offensive and may be upsetting. Biased or toxic speech can be harmful to various demographic groups. Therefore, it is not only important for models to detect these speech, but to also output explanations of why a given text is toxic. Previous literature has mostly focused on classifying and detecting toxic speech, and existing efforts on explaining stereotypes in toxic speech mainly use standard text generation approaches, resulting in generic and repetitive explanations. Building on these prior works, we introduce a novel knowledge-informed encoder-decoder framework to utilize multiple knowledge sources to generate implications of biased text. Experiments show that our knowledge informed models outperform prior state-of-the-art models significantly, and can generate detailed explanations of stereotypes in toxic speech compared to baselines, both quantitatively and qualitatively.

There is an increasing trend in using neural methods for dialogue model evaluation. Lack of a framework to investigate these metrics can cause dialogue models to reflect their biases and cause unforeseen problems during interactions. In this work, we propose an adversarial test-suite which generates problematic variations of various dialogue aspects, e.g. logical entailment, using automatic heuristics. We show that dialogue metrics for both open-domain and task-oriented settings are biased in their assessments of different conversation behaviors and fail to properly penalize problematic conversations, by analyzing their assessments of these problematic examples. We conclude that variability in training methodologies and data-induced biases are some of the main causes of these problems. We also conduct an investigation into the metric behaviors using a black-box interpretability model which corroborates our findings and provides evidence that metrics pay attention to the problematic conversational constructs signaling a misunderstanding of different conversation semantics.

Significance testing—especially the paired-permutation test—has played a vital role in developing NLP systems to provide confidence that the difference in performance between two systems (i.e., the test statistic) is not due to luck. However, practitioners rely on Monte Carlo approximation to perform this test due to a lack of a suitable exact algorithm. In this paper, we provide an efficient exact algorithm for the paired-permutation test for a family of structured test statistics. Our algorithm runs in 𝒪(G N (log GN )(log N)) time where N is the dataset size and G is the range of the test statistic. We found that our exact algorithm was 10x faster than the Monte Carlo approximation with 20000 samples on a common dataset

Interpretability methods are developed to understand the working mechanisms of black-box models, which is crucial to their responsible deployment. Fulfilling this goal requires both that the explanations generated by these methods are correct and that people can easily and reliably understand them. While the former has been addressed in prior work, the latter is often overlooked, resulting in informal model understanding derived from a handful of local explanations. In this paper, we introduce explanation summary (ExSum), a mathematical framework for quantifying model understanding, and propose metrics for its quality assessment. On two domains, ExSum highlights various limitations in the current practice, helps develop accurate model understanding, and reveals easily overlooked properties of the model. We also connect understandability to other properties of explanations such as human alignment, robustness, and counterfactual similarity and plausibility.

Retrieval-augmented generation models have shown state-of-the-art performance across many knowledge-intensive NLP tasks such as open-domain question answering and fact verification. These models are trained to generate a final output given retrieved passages that can be irrelevant to an input query, leading to learning spurious cues or memorization. This work introduces a method to incorporate evidentiality of passages—whether a passage contains correct evidence to support the output—into training the generator. We introduce a multi-task learning framework to jointly generate the final output and predict the evidentiality of each passage. Furthermore, we introduce a new task-agnostic method for obtaining high-quality silver evidentiality labels, addressing the issues of gold evidentiality labels being unavailable in most domains. Our experiments on five datasets across three knowledge-intensive tasks show that our new evidentiality-guided generator significantly outperforms its direct counterpart on all of them, and advances the state of the art on three of them. Our analysis shows that multi-task learning and silver evidentiality mining play key roles. Our code is available at https://github.com/AkariAsai/evidentiality_qa

Event schema depicts the typical structure of complex events, serving as a scaffolding to effectively analyze, predict, and possibly intervene in the ongoing events. To induce event schemas from historical events, previous work uses an event-by-event scheme, ignoring the global structure of the entire schema graph. We propose a new event schema induction framework using double graph autoencoders, which captures the global dependencies among nodes in event graphs. Specifically, we first extract the event skeleton from an event graph and design a variational directed acyclic graph (DAG) autoencoder to learn its global structure. Then we further fill in the event arguments for the skeleton, and use another Graph Convolutional Network (GCN) based autoencoder to reconstruct entity-entity relations as well as to detect coreferential entities. By performing this two-stage induction decomposition, the model can avoid reconstructing the entire graph in one step, allowing it to focus on learning global structures between events. Experimental results on three event graph datasets demonstrate that our method achieves state-of-the-art performance and induces high-quality event schemas with global consistency.

Suicide is a serious problem in every society. Understanding life events of a potential patient is essential for successful suicide-risk assessment and prevention. In this work, we focus on the Event Detection (ED) task to identify event trigger words of suicide-related events in public posts of discussion forums. In particular, we introduce SuicideED: a new dataset for the ED task that features seven suicidal event types to comprehensively capture suicide actions and ideation, and general risk and protective factors. Our experiments with current state-of-the-art ED systems suggest that this domain poses meaningful challenges as there is significant room for improvement of ED models. We will release SuicideED to support future research in this important area.

The power and the potential of deep learning models attract many researchers to design advanced and sophisticated architectures. Nevertheless, the progress is sometimes unreal due to various possible reasons. In this work, through an astonishing example we argue that more efforts should be paid to ensure the progress in developing a new deep learning method. For a highly influential multi-label text classification method XML-CNN, we show that the superior performance claimed in the original paper was mainly due to some unbelievable coincidences. We re-examine XML-CNN and make a re-implementation which reveals some contradictory findings to the claims in the original paper. Our study suggests suitable baselines for multi-label text classification tasks and confirms that the progress on a new architecture cannot be confidently justified without a cautious investigation.

Currently available grammatical error correction (GEC) datasets are compiled using essays or other long-form text written by language learners, limiting the applicability of these datasets to other domains such as informal writing and conversational dialog. In this paper, we present a novel GEC dataset consisting of parallel original and corrected utterances drawn from open-domain chatbot conversations; this dataset is, to our knowledge, the first GEC dataset targeted to a human-machine conversational setting. We also present a detailed annotation scheme which ranks errors by perceived impact on comprehension, making our dataset more representative of real-world language learning applications. To demonstrate the utility of the dataset, we use our annotated data to fine-tune a state-of-the-art GEC model. Experimental results show the effectiveness of our data in improving GEC model performance in a conversational scenario.

Learning representations of entity mentions is a core component of modern entity linking systems for both candidate generation and making linking predictions. In this paper, we present and empirically analyze a novel training approach for learning mention and entity representations that is based on building minimum spanning arborescences (i.e., directed spanning trees) over mentions and entities across documents to explicitly model mention coreference relationships. We demonstrate the efficacy of our approach by showing significant improvements in both candidate generation recall and linking accuracy on the Zero-Shot Entity Linking dataset and MedMentions, the largest publicly available biomedical dataset. In addition, we show that our improvements in candidate generation yield higher quality re-ranking models downstream, setting a new SOTA result in linking accuracy on MedMentions. Finally, we demonstrate that our improved mention representations are also effective for the discovery of new entities via cross-document coreference.

Language models (LMs) are typically trained once on a large-scale corpus and used for years without being updated. However, in a dynamic world, new entities constantly arise. We propose a framework to analyze what LMs can infer about new entities that did not exist when the LMs were pretrained. We derive a dataset of entities indexed by their origination date and paired with their English Wikipedia articles, from which we can find sentences about each entity. We evaluate LMs’ perplexity on masked spans within these sentences. We show that models more informed about the entities, such as those with access to a textual definition of them, achieve lower perplexity on this benchmark. Our experimental results demonstrate that making inferences about new entities remains difficult for LMs. Given its wide coverage on entity knowledge and temporal indexing, our dataset can be used to evaluate LMs and techniques designed to modify or extend their knowledge. Our automatic data collection pipeline can be easily used to continually update our benchmark.

Large language models have achieved high performance on various question answering (QA) benchmarks, but the explainability of their output remains elusive. Structured explanations, called entailment trees, were recently suggested as a way to explain the reasoning behind a QA system’s answer. In order to better generate such entailment trees, we propose an architecture called Iterative Retrieval-Generation Reasoner (IRGR). Our model is able to explain a given hypothesis by systematically generating a step-by-step explanation from textual premises. The IRGR model iteratively searches for suitable premises, constructing a single entailment step at a time. Contrary to previous approaches, our method combines generation steps and retrieval of premises, allowing the model to leverage intermediate conclusions, and mitigating the input size limit of baseline encoder-decoder models. We conduct experiments using the EntailmentBank dataset, where we outperform existing benchmarks on premise retrieval and entailment tree generation, with around 300% gain in overall correctness.

Large-scale pre-trained language models have attracted extensive attentions in the research community and shown promising results on various tasks of natural language processing. However, the attention maps, which record the attention scores between tokens in self-attention mechanism, are sometimes ineffective as they are learned implicitly without the guidance of explicit semantic knowledge. Thus, we aim to infuse explicit external knowledge into pre-trained language models to further boost their performance. Existing works of knowledge infusion largely depend on multi-task learning frameworks, which are inefficient and require large-scale re-training when new knowledge is considered. In this paper, we propose a novel and generic solution, KAM-BERT, which directly incorporates knowledge-generated attention maps into the self-attention mechanism. It requires only a few extra parameters and supports efficient fine-tuning once new knowledge is added. KAM-BERT achieves consistent improvements on various academic datasets for natural language understanding. It also outperforms other state-of-the-art methods which conduct knowledge infusion into transformer-based architectures. Moreover, we apply our model to an industry-scale ad relevance application and show its advantages in the real-world scenario.

Providing conversation models with background knowledge has been shown to make open-domain dialogues more informative and engaging. Existing models treat knowledge selection as a sentence ranking or classification problem where each sentence is handled individually, ignoring the internal semantic connection between sentences. In this work, we propose to automatically convert the background knowledge documents into document semantic graphs and then perform knowledge selection over such graphs. Our document semantic graphs preserve sentence-level information through the use of sentence nodes and provide concept connections between sentences. We apply multi-task learning to perform sentence-level knowledge selection and concept-level knowledge selection, showing that it improves sentence-level selection. Our experiments show that our semantic graph-based knowledge selection improves over sentence selection baselines for both the knowledge selection task and the end-to-end response generation task on HollE and improves generalization on unseen topics in WoW.

This paper introduces a model for incomplete utterance restoration (IUR) called JET (Joint learning token Extraction and Text generation). Different from prior studies that only work on extraction or abstraction datasets, we design a simple but effective model, working for both scenarios of IUR. Our design simulates the nature of IUR, where omitted tokens from the context contribute to restoration. From this, we construct a Picker that identifies the omitted tokens. To support the picker, we design two label creation methods (soft and hard labels), which can work in cases of no annotation data for the omitted tokens. The restoration is done by using a Generator with the help of the Picker on joint learning. Promising results on four benchmark datasets in extraction and abstraction scenarios show that our model is better than the pretrained T5 and non-generative language model methods in both rich and limited training data settings.

In spoken question answering, the systems are designed to answer questions from contiguous text spans within the related speech transcripts. However, the most natural way that human seek or test their knowledge is via human conversations. Therefore, we propose a new Spoken Conversational Question Answering task (SCQA), aiming at enabling the systems to model complex dialogues flow given the speech documents. In this task, our main objective is to build the system to deal with conversational questions based on the audio recordings, and to explore the plausibility of providing more cues from different modalities with systems in information gathering. To this end, instead of directly adopting automatically generated speech transcripts with highly noisy data, we propose a novel unified data distillation approach, DDNet, which effectively ingests cross-modal information to achieve fine-grained representations of the speech and language modalities. Moreover, we propose a simple and novel mechanism, termed Dual Attention, by encouraging better alignments between audio and text to ease the process of knowledge transfer. To evaluate the capacity of SCQA systems in a dialogue-style interaction, we assemble a Spoken Conversational Question Answering (Spoken-CoQA) dataset with more than 40k question-answer pairs from 4k conversations. We first show that the performance of the existing state-of-the-art methods significantly degrade on our dataset, hence demonstrating the necessity of incorporating cross-modal information to achieve good performance gains. Our experimental results demonstrate that our proposed method achieves superior performance in spoken conversational question answering. Codes and datasets will be made publicly available.