EMNLP 2021 Papers

A flaw in QA evaluation is that annotations often only provide one gold answer. Thus, model predictions semantically equivalent to the answer but superficially different are considered incorrect. This work explores mining alias entities from knowledge bases and using them as additional gold answers (i.e., equivalent answers). We incorporate answers for two settings: evaluation with additional answers and model training with equivalent answers. We analyse three QA benchmarks: Natural Questions, TriviaQA, and SQuAD. Answer expansion increases the exact match score on all datasets for evaluation, while incorporating it helps model training over real-world datasets. We ensure the additional answers are valid through a human post hoc evaluation.

The primary paradigm for multi-task training in natural language processing is to represent the input with a shared pre-trained language model, and add a small, thin network (head) per task. Given an input, a target head is the head that is selected for outputting the final prediction. In this work, we examine the behaviour of non-target heads, that is, the output of heads when given input that belongs to a different task than the one they were trained for. We find that non-target heads exhibit emergent behaviour, which may either explain the target task, or generalize beyond their original task. For example, in a numerical reasoning task, a span extraction head extracts from the input the arguments to a computation that results in a number generated by a target generative head. In addition, a summarization head that is trained with a target question answering head, outputs query-based summaries when given a question and a context from which the answer is to be extracted. This emergent behaviour suggests that multi-task training leads to non-trivial extrapolation of skills, which can be harnessed for interpretability and generalization.

We demonstrate that, hidden within one-layer randomly weighted neural networks, there exist subnetworks that can achieve impressive performance, without ever modifying the weight initializations, on machine translation tasks. To find subnetworks for one-layer randomly weighted neural networks, we apply different binary masks to the same weight matrix to generate different layers. Hidden within a one-layer randomly weighted Transformer, we find that subnetworks that can achieve 29.45/17.29 BLEU on IWSLT14/WMT14. Using a fixed pre-trained embedding layer, the previously found subnetworks are smaller than, but can match 98%/92% (34.14/25.24 BLEU) of the performance of, a trained Transformersmall/base on IWSLT14/WMT14. Furthermore, we demonstrate the effectiveness of larger and deeper transformers in this setting, as well as the impact of different initialization methods.

Intermediate task fine-tuning has been shown to culminate in large transfer gains across many NLP tasks. With an abundance of candidate datasets as well as pre-trained language models, it has become infeasible to experiment with all combinations to find the best transfer setting. In this work, we provide a comprehensive comparison of different methods for efficiently identifying beneficial tasks for intermediate transfer learning. We focus on parameter and computationally efficient adapter settings, highlight different data-availability scenarios, and provide expense estimates for each method. We experiment with a diverse set of 42 intermediate and 11 target English classification, multiple choice, question answering, and sequence tagging tasks. Our results demonstrate that efficient embedding based methods, which rely solely on the respective datasets, outperform computational expensive few-shot fine-tuning approaches. Our best methods achieve an average Regret@3 of 1% across all target tasks, demonstrating that we are able to efficiently identify the best datasets for intermediate training.

Previous work has shown that human evaluations in NLP are notoriously under-powered. Here, we argue that there are two common factors which make this problem even worse: NLP studies usually (a) treat ordinal data as interval data and (b) operate under high variance settings while the differences they are hoping to detect are often subtle. We demonstrate through simulation that ordinal mixed effects models are better able to detect small differences between models, especially in high variance settings common in evaluations of generated texts. We release tools for researchers to conduct their own power analysis and test their assumptions. We also make recommendations for improving statistical power.

Machine reading comprehension (MRC) is a challenging NLP task for it requires to carefully deal with all linguistic granularities from word, sentence to passage. For extractive MRC, the answer span has been shown mostly determined by key evidence linguistic units, in which it is a sentence in most cases. However, we recently discovered that sentences may not be clearly defined in many languages to different extents, so that this causes so-called location unit ambiguity problem and as a result makes it difficult for the model to determine which sentence exactly contains the answer span when sentence itself has not been clearly defined at all. Taking Chinese language as a case study, we explain and analyze such a linguistic phenomenon and correspondingly propose a reader with Explicit Span-Sentence Predication to alleviate such a problem. Our proposed reader eventually helps achieve a new state-of-the-art on Chinese MRC benchmark and shows great potential in dealing with other languages.

Sarcasm and sentiment embody intrinsic uncertainty of human cognition, making joint detection of multi-modal sarcasm and sentiment a challenging task. In view of the advantages of quantum probability (QP) in modeling such uncertainty, this paper explores the potential of QP as a mathematical framework and proposes a QP driven multi-task (QPM) learning framework. The QPM framework involves a complex-valued multi-modal representation encoder, a quantum-like fusion subnetwork and a quantum measurement mechanism. Each multi-modal (e.g., textual, visual) utterance is first encoded as a quantum superposition of a set of basis terms using a complex-valued representation. Then, the quantum-like fusion subnetwork leverages quantum state composition and quantum interference to model the contextual interaction between adjacent utterances and the correlations across modalities respectively. Finally, quantum incompatible measurements are performed on the multi-modal representation of each utterance to yield the probabilistic outcomes of sarcasm and sentiment recognition. The experimental results show that our model achieves a state-of-the-art performance.

GPT-3 shows remarkable in-context learning ability of large-scale language models (LMs) trained on hundreds of billion scale data. Here we address some remaining issues less reported by the GPT-3 paper, such as a non-English LM, the performances of different sized models, and the effect of recently introduced prompt optimization on in-context learning. To achieve this, we introduce HyperCLOVA, a Korean variant of 82B GPT-3 trained on a Korean-centric corpus of 560B tokens. Enhanced by our Korean-specific tokenization, HyperCLOVA with our training configuration shows state-of-the-art in-context zero-shot and few-shot learning performances on various downstream tasks in Korean. Also, we show the performance benefits of prompt-based learning and demonstrate how it can be integrated into the prompt engineering pipeline. Then we discuss the possibility of materializing the No Code AI paradigm by providing AI prototyping capabilities to non-experts of ML by introducing HyperCLOVA studio, an interactive prompt engineering interface. Lastly, we demonstrate the potential of our methods with three successful in-house applications.

Dialog is a core building block of human natural language interactions. It contains multi-party utterances used to convey information from one party to another in a dynamic and evolving manner. The ability to compare dialogs is beneficial in many real world use cases, such as conversation analytics for contact center calls and virtual agent design. We propose a novel adaptation of the edit distance metric to the scenario of dialog similarity. Our approach takes into account various conversation aspects such as utterance semantics, conversation flow, and the participants. We evaluate this new approach and compare it to existing document similarity measures on two publicly available datasets. The results demonstrate that our method outperforms the other approaches in capturing dialog flow, and is better aligned with the human perception of conversation similarity.

Web search is an essential way for humans to obtain information, but it’s still a great challenge for machines to understand the contents of web pages. In this paper, we introduce the task of web-based structural reading comprehension. Given a web page and a question about it, the task is to find an answer from the web page. This task requires a system not only to understand the semantics of texts but also the structure of the web page. Moreover, we proposed WebSRC, a novel Web-based Structural Reading Comprehension dataset. WebSRC consists of 400K question-answer pairs, which are collected from 6.4K web pages with corresponding HTML source code, screenshots, and metadata. Each question in WebSRC requires a certain structural understanding of a web page to answer, and the answer is either a text span on the web page or yes/no. We evaluate various strong baselines on our dataset to show the difficulty of our task. We also investigate the usefulness of structural information and visual features. Our dataset and baselines have been publicly available.

Weakly-supervised text classification has received much attention in recent years for it can alleviate the heavy burden of annotating massive data. Among them, keyword-driven methods are the mainstream where user-provided keywords are exploited to generate pseudo-labels for unlabeled texts. However, existing methods treat keywords independently, thus ignore the correlation among them, which should be useful if properly exploited. In this paper, we propose a novel framework called ClassKG to explore keyword-keyword correlation on keyword graph by GNN. Our framework is an iterative process. In each iteration, we first construct a keyword graph, so the task of assigning pseudo labels is transformed to annotating keyword subgraphs. To improve the annotation quality, we introduce a self-supervised task to pretrain a subgraph annotator, and then finetune it. With the pseudo labels generated by the subgraph annotator, we then train a text classifier to classify the unlabeled texts. Finally, we re-extract keywords from the classified texts. Extensive experiments on both long-text and short-text datasets show that our method substantially outperforms the existing ones.

Knowledge-based visual question answering (VQA) requires answering questions with external knowledge in addition to the content of images. One dataset that is mostly used in evaluating knowledge-based VQA is OK-VQA, but it lacks a gold standard knowledge corpus for retrieval. Existing work leverage different knowledge bases (e.g., ConceptNet and Wikipedia) to obtain external knowledge. Because of varying knowledge bases, it is hard to fairly compare models’ performance. To address this issue, we collect a natural language knowledge base that can be used for any VQA system. Moreover, we propose a Visual Retriever-Reader pipeline to approach knowledge-based VQA. The visual retriever aims to retrieve relevant knowledge, and the visual reader seeks to predict answers based on given knowledge. We introduce various ways to retrieve knowledge using text and images and two reader styles: classification and extraction. Both the retriever and reader are trained with weak supervision. Our experimental results show that a good retriever can significantly improve the reader’s performance on the OK-VQA challenge. The code and corpus are provided in https://github.com/luomancs/retriever_reader_for_okvqa.git.

Discourse segmentation, the first step of discourse analysis, has been shown to improve results for text summarization, translation and other NLP tasks. While segmentation models for written text tend to perform well, they are not directly applicable to spontaneous, oral conversation, which has linguistic features foreign to written text. Segmentation is less studied for this type of language, where annotated data is scarce, and existing corpora more heterogeneous. We develop a weak supervision approach to adapt, using minimal annotation, a state of the art discourse segmenter trained on written text to French conversation transcripts. Supervision is given by a latent model bootstrapped by manually defined heuristic rules that use linguistic and acoustic information. The resulting model improves the original segmenter, especially in contexts where information on speaker turns is lacking or noisy, gaining up to 13% in F-score. Evaluation is performed on data like those used to define our heuristic rules, but also on transcripts from two other corpora.

Weakly supervised semantic parsing (WSP) aims at training a parser via utterance-denotation pairs. This task is challenging because it requires (1) searching consistent logical forms in a huge space; and (2) dealing with spurious logical forms. In this work, we propose Learning from Mistakes (LFM), a simple yet effective learning framework for WSP. LFM utilizes the mistakes made by a parser during searching, i.e., generating logical forms that do not execute to correct denotations, for tackling the two challenges. In a nutshell, LFM additionally trains a parser using utterance-logical form pairs created from mistakes, which can quickly bootstrap the parser to search consistent logical forms. Also, it can motivate the parser to learn the correct mapping between utterances and logical forms, thus dealing with the spuriousness of logical forms. We evaluate LFM on WikiTableQuestions, WikiSQL, and TabFact in the WSP setting. The parser trained with LFM outperforms the previous state-of-the-art semantic parsing approaches on the three datasets. Also, we find that LFM can substantially reduce the need for labeled data. Using only 10% of utterance-denotation pairs, the parser achieves 84.2 denotation accuracy on WikiSQL, which is competitive with the previous state-of-the-art approaches using 100% labeled data.

Radiology report generation aims at generating descriptive text from radiology images automatically, which may present an opportunity to improve radiology reporting and interpretation. A typical setting consists of training encoder-decoder models on image-report pairs with a cross entropy loss, which struggles to generate informative sentences for clinical diagnoses since normal findings dominate the datasets. To tackle this challenge and encourage more clinically-accurate text outputs, we propose a novel weakly supervised contrastive loss for medical report generation. Experimental results demonstrate that our method benefits from contrasting target reports with incorrect but semantically-close ones. It outperforms previous work on both clinical correctness and text generation metrics for two public benchmarks.

Standard train-dev-test splits used to benchmark multiple models against each other are ubiquitously used in Natural Language Processing (NLP). In this setup, the train data is used for training the model, the development set for evaluating different versions of the proposed model(s) during development, and the test set to confirm the answers to the main research question(s). However, the introduction of neural networks in NLP has led to a different use of these standard splits; the development set is now often used for model selection during the training procedure. Because of this, comparing multiple versions of the same model during development leads to overestimation on the development data. As an effect, people have started to compare an increasing amount of models on the test data, leading to faster overfitting and “expiration” of our test sets. We propose to use a tune-set when developing neural network methods, which can be used for model picking so that comparing the different versions of a new model can safely be done on the development data.

Unifying acoustic and linguistic representation learning has become increasingly crucial to transfer the knowledge learned on the abundance of high-resource language data for low-resource speech recognition. Existing approaches simply cascade pre-trained acoustic and language models to learn the transfer from speech to text. However, how to solve the representation discrepancy of speech and text is unexplored, which hinders the utilization of acoustic and linguistic information. Moreover, previous works simply replace the embedding layer of the pre-trained language model with the acoustic features, which may cause the catastrophic forgetting problem. In this work, we introduce Wav-BERT, a cooperative acoustic and linguistic representation learning method to fuse and utilize the contextual information of speech and text. Specifically, we unify a pre-trained acoustic model (wav2vec 2.0) and a language model (BERT) into an end-to-end trainable framework. A Representation Aggregation Module is designed to aggregate acoustic and linguistic representation, and an Embedding Attention Module is introduced to incorporate acoustic information into BERT, which can effectively facilitate the cooperation of two pre-trained models and thus boost the representation learning. Extensive experiments show that our Wav-BERT significantly outperforms the existing approaches and achieves state-of-the-art performance on low-resource speech recognition.

Transfer learning (TL) seeks to improve the learning of a data-scarce target domain by using information from source domains. However, the source and target domains usually have different data distributions, which may lead to negative transfer. To alleviate this issue, we propose a Wasserstein Selective Transfer Learning (WSTL) method. Specifically, the proposed method considers a reinforced selector to select helpful data for transfer learning. We further use a Wasserstein-based discriminator to maximize the empirical distance between the selected source data and target data. The TL module is then trained to minimize the estimated Wasserstein distance in an adversarial manner and provides domain invariant features for the reinforced selector. We adopt an evaluation metric based on the performance of the TL module as delayed reward and a Wasserstein-based metric as immediate rewards to guide the reinforced selector learning. Compared with the competing TL approaches, the proposed method selects data samples that are closer to the target domain. It also provides better state features and reward signals that lead to better performance with faster convergence. Extensive experiments on three real-world text mining tasks demonstrate the effectiveness of the proposed method.

Data annotation is a time-consuming and labor-intensive process for many NLP tasks. Although there exist various methods to produce pseudo data labels, they are often task-specific and require a decent amount of labeled data to start with. Recently, the immense language model GPT-3 with 170 billion parameters has achieved tremendous improvement across many few-shot learning tasks. In this paper, we explore ways to leverage GPT-3 as a low-cost data labeler to train other models. We find that to make the downstream model achieve the same performance on a variety of NLU and NLG tasks, it costs 50% to 96% less to use labels from GPT-3 than using labels from humans. Furthermore, we propose a novel framework of combining pseudo labels from GPT-3 with human labels, which leads to even better performance. These results present a cost-effective data labeling methodology that is generalizable to many practical applications.