EMNLP 2021 Papers

I highlight a simple failure mode of state-of-the-art machine reading systems: when contexts do not align with commonly shared beliefs. For example, machine reading systems fail to answer What did Elizabeth want? correctly in the context of ‘My kingdom for a cough drop, cried Queen Elizabeth.’ Biased by co-occurrence statistics in the training data of pretrained language models, systems predict my kingdom, rather than a cough drop. I argue such biases are analogous to human belief biases and present a carefully designed challenge dataset for English machine reading, called Auto-Locke, to quantify such effects. Evaluations of machine reading systems on Auto-Locke show the pervasiveness of belief bias in machine reading.

Although researches on word embeddings have made great progress in recent years, many tasks in natural language processing are on the sentence level. Thus, it is essential to learn sentence embeddings. Recently, Sentence BERT (SBERT) is proposed to learn embeddings on the sentence level, and it uses the inner product (or, cosine similarity) to compute semantic similarity between sentences. However, this measurement cannot well describe the semantic structures among sentences. The reason is that sentences may lie on a manifold in the ambient space rather than distribute in an Euclidean space. Thus, cosine similarity cannot approximate distances on the manifold. To tackle the severe problem, we propose a novel sentence embedding method called Sentence BERT with Locality Preserving (SBERT-LP), which discovers the sentence submanifold from a high-dimensional space and yields a compact sentence representation subspace by locally preserving geometric structures of sentences. We compare the SBERT-LP with several existing sentence embedding approaches from three perspectives: sentence similarity, sentence classification and sentence clustering. Experimental results and case studies demonstrate that our method encodes sentences better in the sense of semantic structures.

Human expertise and the participation of speech communities are essential factors in the success of technologies for low-resource languages. Accordingly, we propose a new computational task which is tuned to the available knowledge and interests in an Indigenous community, and which supports the construction of high quality texts and lexicons. The task is illustrated for Kunwinjku, a morphologically-complex Australian language. We combine a finite state implementation of a published grammar with a partial lexicon, and apply this to a noisy phone representation of the signal. We locate known lexemes in the signal and use the morphological transducer to build these out into hypothetical, morphologically-complex words for human validation. We show that applying a single iteration of this method results in a relative transcription density gain of 17%. Further, we find that 75% of breath groups in the test set receive at least one correct partial or full-word suggestion.

Are pairs of words that tend to occur together also likely to stand in a linguistic dependency? This empirical question is motivated by a long history of literature in cognitive science, psycholinguistics, and NLP. In this work we contribute an extensive analysis of the relationship between linguistic dependencies and statistical dependence between words. Improving on previous work, we introduce the use of large pretrained language models to compute contextualized estimates of the pointwise mutual information between words (CPMI). For multiple models and languages, we extract dependency trees which maximize CPMI, and compare to gold standard linguistic dependencies. Overall, we find that CPMI dependencies achieve an unlabelled undirected attachment score of at most ≈ 0.5. While far above chance, and consistently above a non-contextualized PMI baseline, this score is generally comparable to a simple baseline formed by connecting adjacent words. We analyze which kinds of linguistic dependencies are best captured in CPMI dependencies, and also find marked differences between the estimates of the large pretrained language models, illustrating how their different training schemes affect the type of dependencies they capture.

Text annotation tools assume that their user’s goal is to create a labeled corpus. However,users view annotation as a necessary evil on the way to deliver business value through NLP.Thus an annotation tool should optimize for the throughput of the global NLP process, not only the productivity of individual annotators. LightTag is a text annotation tool designed and built on that principle. This paper shares our design rationale, data modeling choices, and user interface decisions then illustrates how those choices serve the full NLP lifecycle.

Lifelong Learning (LL) black-box models are dynamic in that they keep learning from new tasks and constantly update their parameters. Owing to the need to utilize information from previously seen tasks, and capture commonalities in potentially diverse data, it is hard for automatic explanation methods to explain the outcomes of these models. In addition, existing explanation methods, e.g., LIME, which are computationally expensive when explaining a static black-box model, are even more inefficient in the LL setting. In this paper, we propose a novel Lifelong Explanation (LLE) approach that continuously trains a student explainer under the supervision of a teacher – an arbitrary explanation algorithm – on different tasks undertaken in LL. We also leverage the Experience Replay (ER) mechanism to prevent catastrophic forgetting in the student explainer. Our experiments comparing LLE to three baselines on text classification tasks show that LLE can enhance the stability of the explanations for all seen tasks and maintain the same level of faithfulness to the black-box model as the teacher, while being up to 10ˆ2 times faster at test time. Our ablation study shows that the ER mechanism in our LLE approach enhances the learning capabilities of the student explainer. Our code is available at https://github.com/situsnow/LLE.

Traditional supervised Information Extraction (IE) methods can extract structured knowledge elements from unstructured data, but it is limited to a pre-defined target ontology. In reality, the ontology of interest may change over time, adding emergent new types or more fine-grained subtypes. We propose a new lifelong learning framework to address this challenge. We focus on lifelong event detection as an exemplar case and propose a new problem formulation that is also generalizable to other IE tasks. In event detection and more general IE tasks, rich correlations or semantic relatedness exist among hierarchical knowledge element types. In our proposed framework, knowledge is being transferred between learned old event types and new event types. Specifically, we update old knowledge with new event types’ mentions using a self-training loss. In addition, we aggregate old event types’ representations based on their similarities with new event types to initialize the new event types’ representations. Experimental results show that our framework outperforms competitive baselines with a 5.1% absolute gain in the F1 score. Moreover, our proposed framework can boost the F1 score for over 30% absolute gain on some new long-tail rare event types with few training instances. Our knowledge transfer module improves performance on both learned event types and new event types under the lifelong learning setting, showing that it helps consolidate old knowledge and improve novel knowledge acquisition.

Precise information of word boundary can alleviate the problem of lexical ambiguity to improve the performance of natural language processing (NLP) tasks. Thus, Chinese word segmentation (CWS) is a fundamental task in NLP. Due to the development of pre-trained language models (PLM), pre-trained knowledge can help neural methods solve the main problems of the CWS in significant measure. Existing methods have already achieved high performance on several benchmarks (e.g., Bakeoff-2005). However, recent outstanding studies are limited by the small-scale annotated corpus. To further improve the performance of CWS methods based on fine-tuning the PLMs, we propose a novel neural framework, LBGCN, which incorporates a lexicon-based graph convolutional network into the Transformer encoder. Experimental results on five benchmarks and four cross-domain datasets show the lexicon-based graph convolutional network successfully captures the information of candidate words and helps to improve performance on the benchmarks (Bakeoff-2005 and CTB6) and the cross-domain datasets (SIGHAN-2010). Further experiments and analyses demonstrate that our proposed framework effectively models the lexicon to enhance the ability of basic neural frameworks and strengthens the robustness in the cross-domain scenario.

NLP systems are often challenged by difficulties arising from noisy, non-standard, and domain specific corpora. The task of lexical normalisation aims to standardise such corpora, but currently lacks suitable tools to acquire high-quality annotated data to support deep learning based approaches. In this paper, we present LexiClean, the first open-source web-based annotation tool for multi-task lexical normalisation. LexiClean’s main contribution is support for simultaneous in situ token-level modification and annotation that can be rapidly applied corpus wide. We demonstrate the usefulness of our tool through a case study on two sets of noisy corpora derived from the specialised-domain of industrial mining. We show that LexiClean allows for the rapid and efficient development of high-quality parallel corpora. A demo of our system is available at: https://youtu.be/P7_ooKrQPDU.

In parataxis languages like Chinese, word meanings are constructed using specific word-formations, which can help to disambiguate word senses. However, such knowledge is rarely explored in previous word sense disambiguation (WSD) methods. In this paper, we propose to leverage word-formation knowledge to enhance Chinese WSD. We first construct a large-scale Chinese lexical sample WSD dataset with word-formations. Then, we propose a model FormBERT to explicitly incorporate word-formations into sense disambiguation. To further enhance generalizability, we design a word-formation predictor module in case word-formation annotations are unavailable. Experimental results show that our method brings substantial performance improvement over strong baselines.

Fine-tuning pretrained models for automatically summarizing doctor-patient conversation transcripts presents many challenges: limited training data, significant domain shift, long and noisy transcripts, and high target summary variability. In this paper, we explore the feasibility of using pretrained transformer models for automatically summarizing doctor-patient conversations directly from transcripts. We show that fluent and adequate summaries can be generated with limited training data by fine-tuning BART on a specially constructed dataset. The resulting models greatly surpass the performance of an average human annotator and the quality of previous published work for the task. We evaluate multiple methods for handling long conversations, comparing them to the obvious baseline of truncating the conversation to fit the pretrained model length limit. We introduce a multistage approach that tackles the task by learning two fine-tuned models: one for summarizing conversation chunks into partial summaries, followed by one for rewriting the collection of partial summaries into a complete summary. Using a carefully chosen fine-tuning dataset, this method is shown to be effective at handling longer conversations, improving the quality of generated summaries. We conduct both an automatic evaluation (through ROUGE and two concept-based metrics focusing on medical findings) and a human evaluation (through qualitative examples from literature, assessing hallucination, generalization, fluency, and general quality of the generated summaries).

Tag recommendation relies on either a ranking function for top-k tags or an autoregressive generation method. However, the previous methods neglect one of two seemingly conflicting yet desirable characteristics of a tag set: orderlessness and inter-dependency. While the ranking approach fails to address the inter-dependency among tags when they are ranked, the autoregressive approach fails to take orderlessness into account because it is designed to utilize sequential relations among tokens. We propose a sequence-oblivious generation method for tag recommendation, in which the next tag to be generated is independent of the order of the generated tags and the order of the ground truth tags occurring in training data. Empirical results on two different domains, Instagram and Stack Overflow, show that our method is significantly superior to the previous approaches.

This paper presents an unsupervised extractive approach to summarize scientific long documents based on the Information Bottleneck principle. Inspired by previous work which uses the Information Bottleneck principle for sentence compression, we extend it to document level summarization with two separate steps. In the first step, we use signal(s) as queries to retrieve the key content from the source document. Then, a pre-trained language model conducts further sentence search and edit to return the final extracted summaries. Importantly, our work can be flexibly extended to a multi-view framework by different signals. Automatic evaluation on three scientific document datasets verifies the effectiveness of the proposed framework. The further human evaluation suggests that the extracted summaries cover more content aspects than previous systems.

Integrating knowledge into text is a promising way to enrich text representation, especially in the medical field. However, undifferentiated knowledge not only confuses the text representation but also imports unexpected noises. In this paper, to alleviate this problem, we propose leveraging capsule routing to associate knowledge with medical literature hierarchically (called HiCapsRKL). Firstly, HiCapsRKL extracts two empirically designed text fragments from medical literature and encodes them into fragment representations respectively. Secondly, the capsule routing algorithm is applied to two fragment representations. Through the capsule computing and dynamic routing, each representation is processed into a new representation (denoted as caps-representation), and we integrate the caps-representations as information gain to associate knowledge with medical literature hierarchically. Finally, HiCapsRKL are validated on relevance prediction and medical literature retrieval test sets. The experimental results and analyses show that HiCapsRKLcan more accurately associate knowledge with medical literature than mainstream methods. In summary, HiCapsRKL can efficiently help selecting the most relevant knowledge to the medical literature, which may be an alternative attempt to improve knowledge-based text representation. Source code is released on GitHub.

Recently, sponsored search has become one of the most lucrative channels for marketing. As the fundamental basis of sponsored search, relevance modeling has attracted increasing attention due to the tremendous practical value. Most existing methods solely rely on the query-keyword pairs. However, keywords are usually short texts with scarce semantic information, which may not precisely reflect the underlying advertising intents. In this paper, we investigate the novel problem of advertiser-aware relevance modeling, which leverages the advertisers’ information to bridge the gap between the search intents and advertising purposes. Our motivation lies in incorporating the unsupervised bidding behaviors as the complementary graphs to learn desirable advertiser representations. We further propose a Bidding-Graph augmented Triple-based Relevance model BGTR with three towers to deeply fuse the bidding graphs and semantic textual data. Empirically, we evaluate the BGTR model over a large industry dataset, and the experimental results consistently demonstrate its superiority.

We propose a novel scheme to use the Levenshtein Transformer to perform the task of word-level quality estimation. A Levenshtein Transformer is a natural fit for this task: trained to perform decoding in an iterative manner, a Levenshtein Transformer can learn to post-edit without explicit supervision. To further minimize the mismatch between the translation task and the word-level QE task, we propose a two-stage transfer learning procedure on both augmented data and human post-editing data. We also propose heuristics to construct reference labels that are compatible with subword-level finetuning and inference. Results on WMT 2020 QE shared task dataset show that our proposed method has superior data efficiency under the data-constrained setting and competitive performance under the unconstrained setting.

Dense retrieval requires high-quality text sequence embeddings to support effective search in the representation space. Autoencoder-based language models are appealing in dense retrieval as they train the encoder to output high-quality embedding that can reconstruct the input texts. However, in this paper, we provide theoretical analyses and show empirically that an autoencoder language model with a low reconstruction loss may not provide good sequence representations because the decoder may take shortcuts by exploiting language patterns. To address this, we propose a new self-learning method that pre-trains the autoencoder using a weak decoder, with restricted capacity and attention flexibility to push the encoder to provide better text representations. Our experiments on web search, news recommendation, and open domain question answering show that our pre-trained model significantly boosts the effectiveness and few-shot ability of dense retrieval models. Our code is available at https://github.com/microsoft/SEED-Encoder/.

In this paper, we propose a simple few-shot domain adaptation paradigm for reading comprehension. We first identify the lottery subnetwork structure within the Transformer-based source domain model via gradual magnitude pruning. Then, we only fine-tune the lottery subnetwork, a small fraction of the whole parameters, on the annotated target domain data for adaptation. To obtain more adaptable subnetworks, we introduce self-attention attribution to weigh parameters, beyond simply pruning the smallest magnitude parameters, which can be seen as combining structured pruning and unstructured magnitude pruning softly. Experimental results show that our method outperforms the full model fine-tuning adaptation on four out of five domains when only a small amount of annotated data available for adaptation. Moreover, introducing self-attention attribution reserves more parameters for important attention heads in the lottery subnetwork and improves the target domain model performance. Our further analyses reveal that, besides exploiting fewer parameters, the choice of subnetworks is critical to the effectiveness.

When training most modern reading comprehension models, all the questions associated with a context are treated as being independent from each other. However, closely related questions and their corresponding answers are not independent, and leveraging these relationships could provide a strong supervision signal to a model. Drawing on ideas from contrastive estimation, we introduce several new supervision losses that compare question-answer scores across multiple related instances. Specifically, we normalize these scores across various neighborhoods of closely contrasting questions and/or answers, adding a cross entropy loss term in addition to traditional maximum likelihood estimation. Our techniques require bundles of related question-answer pairs, which we either mine from within existing data or create using automated heuristics. We empirically demonstrate the effectiveness of training with instance bundles on two datasets—HotpotQA and ROPES—showing up to 9% absolute gains in accuracy.