EMNLP 2021 论文列表

Paraphrase generation has benefited extensively from recent progress in the designing of training objectives and model architectures. However, previous explorations have largely focused on supervised methods, which require a large amount of labeled data that is costly to collect. To address this drawback, we adopt a transfer learning approach and propose a training pipeline that enables pre-trained language models to generate high-quality paraphrases in an unsupervised setting. Our recipe consists of task-adaptation, self-supervision, and a novel decoding algorithm named Dynamic Blocking (DB). To enforce a surface form dissimilar from the input, whenever the language model emits a token contained in the source sequence, DB prevents the model from outputting the subsequent source token for the next generation step. We show with automatic and human evaluations that our approach achieves state-of-the-art performance on both the Quora Question Pair (QQP) and the ParaNMT datasets and is robust to domain shift between the two datasets of distinct distributions. We also demonstrate that our model transfers to paraphrasing in other languages without any additional finetuning.

Unsupervised consistency training is a way of semi-supervised learning that encourages consistency in model predictions between the original and augmented data. For Named Entity Recognition (NER), existing approaches augment the input sequence with token replacement, assuming annotations on the replaced positions unchanged. In this paper, we explore the use of paraphrasing as a more principled data augmentation scheme for NER unsupervised consistency training. Specifically, we convert Conditional Random Field (CRF) into a multi-label classification module and encourage consistency on the entity appearance between the original and paraphrased sequences. Experiments show that our method is especially effective when annotations are limited.

Machine translation usually relies on parallel corpora to provide parallel signals for training. The advent of unsupervised machine translation has brought machine translation away from this reliance, though performance still lags behind traditional supervised machine translation. In unsupervised machine translation, the model seeks symmetric language similarities as a source of weak parallel signal to achieve translation. Chomsky’s Universal Grammar theory postulates that grammar is an innate form of knowledge to humans and is governed by universal principles and constraints. Therefore, in this paper, we seek to leverage such shared grammar clues to provide more explicit language parallel signals to enhance the training of unsupervised machine translation models. Through experiments on multiple typical language pairs, we demonstrate the effectiveness of our proposed approaches.

We investigate post-OCR correction in a setting where we have access to different OCR views of the same document. The goal of this study is to understand if a pretrained language model (LM) can be used in an unsupervised way to reconcile the different OCR views such that their combination contains fewer errors than each individual view. This approach is motivated by scenarios in which unconstrained text generation for error correction is too risky. We evaluated different pretrained LMs on two datasets and found significant gains in realistic scenarios with up to 15% WER improvement over the best OCR view. We also show the importance of domain adaptation for post-OCR correction on out-of-domain documents.

Embedding based methods are widely used for unsupervised keyphrase extraction (UKE) tasks. Generally, these methods simply calculate similarities between phrase embeddings and document embedding, which is insufficient to capture different context for a more effective UKE model. In this paper, we propose a novel method for UKE, where local and global contexts are jointly modeled. From a global view, we calculate the similarity between a certain phrase and the whole document in the vector space as transitional embedding based models do. In terms of the local view, we first build a graph structure based on the document where phrases are regarded as vertices and the edges are similarities between vertices. Then, we proposed a new centrality computation method to capture local salient information based on the graph structure. Finally, we further combine the modeling of global and local context for ranking. We evaluate our models on three public benchmarks (Inspec, DUC 2001, SemEval 2010) and compare with existing state-of-the-art models. The results show that our model outperforms most models while generalizing better on input documents with different domains and length. Additional ablation study shows that both the local and global information is crucial for unsupervised keyphrase extraction tasks.

Unsupervised cross-domain dependency parsing is to accomplish domain adaptation for dependency parsing without using labeled data in target domain. Existing methods are often of the pseudo-annotation type, which generates data through self-annotation of the base model and performing iterative training. However, these methods fail to consider the change of model structure for domain adaptation. In addition, the structural information contained in the text cannot be fully exploited. To remedy these drawbacks, we propose a Semantics-Structure Adaptative Dependency Parser (SSADP), which accomplishes unsupervised cross-domain dependency parsing without relying on pseudo-annotation or data selection. In particular, we design two feature extractors to extract semantic and structural features respectively. For each type of features, a corresponding feature adaptation method is utilized to achieve domain adaptation to align the domain distribution, which effectively enhances the unsupervised cross-domain transfer capability of the model. We validate the effectiveness of our model by conducting experiments on the CODT1 and CTB9 respectively, and the results demonstrate that our model can achieve consistent performance improvement. Besides, we verify the structure transfer ability of the proposed model by introducing Weisfeiler-Lehman Test.

Unsupervised Data Augmentation (UDA) is a semisupervised technique that applies a consistency loss to penalize differences between a model’s predictions on (a) observed (unlabeled) examples; and (b) corresponding ‘noised’ examples produced via data augmentation. While UDA has gained popularity for text classification, open questions linger over which design decisions are necessary and how to extend the method to sequence labeling tasks. In this paper, we re-examine UDA and demonstrate its efficacy on several sequential tasks. Our main contribution is an empirical study of UDA to establish which components of the algorithm confer benefits in NLP. Notably, although prior work has emphasized the use of clever augmentation techniques including back-translation, we find that enforcing consistency between predictions assigned to observed and randomly substituted words often yields comparable (or greater) benefits compared to these more complex perturbation models. Furthermore, we find that applying UDA’s consistency loss affords meaningful gains without any unlabeled data at all, i.e., in a standard supervised setting. In short, UDA need not be unsupervised to realize much of its noted benefits, and does not require complex data augmentation to be effective.

Conversation disentanglement aims to separate intermingled messages into detached sessions, which is a fundamental task in understanding multi-party conversations. Existing work on conversation disentanglement relies heavily upon human-annotated datasets, which is expensive to obtain in practice. In this work, we explore training a conversation disentanglement model without referencing any human annotations. Our method is built upon the deep co-training algorithm, which consists of two neural networks: a message-pair classifier and a session classifier. The former is responsible of retrieving local relations between two messages while the latter categorizes a message to a session by capturing context-aware information. Both the two networks are initialized respectively with pseudo data built from the unannotated corpus. During the deep co-training process, we use the session classifier as a reinforcement learning component to learn a session assigning policy by maximizing the local rewards given by the message-pair classifier. For the message-pair classifier, we enrich its training data by retrieving message pairs with high confidence from the disentangled sessions predicted by the session classifier. Experimental results on the large Movie Dialogue Dataset demonstrate that our proposed approach achieves competitive performance compared to previous supervised methods. Further experiments show that the predicted disentangled conversations can promote the performance on the downstream task of multi-party response selection.

In this paper, we address unsupervised chunking as a new task of syntactic structure induction, which is helpful for understanding the linguistic structures of human languages as well as processing low-resource languages. We propose a knowledge-transfer approach that heuristically induces chunk labels from state-of-the-art unsupervised parsing models; a hierarchical recurrent neural network (HRNN) learns from such induced chunk labels to smooth out the noise of the heuristics. Experiments show that our approach largely bridges the gap between supervised and unsupervised chunking.

Complex question answering over knowledge base remains as a challenging task because it involves reasoning over multiple pieces of information, including intermediate entities/relations and other constraints. Previous methods simplify the SPARQL query of a question into such forms as a list or a graph, missing such constraints as “filter” and “order_by”, and present models specialized for generating those simplified forms from a given question. We instead introduce a novel approach that directly generates an executable SPARQL query without simplification, addressing the issue of generating unseen entities. We adapt large scale pre-trained encoder-decoder models and show that our method significantly outperforms the previous methods and also that our method has higher interpretability and computational efficiency than the previous methods.

Much of the world’s population experiences some form of disability during their lifetime. Caution must be exercised while designing natural language processing (NLP) systems to prevent systems from inadvertently perpetuating ableist bias against people with disabilities, i.e., prejudice that favors those with typical abilities. We report on various analyses based on word predictions of a large-scale BERT language model. Statistically significant results demonstrate that people with disabilities can be disadvantaged. Findings also explore overlapping forms of discrimination related to interconnected gender and race identities.

Intermediate layer matching is shown as an effective approach for improving knowledge distillation (KD). However, this technique applies matching in the hidden spaces of two different networks (i.e. student and teacher), which lacks clear interpretability. Moreover, intermediate layer KD cannot easily deal with other problems such as layer mapping search and architecture mismatch (i.e. it requires the teacher and student to be of the same model type). To tackle the aforementioned problems all together, we propose Universal-KD to match intermediate layers of the teacher and the student in the output space (by adding pseudo classifiers on intermediate layers) via the attention-based layer projection. By doing this, our unified approach has three merits: (i) it can be flexibly combined with current intermediate layer distillation techniques to improve their results (ii) the pseudo classifiers of the teacher can be deployed instead of extra expensive teacher assistant networks to address the capacity gap problem in KD which is a common issue when the gap between the size of the teacher and student networks becomes too large; (iii) it can be used in cross-architecture intermediate layer KD. We did comprehensive experiments in distilling BERT-base into BERT-4, RoBERTa-large into DistilRoBERTa and BERT-base into CNN and LSTM-based models. Results on the GLUE tasks show that our approach is able to outperform other KD techniques.

Simultaneous machine translation (SiMT) generates translation before reading the entire source sentence and hence it has to trade off between translation quality and latency. To fulfill the requirements of different translation quality and latency in practical applications, the previous methods usually need to train multiple SiMT models for different latency levels, resulting in large computational costs. In this paper, we propose a universal SiMT model with Mixture-of-Experts Wait-k Policy to achieve the best translation quality under arbitrary latency with only one trained model. Specifically, our method employs multi-head attention to accomplish the mixture of experts where each head is treated as a wait-k expert with its own waiting words number, and given a test latency and source inputs, the weights of the experts are accordingly adjusted to produce the best translation. Experiments on three datasets show that our method outperforms all the strong baselines under different latency, including the state-of-the-art adaptive policy.

This paper presents a novel training method, Conditional Masked Language Modeling (CMLM), to effectively learn sentence representations on large scale unlabeled corpora. CMLM integrates sentence representation learning into MLM training by conditioning on the encoded vectors of adjacent sentences. Our English CMLM model achieves state-of-the-art performance on SentEval, even outperforming models learned using supervised signals. As a fully unsupervised learning method, CMLM can be conveniently extended to a broad range of languages and domains. We find that a multilingual CMLM model co-trained with bitext retrieval (BR) and natural language inference (NLI) tasks outperforms the previous state-of-the-art multilingual models by a large margin, e.g. 10% improvement upon baseline models on cross-lingual semantic search. We explore the same language bias of the learned representations, and propose a simple, post-training and model agnostic approach to remove the language identifying information from the representation while still retaining sentence semantics.

Multilingual and cross-lingual Semantic Role Labeling (SRL) have recently garnered increasing attention as multilingual text representation techniques have become more effective and widely available. While recent work has attained growing success, results on gold multilingual benchmarks are still not easily comparable across languages, making it difficult to grasp where we stand. For example, in CoNLL-2009, the standard benchmark for multilingual SRL, language-to-language comparisons are affected by the fact that each language has its own dataset which differs from the others in size, domains, sets of labels and annotation guidelines. In this paper, we address this issue and propose UniteD-SRL, a new benchmark for multilingual and cross-lingual, span- and dependency-based SRL. UniteD-SRL provides expert-curated parallel annotations using a common predicate-argument structure inventory, allowing direct comparisons across languages and encouraging studies on cross-lingual transfer in SRL. We release UniteD-SRL v1.0 at https://github.com/SapienzaNLP/united-srl.

Effective unimodal representation and complementary crossmodal representation fusion are both important in multimodal representation learning. Prior works often modulate one modal feature to another straightforwardly and thus, underutilizing both unimodal and crossmodal representation refinements, which incurs a bottleneck of performance improvement. In this paper, Unimodal and Crossmodal Refinement Network (UCRN) is proposed to enhance both unimodal and crossmodal representations. Specifically, to improve unimodal representations, a unimodal refinement module is designed to refine modality-specific learning via iteratively updating the distribution with transformer-based attention layers. Self-quality improvement layers are followed to generate the desired weighted representations progressively. Subsequently, those unimodal representations are projected into a common latent space, regularized by a multimodal Jensen-Shannon divergence loss for better crossmodal refinement. Lastly, a crossmodal refinement module is employed to integrate all information. By hierarchical explorations on unimodal, bimodal, and trimodal interactions, UCRN is highly robust against missing modality and noisy data. Experimental results on MOSI and MOSEI datasets illustrated that the proposed UCRN outperforms recent state-of-the-art techniques and its robustness is highly preferred in real multimodal sequence fusion scenarios. Codes will be shared publicly.

Knowledge graph inference has been studied extensively due to its wide applications. It has been addressed by two lines of research, i.e., the more traditional logical rule reasoning and the more recent knowledge graph embedding (KGE). Several attempts have been made to combine KGE and logical rules for better knowledge graph inference. Unfortunately, they either simply treat logical rules as additional constraints into KGE loss or use probabilistic model to approximate the exact logical inference (i.e., MAX-SAT). Even worse, both approaches need to sample ground rules to tackle the scalability issue, as the total number of ground rules is intractable in practice, making them less effective in handling logical rules. In this paper, we propose a novel framework UniKER to address these challenges by restricting logical rules to be definite Horn rules, which can fully exploit the knowledge in logical rules and enable the mutual enhancement of logical rule-based reasoning and KGE in an extremely efficient way. Extensive experiments have demonstrated that our approach is superior to existing state-of-the-art algorithms in terms of both efficiency and effectiveness.

News recommendation techniques can help users on news platforms obtain their preferred news information. Most existing news recommendation methods rely on centrally stored user behavior data to train models and serve users. However, user data is usually highly privacy-sensitive, and centrally storing them in the news platform may raise privacy concerns and risks. In this paper, we propose a unified news recommendation framework, which can utilize user data locally stored in user clients to train models and serve users in a privacy-preserving way. Following a widely used paradigm in real-world recommender systems, our framework contains a stage for candidate news generation (i.e., recall) and a stage for candidate news ranking (i.e., ranking). At the recall stage, each client locally learns multiple interest representations from clicked news to comprehensively model user interests. These representations are uploaded to the server to recall candidate news from a large news pool, which are further distributed to the user client at the ranking stage for personalized news display. In addition, we propose an interest decomposer-aggregator method with perturbation noise to better protect private user information encoded in user interest representations. Besides, we collaboratively train both recall and ranking models on the data decentralized in a large number of user clients in a privacy-preserving way. Experiments on two real-world news datasets show that our method can outperform baseline methods and effectively protect user privacy.

Transformer-based architectures have become the de-facto standard models for a wide range of Natural Language Processing tasks. However, their memory footprint and high latency are prohibitive for efficient deployment and inference on resource-limited devices. In this work, we explore quantization for transformers. We show that transformers have unique quantization challenges – namely, high dynamic activation ranges that are difficult to represent with a low bit fixed-point format. We establish that these activations contain structured outliers in the residual connections that encourage specific attention patterns, such as attending to the special separator token. To combat these challenges, we present three solutions based on post-training quantization and quantization-aware training, each with a different set of compromises for accuracy, model size, and ease of use. In particular, we introduce a novel quantization scheme – per-embedding-group quantization. We demonstrate the effectiveness of our methods on the GLUE benchmark using BERT, establishing state-of-the-art results for post-training quantization. Finally, we show that transformer weights and embeddings can be quantized to ultra-low bit-widths, leading to significant memory savings with a minimum accuracy loss. Our source code is available at https://github.com/qualcomm-ai-research/transformer-quantization.