Papers

Unsupervised sentence representation learning is a fundamental problem in natural language processing. Recently, contrastive learning has made great success on this task. Existing constrastive learning based models usually apply random sampling to select negative examples for training. Previous work in computer vision has shown that hard negative examples help contrastive learning to achieve faster convergency and better optimization for representation learning. However, the importance of hard negatives in contrastive learning for sentence representation is yet to be explored. In this study, we prove that hard negatives are essential for maintaining strong gradient signals in the training process while random sampling negative examples is ineffective for sentence representation. Accordingly, we present a contrastive model, MixCSE, that extends the current state-of-the-art SimCSE by continually constructing hard negatives via mixing both positive and negative features. The superior performance of the proposed approach is demonstrated via empirical studies on Semantic Textual Similarity datasets and Transfer task datasets.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

State of the art neural methods for open information extraction (OpenIE) usually extract triplets (or tuples) iteratively in an autoregressive or predicate-based manner in order not to produce duplicates. In this work, we propose a different approach to the problem that can be equally or more successful. Namely, we present a novel single-pass method for OpenIE inspired by object detection algorithms from computer vision. We use an order-agnostic loss based on bipartite matching that forces unique predictions and a Transformer-based encoder-only architecture for sequence labeling. The proposed approach is faster and shows superior or similar performance in comparison with state of the art models on standard benchmarks in terms of both quality metrics and inference time. Our model sets the new state of the art performance of 67.7% F1 on CaRB evaluated as OIE2016 while being 3.35x faster at inference than previous state of the art. We also evaluate the multilingual version of~our model in~the zero-shot setting for two languages and introduce a~strategy for generating synthetic multilingual data to fine-tune the model for each specific language. In this setting, we show performance improvement of 15\% on multilingual Re-OIE2016, reaching 75% F1 for both Portuguese and Spanish languages. Code and models are available at https://github.com/sberbank-ai/DetIE.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

We introduce VAST, the Valence-Assessing Semantics Test, a novel intrinsic evaluation task for contextualized word embeddings (CWEs). Despite the widespread use of contextualizing language models (LMs), researchers have no intrinsic evaluation task for understanding the semantic quality of CWEs and their unique properties as related to contextualization, the change in the vector representation of a word based on surrounding words; tokenization, the breaking of uncommon words into subcomponents; and LM-specific geometry learned during training. VAST uses valence, the association of a word with pleasantness, to measure the correspondence of word-level LM semantics with widely used human judgments, and examines the effects of contextualization, tokenization, and LM-specific geometry. Because prior research has found that CWEs from OpenAI's 2019 English-language causal LM GPT-2 perform poorly on other intrinsic evaluations, we select GPT-2 as our primary subject, and include results showing that VAST is useful for 7 other LMs, and can be used in 7 languages. GPT-2 results show that the semantics of a word are more similar to the semantics of context in layers closer to model output, such that VAST scores diverge between our contextual settings, ranging from Pearson’s rho of .55 to .77 in layer 11. We also show that multiply tokenized words are not semantically encoded until layer 8, where they achieve Pearson’s rho of .46, indicating the presence of an encoding process for multiply tokenized words which differs from that of singly tokenized words, for which rho is highest in layer 0. We find that a few neurons with values having greater magnitude than the rest mask word-level semantics in GPT-2’s top layer, but that word-level semantics can be recovered by nullifying non-semantic principal components: Pearson’s rho in the top layer improves from .32 to .76. Downstream POS tagging and sentence classification experiments indicate that the GPT-2 uses these principal components for non-semantic purposes, such as to represent sentence-level syntax relevant to next-word prediction. After isolating semantics, we show the utility of VAST for understanding LM semantics via improvements over related work on four word similarity tasks, with a score of .50 on SimLex-999, better than the previous best of .45 for GPT-2. Finally, we show that 8 of 10 WEAT bias tests, which compare differences in word embedding associations between groups of words, exhibit more stereotype-congruent biases after isolating semantics, indicating that non-semantic structures in LMs also mask social biases.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

So far, named entity recognition (NER) has been involved with three major types, including flat, overlapped (aka. nested), and discontinuous NER, which have mostly been studied individually. Recently, a growing interest has been built for unified NER, tackling the above three jobs concurrently with one single model. Current best-performing methods mainly include span-based and sequence-to-sequence models, where unfortunately the former merely focus on boundary identification and the latter may suffer from exposure bias. In this work, we present a novel alternative by modeling the unified NER as word-word relation classification, namely W^2NER. The architecture resolves the kernel bottleneck of unified NER by effectively modeling the neighboring relations between entity words with Next-Neighboring-Word (NNW) and Tail-Head-Word-* (THW-*) relations. Based on the W^2NER scheme we develop a neural framework, in which the unified NER is modeled as a 2D grid of word pairs. We then propose multi-granularity 2D convolutions for better refining the grid representations. Finally, a co-predictor is used to sufficiently reason the word-word relations. We perform extensive experiments on 14 widely-used benchmark datasets for flat, overlapped, and discontinuous NER (8 English and 6 Chinese datasets), where our model beats all the current top-performing baselines, pushing the state-of-the-art performances of unified NER.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

Data-driven methods have achieved notable performance on intent detection, which is a task to comprehend user queries. Nonetheless, they are controversial for over-confident predictions. In some scenarios, users do not only care about the accuracy but also the confidence of model. Unfortunately, mainstream neural networks are poorly calibrated, with a large gap between accuracy and confidence. To handle this problem defined as confidence calibration, we propose a model using the hyperspherical space and rebalanced accuracy-uncertainty loss. Specifically, we project the label vector onto hyperspherical space uniformly to generate a dense label representation matrix, which mitigates over-confident predictions due to overfitting sparse one-hot label matrix. Besides, we rebalance samples of different accuracy and uncertainty to better guide model training. Experiments on the open datasets verify that our model outperforms the existing calibration methods and achieves a significant improvement on the calibration metric.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

Pre-trained Language Models (PLMs) have achieved great success in various Natural Language Processing (NLP) tasks under the pre-training and fine-tuning paradigm. With large quantities of parameters, PLMs are computation-intensive and resource-hungry. Hence, model pruning has been introduced to compress large-scale PLMs. However, most prior approaches only consider task-specific knowledge towards downstream tasks, but ignore the essential task-agnostic knowledge during pruning, which may cause catastrophic forgetting problem and lead to poor generalization ability. To maintain both task-agnostic and task-specific knowledge in our pruned model, we propose ContrAstive Pruning (CAP) under the paradigm of pre-training and fine-tuning. It is designed as a general framework, compatible with both structured and unstructured pruning. Unified in contrastive learn- ing, CAP enables the pruned model to learn from the pre-trained model for task-agnostic knowledge, and fine-tuned model for task-specific knowledge. Besides, to better retain the performance of the pruned model, the snapshots (i.e., the intermediate models at each pruning iteration) also serve as effective supervisions for pruning. Our extensive experiments show that adopting CAP consistently yields significant improvements, especially in extremely high sparsity scenarios. With only 3% model parameters reserved (i.e., 97% sparsity), CAP successfully achieves 99.2% and 96.3% of the original BERT performance in QQP and MNLI tasks. In addition, our probing experiments demonstrate that the model pruned by CAP tends to achieve better generalization ability.

poster session 3 @ red 4, poster session 12 @ red 4, oral session 3 @ red 4, poster session 3, poster session 12, oral session 3

Pre-training language models (LMs) on large-scale unlabeled text data makes the model much easier to achieve exceptional downstream performance than their counterparts directly trained on the downstream tasks. In this work, we study what specific traits in the pre-training data, other than the semantics, make a pre-trained LM superior to their counterparts trained from scratch on downstream tasks. We propose to use artificially constructed datasets as the pre-training data to exclude the effect of semantics, and further control what characteristics the pre-training corpora have. By fine-tuning the pre-trained models on GLUE benchmark, we can learn how beneficial it is to transfer the knowledge from the model trained on the dataset possessing that specific trait. We define and discuss three different characteristics in the artificial dataset: 1) matching the token's uni-gram or bi-gram distribution between pre-training and downstream fine-tuning, 2) the presence of the explicit dependencies among the tokens in a sequence, 3) the length of the implicit dependencies among the tokens in a sequence. Our experiments show that the explicit dependencies in the sequences of the pre-training data are critical to the downstream performance. Our results also reveal that models achieve better downstream performance when pre-trained on a dataset with a longer range of implicit dependencies. Based on our analysis, we find that models pre-trained with artificial datasets are prone to learn spurious correlation in downstream tasks. Our work reveals that even if the LMs are not pre-trained on natural language, they still gain transferability on certain human language downstream tasks once the LMs learn to model the token dependencies in the sequences. This result helps us understand the exceptional transferability of pre-trained LMs.

poster session 3 @ red 4, poster session 12 @ red 4, oral session 12 @ red 4, poster session 3, poster session 12, oral session 12

Neural Module Networks (NMNs) have been quite successful in incorporating explicit reasoning as learnable modules in various question answering tasks, including the most generic form of numerical reasoning over text in Machine Reading Comprehension (MRC). However to achieve this, contemporary Neural Module Networks models obtain strong supervision in form of specialized program annotation from the QA pairs through various heuristic parsing and exhaustive computation of all possible discrete operations on discrete arguments. Consequently they fail to generalize to more open-ended settings without such supervision. Hence, we propose Weakly Supervised Neuro-Symbolic Module Network (WNSMN) trained with answers as the sole supervision for numerical reasoning based MRC. WNSMN learns to execute a noisy heuristic program obtained from the dependency parse of the query, as discrete actions over both neural and symbolic reasoning modules and trains it end-to-end in a reinforcement learning framework with discrete reward from answer matching. On the subset of DROP having numerical answers, WNSMN outperforms NMN by 32% and the reasoning-free generative language model GenBERT by 8% in exact match accuracy under comparable weakly supervised settings. This showcases the effectiveness of modular networks that can handle explicit discrete reasoning over noisy programs in an end-to-end manner.

poster session 3 @ red 4, poster session 12 @ red 4, oral session 12 @ red 4, poster session 3, poster session 12, oral session 12

A riddle is a question or statement with double or veiled meanings, followed by an unexpected answer. Solving riddle is a challenging task for both machine and human, testing the capability of understanding figurative, creative natural language and reasoning with commonsense knowledge. We introduce BiRdQA, a bilingual multiple-choice question answering dataset with 6614 English riddles and 8751 Chinese riddles. For each riddle-answer pair, we provide four distractors with additional information from Wikipedia. The distractors are automatically generated at scale with minimal bias. Existing monolingual and multilingual QA models fail to perform well on our dataset, indicating that there is a long way to go before machine can beat human on solving tricky riddles. The dataset has been released to the community.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

Recently, graph neural networks (GNNs) have been widely used for document classification. However, most existing methods are based on static word co-occurrence graphs without sentence-level information, which poses three challenges:(1) word ambiguity, (2) word synonymity, and (3) dynamic contextual dependency. To address these challenges, we propose a novel GNN-based sparse structure learning model for inductive document classification. Specifically, a document-level graph is initially generated by a disjoint union of sentence-level word co-occurrence graphs. Our model collects a set of trainable edges connecting disjoint words between sentences, and employs structure learning to sparsely select edges with dynamic contextual dependencies. Graphs with sparse structure can jointly exploit local and global contextual information in documents through GNNs.For inductive learning, the refined document graph is further fed into a general readout function for graph-level classification and optimization in an end-to-end manner. Extensive experiments on several real-world datasets demonstrate that the proposed model outperforms most state-of-the-art results, and reveal the necessity to learn sparse structures for each document.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

Recent works have shown explainability and robustness are two crucial ingredients of trustworthy and reliable text classification. However, previous works usually address one of two aspects: i) how to extract accurate rationales for explainability while being beneficial to prediction; ii) how to make the predictive model robust to different types of adversarial attacks. Intuitively, a model that produces helpful explanations should be more robust against adversarial attacks, because we cannot trust the model that outputs explanations but changes its prediction under small perturbations. To this end, we propose a joint classification and rationale extraction model named AT-BMC. It includes two key mechanisms: mixed Adversarial Training (AT) is designed to use various perturbations in discrete and embedding space to improve the model’s robustness, and Boundary Match Constraint (BMC) helps to locate rationales more precisely with the guidance of boundary information. Performances on benchmark datasets demonstrate that the proposed AT-BMC outperforms baselines on both classification and rationale extraction by a large margin. Robustness analysis shows that the proposed AT-BMC decreases the attack success rate effectively by up to 69%. The results indicate that there are connections between robust models and better explanations.

poster session 3 @ red 4, poster session 12 @ red 4, oral session 3 @ red 4, poster session 3, poster session 12, oral session 3

Few-shot text classification has recently been promoted by the meta-learning paradigm which aims to identify target classes with knowledge transferred from source classes with sets of small tasks named episodes. Despite their success, existing works building their meta-learner based on Prototypical Networks are unsatisfactory in learning discriminative text representations between similar classes, which may lead to contradictions during label prediction. In addition, the task-level and instance-level overfitting problems in few-shot text classification caused by a few training examples are not sufficiently tackled. In this work, we propose a contrastive learning framework named ContrastNet to tackle both discriminative representation and overfitting problems in few-shot text classification. ContrastNet learns to pull closer text representations belonging to the same class and push away text representations belonging to different classes, while simultaneously introducing unsupervised contrastive regularization at both task-level and instance-level to prevent overfitting. Experiments on 8 few-shot text classification datasets show that ContrastNet outperforms the current state-of-the-art models.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

Despite the super-human accuracy of recent deep models in NLP tasks, their robustness is reportedly limited due to their reliance on spurious patterns. We thus aim to leverage contrastive learning and counterfactual augmentation for robustness. For augmentation, existing work either requires humans to add counterfactuals to the dataset or machines to automatically matches near-counterfactuals already in the dataset. Unlike existing augmentation is affected by spurious correlations, ours, by synthesizing “a set” of counterfactuals, and making a collective decision on the distribution of predictions on this set, can robustly supervise the causality of each term. Our empirical results show that our approach, by collective decisions, is less sensitive to task model bias of attribution-based synthesis, and thus achieves significant improvements, in diverse dimensions: 1) counterfactual robustness, 2) cross-domain generalization, and 3) generalization from scarce data.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

Understanding events entails recognizing the structural and temporal orders between event mentions to build event structures/graphs for input documents. To achieve this goal, our work addresses the problems of subevent relation extraction (SRE) and temporal event relation extraction (TRE) that aim to predict subevent and temporal relations between two given event mentions/triggers in texts. Recent state-of-the-art methods for such problems have employed transformer-based language models (e.g., BERT) to induce effective contextual representations for input event mention pairs. However, a major limitation of existing transformer-based models for SRE and TRE is that they can only encode input texts of limited length (i.e., up to 512 sub-tokens in BERT), thus unable to effectively capture important context sentences that are farther away in the documents. In this work, we introduce a novel method to better model document-level context with important context sentences for event-event relation extraction. Our method seeks to identify the most important context sentences for a given entity mention pair in a document and pack them into shorter documents to be consume entirely by transformer-based language models for representation learning. The REINFORCE algorithm is employed to train models where novel reward functions are presented to capture model performance, and context-based and knowledge-based similarity between sentences for our problem. Extensive experiments demonstrate the effectiveness of the proposed method with state-of-the-art performance on benchmark datasets.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

Volitionality and subject animacy are fundamental and closely related properties of events. Their classification, however, is challenging because it requires contextual text understanding and a huge amount of labeled data. This paper proposes a novel method that jointly learns volitionality and subject animacy at a low cost, heuristically labeling events in a raw corpus. Volitionality labels are assigned using a small lexicon of volitional and non-volitional adverbs such as deliberately and accidentally; subject animacy labels are assigned using a list of animate and inanimate nouns obtained from ontological knowledge. Since our labeling method assigns labels only to a biased set of events, a classifier is trained with regularization to take into account the property. This paper explores the following two approaches: bias reduction and adversarial representation learning. In bias reduction, the words used for labeling are regarded as bias that should not be over-exploited to make predictions, and their estimated contribution towards predictions is penalized. In adversarial representation learning, the classifier is given unlabeled events as well and makes their latent representations closer to labeled events' ones in an adversarial manner while learning classification on labeled events. We conduct experiments with crowdsourced gold data in Japanese and English and show that our method effectively learns volitionality and subject animacy without manually labeled data.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

Explanations shed light on a machine learning model's rationales and can aid in identifying deficiencies in its reasoning process. Explanation generation models are typically trained in a supervised way given human explanations. When such annotations are not available, explanations are often selected as those portions of the input that maximise a downstream task's performance, which corresponds to optimising an explanation's Faithfulness to a given model. Faithfulness is one of several so-called diagnostic properties, which prior work has identified as useful for gauging the quality of an explanation without requiring annotations. Other diagnostic properties are Data Consistency, which measures how similar explanations are for similar input instances, and Confidence Indication, which shows whether the explanation reflects the confidence of the model. In this work, we show how to directly optimise for these diagnostic properties when training a model to generate sentence-level explanations, which markedly improves explanation quality, agreement with human rationales, and downstream task performance on three complex reasoning tasks.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

We study graph-based approaches to span-based semantic role labeling. This task is difficult due to the need to enumerate all possible predicate-argument pairs and the high degree of imbalance between positive and negative samples. Based on these difficulties, high-order inference that considers interactions between multiple arguments and predicates is often deemed beneficial but has rarely been used in span-based semantic role labeling. Because even for second-order inference, there are already O(n^5) parts for a sentence of length n, and exact high-order inference is intractable. In this paper, we propose a framework consisting of two networks: a predicate-agnostic argument pruning network that reduces the number of candidate arguments to O(n), and a semantic role labeling network with an optional second-order decoder that is unfolded from an approximate inference algorithm. Our experiments show that our framework achieves significant and consistent improvement over previous approaches.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

Traditional text classification requires thousands of annotated data or an additional Neural Machine Translation (NMT) system, which are expensive to obtain in real applications. This paper presents a Contrast-Enhanced Semi-supervised Text Classification (CEST) framework under label-limited settings without incorporating any NMT systems. We propose a certainty-driven sample selection method and a contrast-enhanced similarity graph to utilize data more efficiently in self-training, alleviating the annotation-starving problem. The graph imposes a smoothness constraint on the unlabeled data to improve the coherence and the accuracy of pseudo-labels. Moreover, CEST formulates the training as a “learning from noisy labels” problem and performs the optimization accordingly. A salient feature of this formulation is the explicit suppression of the severe error propagation problem in conventional semi-supervised learning. With solely 30 labeled data per class for both training and validation dataset, CEST outperforms the previous state-of-the-art algorithms by 2.11% accuracy and only falls within the 3.04% accuracy range of fully-supervised pre-training language model fine-tuning on thousands of labeled data.

poster session 3 @ red 4, poster session 12 @ red 4, oral session 3 @ red 4, poster session 3, poster session 12, oral session 3

Acquiring high-quality temporal common sense (TCS) knowledge from free-form text is a crucial but challenging problem for event-centric natural language understanding, due to the language reporting bias problem: people rarely report the obvious events but highlight the special cases. For example, one may rarely report ``I get off the bed in 3 seconds'', but we can observe ``It takes me an hour to get off the bed every morning'' in text. This phenomenon can result in incomplete and biased TCS knowledge. Existing works address this issue mainly by exploiting the interactions among temporal dimensions (e.g., duration, the temporal relation between events) in a multi-task framework. However, this line of work suffers the limitation of implicit, inadequate, and unexplainable interactions. In this paper, we propose a novel Soft Logic Enhanced Event Temporal Reasoning (SLEER) model for acquiring unbiased TCS knowledge, in which the complementary relationship among dimensions are explicitly represented as logic rules and modeled by t-norm fuzzy logics. SLEER can utilize logic rules to regularize its inference process. Experimental results on two intrinsic datasets and two extrinsic datasets show the efficiency of our proposed method.

poster session 3 @ red 4, poster session 12 @ red 4, poster session 3, poster session 12

Question answering over semi-structured tables has attracted significant attention in the NLP community. However, most of the existing work focus on questions that can be answered with short-form answer, i.e. the answer is often a table cell or aggregation of multiple cells. This can mismatch with the intents of users who want to ask more complex questions that require free-form answers such as explanations. To bridge the gap, most recently, pre-trained sequence-to-sequence language models such as T5 are used for generating free-form answers based on the question and table inputs. However, these pre-trained language models have weaker encoding abilities over table cells and schema. To mitigate this issue, in this work, we present an intermediate pre-training framework, Generation-focused Table-based Intermediate Pre-training (GENTAP), that jointly learns representations of natural language questions and tables. GENTAP learns to generate via two training objectives to enhance the question understanding and table representation abilities for complex questions. Based on experimental results, models that leverage GENTAP framework outperform the existing baselines on FETAQA benchmark. The pre-trained models are not only useful for free-form question answering, but also for few-shot data-to-text generation task, thus showing good transfer ability by obtaining new state-of-the-art results.

poster session 3 @ red 4, poster session 12 @ red 4, oral session 12 @ red 4, poster session 3, poster session 12, oral session 12