COLING 2022 Papers

Question answering over knowledge bases (KBQA) for complex questions is a challenging task in natural language processing. Recently, generation-based methods that translate natural language questions to executable logical forms have achieved promising performance. These methods use auxiliary information to augment the logical form generation of questions with unseen KB items or novel combinations, but the noise introduced can also leads to more incorrect results. In this work, we propose GMT-KBQA, a Generation-based KBQA method via Multi-Task learning, to better retrieve and utilize auxiliary information. GMT-KBQA first obtains candidate entities and relations through dense retrieval, and then introduces a multi-task model which jointly learns entity disambiguation, relation classification, and logical form generation. Experimental results show that GMT-KBQA achieves state-of-the-art results on both ComplexWebQuestions and WebQuestionsSP datasets. Furthermore, the detailed evaluation demonstrates that GMT-KBQA benefits from the auxiliary tasks and has a strong generalization capability.

Multi-hop reasoning requires aggregating multiple documents to answer a complex question. Existing methods usually decompose the multi-hop question into simpler single-hop questions to solve the problem for illustrating the explainable reasoning process. However, they ignore grounding on the supporting facts of each reasoning step, which tends to generate inaccurate decompositions. In this paper, we propose an interpretable stepwise reasoning framework to incorporate both single-hop supporting sentence identification and single-hop question generation at each intermediate step, and utilize the inference of the current hop for the next until reasoning out the final result. We employ a unified reader model for both intermediate hop reasoning and final hop inference and adopt joint optimization for more accurate and robust multi-hop reasoning. We conduct experiments on two benchmark datasets HotpotQA and 2WikiMultiHopQA. The results show that our method can effectively boost performance and also yields a better interpretable reasoning process without decomposition supervision.

Explaining the predictions of a deep neural network (DNN) is a challenging problem. Many attempts at interpreting those predictions have focused on attribution-based methods, which assess the contributions of individual features to each model prediction. However, attribution-based explanations do not always provide faithful explanations to the target model, e.g., noisy gradients can result in unfaithful feature attribution for back-propagation methods. We present a method to learn explanations-specific representations while constructing deep network models for text classification. These representations can be used to faithfully interpret black-box predictions, i.e., highlighting the most important input features and their role in any particular prediction. We show that learning specific representations improves model interpretability across various tasks, for both qualitative and quantitative evaluations, while preserving predictive performance.

Summaries, keyphrases, and titles are different ways of concisely capturing the content of a document. While most previous work has released the datasets of keyphrases and summarization separately, in this work, we introduce LipKey, the largest news corpus with human-written abstractive summaries, absent keyphrases, and titles. We jointly use the three elements via multi-task training and training as joint structured inputs, in the context of document summarization. We find that including absent keyphrases and titles as additional context to the source document improves transformer-based summarization models.

Translating between languages where certain features are marked morphologically in one but absent or marked contextually in the other is an important test case for machine translation. When translating into English which marks (in)definiteness morphologically, from Yorùbá which uses bare nouns but marks these features contextually, ambiguities arise. In this work, we perform fine-grained analysis on how an SMT system compares with two NMT systems (BiLSTM and Transformer) when translating bare nouns in Yorùbá into English. We investigate how the systems what extent they identify BNs, correctly translate them, and compare with human translation patterns. We also analyze the type of errors each model makes and provide a linguistic description of these errors. We glean insights for evaluating model performance in low-resource settings. In translating bare nouns, our results show the transformer model outperforms the SMT and BiLSTM models for 4 categories, the BiLSTM outperforms the SMT model for 3 categories while the SMT outperforms the NMT models for 1 category.

This work deploys linguistically motivated features to classify paragraph-level text into fiction and non-fiction genre using a logistic regression model and infers lexical and syntactic properties that distinguish the two genres. Previous works have focused on classifying document-level text into fiction and non-fiction genres, while in this work, we deal with shorter texts which are closer to real-world applications like sentiment analysis of tweets. Going beyond simple POS tag ratios proposed in Qureshi et al.(2019) for document-level classification, we extracted multiple linguistically motivated features belonging to four categories: Lexical features, POS ratio features, Syntactic features and Raw features. For the task of short-text classification, a model containing 28 best-features (selected via Recursive feature elimination with cross-validation; RFECV) confers an accuracy jump of 15.56 % over a baseline model consisting of 2 POS-ratio features found effective in previous work (cited above). The efficacy of the above model containing a linguistically motivated feature set also transfers over to another dataset viz, Baby BNC corpus. We also compared the classification accuracy of the logistic regression model with two deep-learning models. A 1D CNN model gives an increase of 2% accuracy over the logistic Regression classifier on both corpora. And the BERT-base-uncased model gives the best classification accuracy of 97% on Brown corpus and 98% on Baby BNC corpus. Although both the deep learning models give better results in terms of classification accuracy, the problem of interpreting these models remains unsolved. In contrast, regression model coefficients revealed that fiction texts tend to have more character-level diversity and have lower lexical density (quantified using content-function word ratios) compared to non-fiction texts. Moreover, subtle differences in word order exist between the two genres, i.e., in fiction texts Verbs precede Adverbs (inter-alia).

Most NER methods rely on extensive labeled data for model training, which struggles in the low-resource scenarios with limited training data. Existing dominant approaches usually suffer from the challenge that the target domain has different label sets compared with a resource-rich source domain, which can be concluded as class transfer and domain transfer. In this paper, we propose a lightweight tuning paradigm for low-resource NER via pluggable prompting (LightNER). Specifically, we construct the unified learnable verbalizer of entity categories to generate the entity span sequence and entity categories without any label-specific classifiers, thus addressing the class transfer issue. We further propose a pluggable guidance module by incorporating learnable parameters into the self-attention layer as guidance, which can re-modulate the attention and adapt pre-trained weights. Note that we only tune those inserted module with the whole parameter of the pre-trained language model fixed, thus, making our approach lightweight and flexible for low-resource scenarios and can better transfer knowledge across domains. Experimental results show that LightNER can obtain comparable performance in the standard supervised setting and outperform strong baselines in low-resource settings.

Datasets with significant proportions of bias present threats for training a trustworthy model on NLU tasks. Despite yielding great progress, current debiasing methods impose excessive reliance on the knowledge of bias attributes. Definition of the attributes, however, is elusive and varies across different datasets. In addition, leveraging these attributes at input level to bias mitigation may leave a gap between intrinsic properties and the underlying decision rule. To narrow down this gap and liberate the supervision on bias, we suggest extending bias mitigation into feature space. Therefore, a novel model, Recovering Intended-Feature Subspace with Knowledge-Free (RISK) is developed. Assuming that shortcut features caused by various biases are unintended for prediction, RISK views them as redundant features. When delving into a lower manifold to remove redundancies, RISK reveals that an extremely low-dimensional subspace with intended features can robustly represent the highly biased dataset. Empirical results demonstrate our model can consistently improve model generalization to out-of-distribution set, and achieves a new state-of-the-art performance.

In an open-domain dialogue system, the consistent persona is a key factor to generate real and coherent dialogues. Existing methods suffer from the incomprehensive persona tags that have unique and obscure meanings to describe human’s personality. Besides, the addressee information, which is closely related to express personality in multi-party dialogues, has been neglected. In this paper, we construct a multi-party personalized dialogue dataset and propose a graph convolution network model (PersonaTKG) with addressee selecting mechanism that integrates personas, dialogue utterances, and external text knowledge in a unified graph. Extensive experiments have shown that PersonaTKG outperforms the baselines by large margins and effectively improves persona consistency in the generated responses.

Explaining the reasoning of neural models has attracted attention in recent years. Providing highly-accessible and comprehensible explanations in natural language is useful for humans to understand model’s prediction results. In this work, we present a pilot study to investigate explanation generation with a narrative and causal structure for the scenario of health consulting. Our model generates a medical suggestion regarding the patient’s concern and provides an explanation as the outline of the reasoning. To align the generated explanation with the suggestion, we propose a novel discourse-aware mechanism with multi-task learning. Experimental results show that our model achieves promising performances in both quantitative and human evaluation.

Temporal sentence grounding (TSG) is crucial and fundamental for video understanding. Previous works typically model the target activity referred to the sentence query in a video by extracting the appearance information from each whole frame. However, these methods fail to distinguish visually similar background noise and capture subtle details of small objects. Although a few recent works additionally adopt a detection model to filter out the background contents and capture local appearances of foreground objects, they rely on the quality of the detection model and suffer from the time-consuming detection process. To this end, we propose a novel detection-free framework for TSG—Grounding with Learnable Foreground (GLF), which efficiently learns to locate the foreground regions related to the query in consecutive frames for better modelling the target activity. Specifically, we first split each video frame into multiple patch candidates of equal size, and reformulate the foreground detection problem as a patch localization task. Then, we develop a self-supervised coarse-to-fine paradigm to learn to locate the most query-relevant patch in each frame and aggregate them among the video for final grounding. Further, we employ a multi-scale patch reasoning strategy to capture more fine-grained foreground information. Extensive experiments on three challenging datasets (Charades-STA, TACoS, ActivityNet) show that the proposed GLF outperforms state-of-the-art methods.

Target-oriented multimodal sentiment classification (TMSC) is a new subtask of aspect-based sentiment analysis, which aims to determine the sentiment polarity of the opinion target mentioned in a (sentence, image) pair. Recently, dominant works employ the attention mechanism to capture the corresponding visual representations of the opinion target, and then aggregate them as evidence to make sentiment predictions. However, they still suffer from two problems: (1) The granularity of the opinion target in two modalities is inconsistent, which causes visual attention sometimes fail to capture the corresponding visual representations of the target; (2) Even though it is captured, there are still significant differences between the visual representations expressing the same mood, which brings great difficulty to sentiment prediction. To this end, we propose a novel Knowledge-enhanced Framework (KEF) in this paper, which can successfully exploit adjective-noun pairs extracted from the image to improve the visual attention capability and sentiment prediction capability of the TMSC task. Extensive experimental results show that our framework consistently outperforms state-of-the-art works on two public datasets.

Knowledge graph embedding aims to represent entities and relations as low-dimensional vectors, which is an effective way for predicting missing links. It is crucial for knowledge graph embedding models to model and infer various relation patterns, such as symmetry/antisymmetry. However, many existing approaches fail to model semantic hierarchies, which are common in the real world. We propose a new model called HRQE, which represents entities as pure quaternions. The relational embedding consists of two parts: (a) Using unit quaternions to represent the rotation part in 3D space, where the head entities are rotated by the corresponding relations through Hamilton product. (b) Using scale parameters to constrain the modulus of entities to make them have hierarchical distributions. To the best of our knowledge, HRQE is the first model that can encode symmetry/antisymmetry, inversion, composition, multiple relation patterns and learn semantic hierarchies simultaneously. Experimental results demonstrate the effectiveness of HRQE against some of the SOTA methods on four well-established knowledge graph completion benchmarks.

K-Nearest Neighbor Neural Machine Translation (kNNMT) successfully incorporates external corpus by retrieving word-level representations at test time. Generally, kNNMT borrows the off-the-shelf context representation in the translation task, e.g., the output of the last decoder layer, as the query vector of the retrieval task. In this work, we highlight that coupling the representations of these two tasks is sub-optimal for fine-grained retrieval. To alleviate it, we leverage supervised contrastive learning to learn the distinctive retrieval representation derived from the original context representation. We also propose a fast and effective approach to constructing hard negative samples. Experimental results on five domains show that our approach improves the retrieval accuracy and BLEU score compared to vanilla kNNMT.

Dependency tree-based methods might be susceptible to the dependency tree due to that they inevitably introduce noisy information and neglect the rich relation information between words. In this paper, we propose a learnable dependency-based double graph (LD2G) model for aspect-based sentiment classification. We use multi-task learning for domain adaptive pretraining, which combines Biaffine Attention and Mask Language Model by incorporating features such as structure, relations and linguistic features in the sentiment text. Then we utilize the dependency enhanced double graph-based MPNN to deeply fuse structure features and relation features that are affected with each other for ASC. Experiment on four benchmark datasets shows that our model is superior to the state-of-the-art approaches.

Recent work in black-box adversarial attacks for NLP systems has attracted attention. Prior black-box attacks assume that attackers can observe output labels from target models based on selected inputs. In this work, inspired by adversarial transferability, we propose a new type of black-box NLP adversarial attack that an attacker can choose a similar domain and transfer the adversarial examples to the target domain and cause poor performance in target model. Based on domain adaptation theory, we then propose a defensive strategy, called Learn2Weight, which trains to predict the weight adjustments for target model in order to defense the attack of similar-domain adversarial examples. Using Amazon multi-domain sentiment classification dataset, we empirically show that Learn2Weight model is effective against the attack compared to standard black-box defense methods such as adversarial training and defense distillation. This work contributes to the growing literature on machine learning safety.

Pre-trained Language Models (PLMs) are the cornerstone of the modern Natural Language Processing (NLP). However, as PLMs become heavier, fine tuning all their parameters loses their efficiency. Existing parameter-efficient methods generally focus on reducing the trainable parameters in PLMs but neglect the inference speed, which limits the ability to deploy PLMs. In this paper, we propose LayerConnect (hypernetwork-assisted inter-layer connectors) to enhance inference efficiency. Specifically, a light-weight connector with a linear structure is inserted between two Transformer layers, and the parameters inside each connector are tuned by a hypernetwork comprising an interpolator and a down-sampler. We perform extensive experiments on the widely used the GLUE benchmark. The experimental results verify the inference efficiency of our model. Compared to Adapter, our model parameters are reduced to approximately 11.75%, while the performance degradation is kept to less than 5% (2.5 points on average).

The evaluation of recent embedding-based evaluation metrics for text generation is primarily based on measuring their correlation with human evaluations on standard benchmarks. However, these benchmarks are mostly from similar domains to those used for pretraining word embeddings. This raises concerns about the (lack of) generalization of embedding-based metrics to new and noisy domains that contain a different vocabulary than the pretraining data. In this paper, we examine the robustness of BERTScore, one of the most popular embedding-based metrics for text generation. We show that (a) an embedding-based metric that has the highest correlation with human evaluations on a standard benchmark can have the lowest correlation if the amount of input noise or unknown tokens increases, (b) taking embeddings from the first layer of pretrained models improves the robustness of all metrics, and (c) the highest robustness is achieved when using character-level embeddings, instead of token-based embeddings, from the first layer of the pretrained model.

We present LANTERN, a multi-stage transformer architecture for named-entity recognition (NER) designed to operate on indefinitely large text sequences (i.e. > 512 elements). For a given image of a form with structured text, our method uses language and spatial features to predict the entity tags of each text element. It breaks the quadratic computational constraints of the attention mechanism by operating over a learned latent space representation which encodes the input sequence via the cross-attention mechanism while having the multi-stage encoding component as a refinement over the NER predictions. As a proxy task, we propose RADAR, an LSTM classifier operating at character level, which predicts the relevance of a word with respect to the entity-recognition task. Additionally, we formulate a challenging novel NER use case, nutritional information extraction from food product labels. We created a dataset with 11,926 images depicting food product labels entitled TREAT dataset, with fully detailed annotations. Our method achieves superior performance against two competitive models designed for long sequences on the proposed TREAT dataset.