COLING 2022 Papers

Recently, Transformer has achieved great success in Chinese named entity recognition (NER) owing to its good parallelism and ability to model long-range dependencies, which utilizes self-attention to encode context. However, the fully connected way of self-attention may scatter the attention distribution and allow some irrelevant character information to be integrated, leading to entity boundaries being misidentified. In this paper, we propose a data-driven Adaptive Threshold Selective Self-Attention (ATSSA) mechanism that aims to dynamically select the most relevant characters to enhance the Transformer architecture for Chinese NER. In ATSSA, the attention score threshold of each query is automatically generated, and characters with attention score higher than the threshold are selected by the query while others are discarded, so as to address irrelevant attention integration. Experiments on four benchmark Chinese NER datasets show that the proposed ATSSA brings 1.68 average F1 score improvements to the baseline model and achieves state-of-the-art performance.

When a natural language generation (NLG) component is implemented in a real-world task-oriented dialogue system, it is necessary to generate not only natural utterances as learned on training data but also utterances adapted to the dialogue environment (e.g., noise from environmental sounds) and the user (e.g., users with low levels of understanding ability). Inspired by recent advances in reinforcement learning (RL) for language generation tasks, we propose ANTOR, a method for Adaptive Natural language generation for Task-Oriented dialogue via Reinforcement learning. In ANTOR, a natural language understanding (NLU) module, which corresponds to the user’s understanding of system utterances, is incorporated into the objective function of RL. If the NLG’s intentions are correctly conveyed to the NLU, which understands a system’s utterances, the NLG is given a positive reward. We conducted experiments on the MultiWOZ dataset, and we confirmed that ANTOR could generate adaptive utterances against speech recognition errors and the different vocabulary levels of users.

Few-shot text classification aims to classify the text under the few-shot scenario. Most of the previous methods adopt optimization-based meta learning to obtain task distribution. However, due to the neglect of matching between the few amount of samples and complicated models, as well as the distinction between useful and useless task features, these methods suffer from the overfitting issue. To address this issue, we propose a novel Adaptive Meta-learner via Gradient Similarity (AMGS) method to improve the model generalization ability to a new task. Specifically, the proposed AMGS alleviates the overfitting based on two aspects: (i) acquiring the potential semantic representation of samples and improving model generalization through the self-supervised auxiliary task in the inner loop, (ii) leveraging the adaptive meta-learner via gradient similarity to add constraints on the gradient obtained by base-learner in the outer loop. Moreover, we make a systematic analysis of the influence of regularization on the entire framework. Experimental results on several benchmarks demonstrate that the proposed AMGS consistently improves few-shot text classification performance compared with the state-of-the-art optimization-based meta-learning approaches. The code is available at: https://github.com/Tianyi-Lei.

Asymmetric text matching has becoming increasingly indispensable for many downstream tasks (e.g., IR and NLP). Here, asymmetry means that the documents involved for matching hold different amounts of information, e.g., a short query against a relatively longer document. The existing solutions mainly focus on modeling the feature interactions between asymmetric texts, but rarely go one step further to recognize discriminative features and perform feature denoising to enhance relevance learning. In this paper, we propose a novel adaptive feature discrimination and denoising model for asymmetric text matching, called ADDAX. For each asymmetric text pair, ADDAX is devised to explicitly distinguish discriminative features and filter out irrelevant features in a context-aware fashion. Concretely, a matching-adapted gating siamese cell (MAGS) is firstly devised to identify discriminative features and produce the corresponding hybrid representations for a text pair. Afterwards, we introduce a locality-constrained hashing denoiser to perform feature-level denoising by learning a discriminative low-dimensional binary codes for redundantly longer text. Extensive experiments on four real-world datasets from different downstream tasks demostrate that the proposed ADDAX obtains substantial performance gain over 36 up-to-date state-of-the-art alternatives.

Multilingual neural machine translation can translate unseen language pairs during training, i.e. zero-shot translation. However, the zero-shot translation is always unstable. Although prior works attributed the instability to the domination of central language, e.g. English, we supplement this viewpoint with the strict dependence of non-centered languages. In this work, we propose a simple, lightweight yet effective language-specific modeling method by adapting to non-centered languages and combining the shared information and the language-specific information to counteract the instability of zero-shot translation. Experiments with Transformer on IWSLT17, Europarl, TED talks, and OPUS-100 datasets show that our method not only performs better than strong baselines in centered data conditions but also can easily fit non-centered data conditions. By further investigating the layer attribution, we show that our proposed method can disentangle the coupled representation in the correct direction.

Multilingual pre-trained language models (PLMs) have demonstrated impressive performance on several downstream tasks for both high-resourced and low-resourced languages. However, there is still a large performance drop for languages unseen during pre-training, especially African languages. One of the most effective approaches to adapt to a new language is language adaptive fine-tuning (LAFT) — fine-tuning a multilingual PLM on monolingual texts of a language using the pre-training objective. However, adapting to target language individually takes large disk space and limits the cross-lingual transfer abilities of the resulting models because they have been specialized for a single language. In this paper, we perform multilingual adaptive fine-tuning on 17 most-resourced African languages and three other high-resource languages widely spoken on the African continent to encourage cross-lingual transfer learning. To further specialize the multilingual PLM, we removed vocabulary tokens from the embedding layer that corresponds to non-African writing scripts before MAFT, thus reducing the model size by around 50%. Our evaluation on two multilingual PLMs (AfriBERTa and XLM-R) and three NLP tasks (NER, news topic classification, and sentiment classification) shows that our approach is competitive to applying LAFT on individual languages while requiring significantly less disk space. Additionally, we show that our adapted PLM also improves the zero-shot cross-lingual transfer abilities of parameter efficient fine-tuning methods.

News recommender systems face certain challenges. These challenges arise due to evolving users’ preferences over dynamically created news articles. The diversity is necessary for a news recommender system to expose users to a variety of information. We propose a deep neural network based on a two-tower architecture that learns news representation through a news item tower and users’ representations through a query tower. We customize an augmented vector for each query and news item to introduce information interaction between the two towers. We introduce diversity in the proposed architecture by considering a category loss function that aligns items’ representation of uneven news categories. Experimental results on two news datasets reveal that our proposed architecture is more effective compared to the state-of-the-art methods and achieves a balance between accuracy and diversity.

Cross-lingual Abstract Meaning Representation (AMR) parsers are currently evaluated in comparison to gold English AMRs, despite parsing a language other than English, due to the lack of multilingual AMR evaluation metrics. This evaluation practice is problematic because of the established effect of source language on AMR structure. In this work, we present three multilingual adaptations of monolingual AMR evaluation metrics and compare the performance of these metrics to sentence-level human judgments. We then use our most highly correlated metric to evaluate the output of state-of-the-art cross-lingual AMR parsers, finding that Smatch may still be a useful metric in comparison to gold English AMRs, while our multilingual adaptation of S2match (XS2match) is best for comparison with gold in-language AMRs.

Conditional computation algorithms, such as the early exiting (EE) algorithm, can be applied to accelerate the inference of pretrained language models (PLMs) while maintaining competitive performance on resource-constrained devices. However, this approach is only applied to the vertical architecture to decide which layers should be used for inference. Conversely, the operation of the horizontal perspective is ignored, and the determination of which tokens in each layer should participate in the computation fails, leading to a high redundancy for adaptive inference. To address this limitation, a unified horizontal and vertical multi-perspective early exiting (MPEE) framework is proposed in this study to accelerate the inference of transformer-based models. Specifically, the vertical architecture uses recycling EE classifier memory and weighted self-distillation to enhance the performance of the EE classifiers. Then, the horizontal perspective uses recycling class attention memory to emphasize the informative tokens. Conversely, the tokens with less information are truncated by weighted fusion and isolated from the following computation. Based on this, both horizontal and vertical EE are unified to obtain a better tradeoff between performance and efficiency. Extensive experimental results show that MPEE can achieve higher acceleration inference with competent performance than existing competitive methods.

Using style-transfer models to reduce offensiveness of social media comments can help foster a more inclusive environment. However, there are no sizable datasets that contain offensive texts and their inoffensive counterparts, and fine-tuning pretrained models with limited labeled data can lead to the loss of original meaning in the style-transferred text. To address this issue, we provide two major contributions. First, we release the first publicly-available, parallel corpus of offensive Reddit comments and their style-transferred counterparts annotated by expert sociolinguists. Then, we introduce the first discourse-aware style-transfer models that can effectively reduce offensiveness in Reddit text while preserving the meaning of the original text. These models are the first to examine inferential links between the comment and the text it is replying to when transferring the style of offensive Reddit text. We propose two different methods of integrating discourse relations with pretrained transformer models and evaluate them on our dataset of offensive comments from Reddit and their inoffensive counterparts. Improvements over the baseline with respect to both automatic metrics and human evaluation indicate that our discourse-aware models are better at preserving meaning in style-transferred text when compared to the state-of-the-art discourse-agnostic models.

In recent years, multimodal sentiment analysis (MSA) has attracted more and more interest, which aims to predict the sentiment polarity expressed in a video. Existing methods typically 1) treat three modal features (textual, acoustic, visual) equally, without distinguishing the importance of different modalities; and 2) split the video into frames, leading to missing the global acoustic information. In this paper, we propose a global Acoustic feature enhanced Modal-Order-Aware network (AMOA) to address these problems. Firstly, a modal-order-aware network is designed to obtain the multimodal fusion feature. This network integrates the three modalities in a certain order, which makes the modality at the core position matter more. Then, we introduce the global acoustic feature of the whole video into our model. Since the global acoustic feature and multimodal fusion feature originally reside in their own spaces, contrastive learning is further employed to align them before concatenation. Experiments on two public datasets show that our model outperforms the state-of-the-art models. In addition, we also generalize our model to the sentiment with more complex semantics, such as sarcasm detection. Our model also achieves state-of-the-art performance on a widely used sarcasm dataset.

Lexical simplification (LS) is the task of automatically replacing complex words for easier ones making texts more accessible to various target populations (e.g. individuals with low literacy, individuals with learning disabilities, second language learners). To train and test models, LS systems usually require corpora that feature complex words in context along with their potential substitutions. To continue improving the performance of LS systems we introduce ALEXSIS-PT, a novel multi-candidate dataset for Brazilian Portuguese LS containing 9,605 candidate substitutions for 387 complex words. ALEXSIS-PT has been compiled following the ALEXSIS-ES protocol for Spanish opening exciting new avenues for cross-lingual models. ALEXSIS-PT is the first LS multi-candidate dataset that contains Brazilian newspaper articles. We evaluated three models for substitute generation on this dataset, namely mBERT, XLM-R, and BERTimbau. The latter achieved the highest performance across all evaluation metrics.

Artificial agents are nowadays challenged to perform embodied AI tasks. To succeed, agents must understand the meaning of verbs and how their corresponding actions transform the surrounding world. In this work, we propose ACT-Thor, a novel controlled benchmark for embodied action understanding. We use the AI2-THOR simulated environment to produce a controlled setup in which an agent, given a before-image and an associated action command, has to determine what the correct after-image is among a set of possible candidates. First, we assess the feasibility of the task via a human evaluation that resulted in 81.4% accuracy, and very high inter-annotator agreement (84.9%). Second, we design both unimodal and multimodal baselines, using state-of-the-art visual feature extractors. Our evaluation and error analysis suggest that only models that have a very structured representation of the actions together with powerful visual features can perform well on the task. However, they still fall behind human performance in a zero-shot scenario where the model is exposed to unseen (action, object) pairs. This paves the way for a systematic way of evaluating embodied AI agents that understand grounded actions.

Text infilling aims to restore incomplete texts by filling in blanks, which has attracted more attention recently because of its wide application in ancient text restoration and text rewriting. However, attribute- aware text infilling is yet to be explored, and existing methods seldom focus on the infilling length of each blank or the number/location of blanks. In this paper, we propose an Attribute-aware Text Infilling method via a Pre-trained language model (A-TIP), which contains a text infilling component and a plug- and-play discriminator. Specifically, we first design a unified text infilling component with modified attention mechanisms and intra- and inter-blank positional encoding to better perceive the number of blanks and the infilling length for each blank. Then, we propose a plug-and-play discriminator to guide generation towards the direction of improving attribute relevance without decreasing text fluency. Finally, automatic and human evaluations on three open-source datasets indicate that A-TIP achieves state-of- the-art performance compared with all baselines.

In recent years, there has been an increasing interest in claim detection as an important building block for misinformation detection. This involves detecting more fine-grained attributes relating to the claim, such as the claimer, claim topic, claim object pertaining to the topic, etc. Yet, a notable bottleneck of existing claim detection approaches is their portability to emerging events and low-resource training data settings. In this regard, we propose a fine-grained claim detection framework that leverages zero-shot Question Answering (QA) using directed questions to solve a diverse set of sub-tasks such as topic filtering, claim object detection, and claimer detection. We show that our approach significantly outperforms various zero-shot, few-shot and task-specific baselines on the NewsClaims benchmark (Reddy et al., 2021).

Predicting difficulty of questions is crucial for technical interviews. However, such questions are long-form and more open-ended than factoid and multiple choice questions explored so far for question difficulty prediction. Existing models also require large volumes of candidate response data for training. We study weak-supervision and use unsupervised algorithms for both question generation and difficulty prediction. We create a dataset of interview questions with difficulty scores for deep learning and use it to evaluate SOTA models for question difficulty prediction trained using weak supervision. Our analysis brings out the task’s difficulty as well as the promise of weak supervision for it.

Fake news’s quick propagation on social media brings severe social ramifications and economic damage. Previous fake news detection usually learn semantic and structural patterns within a single target propagation tree. However, they are usually limited in narrow signals since they do not consider latent information cross other propagation trees. Motivated by a common phenomenon that most fake news is published around a specific hot event/topic, this paper develops a new concept of propagation forest to naturally combine propagation trees in a semantic-aware clustering. We propose a novel Unified Propagation Forest-based framework (UniPF) to fully explore latent correlations between propagation trees to improve fake news detection. Besides, we design a root-induced training strategy, which encourages representations of propagation trees to be closer to their prototypical root nodes. Extensive experiments on four benchmarks consistently suggest the effectiveness and scalability of UniPF.

Framing is a communication strategy to bias discussion by selecting and emphasizing. Frame detection aims to automatically analyze framing strategy. Previous works on frame detection mainly focus on a single scenario or issue, ignoring the special characteristics of frame detection that new events emerge continuously and policy agenda changes dynamically. To better deal with various context and frame typologies across different issues, we propose a two-stage adaptation framework. In the framing domain adaptation from pre-training stage, we design two tasks based on pivots and prompts to learn a transferable encoder, verbalizer, and prompts. In the downstream scenario generalization stage, the transferable components are applied to new issues and label sets. Experiment results demonstrate the effectiveness of our framework in different scenarios. Also, it shows superiority both in full-resource and low-resource conditions.

Complex Question Understanding (CQU) parses complex questions to Question Decomposition Meaning Representation (QDMR) which is a sequence of atomic operators. Existing works are based on end-to-end neural models which do not explicitly model the intermediate states and lack interpretability for the parsing process. Besides, they predict QDMR in a mismatched granularity and do not model the step-wise information which is an essential characteristic of QDMR. To alleviate the issues, we treat QDMR as a computational graph and propose a transition-based method where a decider predicts a sequence of actions to build the graph node-by-node. In this way, the partial graph at each step enables better representation of the intermediate states and better interpretability. At each step, the decider encodes the intermediate state with specially designed encoders and predicts several candidates of the next action and its confidence. For inference, a searcher seeks the optimal graph based on the predictions of the decider to alleviate the error propagation. Experimental results demonstrate the parsing accuracy of our method against several strong baselines. Moreover, our method has transparent and human-readable intermediate results, showing improved interpretability.