COLING 2022 Papers

Pretrained multilingual language models can help bridge the digital language divide, enabling high-quality NLP models for lower-resourced languages. Studies of multilingual models have so far focused on performance, consistency, and cross-lingual generalisation. However, with their wide-spread application in the wild and downstream societal impact, it is important to put multilingual models under the same scrutiny as monolingual models. This work investigates the group fairness of multilingual models, asking whether these models are equally fair across languages. To this end, we create a new four-way multilingual dataset of parallel cloze test examples (MozArt), equipped with demographic information (balanced with regard to gender and native tongue) about the test participants. We evaluate three multilingual models on MozArt –mBERT, XLM-R, and mT5– and show that across the four target languages, the three models exhibit different levels of group disparity, e.g., exhibiting near-equal risk for Spanish, but high levels of disparity for German.

This paper introduces the problem of determining whether people are located in the places they mention in their tweets. In particular, we investigate the role of text and images to solve this challenging problem. We present a new corpus of tweets that contain both text and images. Our analyses show that this problem is multimodal at its core: human judgments depend on whether annotators have access to the text, the image, or both. Experimental results show that a neural architecture that combines both modalities yields better results. We also conduct an error analysis to provide insights into why and when each modality is beneficial.

Question answering on knowledge bases (KBQA) poses a unique challenge for semantic parsing research due to two intertwined challenges: large search space and ambiguities in schema linking. Conventional ranking-based KBQA models, which rely on a candidate enumeration step to reduce the search space, struggle with flexibility in predicting complicated queries and have impractical running time. In this paper, we present ArcaneQA, a novel generation-based model that addresses both the large search space and the schema linking challenges in a unified framework with two mutually boosting ingredients: dynamic program induction for tackling the large search space and dynamic contextualized encoding for schema linking. Experimental results on multiple popular KBQA datasets demonstrate the highly competitive performance of ArcaneQA in both effectiveness and efficiency.

Without labeled question-answer pairs for necessary training, unsupervised commonsense question-answering (QA) appears to be extremely challenging due to its indispensable unique prerequisite on commonsense source like knowledge bases (KBs), which are usually highly resource consuming in construction. Recently pre-trained language models (PLMs) show effectiveness as an alternative for commonsense clues when they play a role of knowledge generator. However, existing work either relies on large-scale in-domain or out-of-domain labeled data, or fails to generate knowledge of high quality in a general way. Motivated by human thinking experience, we propose an approach of All-round Thinker (ArT) by fully taking association during knowledge generating. In detail, our model first focuses on key parts in the given context, and then generates highly related knowledge on such a basis in an association way like human thinking. Besides, for casual reasoning, a reverse thinking mechanism is especially added to further enhance bidirectional inferring between cause and effect. ArT is totally unsupervised and KBs-free. We evaluate it on three commonsense QA benchmarks: COPA, SocialIQA and SCT. On all scales of PLM backbones, ArT shows its brilliant performance and outperforms previous advanced unsupervised models.

Current practices in building new NLP models for low-resourced languages rely either on Machine Translation of training sets from better resourced languages or on cross-lingual transfer from them. Still we can see a considerable performance gap between the models originally trained within better resourced languages and the models transferred from them. In this study we test the possibility of (1) using natural annotation to build synthetic training sets from resources not initially designed for the target downstream task and (2) employing curriculum learning methods to select the most suitable examples from synthetic training sets. We test this hypothesis across seven Slavic languages and across three curriculum learning strategies on Named Entity Recognition as the downstream task. We also test the possibility of fine-tuning the synthetic resources to reflect linguistic properties, such as the grammatical case and gender, both of which are important for the Slavic languages. We demonstrate the possibility to achieve the mean F1 score of 0.78 across the three basic entities types for Belarusian starting from zero resources in comparison to the baseline of 0.63 using the zero-shot transfer from English. For comparison, the English model trained on the original set achieves the mean F1-score of 0.75. The experimental results are available from https://github.com/ValeraLobov/SlavNER

Social media posts provide a compelling, yet challenging source of data of diverse perspectives from many socially salient groups. Automatic text summarization algorithms make this data accessible at scale by compressing large collections of documents into short summaries that preserve salient information from the source text. In this work, we take a complementary approach to analyzing and improving the quality of summaries generated from social media data in terms of their ability to represent salient as well as diverse perspectives. We introduce a novel dataset, DivSumm, of dialect diverse tweets and human-written extractive and abstractive summaries. Then, we study the extent of dialect diversity reflected in human-written reference summaries as well as system-generated summaries. The results of our extensive experiments suggest that humans annotate fairly well-balanced dialect diverse summaries, and that cluster-based pre-processing approaches seem beneficial in improving the overall quality of the system-generated summaries without loss in diversity.

Existing studies on the analysis of persuasion in online discussions focus on investigating the effectiveness of comments in discussions and ignore the analysis of the effectiveness of debaters over multiple discussions. In this paper, we propose to quantify debaters effectiveness in the online discussion platform: “ChangeMyView” in order to explore diverse insights into their persuasion strategies. In particular, targeting debaters with different levels of effectiveness (e.g., good vs. bad), various behavioral characteristics (e..g, engagement) and text stylistic features (e.g., used frames) of debaters are carefully examined, leading to several outcomes that can be the backbone of writing assistants and persuasive text generation.

Essay exams have been attracting attention as a way of measuring the higher-order abilities of examinees, but they have two major drawbacks in that grading them is expensive and raises questions about fairness. As an approach to overcome these problems, automated essay scoring (AES) is in increasing need. Many AES models based on deep neural networks have been proposed in recent years and have achieved high accuracy, but most of these models are designed to predict only a single overall score. However, to provide detailed feedback in practical situations, we often require not only the overall score but also analytic scores corresponding to various aspects of the essay. Several neural AES models that can predict both the analytic scores and the overall score have also been proposed for this very purpose. However, conventional models are designed to have complex neural architectures for each analytic score, which makes interpreting the score prediction difficult. To improve the interpretability of the prediction while maintaining scoring accuracy, we propose a new neural model for automated analytic scoring that integrates a multidimensional item response theory model, which is a popular psychometric model.

Semantic Role Labeling (SRL) aims at recognizing the predicate-argument structure of a sentence and can be decomposed into two subtasks: predicate disambiguation and argument labeling. Prior work deals with these two tasks independently, which ignores the semantic connection between the two tasks. In this paper, we propose to use the machine reading comprehension (MRC) framework to bridge this gap. We formalize predicate disambiguation as multiple-choice machine reading comprehension, where the descriptions of candidate senses of a given predicate are used as options to select the correct sense. The chosen predicate sense is then used to determine the semantic roles for that predicate, and these semantic roles are used to construct the query for another MRC model for argument labeling. In this way, we are able to leverage both the predicate semantics and the semantic role semantics for argument labeling. We also propose to select a subset of all the possible semantic roles for computational efficiency. Experiments show that the proposed framework achieves state-of-the-art or comparable results to previous work.

Unsupervised sentence embeddings learning has been recently dominated by contrastive learning methods (e.g., SimCSE), which keep positive pairs similar and push negative pairs apart. The contrast operation aims to keep as much information as possible by maximizing the mutual information between positive instances, which leads to redundant information in sentence embedding. To address this problem, we present an information minimization based contrastive learning InforMin-CL model to retain the useful information and discard the redundant information by maximizing the mutual information and minimizing the information entropy between positive instances meanwhile for unsupervised sentence representation learning. Specifically, we find that information minimization can be achieved by simple contrast and reconstruction objectives. The reconstruction operation reconstitutes the positive instance via the other positive instance to minimize the information entropy between positive instances. We evaluate our model on fourteen downstream tasks, including both supervised and unsupervised (semantic textual similarity) tasks. Extensive experimental results show that our InforMin-CL obtains a state-of-the-art performance.

Unsupervised summarization methods have achieved remarkable results by incorporating representations from pre-trained language models. However, existing methods fail to consider efficiency and effectiveness at the same time when the input document is extremely long. To tackle this problem, in this paper, we proposed an efficient Coarse-to-Fine Facet-Aware Ranking (C2F-FAR) framework for unsupervised long document summarization, which is based on the semantic block. The semantic block refers to continuous sentences in the document that describe the same facet. Specifically, we address this problem by converting the one-step ranking method into the hierarchical multi-granularity two-stage ranking. In the coarse-level stage, we proposed a new segment algorithm to split the document into facet-aware semantic blocks and then filter insignificant blocks. In the fine-level stage, we select salient sentences in each block and then extract the final summary from selected sentences. We evaluate our framework on four long document summarization datasets: Gov-Report, BillSum, arXiv, and PubMed. Our C2F-FAR can achieve new state-of-the-art unsupervised summarization results on Gov-Report and BillSum. In addition, our method speeds up 4-28 times more than previous methods.

Automatic math problem solving has attracted much attention of NLP researchers recently. However, most of the works focus on the solving of Math Word Problems (MWPs). In this paper, we study on the Geometric Problem Solving based on neural networks. Solving geometric problems requires the integration of text and diagram information as well as the knowledge of the relevant theorems. The lack of high-quality datasets and efficient neural geometric solvers impedes the development of automatic geometric problems solving. Based on GeoQA, we newly annotate 2,518 geometric problems with richer types and greater difficulty to form an augmented benchmark dataset GeoQA+, containing 6,027 problems in training set and 7,528 totally. We further perform data augmentation method to expand the training set to 12,054. Besides, we design a Dual Parallel text Encoder DPE to efficiently encode long and medium-length problem text. The experimental results validate the effectiveness of GeoQA+ and DPE module, and the accuracy of automatic geometric problem solving is improved to 66.09%.

Recent studies show that the attention heads in Transformer are not equal. We relate this phenomenon to the imbalance training of multi-head attention and the model dependence on specific heads. To tackle this problem, we propose a simple masking method: HeadMask, in two specific ways. Experiments show that translation improvements are achieved on multiple language pairs. Subsequent empirical analyses also support our assumption and confirm the effectiveness of the method.

Multilingual pretrained models, while effective on monolingual data, need additional training to work well with code-switched text. In this work, we present a novel idea of training multilingual models with alignment objectives using parallel text so as to explicitly align word representations with the same underlying semantics across languages. Such an explicit alignment step has a positive downstream effect and improves performance on multiple code-switched NLP tasks. We explore two alignment strategies and report improvements of up to 7.32%, 0.76% and 1.9% on Hindi-English Sentiment Analysis, Named Entity Recognition and Question Answering tasks compared to a competitive baseline model.

To automatically correct handwritten assignments, the traditional approach is to use an OCR model to recognize characters and compare them to answers. The OCR model easily gets confused on recognizing handwritten Chinese characters, and the textual information of the answers is missing during the model inference. However, teachers always have these answers in mind to review and correct assignments. In this paper, we focus on the Chinese cloze tests correction and propose a multimodal approach(named AiM). The encoded representations of answers interact with the visual information of students’ handwriting. Instead of predicting ‘right’ or ‘wrong’, we perform the sequence labeling on the answer text to infer which answer character differs from the handwritten content in a fine-grained way. We take samples of OCR datasets as the positive samples for this task, and develop a negative sample augmentation method to scale up the training data. Experimental results show that AiM outperforms OCR-based methods by a large margin. Extensive studies demonstrate the effectiveness of our multimodal approach.

We propose a method to distill language-agnostic meaning embeddings from multilingual sentence encoders for unsupervised quality estimation of machine translation. Our method facilitates that the meaning embeddings focus on semantics by adversarial training that attempts to eliminate language-specific information. Experimental results on unsupervised quality estimation reveal that our method achieved higher correlations with human evaluations.

Most attempts on Text-to-SQL task using encoder-decoder approach show a big problem of dramatic decline in performance for new databases. For the popular Spider dataset, despite models achieving 70% accuracy on its development or test sets, the same models show a huge decline below 20% accuracy for unseen databases. The root causes for this problem are complex and they cannot be easily fixed by adding more manually created training. In this paper we address the problem and propose a solution that is a hybrid system using automated training-data augmentation technique. Our system consists of a rule-based and a deep learning components that interact to understand crucial information in a given query and produce correct SQL as a result. It achieves double-digit percentage improvement for databases that are not part of the Spider corpus.

We address contextualized code retrieval, the search for code snippets helpful to fill gaps in a partial input program. Our approach facilitates a large-scale self-supervised contrastive training by splitting source code randomly into contexts and targets. To combat leakage between the two, we suggest a novel approach based on mutual identifier masking, dedentation, and the selection of syntax-aligned targets. Our second contribution is a new dataset for direct evaluation of contextualized code retrieval, based on a dataset of manually aligned subpassages of code clones. Our experiments demonstrate that the proposed approach improves retrieval substantially, and yields new state-of-the-art results for code clone and defect detection.

Multilingual neural machine translation (NMT) enables positive knowledge transfer among multiple translation tasks with a shared underlying model, but a unified multilingual model usually suffers from capacity bottleneck when tens or hundreds of languages are involved. A possible solution is to cluster languages and train individual model for each cluster. However, the existing clustering methods based on language similarity cannot handle the asymmetric problem in multilingual NMT, i.e., one translation task A can benefit from another translation task B but task B will be harmed by task A. To address this problem, we propose a fuzzy task clustering method for multilingual NMT. Specifically, we employ task affinity, defined as the loss change of one translation task caused by the training of another, as the clustering criterion. Next, we cluster the translation tasks based on the task affinity, such that tasks from the same cluster can benefit each other. For each cluster, we further find out a set of auxiliary translation tasks that benefit the tasks in this cluster. In this way, the model for each cluster is trained not only on the tasks in the cluster but also on the auxiliary tasks. We conduct extensive experiments for one-to-many, manyto-one, and many-to-many translation scenarios to verify the effectiveness of our method.