COLING 2022 Papers

As AI is more and more pervasive in everyday life, humans have an increasing demand to understand its behavior and decisions. Most research on explainable AI builds on the premise that there is one ideal explanation to be found. In fact, however, everyday explanations are co-constructed in a dialogue between the person explaining (the explainer) and the specific person being explained to (the explainee). In this paper, we introduce a first corpus of dialogical explanations to enable NLP research on how humans explain as well as on how AI can learn to imitate this process. The corpus consists of 65 transcribed English dialogues from the Wired video series 5 Levels, explaining 13 topics to five explainees of different proficiency. All 1550 dialogue turns have been manually labeled by five independent professionals for the topic discussed as well as for the dialogue act and the explanation move performed. We analyze linguistic patterns of explainers and explainees, and we explore differences across proficiency levels. BERT-based baseline results indicate that sequence information helps predicting topics, acts, and moves effectively.

The task of Chinese Spelling Check (CSC) is aiming to detect and correct spelling errors that can be found in the text. While manually annotating a high-quality dataset is expensive and time-consuming, thus the scale of the training dataset is usually very small (e.g., SIGHAN15 only contains 2339 samples for training), therefore supervised-learning based models usually suffer the data sparsity limitation and over-fitting issue, especially in the era of big language models. In this paper, we are dedicated to investigating the unsupervised paradigm to address the CSC problem and we propose a framework named uChecker to conduct unsupervised spelling error detection and correction. Masked pretrained language models such as BERT are introduced as the backbone model considering their powerful language diagnosis capability. Benefiting from the various and flexible MASKing operations, we propose a Confusionset-guided masking strategy to fine-train the masked language model to further improve the performance of unsupervised detection and correction. Experimental results on standard datasets demonstrate the effectiveness of our proposed model uChecker in terms of character-level and sentence-level Accuracy, Precision, Recall, and F1-Measure on tasks of spelling error detection and correction respectively.

Recently, a new training oaxe loss has proven effective to ameliorate the effect of multimodality for non-autoregressive translation (NAT), which removes the penalty of word order errors in the standard cross-entropy loss. Starting from the intuition that reordering generally occurs between phrases, we extend oaxe by only allowing reordering between ngram phrases and still requiring a strict match of word order within the phrases. Extensive experiments on NAT benchmarks across language pairs and data scales demonstrate the effectiveness and universality of our approach. Further analyses show that ngram noaxe indeed improves the translation of ngram phrases, and produces more fluent translation with a better modeling of sentence structure.

Driven by deep learning breakthroughs, natural language generation (NLG) models have been at the center of steady progress in the last few years, with a ubiquitous task influence. However, since our ability to generate human-indistinguishable artificial text lags behind our capacity to assess it, it is paramount to develop and apply even better automatic evaluation metrics. To facilitate researchers to judge the effectiveness of their models broadly, we introduce NLG-Metricverse—an end-to-end open-source library for NLG evaluation based on Python. Our framework provides a living collection of NLG metrics in a unified and easy-to-use environment, supplying tools to efficiently apply, analyze, compare, and visualize them. This includes (i) the extensive support to heterogeneous automatic metrics with n-arity management, (ii) the meta-evaluation upon individual performance, metric-metric and metric-human correlations, (iii) graphical interpretations for helping humans better gain score intuitions, (iv) formal categorization and convenient documentation to accelerate metrics understanding. NLG-Metricverse aims to increase the comparability and replicability of NLG research, hopefully stimulating new contributions in the area.

Script knowledge is useful to a variety of NLP tasks. However, existing resources only cover a small number of activities, limiting its practical usefulness. In this work, we propose a zero-shot learning approach to script parsing, the task of tagging texts with scenario-specific event and participant types, which enables us to acquire script knowledge without domain-specific annotations. We (1) learn representations of potential event and participant mentions by promoting cluster consistency according to the annotated data; (2) perform clustering on the event / participant candidates from unannotated texts that belongs to an unseen scenario. The model achieves 68.1/74.4 average F1 for event / participant parsing, respectively, outperforming a previous CRF model that, in contrast, has access to scenario-specific supervision. We also evaluate the model by testing on a different corpus, where it achieved 55.5/54.0 average F1 for event / participant parsing.

Disfluencies that appear in the transcriptions from automatic speech recognition systems tend to impair the performance of downstream NLP tasks. Disfluency correction models can help alleviate this problem. However, the unavailability of labeled data in low-resource languages impairs progress. We propose using a pretrained multilingual model, finetuned only on English disfluencies, for zero-shot disfluency detection in Indian languages. We present a detailed pipeline to synthetically generate disfluent text and create evaluation datasets for four Indian languages: Bengali, Hindi, Malayalam, and Marathi. Even in the zero-shot setting, we obtain F1 scores of 75 and higher on five disfluency types across all four languages. We also show the utility of synthetically generated disfluencies by evaluating on real disfluent text in Bengali, Hindi, and Marathi. Finetuning the multilingual model on additional synthetic Hindi disfluent text nearly doubles the number of exact matches and yields a 20-point boost in F1 scores when evaluated on real Hindi disfluent text, compared to training with only English disfluent text.

In this study, we propose a morpheme-based scheme for Korean dependency parsing and adopt the proposed scheme to Universal Dependencies. We present the linguistic rationale that illustrates the motivation and the necessity of adopting the morpheme-based format, and develop scripts that convert between the original format used by Universal Dependencies and the proposed morpheme-based format automatically. The effectiveness of the proposed format for Korean dependency parsing is then testified by both statistical and neural models, including UDPipe and Stanza, with our carefully constructed morpheme-based word embedding for Korean. morphUD outperforms parsing results for all Korean UD treebanks, and we also present detailed error analysis.

In the effort to minimize the risk of extinction of a language, linguistic resources are fundamental. Quechua, a low-resource language from South America, is a language spoken by millions but, despite several efforts in the past, still lacks the resources necessary to build high-performance computational systems. In this article, we present WordNet-QU which signifies the inclusion of Quechua in a well-known lexical database called wordnet. We propose WordNet-QU to be included as an extension to wordnet after demonstrating a manually-curated collection of multiple digital resources for lexical use in Quechua. Our work uses the synset alignment algorithm to compare Quechua to its geographically nearest high-resource language, Spanish. Altogether, we propose a total of 28,582 unique synset IDs divided according to region like so: 20510 for Southern Quechua, 5993 for Central Quechua, 1121 for Northern Quechua, and 958 for Amazonian Quechua.

Word sense disambiguation (WSD), identifying the most suitable meaning of ambiguous words in the given contexts according to a predefined sense inventory, is one of the most classical and challenging tasks in natural language processing. Benefiting from the powerful ability of deep neural networks, WSD has achieved a great advancement in recent years. Reformulating WSD as a text span extraction task is an effective approach, which accepts a sentence context of an ambiguous word together with all definitions of its candidate senses simultaneously, and requires to extract the text span corresponding with the right sense. However, the approach merely depends on a short definition to learn sense representation, which neglects abundant semantic knowledge from related senses and leads to data-inefficient learning and suboptimal WSD performance. To address the limitations, we propose a novel WSD method with Knowledge-Enhanced and Local Self-Attention-based Extractive Sense Comprehension (KELESC). Specifically, a knowledge-enhanced method is proposed to enrich semantic representation by incorporating additional examples and definitions of the related senses in WordNet. Then, in order to avoid the huge computing complexity induced by the additional information, a local self-attention mechanism is utilized to constrain attention to be local, which allows longer input texts without large-scale computing burdens. Extensive experimental results demonstrate that KELESC achieves better performance than baseline models on public benchmark datasets.

Conversational Task Assistants (CTAs) are conversational agents whose goal is to help humans perform real-world tasks. CTAs can help in exploring available tasks, answering task-specific questions and guiding users through step-by-step instructions. In this work, we present Wizard of Tasks, the first corpus of such conversations in two domains: Cooking and Home Improvement. We crowd-sourced a total of 549 conversations (18,077 utterances) with an asynchronous Wizard-of-Oz setup, relying on recipes from WholeFoods Market for the cooking domain, and WikiHow articles for the home improvement domain. We present a detailed data analysis and show that the collected data can be a valuable and challenging resource for CTAs in two tasks: Intent Classification (IC) and Abstractive Question Answering (AQA). While on IC we acquired a high performing model (>85% F1), on AQA the performance is far from being satisfactory (~27% BertScore-F1), suggesting that more work is needed to solve the task of low-resource AQA.

Most comparative datasets of Chinese varieties are not digital; however, Wiktionary includes a wealth of transcriptions of words from these varieties. The usefulness of these data is limited by the fact that they use a wide range of variety-specific romanizations, making data difficult to compare. The current work collects this data into a single constituent (IPA, or International Phonetic Alphabet) and structured form (TSV) for use in comparative linguistics and Chinese NLP. At the time of writing, the dataset contains 67,943 entries across 8 varieties and Middle Chinese. The dataset is validated on a protoform reconstruction task using an encoder-decoder cross-attention architecture (Meloni et al 2021), achieving an accuracy of 54.11%, a PER (phoneme error rate) of 17.69%, and a FER (feature error rate) of 6.60%.

In a hate speech detection model, we should consider two critical aspects in addition to detection performance–bias and explainability. Hate speech cannot be identified based solely on the presence of specific words; the model should be able to reason like humans and be explainable. To improve the performance concerning the two aspects, we propose Masked Rationale Prediction (MRP) as an intermediate task. MRP is a task to predict the masked human rationales–snippets of a sentence that are grounds for human judgment–by referring to surrounding tokens combined with their unmasked rationales. As the model learns its reasoning ability based on rationales by MRP, it performs hate speech detection robustly in terms of bias and explainability. The proposed method generally achieves state-of-the-art performance in various metrics, demonstrating its effectiveness for hate speech detection. Warning: This paper contains samples that may be upsetting.

Nowadays, deep-learning based NLP models are usually trained with large-scale third-party data which can be easily injected with malicious backdoors. Thus, BackDoor Attack (BDA) study has become a trending research to help promote the robustness of an NLP system. Text-based BDA aims to train a poisoned model with both clean and poisoned texts to perform normally on clean inputs while being misled to predict those trigger-embedded texts as target labels set by attackers. Previous works usually choose fixed Positions-to-Poison (P2P) first, then add triggers upon those positions such as letter insertion or deletion. However, considering the positions of words with important semantics may vary in different contexts, fixed P2P models are severely limited in flexibility and performance. We study the text-based BDA from the perspective of automatically and dynamically selecting P2P from contexts. We design a novel Locator model which can predict P2P dynamically without human intervention. Based on the predicted P2P, four effective strategies are introduced to show the BDA performance. Experiments on two public datasets show both tinier test accuracy gap on clean data and higher attack success rate on poisoned ones. Human evaluation with volunteers also shows the P2P predicted by our model are important for classification. Source code is available at https://github.com/jncsnlp/LocatorModel

Probing studies have extensively explored where in neural language models linguistic information is located. The standard approach to interpreting the results of a probing classifier is to focus on the layers whose representations give the highest performance on the probing task. We propose an alternative method that asks where the task-relevant information emerges in the model. Our framework consists of a family of metrics that explicitly model local information gain relative to the previous layer and each layer’s contribution to the model’s overall performance. We apply the new metrics to two pairs of syntactic probing tasks with different degrees of complexity and find that the metrics confirm the expected ordering only for one of the pairs. Our local metrics show a massive dominance of the first layers, indicating that the features that contribute the most to our probing tasks are not as high-level as global metrics suggest.

Prerecorded laughter accompanying dialog in comedy TV shows encourages the audience to laugh by clearly marking humorous moments in the show. We present an approach for automatically detecting humor in the Friends TV show using multimodal data. Our model is capable of recognizing whether an utterance is humorous or not and assess the intensity of it. We use the prerecorded laughter in the show as annotation as it marks humor and the length of the audience’s laughter tells us how funny a given joke is. We evaluate the model on episodes the model has not been exposed to during the training phase. Our results show that the model is capable of correctly detecting whether an utterance is humorous 78% of the time and how long the audience’s laughter reaction should last with a mean absolute error of 600 milliseconds.

In this research, we present pilot experiments to distil monolingual models from a jointly trained model for 102 languages (mBERT). We demonstrate that it is possible for the target language to outperform the original model, even with a basic distillation setup. We evaluate our methodology for 6 languages with varying amounts of resources and language families.

The world of pronouns is changing – from a closed word class with few members to an open set of terms to reflect identities. However, Natural Language Processing (NLP) barely reflects this linguistic shift, resulting in the possible exclusion of non-binary users, even though recent work outlined the harms of gender-exclusive language technology. The current modeling of 3rd person pronouns is particularly problematic. It largely ignores various phenomena like neopronouns, i.e., novel pronoun sets that are not (yet) widely established. This omission contributes to the discrimination of marginalized and underrepresented groups, e.g., non-binary individuals. It thus prevents gender equality, one of the UN’s sustainable development goals (goal 5). Further, other identity-expressions beyond gender are ignored by current NLP technology. This paper provides an overview of 3rd person pronoun issues for NLP. Based on our observations and ethical considerations, we define a series of five desiderata for modeling pronouns in language technology, which we validate through a survey. We evaluate existing and novel modeling approaches w.r.t. these desiderata qualitatively and quantify the impact of a more discrimination-free approach on an established benchmark dataset.

Mathematical reasoning task is a subset of the natural language question answering task. Existing work suggested solving this task with a two-phase approach, where the model first predicts formulas from questions and then calculates answers from such formulas. This approach achieved desirable performance in existing work. However, its reliance on annotated formulas as intermediate labels throughout its training limited its application. In this work, we put forward the idea to enable models to learn optimal formulas autonomously. We proposed Weakly Supervised Formula Learner, a learning framework that drives the formula exploration with weak supervision from the final answers to mathematical problems. Our experiments are conducted on two representative mathematical reasoning datasets MathQA and Math23K. On MathQA, our method outperformed baselines trained on complete yet imperfect formula annotations. On Math23K, our method outperformed other weakly supervised learning methods.

Solving math word problems (MWPs) is an important and challenging problem in natural language processing. Existing approaches to solving MWPs require full supervision in the form of intermediate equations. However, labeling every MWP with its corresponding equations is a time-consuming and expensive task. In order to address this challenge of equation annotation, we propose a weakly supervised model for solving MWPs by requiring only the final answer as supervision. We approach this problem by first learning to generate the equation using the problem description and the final answer, which we subsequently use to train a supervised MWP solver. We propose and compare various weakly supervised techniques to learn to generate equations directly from the problem description and answer. Through extensive experiments, we demonstrate that without using equations for supervision, our approach achieves accuracy gains of 4.5% and 32% over the current state-of-the-art weakly-supervised approach, on the standard Math23K and AllArith datasets respectively. Additionally, we curate and release new datasets of roughly 10k MWPs each in English and in Hindi (a low-resource language). These datasets are suitable for training weakly supervised models. We also present an extension of our model to semi-supervised learning and present further improvements on results, along with insights.