AACL 2023 论文列表

significantly depends on their ability to effectively communicate with the users. Thus, the research Politeness is a key component that can assist in in recent times has been inclined towards modulatbuilding a strong customer-agent relationship. ing biases, styles, and control in text generation to With the ongoing increase in customer-care sysimprove these interactions. tems, it is crucial to have healthy relations with the users providing satisfaction and a better cusCustomer care is a typical application area for tomer experience. In this regard, it is significant automated task-oriented conversational systems. to model the different polite behaviors in an These systems provide potentially cost-effective agent to help the user in reaching the intended and reliable solutions for customer care. Nowadays, objectives. In our current work, we propose companies rely heavily on customer care service to the task of polite behavior-aware generation successfully guide and assist customers and estabconsidering the affective state of the user and lish long-term and healthy relationships with them. the conversational context. We design a Transformer based encoder-decoder framework with Customers generally interact with conversational three major components i.e., Affective tracker, agents with diverse expectations and attitudes (FølBehaviour-aware generators, and Polite genstad and Skjuve, 2019). In such situation, if the erator. The affective tracker is a context enagent fails to satisfy their expectations or fix their coder that captures the contextual information issue as desired, they may get frustrated or furialong with the affective information in the utous quickly (Jain et al., 2018). To avoid this, it terances; the behavior-aware generators indeis imperative for customer care agents to mimic pendently attends to the context information to compute behavior-aware polite representahuman-like (Fink, 2012) behaviour during the contions and finally, polite generator generates the versation. final polite response considering the represenPoliteness and empathy, which encompass sotations from different generators. Experimental cial belongingness are a few fundamental socioresults on the CYCCD dataset prove that our linguistic cues of humans (Brown et al., 1987; approach generates contextually correct and relMaslow, 1943). The use of social language by the evant responses compared to the state-of-the-art agents facilitates building rapport and emotional approaches and the baselines. connection with the customers. Politeness has been investigated recently (Golchha et al., 2019; Firdaus

Prompt-based classifiers are an attractive approach for zero-shot classification. However, the precise choice of the prompt template and label words can largely influence performance, with semantically equivalent settings often showing notable performance difference. This discrepancy can be partly attributed to word biases, where the classifier may be biased towards classes. To address this problem, it is possible to optimise classification thresholds on a labelled data set, however, this mitigates some of the advantages of prompt-based classifiers. This paper instead approaches this problem by examining the expected marginal probabilities of the classes. Here, probabilities are reweighted to have a uniform prior over classes, in an unsupervised fashion. Further, we draw a theoretical connection between the class priors and the language models’ word prior, and offer the ability to set a threshold in a zero-resource fashion. We show that matching class priors correlates strongly with the oracle upper bound performance and demonstrate large consistent performance gains for prompt settings over a range of NLP tasks.1

Open Information Extraction (OIE) is the process of extracting informative facts from opendomain natural language text. A multilingual OIE tool, IndIE, has been proposed, which performs chunking, creates a Merged-phrase Dependency Tree (MDT), and generates triples using hand-crafted rules. It is observed that fine-tuned transformer-based chunker outperforms other traditional methods of chunking. A benchmark called Hindi-BenchIE has also been developed for automatically evaluating Hindi triples. The developed OIE tool, IndIE, has been automatically evaluated on the goldentriples of 112 Hindi sentences. Compared to other multilingual methods, the IndIE method generates more meaningful triples with 0.51 F1-score. It is observed that IndIE generates more fine-grained triples than other methods. It is conjectured that IndIE has the ability to generate meaningful triples for Urdu, Tamil, and Telugu sentences as well because the developed chunker is shown to generalize across various natural languages, and the triple generation rules are based on dependency relations that are common to the aforementioned Indic languages.

One concept underlying methods that aim to exploit social information for identifying the facThe rise of social media has enabled the tuality of news sources is the social homophily widespread propagation of fake news, text that principle (McPherson et al., 2001), which capis published with an intent to spread misinformation and sway beliefs. Rapidly detecting tures the tendency of members of the same social fake news, especially as new events arise, is group to hold similar views and content preferimportant to prevent misinformation. ences. This often leads to the formation of “echo While prior works have tackled this problem uschambers” (Jamieson and Cappella, 2008; Quating supervised learning systems, automatedly trociocchi et al., 2016), tightly-knit “information modeling the complexities of the social media communities” that have little interaction with other landscape that enables the spread of fake news communities holding different views. Prior work is challenging. On the contrary, having humans shows how similar news, particularly misinformafact check all news is not scalable. Thus, in this tion, tends to spread more in some of these tightlypaper, we propose to approach this problem inknit information communities (Bessi et al., 2016). teractively, where humans can interact to help an automated system learn a better social media Thus, identifying them can provide the needed inrepresentation quality. On real world events, formation for capturing the factuality of sources our experiments show performance improve(communities spreading mostly low factuality conments in detecting factuality of news sources, tent in the past are likely to spread low factuality even after few human interactions. content in the future). In this work, we first capture

Grammatical Error Detection and Correction (GEC) tools have proven useful for native speakers and second language learners. Developing such tools requires a large amount of parallel, annotated data, which is unavailable for most languages. Synthetic data generation is a common practice to overcome the scarcity of such data. However, it is not straightforward for morphologically rich languages like Turkish due to complex writing rules that require phonological, morphological, and syntactic information. In this work, we present a flexible and extensible synthetic data generation pipeline for Turkish covering more than 20 expert-curated grammar and spelling rules (a.k.a., writing rules) implemented through complex transformation functions. Using this pipeline, we derive 130,000 high-quality parallel sentences from professionally edited articles. Additionally, we create a more realistic test set by manually annotating a set of movie reviews. We implement three baselines formulating the task as i) neural machine translation, ii) sequence tagging, and iii) prefix tuning with a pretrained decoder-only model, achieving strong results. Furthermore, we perform exhaustive experiments on out-of-domain datasets to gain insights on the transferability and robustness of the proposed approaches. Our results suggest that our corpus, GECTurk, is high-quality and allows knowledge transfer for the out-of-domain setting. To encourage further research on Turkish GEC, we release our datasets, baseline models, and the synthetic data generation pipeline at https: //github.com/GGLAB-KU/gecturk.

tems to remove such sensitive information before being used for retraining purposes. Industry apRegulations introduced General Data Protecplications also adopt rigorous approval processes tion Regulation (GDPR) in the EU or Caliwith opt-in or opt-out capabilities to ensure that fornia Consumer Privacy Act (CCPA) in the user generated content is used in a safe manner US have included provisions on the right to be while raising awareness among users on how their forgotten that mandates industry applications to remove data related to an individual from generated content will be used. their systems. In several real world industry apMore recently, the GDPR and CCPA provisioned plications that use Machine Learning to build the right to be forgotten, introducing new regulamodels on user data, such mandates require sigtions that mandates industry applications to support nificant effort both in terms of data cleansing deletion of user generated content when requested as well as model retraining while ensuring the by users. Although privacy concerns have gained models do not deteriorate in prediction quality significant importance now, the scale at which these due to removal of data. As a result, continuous removal of data and model retraining steps do models are built make it extremely expensive and not scale if these applications receive such retime-consuming to remove data efficiently while quests at a very high frequency. Recently, a few providing complete removal guarantees. researchers proposed the idea of Machine UnUnlearning assures the users that their data has learning to tackle this challenge. Despite the been completely removed when they prefer to have significant importance of this task, the area of their data erased (Bourtoule et al., 2019). This Machine Unlearning is under-explored in Natural Language Processing (NLP) tasks. In this task is different from Differential Privacy that fopaper, we explore the Unlearning framework on cuses on models that ensure that the users’ inforvarious GLUE tasks (Wang et al., 2018), such mation in the training data cannot be inferred and as, QQP, SST and MNLI. We propose compuprovides guarantees on the effect of an individual tationally efficient approaches (SISA-FC and examples. Such a task is challenging as removal SISA-A) to perform guaranteed Unlearning of data points can have significant deterioration on that provides significant reduction in terms of the performance of the models aka catastrophic both memory (90-95%), time (100x) and space consumption (99%) in comparison to the baseunlearning (Du et al., 2019; Golatkar et al., 2019; lines while having minimal impact on model Nguyen et al., 2020, 2022a). performance.1 . Approaches that guarantee complete removal of users’ data points have recently gained momentum

Temporal Relation Classification (TRC) is a fundamental task in natural language processing (NLP) and is essential for achieving a comprehensive understanding of a natural language. Given a document containing two event mentions, the objective of this task is to discern which of the two events happened first. Existing TRC datasets predominantly consist of texts written in English. To accommodate the growing interest in relevant NLP applications for Hebrew, we introduce a new TRC dataset for Hebrew. Professional annotators labeled Hebrew documents with TRC labels, adhering to guidelines adapted from a similar project on English and with some changes required to address some unique aspects of the Hebrew language. Overall, we annotated a corpus of 28,757 words, corresponding to 7,260 pairs of events. In addition to releasing the new dataset, which can be accessed at https://github.com/ shahafp/TRC-Hebrew, we train several baseline models for TRC and report their performance. 1 Kfir Bar Consider the following Hebrew text as an example: "‫א כ ל תי א ת ה ת פו ח א ח רי ש ש ט פ תי או תו‬ "I ate the apple after washing it" The event detection subtask is about detecting only the relevant events for our domain of interest; in this example: ‫“ א כ ל ת י‬I ate” and ‫“ ש ט פ ת י‬I washed”. In this case, both words are verbs representing actions which we subsequently mark as events. In the TRC subtask, we classify each pair of events using a closed set of labels to establish their chronological order. In this example, the events ‫“ א כ ל ת י‬I ate” and ‫“ ש ט פ ת י‬I washed” should be assigned the label AFTER to indicate that “I ate” occurs after “I washed” in chronological order. This study focuses on creating a Hebrew dataset for the second subtask, TRC, which is generally defined as a classification problem involving two events accompanied by their contextual information. Utilizing the annotated data, we train multiple TRC baseline models using all available Hebrew language models.

Commonsense QA Sample Large-scale sequence-to-sequence models have shown to be adept at both multiple-choice and open-domain commonsense reasoning tasks. However, the current formulations do not provide the ability to control the various attributes of the reasoning chain. To enable better controllability, we propose to study the commonsense reasoning as a template filling task (TemplateCSR) — where the language models fills reasoning templates with the given constraints as control factors. As an approach to TemplateCSR, we (i) propose a dataset of commonsense reasoning templateexpansion pairs for healthcare and well-being domain and (ii) introduce ITO, an instruction fine-tuned sequence-to-sequence model that performs commonsense reasoning across concepts in the template. Our experiments show that our approach outperforms baseline both in generation metrics and factuality metrics. We also present a detailed error analysis on our approach’s ability to reliably perform template based commonsense reasoning1 . Input: Are people who smoke at higher risk for lung cancer ?

The research on code-mixed data is limited due to the unavailability of dedicated code-mixed datasets and pre-trained language models. In this work, we focus on the low-resource Indian language Marathi which lacks any prior work in code-mixing. We present L3Cube-MeCorpus, a large code-mixed Marathi-English (Mr-En) corpus with 10 million social media sentences for pretraining. We also release L3Cube-MeBERT and MeRoBERTa, code-mixed BERT-based transformer models pre-trained on MeCorpus. Furthermore, for benchmarking, we present three supervised datasets MeHate, MeSent, and MeLID for downstream tasks like codemixed Mr-En hate speech detection, sentiment analysis, and language identification respectively. These evaluation datasets individually consist of manually annotated ~12,000 Marathi-English code-mixed tweets. Ablations show that the models trained on this novel corpus significantly outperform the existing state-of-the-art BERT models. This is the first work that presents artifacts for code-mixed Marathi research. All datasets and models are publicly released at https://github.com/ l3cube-pune/MarathiNLP.

on the target task (Dwivedi and Roig, 2019; Song et al., 2019). Nevertheless, devising an effective Transfer learning is frequently utilized in scemethod to identify a suitable similar task remains a narios with limited labeled examples, where a challenging endeavor (Ramirez et al., 2023). crucial step is to identify a related task to the tarTo evaluate the relatedness between tasks, difget task. CogTaskonomy (Luo et al., 2022) was ferent methods have been employed. The first catproposed to acquire a taxonomy of NLP tasks, specifically focusing on assessing the similaregory of method is task embedding, which learns ities between tasks. This method, inspired by a dedicated high-dimensional representation for cognitive processes, exhibits notable time effieach task using a task encoder (James et al., 2018; ciency. Nevertheless, it does not fully exploit Lan et al., 2019; Achille et al., 2019; Vu et al., the task-related information present in cogni2020). Despite the low time complexity, modtive data and lacks a comprehensive evaluation ulating the model to adapt for a new task using of various types of cognitive data. To address this method is challenging. Another approach is these limitations, this paper proposes a comprehensive neural and behavioral method to invesTaskonomy method (Zamir et al., 2018), which tigate the relationship among NLP tasks. Our fine-tunes a model on each task, and transfers each approach utilizes cognitive data, encompassing fine-tuned model to other tasks in a fully superboth neural data such as fMRI and EEG, as well vised manner. This process can effectively capture as behavioral data including eye-tracking and task similarity, while it is demanding and timesemantic feature ratings. Each data modality is consuming. The third category encompasses the employed to establish a common representation cognitively inspired CogTaskonomy proposed by space with Representation Similarity Analysis Luo et al. (2022). This method projects task-related for projecting task-related representations. To fully leverage the cognitive information, we representations into a shared space based on fMRI effectively extract the task-relevant informaand subsequently evaluates the similarity between tion extracted from neural data through feature tasks. It only needs to fine-tune a model for each ranking. Experimental results on 12 NLP tasks task separately, which exhibits notable time effidemonstrate that our proposed method outperciency. However, shared spaces in CogTaskonomy forms state-of-the-art methods on evaluating are not strict shared spaces, and this method does task similarity. not effectively leverage task-related information from neural data. Our study falls under the third

Extractive Question Answering (EQA) is a fundamental problem in Natural Language Understanding, aiming at answering given questions via extracting a contiguous sequence or span of words from a passage. Recent work on EQA has achieved promising performance with the help of pre-trained language models, for which Masked Language Modeling (MLM) is usually adopted as a pre-training task to predict masked tokens. This paper revisits MLM and proposes a simple yet effective method to improve the EQA performance, termed the [Mask]-for-Answering method (M4A). Specifically, three masking strategies are first introduced, which produce masked copies of the original passages. Instead of predicting masked tokens as in MLM, both original samples and masked copies are utilized simultaneously for training the EQA model. Importantly, a discrepancy loss is further incorporated to ensure that masked copies remain semantically close to the originals. As such, M4A is able to produce robust embeddings for both original and masked samples and infer correct answers even with masked context. Experimental study on several highly-competitive benchmarks consistently demonstrates the superiority of our proposed method over existing methods. M4A also achieves strong performance in low-resource settings and out-of-domain generalization. Limit Context: Portugal has the largest aquarium in Europe, the Lisbon Oceanarium, and the Portuguese have several other notable organizations focused on science-related exhibits and divulgation, like the state agency Ciência Viva, a programme of the Portuguese Ministry of Science and Technology to the promotion of a scientific and technological culture among the Portuguese population, the Science Museum of the University of Coimbra, the National Museum of Natural History at the University of Lisbon, and the Visionarium. Question: What is the name of the largest European aquarium? Answer: Lisbon Oceanarium Figure 1: An illustration case from SQuAD (1.1) (Fisch et al., 2019), using the proposed [Mask]-for-Answering algorithm (M4A) to infer answer(s) from samples with limit (unmasked) context. question-passage pair. Abundant evidences indicate that the strong encoding capability of PLMs has rapidly advanced the EQA performance, compared to traditional word embedding methods, such as GloVe and Word2Vec (Devlin et al., 2019; Joshi et al., 2020; Liu et al., 2019). Recently, how humans approaching reading comprehension becomes a major source of inspiration for enhancing EQA. There is a rich literature to incorporate the human-like reading comprehension strategy with PLMs and achieve remarkable success beyond the vanilla models, as later shown in Section 2. Notably, when experienced human readers perform question reasoning, they could infer the correct answer using only few sentences (even some parts of one sentence), instead of the entire

Semantic parsing aims at mapping natural language utterances into machine-interpretable meaning representations, facilitating user accesses to knowledge bases. However, knowledge in real-world scenarios is often duplicated in multiple storages and different representations. Although researchers have made great success by improving neural semantic parsers, existing works can only handle a specific kind of meaning representation, i.e., the single-target semantic parsing. In this paper, we introduce a multi-target semantic parsing model based on a collaborative deliberation network, which can not only decode multiple meaning representations simultaneously but also allow meaning representations to make use of information from each other while decoding. Experiments show that the proposed model improves the EM accuracy of four MRs averagely by 2.48% to 5.05% on three public datasets 1 .

significant advancements, and areas of improvement within the field. We synthesize and critically evaluate the findings from various studies, providing valuable insights for researchers and practitioners alike. This paper presents a comprehensive review of Natural Language Processing (NLP) research on Armenian, a language that, despite its rich history and unique linguistic characteristics, is currently low-resource in the field of NLP. We critically synthesize and evaluate various studies in Armenian NLP, highlighting key advancements, challenges, and areas for improvement. A notable aspect of our work is the underlined lack of application of Large Language Models (LLMs) in Armenian NLP, signifying an area of potential exploration and development. Identifying and discussing these challenges and opportunities lays the groundwork for future research directions in Armenian NLP. The emphasis on Armenian also advocates for increased attention to low-resource languages in NLP research, stressing the importance of linguistic diversity and equity. To the best of our knowledge, this is the first paper providing such an extensive review of Armenian NLP, marking a significant contribution to the field. 1 2. Identifying Challenges and Opportunities: Our paper highlights the challenges of developing NLP technologies for Armenian, a lowresource language. It elucidates the issues related to the lack of linguistic resources and tools, gaps in machine learning methodologies, and the specific challenges associated with Armenian. This knowledge is essential for future research and resource allocation. 3. Emphasis on Large Language Models (LLMs): Our paper underscores the underutilization of LLMs in Armenian NLP. It examines the potential benefits and challenges of deploying these models in low-resource settings and suggests avenues for future research.

One of the pursued objectives of deep learning is to provide tools that learn abstract representations of reality from the observation of multiple contextual situations. More precisely, one wishes to extract disentangled representations which are (i) low dimensional and (ii) whose components are independent and correspond to concepts capturing the essence of the objects under consideration (Locatello et al., 2019b). One step towards this ambitious project consists in learning disentangled representations with respect to a predefined (sensitive) attribute, e.g., the gender or age of the writer. Perhaps one of the main application for such disentangled representations is fair classification. Existing methods extract the last layer of a neural network trained with a loss that is composed of a cross-entropy objective and a disentanglement regularizer. In this work, we adopt an information-theoretic view of this problem which motivates a novel family of regularizers that minimizes the mutual information between the latent representation and the sensitive attribute conditional to the target. The resulting set of losses, called CLINIC, is parameter free and thus, it is easier and faster to train. CLINIC losses are studied through extensive numerical experiments by training over 2k neural networks. We demonstrate that our methods offer a better disentanglement/accuracy trade-off than previous techniques, and generalize better than training with cross-entropy loss solely provided that the disentanglement task is not too constraining.

Figure 1: Correlation with humans over time considering all existing metrics combined. On the x-axis: evaluation metrics ordered by their release time; y-axis: utterance-level Kendall’s τ with human when training a model to fit human judgments with all metrics available at the time (5-Fold cross-validation with XGBoost regressor). The dotted lines represent different human annotations and datasets. Different variants of the same metrics (like ROUGE-1 and ROUGE-2) are averaged. The datasets and metrics are described in Sec. 2. judgments globally. We do not make metric recommendation but reflect upon the global progress in the field of automatic evaluation. Our work is motivated by the findings of Fig. 1. It depicts the

sparse dot product operation. Their results by using a benchmark dataset were beyond 80%. Supervised Word Sense Disambiguation (WSD) has been studied intensively for over three decades. However, disentangling diverse contexts is still a challenging problem. This paper addresses the problem and proposes a Perturbation-based constrained attention network (Pconan) for injecting lexical knowledge derived from the WordNet. The Pconan allows modeling beneficial dependencies between the segments/words within the input sequence with the mask-attention technique. We incorporate a perturbation method into our model to mitigate the overfitting problem resulting from intensive learning. The experimental results by using a benchmark dataset show that our method is comparable to the SOTA WSD methods. Our source codes are available online1 .

Multilingual neural machine translation (MNMT) aims at using one single model for multiple translation directions. Recent work applies non-autoregressive Transformers to improve the efficiency of MNMT, but requires expensive knowledge distillation (KD) processes. To this end, we propose an M-DAT approach to non-autoregressive multilingual machine translation. Our system leverages the recent advance of the directed acyclic Transformer (DAT), which does not require KD. We further propose a pivot back-translation (PivotBT) approach to improve the generalization to unseen translation directions. Experiments show that our M-DAT achieves state-of-the-art performance in non-autoregressive MNMT.

This paper introduces the problem of determining whether people are located in the locations they mention in their Twitter streams. In particular, we investigate the role of context—tweets published before and after a tweet mentioning a location—in this challenging problem. We present a new corpus of Twitter streams with spatial information. Our analyses show that context is key to define the ground truth, as human judgments depend on whether annotators have access to context. Experimental results show that a neural architecture that takes into account context in addition to the tweet mentioning a location yields better results. We also conduct an error analysis to provide insights into the errors made by our best model.

In this paper, we explore the ability of the generative transformers to predict objects in IS-A (hypo-hypernym) relations. We solve the task for both directions of the relations: we learn to predict hypernyms given the input word and hyponyms, given the input concept and its neighbourhood from the taxonomy. To the best of our knowledge, this is the first paper which provides a comprehensive analysis of transformerbased models for the task of hypernymy extraction. Apart from the standard finetuning of various generative models, we experiment with different input formats and prefixes, zeroand few-shot learning strategies, and generation parameters. Results show that higher performance on both subtasks can be achieved by generative transformers with no additional data (like definitions or lemma names). Such models have phenomenally high abilities at the task given a little training and proper prompts in comparison to specialized rule-based and statistical methods as well as encoder-based transformer models. toy_dog.n.01 ym pern hy king charles spaniel n ictio pred hyp ony toy_spaniel.n.01 mp english toy spaniel red icti on papillon Figure 1: Two formulations of taxonomy enrichment task: attaching new candidates to the existing nodes (red) and generating new candidates at the specified place in the taxonomy (blue). et al., 2022), hypernymy (Ravichander et al., 2020), negation (Ettinger, 2020), etc. Applying modern transformers for identifying IS-A relations would allow not only to check the capacities of large language models (LLMs) in linguistics, but also to perform automatic extension of the lexical taxonomic structures with such kinds of