AACL 2023 论文列表

Zero-shot cross-lingual transfer is a central task in multilingual NLP, allowing models trained in languages with more sufficient training resources to generalize to other low-resource languages. Earlier efforts on this task use parallel corpora, bilingual dictionaries, or other annotated alignment data to improve cross-lingual transferability, which are typically expensive to obtain. In this paper, we propose a simple yet effective method, SALT, to improve the zero-shot cross-lingual transfer of the multilingual pretrained language models without the help of such external data. By incorporating code-switching and embedding mixup with self-augmentation, SALT effectively distills cross-lingual knowledge from the multilingual PLM and enhances its transferability on downstream tasks. Experimental results on XNLI and PAWS-X show that our method is able to improve zero-shot cross-lingual transferability without external data. Our code is available at https://github.com/luka-group/SALT.

Bangla (or Bengali) is the fifth most spoken language globally; yet, the state-of-the-art NLP in Bangla is lagging for even simple tasks such as lemmatization, POS tagging, etc. This is partly due to lack of a varied quality corpus. To alleviate this need, we build Vacaspati, a diverse corpus of Bangla literature. The literary works are collected from various websites; only those works that are publicly available without copyright violations or restrictions are collected. We believe that published literature captures the features of a language much better than newspapers, blogs or social media posts which tend to follow only a certain literary pattern and, therefore, miss out on language variety. Our corpus Vacaspati is varied from multiple aspects, including type of composition, topic, author, time, space, etc. It contains more than 11 million sentences and 115 million words. We also built a word embedding model, Vac-FT, using FastText from Vacaspati as well as trained an Electra model, Vac-BERT, using the corpus. Vac-BERT has far fewer parameters and requires only a fraction of resources compared to other state-of-the-art transformer models and yet performs either better or similar on various downstream tasks. On multiple downstream tasks, Vac-FT outperforms other FastText-based models. We also demonstrate the efficacy of Vacaspati as a corpus by showing that similar models built from other corpora are not as effective. The models are available at https://bangla.iitk.ac.in/.

We present PACT ( P retraining with A dversarial C ontrastive Learning for T ext Classification), a novel self-supervised framework for text classification. Instead of contrasting against in-batch negatives, a popular approach in the literature, PACT mines negatives closer to the anchor representation. PACT operates by en-dowing the standard pretraining mechanisms of BERT with adversarial contrastive learning objectives, allowing for effective joint optimization of token-and sentence-level pretraining of the BERT model. Our experiments on 13 diverse datasets including token-level, single-sentence, and sentence-pair text classification tasks show that PACT achieves consistent improvements over SOTA baselines. We further show that PACT regularizes both token-level and sentence-level embedding spaces into more uniform representations, thereby alleviating the undesirable anisotropic phenomenon of language models. 1

Text detoxification is the task of transferring the style of text from toxic to neutral. While here are approaches yielding promising results in monolingual setup, e.g., (Dale et al., 2021; Hallinan et al., 2022), cross-lingual transfer for this task remains a challenging open problem (Moskovskiy et al., 2022). In this work, we present a large-scale study of strategies for cross-lingual text detoxification -- given a parallel detoxification corpus for one language; the goal is to transfer detoxification ability to another language for which we do not have such a corpus. Moreover, we are the first to explore a new task where text translation and detoxification are performed simultaneously, providing several strong baselines for this task. Finally, we introduce new automatic detoxification evaluation metrics with higher correlations with human judgments than previous benchmarks. We assess the most promising approaches also with manual markup, determining the answer for the best strategy to transfer the knowledge of text detoxification between languages.

Large language models (LLMs) have demonstrated impressive capabilities in natural language generation. However, their output quality can be inconsistent, posing challenges for generating natural language from logical forms (LFs). This task requires the generated outputs to embody the exact semantics of LFs, without missing any LF semantics or creating any hallucinations. In this work, we tackle this issue by proposing a novel generate-and-rerank approach. Our approach involves initially generating a set of candidate outputs by prompting an LLM and subsequently reranking them using a task-specific reranker model. In addition, we curate a manually collected dataset to evaluate the alignment between different ranking metrics and human judgements. The chosen ranking metrics are utilized to enhance the training and evaluation of the reranker model. By conducting extensive experiments on three diverse datasets, we demonstrate that the candidates selected by our reranker outperform those selected by baseline methods in terms of semantic consistency and fluency, as measured by three comprehensive metrics. Our findings provide strong evidence for the effectiveness of our approach in improving the quality of generated outputs.

This work presents KoBigBird-large, a large size of Korean BigBird that achieves state-of-the-art performance and allows long sequence processing for Korean language understanding. Without further pretraining, we only transform the architecture and extend the positional encoding with our proposed Tapered Absolute Positional Encoding Representations (TAPER). In experiments, KoBigBird-large shows state-of-the-art overall performance on Korean language understanding benchmarks and the best performance on document classification and question answering tasks for longer sequences against the competitive baseline models. We publicly release our model here.

Abduction has long been seen as crucial for narrative comprehension and reasoning about everyday situations. The abductive natural language inference ($\alpha$NLI) task has been proposed, and this narrative text-based task aims to infer the most plausible hypothesis from the candidates given two observations. However, the inter-sentential coherence and the model consistency have not been well exploited in the previous works on this task. In this work, we propose a prompt tuning model $\alpha$-PACE, which takes self-consistency and inter-sentential coherence into consideration. Besides, we propose a general self-consistent framework that considers various narrative sequences (e.g., linear narrative and reverse chronology) for guiding the pre-trained language model in understanding the narrative context of input. We conduct extensive experiments and thorough ablation studies to illustrate the necessity and effectiveness of $\alpha$-PACE. The performance of our method shows significant improvement against extensive competitive baselines.

Given an image depicting multiple individuals, humans are capable of inferring each individual’s emotions, intentions, and social norms based on commonsense understanding. However, a machine’s ability of commonsense reasoning about distinct individuals in images remains underexplored. In this study, we examine the consistency of visual commonsense reasoning based on person grounding. We introduce a novel test dataset called V isual C ommonsense R easoning-C ontrast S ets (VCR-CS) to evaluate whether models can reason about individual people in an image by changing the person tags in the questions and answers. We benchmark various vision-language models on VCR-CS and observe that they fail in consistent common-sense reasoning about different people in one image, showing a performance decrease of up to 31.5%. To mitigate such failures, we propose a multi-task learning framework called P erson-centric ground I ng e N hanced T uning (PINT). Our framework enhances a model’s ability to perform person-grounded commonsense reasoning by leveraging two novel person-centric pretraining tasks: Image Person-based Text Matching and Person-Masked Language Modeling. The experimental results revealed the effectiveness of PINT by showing the lowest performance degradation on VCR-CS and the improvements in consistency and sensitivity metrics. Our dataset and code are publicly available 1 .

Knowledge-intensive generation tasks like generative question answering require models to retrieve appropriate passages from external knowledge sources to support answer generation. The generation quality relies heavily on the retrieved passages, which serve as contextual information. State-of-the-art Retrieval Augmented Generation models with marginalized output dominate this area but focus too much on label-relevant passages, rather than question-relevant passages and answers. This work addresses this issue by incorporating rich answer encoding through Dense Knowledge Similarity (DKS) and Retriever as Answer Classifier (RAC). We demonstrate the advantages of our proposed approach in open domain question answering (MSMARCO) and conversation (Wizard of Wikipedia) datasets, reporting both generation and retrieval metrics. In the MSMARCO development set, our best model achieves 12.1% relative improvement 1 on Recall@1 and 4.5% relative improvement on BLEU-4 compared to the baseline model. In the KILT-WoW leaderboard, our best model achieves 8.9% relative improvement on R-Precision and 13.3% relative improvement on KILT-RL compared to the baseline model. Our codes and models are available at https://github.com/hwy9855/rag-ae .

While large pre-trained language models (LLMs) have shown their impressive capabilities in various NLP tasks, they are still under-explored in the misinformation domain. In this paper, we examine LLMs with in-context learning (ICL) for news claim verification, and find that only with 4-shot demonstration examples, the performance of several prompting methods can be comparable with previous supervised models. To further boost performance, we introduce a Hierarchical Step-by-Step (HiSS) prompting method which directs LLMs to separate a claim into several subclaims and then verify each of them via multiple questions-answering steps progressively. Experiment results on two public misinformation datasets show that HiSS prompting outperforms state-of-the-art fully-supervised approach and strong few-shot ICL-enabled baselines.

Grounding dialogue response generation on external knowledge is proposed to produce informative and engaging responses. However, current knowledge-grounded dialogue (KGD) systems often fail to align the generated responses with human-preferred qualities due to several issues like hallucination and the lack of coherence. Upon analyzing multiple language model generations, we observe the presence of alternative generated responses within a single decoding process. These alternative responses are more faithful and exhibit a comparable or higher level of relevance to prior conversational turns compared to the optimal responses prioritized by the decoding processes. To address these challenges and driven by these observations, we propose Polished \&Informed Candidate Scoring (PICK), a generation re-scoring framework that empowers models to generate faithful and relevant responses without requiring additional labeled data or model tuning. Through comprehensive automatic and human evaluations, we demonstrate the effectiveness of PICK in generating responses that are more faithful while keeping them relevant to the dialogue history. Furthermore, PICK consistently improves the system's performance with both oracle and retrieved knowledge in all decoding strategies. We provide the detailed implementation in https://github.com/bryanwilie/pick .

Warning: This paper contains content that can be offensive or upsetting. The research for detecting offensive language on online platforms has much advanced. However, the majority of these studies have primarily focused on English. Given the unique characteristics of offensive language, where social and cultural contexts significantly influence content understanding, language-specific datasets are essential. Acquiring comprehensive datasets in Korean, a less-resourced language, has mostly relied on human annotations, suffering from inherent limitations in terms of labor intensity and potential annotator bias. Automatic generation of datasets using generative methods offers an alternative approach to address these limitations, yet faces challenges in capturing linguistic and cultural di-versities while maintaining native-level fluency. To address these challenges, we introduce a prompt design methodology, Korean Offensive language Machine Generation (K-OMG), using large language models. By manipulating three prompt factors, we find an effective prompt de-sign to generate culturally aligned offensive language with fluent expressions. Experimental results demonstrate the high quality and utility of our automatically generated dataset. Our detailed analysis shows that the proposed approach achieves exceptional fluency in generating texts while effectively incorporating social and cultural diversities.

Persuasive meme identification is a crucial task in automatically categorizing memes based on their persuasive nature. Memes, being highly influential in online communication, have the ability to shape individuals’ attitudes, behaviors, and beliefs, both positively and negatively. They can be utilized to promote positive actions, challenge social norms, and raise awareness, but they can also perpetuate harmful ide-ologies, spread misinformation, stereotype, and manipulate emotions. In this paper, we are addressing this challenge by empirically investigating three novel tasks, viz. (i) Task 1: Persua-sive meme detection, (ii) Task 2: Identification of the effectiveness of persuasive memes, and (iii) Task 3: Identification of persuasion techniques used in persuasive memes. To this end, we make the very first attempt to release a high-quality, large-scale dataset, Persuasive_meme 1 , since there is no publicly available such dataset for the Hindi-English code-mixed (Hinglish) domain. 2 We further developed several base-line unimodal and multimodal models for these tasks. Empirical evaluation with respect to both, qualitative and quantitative analysis, on the Persuasive_meme dataset highlight the significance of multimodality in addressing these tasks effectively. Additionally, we discuss the limitations of the current models and emphasize

Democratizing access to natural language processing (NLP) technology is crucial, especially for underrepresented and extremely low-resource languages. Previous research has focused on developing labeled and unlabeled corpora for these languages through online scraping and document translation. While these methods have proven effective and cost-efficient, we have identified limitations in the resulting corpora, including a lack of lexical diversity and cultural relevance to local communities. To address this gap, we conduct a case study on Indonesian local languages. We compare the effectiveness of online scraping, human translation, and paragraph writing by native speakers in constructing datasets. Our findings demonstrate that datasets generated through paragraph writing by native speakers exhibit superior quality in terms of lexical diversity and cultural content. In addition, we present the \datasetname{} benchmark, encompassing 12 underrepresented and extremely low-resource languages spoken by millions of individuals in Indonesia. Our empirical experiment results using existing multilingual large language models conclude the need to extend these models to more underrepresented languages. We release the NusaWrites dataset at https://github.com/IndoNLP/nusa-writes.

Subsampling is effective in Knowledge Graph Embedding (KGE) for reducing overfitting caused by the sparsity in Knowledge Graph (KG) datasets. However, current subsampling approaches consider only frequencies of queries that consist of entities and their relations. Thus, the existing subsampling potentially underestimates the appearance probabilities of infrequent queries even if the frequencies of their entities or relations are high. To address this problem, we propose Model-based Subsampling (MBS) and Mixed Subsampling (MIX) to estimate their appearance probabilities through predictions of KGE models. Evaluation results on datasets FB15k-237, WN18RR, and YAGO3-10 showed that our proposed subsampling methods actually improved the KG completion performances for popular KGE models, RotatE, TransE, HAKE, ComplEx, and DistMult.

,

We study the effect of tokenization on gender bias in machine translation, an aspect that has been largely overlooked in previous works. Specifically, we focus on the interactions between the frequency of gendered profession names in training data, their representation in the subword tokenizer's vocabulary, and gender bias. We observe that female and non-stereotypical gender inflections of profession names (e.g., Spanish"doctora"for"female doctor") tend to be split into multiple subword tokens. Our results indicate that the imbalance of gender forms in the model's training corpus is a major factor contributing to gender bias and has a greater impact than subword splitting. We show that analyzing subword splits provides good estimates of gender-form imbalance in the training data and can be used even when the corpus is not publicly available. We also demonstrate that fine-tuning just the token embedding layer can decrease the gap in gender prediction accuracy between female and male forms without impairing the translation quality.

The Natural Language Inference (NLI) task often requires reasoning over multiple steps to reach the conclusion. While the necessity of generating such intermediate steps (instead of a summary explanation) has gained popular support, it is unclear how to generate such steps without complete end-to-end supervision and how such generated steps can be further utilized . In this work, we train and enhance a sequence-to-sequence next-step prediction model with external commonsense knowledge and search to generate intermediate steps with limited next-step supervision. We show the correctness of such generated steps through human verification, on MNLI and MED datasets (and discuss the limitations through qualitative examples). We show that such generated steps can help improve end-to-end NLI task performance using simple data augmentation strategies. Using a C HECK L IST dataset for NLI, we also explore the effect of augmentation on specific reasoning types. The code and human-evaluation dataset is available at https://github.com/deepanwayx/prover.

Affordance knowledge is a fundamental aspect of commonsense knowledge. Recent findings indicate that world knowledge emerges through large-scale self-supervised pretraining, motivating our exploration of acquiring affordance knowledge from the visual domain. To this end, we augment an existing instructional video resource to create the new Causal Action-Effect (CAE) dataset and design two novel pretraining tasks -- Masked Action Modeling (MAM) and Masked Effect Modeling (MEM) -- promoting the acquisition of two affordance properties in models: behavior and entity equivalence, respectively. We empirically demonstrate the effectiveness of our proposed methods in learning affordance properties. Furthermore, we show that a model pretrained on both tasks outperforms a strong image-based visual-linguistic foundation model (FLAVA) as well as pure linguistic models on a zero-shot physical reasoning probing task.