Automated essay scoring (AES) is a useful tool in English as a Foreign Language (EFL) writing education, offering real-time essay scores for students and instructors. However, previous AES models were trained on essays and scores irrelevant to the practical scenarios of EFL writing education and usually provided a single holistic score due to the lack of appropriate datasets. In this paper, we release DREsS, a large-scale, standard dataset for rubric-based automated essay scoring with 48.9K samples in total. DREsS comprises three sub-datasets: DREsS_New, DREsS_Std., and DREsS_CASE. We collect DREsS_New, a real-classroom dataset with 2.3K essays authored by EFL undergraduate students and scored by English education experts. We also standardize existing rubric-based essay scoring datasets as DREsS_Std. We suggest CASE, a corruption-based augmentation strategy for essays, which generates 40.1K synthetic samples of DREsS_CASE and improves the baseline results by 45.44%. DREsS will enable further research to provide a more accurate and practical AES system for EFL writing education.

Reliable multilingual evaluation is difficult, and culturally appropriate evaluation is even harder to achieve.A common practice to fill this gap is to machine-translate English evaluation sets. However, translation introduces language bias and carries over cultural and regional assumptions from the original questions – often testing knowledge irrelevant to the target audience. In this work, we highlight the extent and impact of these biases and present a multilingual evaluation framework that aims to mitigate them through improved translations and annotation practices.Through a large-scale study involving professional and community translators and annotators, we show that state-of-the-art models excel primarily by learning Western-centric concepts. Notably, we find that model rankings on the full MMLU change when evaluated on a subset of questions explicitly marked as culturally sensitive.We release Global MMLU, a multilingual extension of MMLU across 42 languages, featuring improved translation quality, expanded language coverage, and designated subsets labeled as culturally sensitive and culturally agnostic to enable a more comprehensive and equitable benchmark for evaluating language models across diverse linguistic and cultural contexts.

Large language models (LLMs) generate human-like text, raising concerns about their misuse in creating deceptive content. Detecting LLM-generated comments (LGC) in online news is essential for preserving online discourse integrity and preventing opinion manipulation. However, effective detection faces two key challenges; the brevity and informality of news comments limit traditional methods, and the absence of a publicly available LGC dataset hinders model training, especially for languages other than English. To address these challenges, we propose a twofold approach. First, we develop an LGC generation framework to construct a high-quality dataset with diverse and complex examples. Second, we introduce XDAC (XAI-Driven Detection and Attribution of LLM-Generated Comments), a framework utilizing explainable AI, designed for the detection and attribution of short-form LGC in Korean news articles. XDAC leverages XAI to uncover distinguishing linguistic patterns at both token and character levels. We present the first large-scale benchmark dataset, comprising 1.3M human-written comments from Korean news platforms and 1M LLM-generated comments from 14 distinct models. XDAC outperforms existing methods, achieving a 98.5% F1 score in LGC detection with a relative improvement of 68.1%, and an 84.3% F1 score in attribution. To validate real-world applicability, we analyze 5.24M news comments from Naver, South Korea’s leading online news platform, identifying 27,029 potential LLM-generated comments.

Text-to-image diffusion models have recently enabled the creation of visually compelling, detailed images from textual prompts. However, their ability to accurately represent various cultural nuances remains an open question. In our work, we introduce CULTDIFF benchmark, evaluating whether state-of-the-art diffusion models can generate culturally specific images spanning ten countries. We show that these models often fail to generate cultural artifacts in architecture, clothing, and food, especially for underrepresented country regions, by conducting a fine-grained analysis of different similarity aspects, revealing significant disparities in cultural relevance, description fidelity, and realism compared to real-world reference images. With the collected human evaluations, we develop a neural-based image-image similarity metric, namely, CULTDIFF-S, to predict human judgment on real and generated images with cultural artifacts. Our work highlights the need for more inclusive generative AI systems and equitable dataset representation over a wide range of cultures.

Content moderation platforms concentrate resources on English content despite serving predominantly non-English speaking users.Also, given the scarcity of native moderators for low-resource languages, non-native moderators must bridge this gap in moderation tasks such as hate speech moderation.Through a user study, we identify that non-native moderators struggle with understanding culturally-specific knowledge, sentiment, and internet culture in the hate speech.To assist non-native moderators, we present LLM-C3MOD, a human-LLM collaborative pipeline with three steps: (1) RAG-enhanced cultural context annotations; (2) initial LLM-based moderation; and (3) targeted human moderation for cases lacking LLM consensus.Evaluated on Korean hate speech dataset with Indonesian and German participants, our system achieves 78% accuracy (surpassing GPT-4o’s 71% baseline) while reducing human workload by 83.6%.In addition, cultural context annotations improved non-native moderator accuracy from 22% to 61%, with humans notably excelling at nuanced tasks where LLMs struggle.Our findings demonstrate that non-native moderators, when properly supported by LLMs, can effectively contribute to cross-cultural hate speech moderation.

In a highly globalized world, it is important for multi-modal large language models (MLLMs) to recognize and respond correctly to mixed-cultural inputs.For example, a model should correctly identify kimchi (Korean food) in an image both when an Asian woman is eating it, as well as an African man is eating it.However, current MLLMs show an over-reliance on the visual features of the person, leading to misclassification of the entities. To examine the robustness of MLLMs to different ethnicity, we introduce MIXCUBE, a cross-cultural bias benchmark, and study elements from five countries and four ethnicities. Our findings reveal that MLLMs achieve both higher accuracy and lower sensitivity to such perturbation for high-resource cultures, but not for low-resource cultures. GPT-4o, the best-performing model overall, shows up to 58% difference in accuracy between the original and perturbed cultural settings in low-resource cultures

Large language models (LLMs) now exhibit near human-level performance in various tasks, but their performance drops drastically after a handful of high-resource languages due to the imbalance in pre-training data. Inspired by the human process of second language acquisition, particularly code-switching—the practice of language alternation in a conversation—we propose code-switching curriculum learning (CSCL) to enhance cross-lingual transfer for LLMs. CSCL mimics the stages of human language learning by progressively training models with a curriculum consisting of 1) token-level code-switching, 2) sentence-level code-switching, and 3) monolingual corpora. Using Qwen 2 as our underlying model, we demonstrate the efficacy of the CSCL in improving language transfer to Korean, achieving significant performance gains compared to monolingual continual pre-training methods. Ablation studies reveal that both token- and sentence-level code-switching significantly enhance cross-lingual transfer and that curriculum learning amplifies these effects. We also extend our findings into various languages, including Japanese (high-resource) and Indonesian (low-resource), and using two additional models (Gemma 2 and Phi 3.5). We further show that CSCL mitigates spurious correlations between language resources and safety alignment, presenting a robust, efficient framework for more equitable language transfer in LLMs. We observe that CSCL is effective for low-resource settings where high-quality, monolingual corpora for language transfer are hardly available.

Measuring social bias in large language models (LLMs) is crucial, but existing bias evaluation methods struggle to assess bias in long-form generation. We propose a Bias Benchmark for Generation (BBG), an adaptation of the Bias Benchmark for QA (BBQ), designed to evaluate social bias in long-form generation by having LLMs generate continuations of story prompts. Building our benchmark in English and Korean, we measure the probability of neutral and biased generations across ten LLMs. We also compare our long-form story generation evaluation results with multiple-choice BBQ evaluation, showing that the two approaches produce inconsistent results.

Ensuring persona fidelity in large language models (LLMs) is essential for maintaining coherent and engaging human-AI interactions. However, LLMs often exhibit Out-of-Character (OOC) behavior, where generated responses deviate from an assigned persona, leading to inconsistencies that affect model reliability. Existing evaluation methods typically assign single scores to entire responses, struggling to capture subtle persona misalignment, particularly in long-form text generation. To address this limitation, we propose an atomic-level evaluation framework that quantifies persona fidelity at a finer granularity. Our three key metrics measure the degree of persona alignment and consistency within and across generations. Our approach enables a more precise and realistic assessment of persona fidelity by identifying subtle deviations that real users would encounter. Through our experiments, we demonstrate that our framework effectively detects persona inconsistencies that prior methods overlook. By analyzing persona fidelity across diverse tasks and personality types, we reveal how task structure and persona desirability influence model adaptability, highlighting challenges in maintaining consistent persona expression.

We introduce LLM-as-an-Interviewer, a novel paradigm for evaluating large language models (LLMs). This approach leverages multi-turn interactions where the LLM interviewer actively provides feedback on responses and poses follow-up questions to the evaluated LLM. At the start of the interview, the LLM interviewer dynamically modifies datasets to generate initial questions, mitigating data contamination. We apply the LLM-as-an-Interviewer framework to evaluate six models on the reasoning, factuality and instruction-following tasks. Our results show that the framework effectively provides insights into LLM performance, including the quality of initial responses, adaptability to feedback, and ability to address follow-up queries like clarification or additional knowledge requests. The framework also addresses key limitations of conventional methods like LLM-as-a-Judge, including verbosity bias and inconsistency across runs. Finally, we propose the Interview Report, which aggregates insights from the interview process, providing examples and a comprehensive analysis of the LLM’s strengths and weaknesses. This report offers a detailed snapshot of the model’s real-world applicability.

Vision Language Models (VLMs) often struggle with culture-specific knowledge, particularly in languages other than English and in underrepresented cultural contexts. To evaluate their understanding of such knowledge, we introduce WorldCuisines, a massive-scale benchmark for multilingual and multicultural, visually grounded language understanding. This benchmark includes a visual question answering (VQA) dataset with text-image pairs across 30 languages and dialects, spanning 9 language families and featuring over 1 million data points, making it the largest multicultural VQA benchmark to date. It includes tasks for identifying dish names and their origins. We provide evaluation datasets in two sizes (12k and 60k instances) alongside a training dataset (1 million instances). Our findings show that while VLMs perform better with correct location context, they struggle with adversarial contexts and predicting specific regional cuisines and languages. To support future research, we release a knowledge base with annotated food entries and images along with the VQA data.

In the context of English as a Foreign Language (EFL) writing education, LLM-as-a-tutor can assist students by providing real-time feedback on their essays. However, challenges arise in assessing LLM-as-a-tutor due to differing standards between educational and general use cases. To bridge this gap, we integrate pedagogical principles to assess student-LLM interaction. First, we explore how LLMs can function as English tutors, providing effective essay feedback tailored to students. Second, we propose three criteria to evaluate LLM-as-a-tutor specifically designed for EFL writing education, emphasizing pedagogical aspects. In this process, EFL experts evaluate the feedback from LLM-as-a-tutor regarding (1) quality and (2) characteristics. On the other hand, EFL learners assess their (3) learning outcomes from interaction with LLM-as-a-tutor. This approach lays the groundwork for developing LLMs-as-a-tutor tailored to the needs of EFL learners, advancing the effectiveness of writing education in this context.

This paper explores the assumption that Large Language Models (LLMs) skilled in generation tasks are equally adept as evaluators. We assess the performance of three LLMs and one open-source LM in Question-Answering (QA) and evaluation tasks using the TriviaQA (Joshi et al., 2017) dataset. Results indicate a significant disparity, with LLMs exhibiting lower performance in evaluation tasks compared to generation tasks. Intriguingly, we discover instances of unfaithful evaluation where models accurately evaluate answers in areas where they lack competence, underscoring the need to examine the faithfulness and trustworthiness of LLMs as evaluators. This study contributes to the understanding of “the Generative AI Paradox” (West et al., 2023), highlighting a need to explore the correlation between generative excellence and evaluation proficiency, and the necessity to scrutinize the faithfulness aspect in model evaluations.

While humans naturally develop theory of mind (ToM), the capability to understand other people’s mental states and beliefs, state-of-the-art large language models (LLMs) underperform on simple ToM benchmarks. We posit that we can extend our understanding of LLMs’ ToM abilities by evaluating key human ToM precursors-perception inference and perception-to-belief inference-in LLMs. We introduce two datasets, Percept-ToMi and Percept-FANToM, to evaluate these precursory inferences for ToM in LLMs by annotating characters’ perceptions on ToMi and FANToM, respectively.Our evaluation of eight state-of-the-art LLMs reveals that the models generally perform well in perception inference while exhibiting limited capability in perception-to-belief inference (e.g., lack of inhibitory control).Based on these results, we present PercepToM, a novel ToM method leveraging LLMs’ strong perception inference capability while supplementing their limited perception-to-belief inference. Experimental results demonstrate that PercepToM significantly enhances LLM’s performance, especially in false belief scenarios.

Large Language Models (LLMs) are increasingly being used to generate synthetic data for training and evaluating models. However, it is unclear whether they can generate a good quality of question answering (QA) dataset that incorporates knowledge and cultural nuance embedded in a language, especially for low-resource languages. In this study, we investigate the effectiveness of using LLMs in generating culturally relevant commonsense QA datasets for Indonesian and Sundanese languages. To do so, we create datasets for these languages using various methods involving both LLMs and human annotators, resulting in 4.5K questions per language (9K in total), making our dataset the largest of its kind. Our experiments show that automatic data adaptation from an existing English dataset is less effective for Sundanese. Interestingly, using the direct generation method on the target language, GPT-4 Turbo can generate questions with adequate general knowledge in both languages, albeit not as culturally ‘deep’ as humans. We also observe a higher occurrence of fluency errors in the Sundanese dataset, highlighting the discrepancy between medium- and lower-resource languages.

In this study, we introduce BEnQA, a dataset comprising parallel Bengali and English exam questions for middle and high school levels in Bangladesh. Our dataset consists of approximately 5K questions covering several subjects in science with different types of questions, including factual, application, and reasoning-based questions. We benchmark several Large Language Models (LLMs) with our parallel dataset and observe a notable performance disparity between the models in Bengali and English. We also investigate some prompting methods, and find that Chain-of-Thought prompting is beneficial mostly on reasoning questions, but not so much on factual ones. We also find that appending English translation helps to answer questions in Bengali. Our findings point to promising future research directions for improving the performance of LLMs in Bengali and more generally in low-resource languages.

Application of LLM to database queries on natural language sentences has demonstrated impressive results in both single and multi-hop scenarios.In the existing methodologies, the requirement to re-encode query vectors at each stage for processing multi-hop queries presents a significant bottleneck to the inference speed.This paper proposes VKGFR (Virtual Knowledge Graph based Fact Retriever) that leverages large language models to extract representations corresponding to a sentence’s knowledge graph, significantly enhancing inference speed for multi-hop reasoning without performance loss.Given that both the queries and natural language database sentences can be structured as a knowledge graph, we suggest extracting a Virtual Knowledge Graph (VKG) representation from sentences with LLM.Over the pre-constructed VKG, our VKGFR conducts retrieval with a tiny model structure, showing performance improvements with higher computational efficiency. We evaluate VKGFR on the WikiNLDB and MetaQA dataset, designed for multi-hop database reasoning over text. The results indicate 13x faster inference speed on the WikiNLDB dataset without performance loss.

Despite the rapid development of large language models (LLMs) for the Korean language, there remains an obvious lack of benchmark datasets that test the requisite Korean cultural and linguistic knowledge. Because many existing Korean benchmark datasets are derived from the English counterparts through translation, they often overlook the different cultural contexts. For the few benchmark datasets that are sourced from Korean data capturing cultural knowledge, only narrow tasks such as hate speech detection are offered. To address this gap, we introduce a benchmark of Cultural and Linguistic Intelligence in Korean (CLIcK), a dataset comprising 1,995 QA pairs. CLIcK sources its data from official Korean exams and textbooks, partitioning the questions into eleven categories under the two main categories of language and culture. For each instance in click, we provide fine-grained annotation of which cultural and linguistic knowledge is required to correctly answer the question. Using CLIcK, we test 13 language models to assess their performance. Our evaluation uncovers insights into their performances across the categories, as well as the diverse factors affecting their comprehension. CLIcK offers the first large-scale comprehensive Korean-centric analysis of LLMs’ proficiency in Korean language and culture.

The integration of generative AI in education is expanding, yet empirical analyses of large-scale and real-world interactions between students and AI systems still remain limited. Addressing this gap, we present RECIPE4U (RECIPE for University), a dataset sourced from a semester-long experiment with 212 college students in English as Foreign Language (EFL) writing courses. During the study, students engaged in dialogues with ChatGPT to revise their essays. RECIPE4U includes comprehensive records of these interactions, including conversation logs, students’ intent, students’ self-rated satisfaction, and students’ essay edit histories. In particular, we annotate the students’ utterances in RECIPE4U with 13 intention labels based on our coding schemes. We establish baseline results for two subtasks in task-oriented dialogue systems within educational contexts: intent detection and satisfaction estimation. As a foundational step, we explore student-ChatGPT interaction patterns through RECIPE4U and analyze them by focusing on students’ dialogue, essay data statistics, and students’ essay edits. We further illustrate potential applications of RECIPE4U dataset for enhancing the incorporation of LLMs in educational frameworks. RECIPE4U is publicly available at https://zeunie.github.io/RECIPE4U/.

Most hate speech datasets neglect the cultural diversity within a single language, resulting in a critical shortcoming in hate speech detection. To address this, we introduce CREHate, a CRoss-cultural English Hate speech dataset. To construct CREHate, we follow a two-step procedure: 1) cultural post collection and 2) cross-cultural annotation. We sample posts from the SBIC dataset, which predominantly represents North America, and collect posts from four geographically diverse English-speaking countries (Australia, United Kingdom, Singapore, and South Africa) using culturally hateful keywords we retrieve from our survey. Annotations are collected from the four countries plus the United States to establish representative labels for each country. Our analysis highlights statistically significant disparities across countries in hate speech annotations. Only 56.2% of the posts in CREHate achieve consensus among all countries, with the highest pairwise label difference rate of 26%. Qualitative analysis shows that label disagreement occurs mostly due to different interpretations of sarcasm and the personal bias of annotators on divisive topics. Lastly, we evaluate large language models (LLMs) under a zero-shot setting and show that current LLMs tend to show higher accuracies on Anglosphere country labels in CREHate.Our dataset and codes are available at: https://github.com/nlee0212/CREHate

Warning: This paper contains examples of stereotypes and biases. The Bias Benchmark for Question Answering (BBQ) is designed to evaluate social biases of language models (LMs), but it is not simple to adapt this benchmark to cultural contexts other than the US because social biases depend heavily on the cultural context. In this paper, we present KoBBQ, a Korean bias benchmark dataset, and we propose a general framework that addresses considerations for cultural adaptation of a dataset. Our framework includes partitioning the BBQ dataset into three classes—Simply-Transferred (can be used directly after cultural translation), Target-Modified (requires localization in target groups), and Sample-Removed (does not fit Korean culture)—and adding four new categories of bias specific to Korean culture. We conduct a large-scale survey to collect and validate the social biases and the targets of the biases that reflect the stereotypes in Korean culture. The resulting KoBBQ dataset comprises 268 templates and 76,048 samples across 12 categories of social bias. We use KoBBQ to measure the accuracy and bias scores of several state-of-the-art multilingual LMs. The results clearly show differences in the bias of LMs as measured by KoBBQ and a machine-translated version of BBQ, demonstrating the need for and utility of a well-constructed, culturally aware social bias benchmark.

The potential social harms that large language models pose, such as generating offensive content and reinforcing biases, are steeply rising. Existing works focus on coping with this concern while interacting with ill-intentioned users, such as those who explicitly make hate speech or elicit harmful responses. However, discussions on sensitive issues can become toxic even if the users are well-intentioned. For safer models in such scenarios, we present the Sensitive Questions and Acceptable Response (SQuARe) dataset, a large-scale Korean dataset of 49k sensitive questions with 42k acceptable and 46k non-acceptable responses. The dataset was constructed leveraging HyperCLOVA in a human-in-the-loop manner based on real news headlines. Experiments show that acceptable response generation significantly improves for HyperCLOVA and GPT-3, demonstrating the efficacy of this dataset.

Research on Korean grammatical error correction (GEC) is limited, compared to other major languages such as English. We attribute this problematic circumstance to the lack of a carefully designed evaluation benchmark for Korean GEC. In this work, we collect three datasets from different sources (Kor-Lang8, Kor-Native, and Kor-Learner) that covers a wide range of Korean grammatical errors. Considering the nature of Korean grammar, We then define 14 error types for Korean and provide KAGAS (Korean Automatic Grammatical error Annotation System), which can automatically annotate error types from parallel corpora. We use KAGAS on our datasets to make an evaluation benchmark for Korean, and present baseline models trained from our datasets. We show that the model trained with our datasets significantly outperforms the currently used statistical Korean GEC system (Hanspell) on a wider range of error types, demonstrating the diversity and usefulness of the datasets. The implementations and datasets are open-sourced.

Researchers have traditionally recruited native speakers to provide annotations for the widely used benchmark datasets. But there are languages for which recruiting native speakers is difficult, and it would help to get learners of those languages to annotate the data. In this paper, we investigate whether language learners can contribute annotations to the benchmark datasets. In a carefully controlled annotation experiment, we recruit 36 language learners, provide two types of additional resources (dictionaries and machine-translated sentences), and perform mini-tests to measure their language proficiency. We target three languages, English, Korean, and Indonesian, and four NLP tasks, sentiment analysis, natural language inference, named entity recognition, and machine reading comprehension. We find that language learners, especially those with intermediate or advanced language proficiency, are able to provide fairly accurate labels with the help of additional resources. Moreover, we show that data annotation improves learners’ language proficiency in terms of vocabulary and grammar. The implication of our findings is that broadening the annotation task to include language learners can open up the opportunity to build benchmark datasets for languages for which it is difficult to recruit native speakers.

Unsupervised sentence representation learning has progressed through contrastive learning and data augmentation methods such as dropout masking. Despite this progress, sentence encoders are still limited to using only an input sentence when predicting its semantic vector. In this work, we show that the semantic meaning of a sentence is also determined by nearest-neighbor sentences that are similar to the input sentence. Based on this finding, we propose a novel unsupervised sentence encoder, RankEncoder. RankEncoder predicts the semantic vector of an input sentence by leveraging its relationship with other sentences in an external corpus, as well as the input sentence itself. We evaluate RankEncoder on semantic textual benchmark datasets. From the experimental results, we verify that 1) RankEncoder achieves 80.07% Spearman’s correlation, a 1.1% absolute improvement compared to the previous state-of-the-art performance, 2) RankEncoder is universally applicable to existing unsupervised sentence embedding methods, and 3) RankEncoder is specifically effective for predicting the similarity scores of similar sentence pairs.

Increasingly, language models and machine translation are becoming valuable tools to help people communicate with others from diverse cultural backgrounds. However, current language models lack cultural awareness because they are trained on data representing only the culture within the dataset. This presents a problem in the context of hate speech classification, where cultural awareness is especially critical. This study aims to quantify the cultural insensitivity of three monolingual (Korean, English, Arabic) hate speech classifiers by evaluating their performance on translated datasets from the other two languages. Our research has revealed that hate speech classifiers evaluated on datasets from other cultures yield significantly lower F1 scores, up to almost 50%. In addition, they produce considerably higher false negative rates, with a magnitude up to five times greater, demonstrating the extent of the cultural gap. The study highlights the severity of cultural insensitivity of language models in hate speech classification.

Time is one of the crucial factors in real-world question answering (QA) problems. However, language models have difficulty understanding the relationships between time specifiers, such as ‘after’ and ‘before’, and numbers, since existing QA datasets do not include sufficient time expressions. To address this issue, we propose a Time-Context aware Question Answering (TCQA) framework. We suggest a Time-Context dependent Span Extraction (TCSE) task, and build a time-context dependent data generation framework for model training. Moreover, we present a metric to evaluate the time awareness of the QA model using TCSE. The TCSE task consists of a question and four sentence candidates classified as correct or incorrect based on time and context. The model is trained to extract the answer span from the sentence that is both correct in time and context. The model trained with TCQA outperforms baseline models up to 8.5 of the F1-score in the TimeQA dataset. Our dataset and code are available at https://github.com/sonjbin/TCQA

Domain-specific documents cover terminologies and specialized knowledge. This has been the main challenge of domain-specific document retrieval systems. Previous approaches propose domain-adaptation and transfer learning methods to alleviate this problem. However, these approaches still follow the same document representation method in previous approaches; a document is embedded into a single vector. In this study, we propose VKGDR. VKGDR represents a given corpus into a graph of entities and their relations (known as a virtual knowledge graph) and computes the relevance between queries and documents based on the graph representation. We conduct three experiments 1) domain-specific document retrieval, 2) comparison of our virtual knowledge graph construction method with previous approaches, and 3) ablation study on each component of our virtual knowledge graph. From the results, we see that unsupervised VKGDR outperforms baselines in a zero-shot setting and even outperforms fully-supervised bi-encoder. We also verify that our virtual knowledge graph construction method results in better retrieval performance than previous approaches.

Machine Reading Comprehension (MRC) has become one of the essential tasks in Natural Language Understanding (NLU) as it is often included in several NLU benchmarks (Liang et al., 2020; Wilie et al., 2020). However, most MRC datasets only have answerable question type, overlooking the importance of unanswerable questions. MRC models trained only on answerable questions will select the span that is most likely to be the answer, even when the answer does not actually exist in the given passage (Rajpurkar et al., 2018). This problem especially remains in medium- to low-resource languages like Indonesian. Existing Indonesian MRC datasets (Purwarianti et al., 2007; Clark et al., 2020) are still inadequate because of the small size and limited question types, i.e., they only cover answerable questions. To fill this gap, we build a new Indonesian MRC dataset called I(n)don’tKnow- MRC (IDK-MRC) by combining the automatic and manual unanswerable question generation to minimize the cost of manual dataset construction while maintaining the dataset quality. Combined with the existing answerable questions, IDK-MRC consists of more than 10K questions in total. Our analysis shows that our dataset significantly improves the performance of Indonesian MRC models, showing a large improvement for unanswerable questions.

Recent directions for offensive language detection are hierarchical modeling, identifying the type and the target of offensive language, and interpretability with offensive span annotation and prediction. These improvements are focused on English and do not transfer well to other languages because of cultural and linguistic differences. In this paper, we present the Korean Offensive Language Dataset (KOLD) comprising 40,429 comments, which are annotated hierarchically with the type and the target of offensive language, accompanied by annotations of the corresponding text spans. We collect the comments from NAVER news and YouTube platform and provide the titles of the articles and videos as the context information for the annotation process. We use these annotated comments as training data for Korean BERT and RoBERTa models and find that they are effective at offensiveness detection, target classification, and target span detection while having room for improvement for target group classification and offensive span detection. We discover that the target group distribution differs drastically from the existing English datasets, and observe that providing the context information improves the model performance in offensiveness detection (+0.3), target classification (+1.5), and target group classification (+13.1). We publicly release the dataset and baseline models.

The retriever-reader pipeline has shown promising performance in open-domain QA but suffers from a very slow inference speed. Recently proposed question retrieval models tackle this problem by indexing question-answer pairs and searching for similar questions. These models have shown a significant increase in inference speed, but at the cost of lower QA performance compared to the retriever-reader models. This paper proposes a two-step question retrieval model, SQuID (Sequential Question-Indexed Dense retrieval) and distant supervision for training. SQuID uses two bi-encoders for question retrieval. The first-step retriever selects top-k similar questions, and the second-step retriever finds the most similar question from the top-k questions. We evaluate the performance and the computational efficiency of SQuID. The results show that SQuID significantly increases the performance of existing question retrieval models with a negligible loss on inference speed.

Historical records in Korea before the 20th century were primarily written in Hanja, an extinct language based on Chinese characters and not understood by modern Korean or Chinese speakers. Historians with expertise in this time period have been analyzing the documents, but that process is very difficult and time-consuming, and language models would significantly speed up the process. Toward building and evaluating language models for Hanja, we release the Hanja Understanding Evaluation dataset consisting of chronological attribution, topic classification, named entity recognition, and summary retrieval tasks. We also present BERT-based models continued training on the two major corpora from the 14th to the 19th centuries: the Annals of the Joseon Dynasty and Diaries of the Royal Secretariats. We compare the models with several baselines on all tasks and show there are significant improvements gained by training on the two corpora. Additionally, we run zero-shot experiments on the Daily Records of the Royal Court and Important Officials (DRRI). The DRRI dataset has not been studied much by the historians, and not at all by the NLP community.

The Annals of Joseon Dynasty (AJD) contain the daily records of the Kings of Joseon, the 500-year kingdom preceding the modern nation of Korea.The Annals were originally written in an archaic Korean writing system, ‘Hanja’, and were translated into Korean from 1968 to 1993.The resulting translation was however too literal and contained many archaic Korean words; thus, a new expert translation effort began in 2012. Since then, the records of only one king have been completed in a decade.In parallel, expert translators are working on English translation, also at a slow pace and produced only one king’s records in English so far.Thus, we propose H2KE, a neural machine translation model, that translates historical documents in Hanja to more easily understandable Korean and to English.Built on top of multilingual neural machine translation, H2KE learns to translate a historical document written in Hanja, from both a full dataset of outdated Korean translation and a small dataset of more recently translated contemporary Korean and English.We compare our method against two baselines:a recent model that simultaneously learns to restore and translate Hanja historical documentand a Transformer based model trained only on newly translated corpora.The experiments reveal that our method significantly outperforms the baselines in terms of BLEU scores for both contemporary Korean and English translations.We further conduct extensive human evaluation which shows that our translation is preferred over the original expert translations by both experts and non-expert Korean speakers.

Knowledge distillation is widely used to transfer the language understanding of a large model to a smaller model. However, after knowledge distillation, it was found that the smaller model is more biased by gender compared to the source large model. This paper studies what causes gender bias to increase after the knowledge distillation process. Moreover, we suggest applying a variant of the mixup on knowledge distillation, which is used to increase generalizability during the distillation process, not for augmentation. By doing so, we can significantly reduce the gender bias amplification after knowledge distillation. We also conduct an experiment on the GLUE benchmark to demonstrate that even if the mixup is applied, it does not have a significant adverse effect on the model’s performance.

We introduce CS1QA, a dataset for code-based question answering in the programming education domain. CS1QA consists of 9,237 question-answer pairs gathered from chat logs in an introductory programming class using Python, and 17,698 unannotated chat data with code. Each question is accompanied with the student’s code, and the portion of the code relevant to answering the question. We carefully design the annotation process to construct CS1QA, and analyze the collected dataset in detail. The tasks for CS1QA are to predict the question type, the relevant code snippet given the question and the code and retrieving an answer from the annotated corpus. Results for the experiments on several baseline models are reported and thoroughly analyzed. The tasks for CS1QA challenge models to understand both the code and natural language. This unique dataset can be used as a benchmark for source code comprehension and question answering in the educational setting.

In this paper, we study ethnic bias and how it varies across languages by analyzing and mitigating ethnic bias in monolingual BERT for English, German, Spanish, Korean, Turkish, and Chinese. To observe and quantify ethnic bias, we develop a novel metric called Categorical Bias score. Then we propose two methods for mitigation; first using a multilingual model, and second using contextual word alignment of two monolingual models. We compare our proposed methods with monolingual BERT and show that these methods effectively alleviate the ethnic bias. Which of the two methods works better depends on the amount of NLP resources available for that language. We additionally experiment with Arabic and Greek to verify that our proposed methods work for a wider variety of languages.

We introduce EfficientCL, a memory-efficient continual pretraining method that applies contrastive learning with novel data augmentation and curriculum learning. For data augmentation, we stack two types of operation sequentially: cutoff and PCA jittering. While pretraining steps proceed, we apply curriculum learning by incrementing the augmentation degree for each difficulty step. After data augmentation is finished, contrastive learning is applied on projected embeddings of original and augmented examples. When finetuned on GLUE benchmark, our model outperforms baseline models, especially for sentence-level tasks. Additionally, this improvement is capable with only 70% of computational memory compared to the baseline model.

We present a model to predict fine-grained emotions along the continuous dimensions of valence, arousal, and dominance (VAD) with a corpus with categorical emotion annotations. Our model is trained by minimizing the EMD (Earth Mover’s Distance) loss between the predicted VAD score distribution and the categorical emotion distributions sorted along VAD, and it can simultaneously classify the emotion categories and predict the VAD scores for a given sentence. We use pre-trained RoBERTa-Large and fine-tune on three different corpora with categorical labels and evaluate on EmoBank corpus with VAD scores. We show that our approach reaches comparable performance to that of the state-of-the-art classifiers in categorical emotion classification and shows significant positive correlations with the ground truth VAD scores. Also, further training with supervision of VAD labels leads to improved performance especially when dataset is small. We also present examples of predictions of appropriate emotion words that are not part of the original annotations.

Bill writing is a critical element of representative democracy. However, it is often overlooked that most legislative bills are derived, or even directly copied, from other bills. Despite the significance of bill-to-bill linkages for understanding the legislative process, existing approaches fail to address semantic similarities across bills, let alone reordering or paraphrasing which are prevalent in legal document writing. In this paper, we overcome these limitations by proposing a 5-class classification task that closely reflects the nature of the bill generation process. In doing so, we construct a human-labeled dataset of 4,721 bill-to-bill relationships at the subsection-level and release this annotated dataset to the research community. To augment the dataset, we generate synthetic data with varying degrees of similarity, mimicking the complex bill writing process. We use BERT variants and apply multi-stage training, sequentially fine-tuning our models with synthetic and human-labeled datasets. We find that the predictive performance significantly improves when training with both human-labeled and synthetic data. Finally, we apply our trained model to infer section- and bill-level similarities. Our analysis shows that the proposed methodology successfully captures the similarities across legal documents at various levels of aggregation.

Finding counterevidence to statements is key to many tasks, including counterargument generation. We build a system that, given a statement, retrieves counterevidence from diverse sources on the Web. At the core of this system is a natural language inference (NLI) model that determines whether a candidate sentence is valid counterevidence or not. Most NLI models to date, however, lack proper reasoning abilities necessary to find counterevidence that involves complex inference. Thus, we present a knowledge-enhanced NLI model that aims to handle causality- and example-based inference by incorporating knowledge graphs. Our NLI model outperforms baselines for NLI tasks, especially for instances that require the targeted inference. In addition, this NLI model further improves the counterevidence retrieval system, notably finding complex counterevidence better.

Automatic evaluation of open-domain dialogue response generation is very challenging because there are many appropriate responses for a given context. Existing evaluation models merely compare the generated response with the ground truth response and rate many of the appropriate responses as inappropriate if they deviate from the ground truth. One approach to resolve this problem is to consider the similarity of the generated response with the conversational context. In this paper, we propose an automatic evaluation model based on that idea and learn the model parameters from an unlabeled conversation corpus. Our approach considers the speakers in defining the different levels of similar context. We use a Twitter conversation corpus that contains many speakers and conversations to test our evaluation model. Experiments show that our model outperforms the other existing evaluation metrics in terms of high correlation with human annotation scores. We also show that our model trained on Twitter can be applied to movie dialogues without any additional training. We provide our code and the learned parameters so that they can be used for automatic evaluation of dialogue response generation models.

Extractive QA models have shown very promising performance in predicting the correct answer to a question for a given passage. However, they sometimes result in predicting the correct answer text but in a context irrelevant to the given question. This discrepancy becomes especially important as the number of occurrences of the answer text in a passage increases. To resolve this issue, we propose BLANC (BLock AttentioN for Context prediction) based on two main ideas: context prediction as an auxiliary task in multi-task learning manner, and a block attention method that learns the context prediction task. With experiments on reading comprehension, we show that BLANC outperforms the state-of-the-art QA models, and the performance gap increases as the number of answer text occurrences increases. We also conduct an experiment of training the models using SQuAD and predicting the supporting facts on HotpotQA and show that BLANC outperforms all baseline models in this zero-shot setting.

We analyze social media for detecting the suicidal risk of military personnel, which is especially crucial for countries with compulsory military service such as the Republic of Korea. From a widely-used Korean social Q&A site, we collect posts containing military-relevant content written by active-duty military personnel. We then annotate the posts with two groups of experts: military experts and mental health experts. Our dataset includes 2,791 posts with 13,955 corresponding expert annotations of suicidal risk levels, and this dataset is available to researchers who consent to research ethics agreement. Using various fine-tuned state-of-the-art language models, we predict the level of suicide risk, reaching .88 F1 score for classifying the risks.

Generating appropriate conversation responses requires careful modeling of the utterances and speakers together. Some recent approaches to response generation model both the utterances and the speakers, but these approaches tend to generate responses that are overly tailored to the speakers. To overcome this limitation, we propose a new model with a stochastic variable designed to capture the speaker information and deliver it to the conversational context. An important part of this model is the network of speakers in which each speaker is connected to one or more conversational partner, and this network is then used to model the speakers better. To test whether our model generates more appropriate conversation responses, we build a new conversation corpus containing approximately 27,000 speakers and 770,000 conversations. With this corpus, we run experiments of generating conversational responses and compare our model with other state-of-the-art models. By automatic evaluation metrics and human evaluation, we show that our model outperforms other models in generating appropriate responses. An additional advantage of our model is that it generates better responses for various new user scenarios, for example when one of the speakers is a known user in our corpus but the partner is a new user. For replicability, we make available all our code and data.

Additive compositionality of word embedding models has been studied from empirical and theoretical perspectives. Existing research on justifying additive compositionality of existing word embedding models requires a rather strong assumption of uniform word distribution. In this paper, we relax that assumption and propose more realistic conditions for proving additive compositionality, and we develop a novel word and sub-word embedding model that satisfies additive compositionality under those conditions. We then empirically show our model’s improved semantic representation performance on word similarity and noisy sentence similarity.

The recent surge of text-based online counseling applications enables us to collect and analyze interactions between counselors and clients. A dataset of those interactions can be used to learn to automatically classify the client utterances into categories that help counselors in diagnosing client status and predicting counseling outcome. With proper anonymization, we collect counselor-client dialogues, define meaningful categories of client utterances with professional counselors, and develop a novel neural network model for classifying the client utterances. The central idea of our model, ConvMFiT, is a pre-trained conversation model which consists of a general language model built from an out-of-domain corpus and two role-specific language models built from unlabeled in-domain dialogues. The classification result shows that ConvMFiT outperforms state-of-the-art comparison models. Further, the attention weights in the learned model confirm that the model finds expected linguistic patterns for each category.

Styles of leaders when they make decisions in groups vary, and the different styles affect the performance of the group. To understand the key words and speakers associated with decisions, we initially formalize the problem as one of predicting leaders’ decisions from discussion with group members. As a dataset, we introduce conversational meeting records from a historical corpus, and develop a hierarchical RNN structure with attention and pre-trained speaker embedding in the form of a, Conversational Decision Making Model (CDMM). The CDMM outperforms other baselines to predict leaders’ final decisions from the data. We explain why CDMM works better than other methods by showing the key words and speakers discovered from the attentions as evidence.

In news and discussions, many articles and posts are provided without their related previous articles or posts. Hence, it is difficult to understand the context from which the articles and posts have occurred. In this paper, we propose the Hierarchical Dirichlet Gaussian Marked Hawkes process (HD-GMHP) for reconstructing the narratives and thread structures of news articles and discussion posts. HD-GMHP unifies three modeling strategies in previous research: temporal characteristics, triggering event relations, and meta information of text in news articles and discussion threads. To show the effectiveness of the model, we perform experiments in narrative reconstruction and thread reconstruction with real world datasets: articles from the New York Times and a corpus of Wikipedia conversations. The experimental results show that HD-GMHP outperforms the baselines of LDA, HDP, and HDHP for both tasks.

Research on distributed word representations is focused on widely-used languages such as English. Although the same methods can be used for other languages, language-specific knowledge can enhance the accuracy and richness of word vector representations. In this paper, we look at improving distributed word representations for Korean using knowledge about the unique linguistic structure of Korean. Specifically, we decompose Korean words into the jamo-level, beyond the character-level, allowing a systematic use of subword information. To evaluate the vectors, we develop Korean test sets for word similarity and analogy and make them publicly available. The results show that our simple method outperforms word2vec and character-level Skip-Grams on semantic and syntactic similarity and analogy tasks and contributes positively toward downstream NLP tasks such as sentiment analysis.

Vector representation of words improves performance in various NLP tasks, but the high dimensional word vectors are very difficult to interpret. We apply several rotation algorithms to the vector representation of words to improve the interpretability. Unlike previous approaches that induce sparsity, the rotated vectors are interpretable while preserving the expressive performance of the original vectors. Furthermore, any prebuilt word vector representation can be rotated for improved interpretability. We apply rotation to skipgrams and glove and compare the expressive power and interpretability with the original vectors and the sparse overcomplete vectors. The results show that the rotated vectors outperform the original and the sparse overcomplete vectors for interpretability and expressiveness tasks.

Much of scientific progress stems from previously published findings, but searching through the vast sea of scientific publications is difficult. We often rely on metrics of scholarly authority to find the prominent authors but these authority indices do not differentiate authority based on research topics. We present Latent Topical-Authority Indexing (LTAI) for jointly modeling the topics, citations, and topical authority in a corpus of academic papers. Compared to previous models, LTAI differs in two main aspects. First, it explicitly models the generative process of the citations, rather than treating the citations as given. Second, it models each author’s influence on citations of a paper based on the topics of the cited papers, as well as the citing papers. We fit LTAI into four academic corpora: CORA, Arxiv Physics, PNAS, and Citeseer. We compare the performance of LTAI against various baselines, starting with the latent Dirichlet allocation, to the more advanced models including author-link topic model and dynamic author citation topic model. The results show that LTAI achieves improved accuracy over other similar models when predicting words, citations and authors of publications.