Hidetaka Kamigaito

Riza Setiawan Soetedjo, Yusuke Sakai, Hidetaka Kamigaito et al.

Improving the quality of model-generated summaries, especially factuality, the accuracy of a summary with respect to its source content, remains a challenge. While reranking could select the optimal output from multiple generated candidates, it is limited to only using the source as guidance, resulting in unreliable summaries. To address this limitation, we propose ConSUM that reranks candidate summaries by considering two factors: consistency to the source document and consensus among the other candidates. Consensus is established using Minimum Bayes Risk (MBR) decoding over the set of generated summaries, while ensuring consistency by employing factuality-aware metrics that compare the summary against the source. Rigorous testing demonstrates that our system is competitive with existing methods, with human evaluations further confirming that its generated summaries are preferred over those from other systems. Our code is available at https://github.com/naist-nlp/ConSUM .

Haruki Sakajo, Frederikus Hudi, Yusuke Sakai et al.

Language exhibits inherent structures, a property that explains both language acquisition and language change. Given this characteristic, we expect language models to manifest their own internal structures as well. While interpretability research has investigated how models compute representations mechanistically through attention patterns and Sparse AutoEncoders, the organization of the resulting representations is overlooked. To address this gap, we introduce StructLens, a framework to analyze representations through a holistic structural view. StructLens constructs maximum spanning trees based on the semantic representations in residual streams, inspired by tree representation in dependency parsing, and provides summaries of token relationships in representation space. We analyze how contiguous tokens are also nearby in representation space and find that middle layers show the strongest local-span organization. Moreover, analysis of pre-training checkpoints reveals that smaller local units become detectable earlier in pre-training, and larger units later. Our findings demonstrate that StructLens provides insights into how models organize token representations across layers and training. Our code is available at https://github.com/naist-nlp/structlens.

Haruki Sakajo, Yusuke Sakai, Hidetaka Kamigaito et al.

Large language models (LLMs) have excelled in processing multiple languages through pre- and post-training on multilingual data, even though English dominates the training data. Prior work focusing on token representations has revealed how those LLMs process non-English text. Although these analyses have provided insightful findings, they fail to capture a structural view, which is an inherent property of language. In this study, we explore the multilinguality of LLMs through representational structural analysis. Our findings reveal that low-resource languages are structurally more different from English than high- and mid-resource languages, and that language-specific post-training alters their structures while preserving inter-language relationships.

Hiroyuki Deguchi, Yusuke Sakai, Hidetaka Kamigaito et al.

Minimum Bayes risk (MBR) decoding is a decision rule of text generation tasks that outperforms conventional maximum a posterior (MAP) decoding using beam search by selecting high-quality outputs based on a utility function rather than those with high-probability. Typically, it finds the most suitable hypothesis from the set of hypotheses under the sampled pseudo-references. mbrs is a library of MBR decoding, which can flexibly combine various metrics, alternative expectation estimations, and algorithmic variants. It is designed with a focus on speed measurement and calling count of code blocks, transparency, reproducibility, and extensibility, which are essential for researchers and developers. We published our mbrs as an MIT-licensed open-source project, and the code is available on GitHub. GitHub: https://github.com/naist-nlp/mbrs

Shintaro Ozaki, Kazuki Hayashi, Yusuke Sakai et al.

As the performance of Large-scale Vision Language Models (LVLMs) improves, they are increasingly capable of responding in multiple languages, and there is an expectation that the demand for explanations generated by LVLMs will grow. However, pre-training of Vision Encoder and the integrated training of LLMs with Vision Encoder are mainly conducted using English training data, leaving it uncertain whether LVLMs can completely handle their potential when generating explanations in languages other than English. In addition, multilingual QA benchmarks that create datasets using machine translation have cultural differences and biases, remaining issues for use as evaluation tasks. To address these challenges, this study created an extended dataset in multiple languages without relying on machine translation. This dataset that takes into account nuances and country-specific phrases was then used to evaluate the generation explanation abilities of LVLMs. Furthermore, this study examined whether Instruction-Tuning in resource-rich English improves performance in other languages. Our findings indicate that LVLMs perform worse in languages other than English compared to English. In addition, it was observed that LVLMs struggle to effectively manage the knowledge learned from English data. Our dataset is available at https://huggingface.co/datasets/naist-nlp/MultiExpArt

Naoki Kobayashi, Tsutomu Hirao, Hidetaka Kamigaito et al.

To promote and further develop RST-style discourse parsing models, we need a strong baseline that can be regarded as a reference for reporting reliable experimental results. This paper explores a strong baseline by integrating existing simple parsing strategies, top-down and bottom-up, with various transformer-based pre-trained language models. The experimental results obtained from two benchmark datasets demonstrate that the parsing performance strongly relies on the pretrained language models rather than the parsing strategies. In particular, the bottom-up parser achieves large performance gains compared to the current best parser when employing DeBERTa. We further reveal that language models with a span-masking scheme especially boost the parsing performance through our analysis within intra- and multi-sentential parsing, and nuclearity prediction.

Soichiro Murakami, Peinan Zhang, Sho Hoshino et al.

Writing an ad text that attracts people and persuades them to click or act is essential for the success of search engine advertising. Therefore, ad creators must consider various aspects of advertising appeals (A$^3$) such as the price, product features, and quality. However, products and services exhibit unique effective A$^3$ for different industries. In this work, we focus on exploring the effective A$^3$ for different industries with the aim of assisting the ad creation process. To this end, we created a dataset of advertising appeals and used an existing model that detects various aspects for ad texts. Our experiments demonstrated that different industries have their own effective A$^3$ and that the identification of the A$^3$ contributes to the estimation of advertising performance.

Hidetaka Kamigaito, Katsuhiko Hayashi

Negative sampling (NS) loss plays an important role in learning knowledge graph embedding (KGE) to handle a huge number of entities. However, the performance of KGE degrades without hyperparameters such as the margin term and number of negative samples in NS loss being appropriately selected. Currently, empirical hyperparameter tuning addresses this problem at the cost of computational time. To solve this problem, we theoretically analyzed NS loss to assist hyperparameter tuning and understand the better use of the NS loss in KGE learning. Our theoretical analysis showed that scoring methods with restricted value ranges, such as TransE and RotatE, require appropriate adjustment of the margin term or the number of negative samples different from those without restricted value ranges, such as RESCAL, ComplEx, and DistMult. We also propose subsampling methods specialized for the NS loss in KGE studied from a theoretical aspect. Our empirical analysis on the FB15k-237, WN18RR, and YAGO3-10 datasets showed that the results of actually trained models agree with our theoretical findings.

Toshiki Kawamoto, Hidetaka Kamigaito, Kotaro Funakoshi et al.

A repetition is a response that repeats words in the previous speaker's utterance in a dialogue. Repetitions are essential in communication to build trust with others, as investigated in linguistic studies. In this work, we focus on repetition generation. To the best of our knowledge, this is the first neural approach to address repetition generation. We propose Weighted Label Smoothing, a smoothing method for explicitly learning which words to repeat during fine-tuning, and a repetition scoring method that can output more appropriate repetitions during decoding. We conducted automatic and human evaluations involving applying these methods to the pre-trained language model T5 for generating repetitions. The experimental results indicate that our methods outperformed baselines in both evaluations.

Yusuke Ide, Yuto Nishida, Justin Vasselli et al.

The grammatical knowledge of language models (LMs) is often measured using a benchmark of linguistic minimal pairs, where the LMs are presented with a pair of acceptable and unacceptable sentences and required to judge which is more acceptable. Conventional approaches directly compare sentence probabilities assigned by LMs, but recent large language models (LLMs) are trained to perform tasks via prompting, and thus, the raw probabilities they assign may not fully reflect their grammatical knowledge. In this study, we attempt to derive more accurate acceptability judgments from LLMs using prompts and templates. Through extensive experiments in English and Chinese, we compare nine judgment methods and find two of them, a probability readout method -- in-template LP and a prompt-based method -- Yes/No probability computing, achieve higher accuracy than the conventional ones. Our analysis reveals that these methods excel in different linguistic phenomena, suggesting they access different aspects of LLMs' knowledge. We also find that ensembling the two methods outperforms single methods. Consequently, we recommend these techniques, either individually or ensembled, as more effective alternatives to conventional approaches for assessing grammatical knowledge in LLMs.

Xincan Feng, Hidetaka Kamigaito, Katsuhiko Hayashi et al.

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.

Yusuke Sakai, Adam Nohejl, Jiangnan Hang et al.

The natural language understanding (NLU) performance of large language models (LLMs) has been evaluated across various tasks and datasets. The existing evaluation methods, however, do not take into account the variance in scores due to differences in prompts, which leads to unfair evaluation and comparison of NLU performance. Moreover, evaluation designed for specific prompts is inappropriate for instruction tuning, which aims to perform well with any prompt. It is therefore necessary to find a way to measure NLU performance in a fair manner, considering score variance between different instruction templates. In this study, we provide English and Japanese cross-lingual datasets for evaluating the NLU performance of LLMs, which include multiple instruction templates for fair evaluation of each task, along with regular expressions to constrain the output format. Furthermore, we propose the Sharpe score as an evaluation metric that takes into account the variance in scores between templates. Comprehensive analysis of English and Japanese LLMs reveals that the high variance among templates has a significant impact on the fair evaluation of LLMs.

Hidetaka Kamigaito, Katsuhiko Hayashi, Taro Watanabe

In this paper, we propose a table and image generation task to verify how the knowledge about entities acquired from natural language is retained in Vision & Language (V&L) models. This task consists of two parts: the first is to generate a table containing knowledge about an entity and its related image, and the second is to generate an image from an entity with a caption and a table containing related knowledge of the entity. In both tasks, the model must know the entities used to perform the generation properly. We created the Wikipedia Table and Image Generation (WikiTIG) dataset from about 200,000 infoboxes in English Wikipedia articles to perform the proposed tasks. We evaluated the performance on the tasks with respect to the above research question using the V&L model OFA, which has achieved state-of-the-art results in multiple tasks. Experimental results show that OFA forgets part of its entity knowledge by pre-training as a complement to improve the performance of image related tasks.

Wataru Hashimoto, Hidetaka Kamigaito, Taro Watanabe

This work investigates the impact of data augmentation on confidence calibration and uncertainty estimation in Named Entity Recognition (NER) tasks. For the future advance of NER in safety-critical fields like healthcare and finance, it is essential to achieve accurate predictions with calibrated confidence when applying Deep Neural Networks (DNNs), including Pre-trained Language Models (PLMs), as a real-world application. However, DNNs are prone to miscalibration, which limits their applicability. Moreover, existing methods for calibration and uncertainty estimation are computational expensive. Our investigation in NER found that data augmentation improves calibration and uncertainty in cross-genre and cross-lingual setting, especially in-domain setting. Furthermore, we showed that the calibration for NER tends to be more effective when the perplexity of the sentences generated by data augmentation is lower, and that increasing the size of the augmentation further improves calibration and uncertainty.

Soichiro Murakami, Hidetaka Kamigaito, Hiroya Takamura et al.

Humor is a salient testbed for human-like creative thinking in large language models (LLMs). We study humor using the Japanese creative response game Oogiri, in which participants produce witty responses to a given prompt, and ask the following research question: What makes such responses funny to humans? Previous work has offered only limited reliable means to answer this question. Existing datasets contain few candidate responses per prompt, expose popularity signals during ratings, and lack objective and comparable metrics for funniness. Thus, we introduce Oogiri-Master and Oogiri-Corpus, which are a benchmark and dataset designed to enable rigorous evaluation of humor understanding in LLMs. Each prompt is paired with approximately 100 diverse candidate responses, and funniness is rated independently by approximately 100 human judges without access to others' ratings, reducing popularity bias and enabling robust aggregation. Using Oogiri-Corpus, we conduct a quantitative analysis of the linguistic factors associated with funniness, such as text length, ambiguity, and incongruity resolution, and derive objective metrics for predicting human judgments. Subsequently, we benchmark a range of LLMs and human baselines in Oogiri-Master, demonstrating that state-of-the-art models approach human performance and that insight-augmented prompting improves the model performance. Our results provide a principled basis for evaluating and advancing humor understanding in LLMs.

Hiroshi Takato, Hiroshi Tsutsui, Komei Soda et al.

Identifying risky driving behavior in real-world situations is essential for the safety of both drivers and pedestrians. However, integrating natural language models in this field remains relatively untapped. To address this, we created a novel multi-modal instruction tuning dataset and driver coaching inference system. Our primary use case is dashcam-based coaching for commercial drivers. The North American Dashcam Market is expected to register a CAGR of 15.4 percent from 2022 to 2027. Our dataset enables language models to learn visual instructions across various risky driving scenarios, emphasizing detailed reasoning crucial for effective driver coaching and managerial comprehension. Our model is trained on road-facing and driver-facing RGB camera footage, capturing the comprehensive scope of driving behavior in vehicles equipped with dashcams.

Zhe Cao, Zhi Qu, Hidetaka Kamigaito et al.

Multilingual neural machine translation models support fine-tuning hundreds of languages simultaneously. However, fine-tuning on full parameters solely is inefficient potentially leading to negative interactions among languages. In this work, we demonstrate that the fine-tuning for a language occurs in its intrinsic language-specific subspace with a tiny fraction of entire parameters. Thus, we propose language-specific LoRA to isolate intrinsic language-specific subspaces. Furthermore, we propose architecture learning techniques and introduce a gradual pruning schedule during fine-tuning to exhaustively explore the optimal setting and the minimal intrinsic subspaces for each language, resulting in a lightweight yet effective fine-tuning procedure. The experimental results on a 12-language subset and a 30-language subset of FLORES-101 show that our methods not only outperform full-parameter fine-tuning up to 2.25 spBLEU scores but also reduce trainable parameters to $0.4\%$ for high and medium-resource languages and $1.6\%$ for low-resource ones.

Xincan Feng, Hidetaka Kamigaito, Katsuhiko Hayashi et al.

Knowledge Graphs (KGs) are fundamental resources in knowledge-intensive tasks in NLP. Due to the limitation of manually creating KGs, KG Completion (KGC) has an important role in automatically completing KGs by scoring their links with KG Embedding (KGE). To handle many entities in training, KGE relies on Negative Sampling (NS) loss that can reduce the computational cost by sampling. Since the appearance frequencies for each link are at most one in KGs, sparsity is an essential and inevitable problem. The NS loss is no exception. As a solution, the NS loss in KGE relies on smoothing methods like Self-Adversarial Negative Sampling (SANS) and subsampling. However, it is uncertain what kind of smoothing method is suitable for this purpose due to the lack of theoretical understanding. This paper provides theoretical interpretations of the smoothing methods for the NS loss in KGE and induces a new NS loss, Triplet Adaptive Negative Sampling (TANS), that can cover the characteristics of the conventional smoothing methods. Experimental results of TransE, DistMult, ComplEx, RotatE, HAKE, and HousE on FB15k-237, WN18RR, and YAGO3-10 datasets and their sparser subsets show the soundness of our interpretation and performance improvement by our TANS.

Yusuke Sakai, Hidetaka Kamigaito, Katsuhiko Hayashi et al.

Knowledge graphs (KGs) consist of links that describe relationships between entities. Due to the difficulty of manually enumerating all relationships between entities, automatically completing them is essential for KGs. Knowledge Graph Completion (KGC) is a task that infers unseen relationships between entities in a KG. Traditional embedding-based KGC methods, such as RESCAL, TransE, DistMult, ComplEx, RotatE, HAKE, HousE, etc., infer missing links using only the knowledge from training data. In contrast, the recent Pre-trained Language Model (PLM)-based KGC utilizes knowledge obtained during pre-training. Therefore, PLM-based KGC can estimate missing links between entities by reusing memorized knowledge from pre-training without inference. This approach is problematic because building KGC models aims to infer unseen links between entities. However, conventional evaluations in KGC do not consider inference and memorization abilities separately. Thus, a PLM-based KGC method, which achieves high performance in current KGC evaluations, may be ineffective in practical applications. To address this issue, we analyze whether PLM-based KGC methods make inferences or merely access memorized knowledge. For this purpose, we propose a method for constructing synthetic datasets specified in this analysis and conclude that PLMs acquire the inference abilities required for KGC through pre-training, even though the performance improvements mostly come from textual information of entities and relations.

Hidetaka Kamigaito, Katsuhiko Hayashi

In this article, we explain the recent advance of subsampling methods in knowledge graph embedding (KGE) starting from the original one used in word2vec.

Wataru Hashimoto, Hidetaka Kamigaito, Taro Watanabe

Trustworthiness in model predictions is crucial for safety-critical applications in the real world. However, deep neural networks often suffer from the issues of uncertainty estimation, such as miscalibration. In this study, we propose $k$-Nearest Neighbor Uncertainty Estimation ($k$NN-UE), which is a new uncertainty estimation method that uses not only the distances from the neighbors, but also the ratio of labels in the neighbors. Experiments on sentiment analysis, natural language inference, and named entity recognition show that our proposed method outperforms the baselines and recent density-based methods in several calibration and uncertainty metrics. Moreover, our analyses indicate that approximate nearest neighbor search techniques reduce the inference overhead without significantly degrading the uncertainty estimation performance when they are appropriately combined.

Jingxuan Liu, Zhi Qu, Jin Tei et al.

Automatic evaluation metrics are essential for building multilingual translation systems. The common practice of evaluating these systems is averaging metric scores across languages, yet this is suspicious since metrics may suffer from cross-lingual scoring bias, where translations of equal quality receive different scores across languages. This problem has not been systematically studied because no benchmark exists that provides parallel-quality instances across languages, and expert annotation is not realistic. In this work, we propose XQ-MEval, a semi-automatically built dataset covering nine translation directions, to benchmark translation metrics. Specifically, we inject MQM-defined errors into gold translations automatically, filter them by native speakers for reliability, and merge errors to generate pseudo translations with controllable quality. These pseudo translations are then paired with corresponding sources and references to form triplets used in assessing the qualities of translation metrics. Using XQ-MEval, our experiments on nine representative metrics reveal the inconsistency between averaging and human judgment and provide the first empirical evidence of cross-lingual scoring bias. Finally, we propose a normalization strategy derived from XQ-MEval that aligns score distributions across languages, improving the fairness and reliability of multilingual metric evaluation.

Yusuke Sakai, Hidetaka Kamigaito, Taro Watanabe

Recently, we have often observed hallucinated citations or references that do not correspond to any existing work in papers under review, preprints, or published papers. Such hallucinated citations pose a serious concern to scientific reliability. When they appear in accepted papers, they may also negatively affect the credibility of conferences. In this study, we refer to hallucinated citations as "HalluCitation" and systematically investigate their prevalence and impact. We analyze all papers published at ACL, NAACL, and EMNLP in 2024 and 2025, including main conference, Findings, and workshop papers. Our analysis reveals that nearly 300 papers contain at least one HalluCitation, most of which were published in 2025. Notably, half of these papers were identified at EMNLP 2025, the most recent conference, indicating that this issue is rapidly increasing. Moreover, more than 100 such papers were accepted as main conference and Findings papers at EMNLP 2025, affecting the credibility.

Ken Nishida, Kojiro Machi, Kazuma Onishi et al.

The extreme multi-label classification~(XMC) task involves learning a classifier that can predict from a large label set the most relevant subset of labels for a data instance. While deep neural networks~(DNNs) have demonstrated remarkable success in XMC problems, the task is still challenging because it must deal with a large number of output labels, which make the DNN training computationally expensive. This paper addresses the issue by exploring the use of random circular vectors, where each vector component is represented as a complex amplitude. In our framework, we can develop an output layer and loss function of DNNs for XMC by representing the final output layer as a fully connected layer that directly predicts a low-dimensional circular vector encoding a set of labels for a data instance. We conducted experiments on synthetic datasets to verify that circular vectors have better label encoding capacity and retrieval ability than normal real-valued vectors. Then, we conducted experiments on actual XMC datasets and found that these appealing properties of circular vectors contribute to significant improvements in task performance compared with a previous model using random real-valued vectors, while reducing the size of the output layers by up to 99%.

Xuefeng Wei, Zhixuan Wang, Xuan Zhou et al.

We introduce CARTBENCH, a museum-grounded benchmark for evaluating vision-language models (VLMs) on Chinese artworks beyond short-form recognition and QA. CARTBENCH comprises four subtasks: CURATORQA for evidence-grounded recognition and reasoning, CATALOGCAPTION for structured four-section expert-style appreciation, REINTERPRET for defensible reinterpretation with expert ratings, and CONNOISSEURPAIRS for diagnostic authenticity discrimination under visually similar confounds. CARTBENCH is built by aligning image-bearing Palace Museum objects from Wikidata with authoritative catalog pages, spanning five art categories across multiple dynasties. Across nine representative VLMs, we find that high overall CURATORQA accuracy can mask sharp drops on hard evidence linking and style-to-period inference; long-form appreciation remains far from expert references; and authenticity-oriented diagnostic discrimination stays near chance, underscoring the difficulty of connoisseur-level reasoning for current models.

Yusuke Sakai, Mana Makinae, Hidetaka Kamigaito et al.

In Simultaneous Machine Translation (SiMT) systems, training with a simultaneous interpretation (SI) corpus is an effective method for achieving high-quality yet low-latency systems. However, it is very challenging to curate such a corpus due to limitations in the abilities of annotators, and hence, existing SI corpora are limited. Therefore, we propose a method to convert existing speech translation corpora into interpretation-style data, maintaining the original word order and preserving the entire source content using Large Language Models (LLM-SI-Corpus). We demonstrate that fine-tuning SiMT models in text-to-text and speech-to-text settings with the LLM-SI-Corpus reduces latencies while maintaining the same level of quality as the models trained with offline datasets. The LLM-SI-Corpus is available at \url{https://github.com/yusuke1997/LLM-SI-Corpus}.

Soichiro Murakami, Peinan Zhang, Hidetaka Kamigaito et al.

Effective linguistic choices that attract potential customers play crucial roles in advertising success. This study aims to explore the linguistic features of ad texts that influence human preferences. Although the creation of attractive ad texts is an active area of research, progress in understanding the specific linguistic features that affect attractiveness is hindered by several obstacles. First, human preferences are complex and influenced by multiple factors, including their content, such as brand names, and their linguistic styles, making analysis challenging. Second, publicly available ad text datasets that include human preferences are lacking, such as ad performance metrics and human feedback, which reflect people's interests. To address these problems, we present AdParaphrase, a paraphrase dataset that contains human preferences for pairs of ad texts that are semantically equivalent but differ in terms of wording and style. This dataset allows for preference analysis that focuses on the differences in linguistic features. Our analysis revealed that ad texts preferred by human judges have higher fluency, longer length, more nouns, and use of bracket symbols. Furthermore, we demonstrate that an ad text-generation model that considers these findings significantly improves the attractiveness of a given text. The dataset is publicly available at: https://github.com/CyberAgentAILab/AdParaphrase.

Soichiro Murakami, Peinan Zhang, Hidetaka Kamigaito et al.

Identifying factors that make ad text attractive is essential for advertising success. This study proposes AdParaphrase v2.0, a dataset for ad text paraphrasing, containing human preference data, to enable the analysis of the linguistic factors and to support the development of methods for generating attractive ad texts. Compared with v1.0, this dataset is 20 times larger, comprising 16,460 ad text paraphrase pairs, each annotated with preference data from ten evaluators, thereby enabling a more comprehensive and reliable analysis. Through the experiments, we identified multiple linguistic features of engaging ad texts that were not observed in v1.0 and explored various methods for generating attractive ad texts. Furthermore, our analysis demonstrated the relationships between human preference and ad performance, and highlighted the potential of reference-free metrics based on large language models for evaluating ad text attractiveness. The dataset is publicly available at: https://github.com/CyberAgentAILab/AdParaphrase-v2.0.

Koki Natsumi, Hiroyuki Deguchi, Yusuke Sakai et al.

Minimum Bayes risk (MBR) decoding generates high-quality translations by maximizing the expected utility of output candidates, but it evaluates all pairwise scores over the candidate set; hence, it takes quadratic time with respect to the number of candidates. To reduce the number of utility function calls, probabilistic MBR (PMBR) decoding partially evaluates quality scores using sampled pairs of candidates and completes the missing scores with a matrix completion algorithm. Nevertheless, it degrades the translation quality as the number of utility function calls is reduced. Therefore, to improve the trade-off between quality and cost, we propose agreement-constrained PMBR (AC-PMBR) decoding, which leverages a knowledge distilled model to guide the completion of the score matrix. Our AC-PMBR decoding improved approximation errors of matrix completion by up to 3 times and achieved higher translation quality compared with PMBR decoding at a comparable computational cost on the WMT'23 En$\leftrightarrow$De translation tasks.

Kazuki Hayashi, Yusuke Sakai, Hidetaka Kamigaito et al.

Large-scale vision-language models (LVLMs) output text from images and instructions, demonstrating advanced capabilities in text generation and comprehension. However, it has not been clarified to what extent LVLMs understand the knowledge necessary for explaining images, the complex relationships between various pieces of knowledge, and how they integrate these understandings into their explanations. To address this issue, we propose a new task: the artwork explanation generation task, along with its evaluation dataset and metric for quantitatively assessing the understanding and utilization of knowledge about artworks. This task is apt for image description based on the premise that LVLMs are expected to have pre-existing knowledge of artworks, which are often subjects of wide recognition and documented information. It consists of two parts: generating explanations from both images and titles of artworks, and generating explanations using only images, thus evaluating the LVLMs' language-based and vision-based knowledge. Alongside, we release a training dataset for LVLMs to learn explanations that incorporate knowledge about artworks. Our findings indicate that LVLMs not only struggle with integrating language and visual information but also exhibit a more pronounced limitation in acquiring knowledge from images alone. The datasets (ExpArt=Explain Artworks) are available at https://huggingface.co/datasets/naist-nlp/ExpArt.

Haruki Sakajo, Yusuke Sakai, Hidetaka Kamigaito et al.

Vowels are primarily characterized by tongue position. Humans have discovered these features of vowel articulation through their own experience and explicit objective observation such as using MRI. With this knowledge and our experience, we can explain and understand the relationship between tongue positions and vowels, and this knowledge is helpful for language learners to learn pronunciation. Since language models (LMs) are trained on a large amount of data that includes linguistic and medical fields, our preliminary studies indicate that an LM is able to explain the pronunciation mechanisms of vowels. However, it is unclear whether multi-modal LMs, such as vision LMs, align textual information with visual information. One question arises: do LMs associate real tongue positions with vowel articulation? In this study, we created video and image datasets from the existing real-time MRI dataset and investigated whether LMs can understand vowel articulation based on tongue positions using vision-based information. Our findings suggest that LMs exhibit potential for understanding vowels and tongue positions when reference examples are provided while they have difficulties without them. Our code for dataset building is available on GitHub.

Hidetaka Kamigaito, Hiroyuki Deguchi, Yusuke Sakai et al.

Inference methods play an important role in eliciting the performance of large language models (LLMs). Currently, LLMs use inference methods utilizing generated multiple samples, which can be derived from Minimum Bayes Risk (MBR) Decoding. Previous studies have conducted empirical analyses to clarify the improvements in generation performance achieved by MBR decoding and have reported various observations. However, the theoretical underpinnings of these findings remain uncertain. To address this, we offer a new theoretical interpretation of MBR decoding from the perspective of bias-diversity decomposition. In this interpretation, the error in the quality estimation of hypotheses by MBR decoding is decomposed into two main factors: bias, which considers the closeness between the utility function and human evaluation, and diversity, which represents the variability in the quality estimation of the utility function. The theoretical analysis reveals the difficulty of simultaneously improving bias and diversity, confirming the validity of enhancing MBR decoding performance by increasing diversity. Furthermore, we reveal that diversity can explain one aspect of inference scaling laws that describe performance improvement by increasing sample size. Moreover, experiments across multiple NLP tasks yielded results consistent with these theoretical characteristics. Our code is available at https://github.com/naist-nlp/mbr-bias-diversity.

Huayang Li, Deng Cai, Zhi Qu et al.

Phrase-level dense retrieval has shown many appealing characteristics in downstream NLP tasks by leveraging the fine-grained information that phrases offer. In our work, we propose a new task formulation of dense retrieval, cross-lingual contextualized phrase retrieval, which aims to augment cross-lingual applications by addressing polysemy using context information. However, the lack of specific training data and models are the primary challenges to achieve our goal. As a result, we extract pairs of cross-lingual phrases using word alignment information automatically induced from parallel sentences. Subsequently, we train our Cross-lingual Contextualized Phrase Retriever (CCPR) using contrastive learning, which encourages the hidden representations of phrases with similar contexts and semantics to align closely. Comprehensive experiments on both the cross-lingual phrase retrieval task and a downstream task, i.e, machine translation, demonstrate the effectiveness of CCPR. On the phrase retrieval task, CCPR surpasses baselines by a significant margin, achieving a top-1 accuracy that is at least 13 points higher. When utilizing CCPR to augment the large-language-model-based translator, it achieves average gains of 0.7 and 1.5 in BERTScore for translations from X=>En and vice versa, respectively, on WMT16 dataset. Our code and data are available at \url{https://github.com/ghrua/ccpr_release}.

Boxuan Lyu, Haiyue Song, Hidetaka Kamigaito et al.

Error Span Detection (ESD) is a subtask of automatic machine translation evaluation that localizes error spans in translations and labels their severity. State-of-the-art generative ESD methods typically decode using Maximum a Posteriori (MAP), assuming that model-estimated probabilities are perfectly correlated with similarity to human annotation. However, we observed that annotations dissimilar to the human annotation could achieve a higher model likelihood than the human annotation. We address this issue by applying Minimum Bayes Risk (MBR) decoding to generative ESD models. Specifically, we employ sentence- and span-level similarity metrics as utility functions to select candidate hypotheses based on their approximate similarity to the human annotation. Extensive experimental results show that our MBR decoding outperforms the MAP baseline at the system, sentence, and span-levels. Furthermore, to mitigate the computational cost of MBR decoding, we demonstrate that applying MBR distillation enables a standard greedy model to match MBR decoding performance, effectively eliminating the inference-time latency bottleneck.

Jesse Atuhurra, Iqra Ali, Tomoya Iwakura et al.

Vision Language Models (VLMs) are pivotal for advancing perception in intelligent agents. Yet, evaluation of VLMs remains limited to predominantly English-centric benchmarks in which the image-text pairs comprise short texts. To evaluate VLM fine-grained abilities, in four languages under long-text settings, we introduce a novel multilingual benchmark VLURes featuring eight vision-and-language tasks, and a pioneering unrelatedness task, to probe the fine-grained Visual and Linguistic Understanding capabilities of VLMs across English, Japanese, and low-resource languages, Swahili, and Urdu. Our datasets, curated from web resources in the target language, encompass ten diverse image categories and rich textual context, introducing valuable vision-language resources for Swahili and Urdu. By prompting VLMs to generate responses and rationales, evaluated automatically and by native speakers, we uncover performance disparities across languages and tasks critical to intelligent agents, such as object recognition, scene understanding, and relationship understanding. We conducted evaluations of ten VLMs with VLURes. The best performing model, GPT-4o, achieves an overall accuracy of 90.8% and lags human performance by 6.7%, though the gap is larger for open-source models. The gap highlights VLURes' critical role in developing intelligent agents to tackle multi-modal visual reasoning.

Jesse Atuhurra, Hidetaka Kamigaito, Taro Watanabe et al.

We introduce J-ORA, a novel multimodal dataset that bridges the gap in robot perception by providing detailed object attribute annotations within Japanese human-robot dialogue scenarios. J-ORA is designed to support three critical perception tasks, object identification, reference resolution, and next-action prediction, by leveraging a comprehensive template of attributes (e.g., category, color, shape, size, material, and spatial relations). Extensive evaluations with both proprietary and open-source Vision Language Models (VLMs) reveal that incorporating detailed object attributes substantially improves multimodal perception performance compared to without object attributes. Despite the improvement, we find that there still exists a gap between proprietary and open-source VLMs. In addition, our analysis of object affordances demonstrates varying abilities in understanding object functionality and contextual relationships across different VLMs. These findings underscore the importance of rich, context-sensitive attribute annotations in advancing robot perception in dynamic environments. See project page at https://jatuhurrra.github.io/J-ORA/.

Katsuhiko Hayashi, Hidetaka Kamigaito

We prove that all standard subregular language classes are linearly separable when represented by their deciding predicates. This establishes finite observability and guarantees learnability with simple linear models. Synthetic experiments confirm perfect separability under noise-free conditions, while real-data experiments on English morphology show that learned features align with well-known linguistic constraints. These results demonstrate that the subregular hierarchy provides a rigorous and interpretable foundation for modeling natural language structure. Our code used in real-data experiments is available at https://github.com/UTokyo-HayashiLab/subregular.

Hyeyeon Kim, Sungwoo Han, Jingun Kwon et al.

In this study, we introduce a novel cover image generation task that produces both a concise summary and a visually corresponding image from a given text-only document. Because no existing datasets are available for this task, we propose a multimodal pseudo-labeling method to construct high-quality datasets at low cost. We first collect documents that contain multiple images with their captions, and their summaries by excluding factually inconsistent instances. Our approach selects one image from the multiple images accompanying the documents. Using the gold summary, we independently rank both the images and their captions. Then, we annotate a pseudo-label for an image when both the image and its corresponding caption are ranked first in their respective rankings. Finally, we remove documents that contain direct image references within texts. Experimental results demonstrate that the proposed multimodal pseudo-labeling method constructs more precise datasets and generates higher quality images than text- and image-only pseudo-labeling methods, which consider captions and images separately. We release our code at: https://github.com/HyeyeeonKim/MMCIG

Yusuke Sakai, Hidetaka Kamigaito, Taro Watanabe

It is very challenging to curate a dataset for language-specific knowledge and common sense in order to evaluate natural language understanding capabilities of language models. Due to the limitation in the availability of annotators, most current multilingual datasets are created through translation, which cannot evaluate such language-specific aspects. Therefore, we propose Multilingual CommonsenseQA (mCSQA) based on the construction process of CSQA but leveraging language models for a more efficient construction, e.g., by asking LM to generate questions/answers, refine answers and verify QAs followed by reduced human efforts for verification. Constructed dataset is a benchmark for cross-lingual language-transfer capabilities of multilingual LMs, and experimental results showed high language-transfer capabilities for questions that LMs could easily solve, but lower transfer capabilities for questions requiring deep knowledge or commonsense. This highlights the necessity of language-specific datasets for evaluation and training. Finally, our method demonstrated that multilingual LMs could create QA including language-specific knowledge, significantly reducing the dataset creation cost compared to manual creation. The datasets are available at https://huggingface.co/datasets/yusuke1997/mCSQA.

Boxuan Lyu, Soichiro Murakami, Hidetaka Kamigaito et al.

Mixture-of-Experts (MoE) architectures scale large language models efficiently by employing a parametric "router" to dispatch tokens to a sparse subset of experts. Typically, this router is trained once and then frozen, rendering routing decisions brittle under distribution shifts. We address this limitation by introducing kNN-MoE, a retrieval-augmented routing framework that reuses optimal expert assignments from a memory of similar past cases. This memory is constructed offline by directly optimizing token-wise routing logits to maximize the likelihood on a reference set. Crucially, we use the aggregate similarity of retrieved neighbors as a confidence-driven mixing coefficient, thus allowing the method to fall back to the frozen router when no relevant cases are found. Experiments show kNN-MoE outperforms zero-shot baselines and rivals computationally expensive supervised fine-tuning.

Yusuke Sakai, Hidetaka Kamigaito, Taro Watanabe

We introduce HalluCiteChecker, a toolkit for detecting and verifying hallucinated citations in scientific papers. While AI assistant technologies have transformed the academic writing process, including citation recommendation, they have also led to the emergence of hallucinated citations that do not correspond to any existing work. Such citations not only undermine the credibility of scientific papers but also impose an additional burden on reviewers and authors, who must manually verify their validity during the review process. In this study, we formalize hallucinated citation detection as an NLP task and provide a corresponding toolkit as a practical foundation for addressing this problem. Our package is lightweight and can perform verification in seconds on a standard laptop. It can also be executed entirely offline and runs efficiently using only CPUs. We hope that HalluCiteChecker will help reduce reviewer workload and support organizers by enabling systematic pre-review and publication checks. Our code is released under the Apache 2.0 license on GitHub and is distributed as an installable package via PyPI. A demonstration video is available on YouTube.

Hiroyuki Deguchi, Yusuke Sakai, Hidetaka Kamigaito et al.

Minimum Bayes risk (MBR) decoding achieved state-of-the-art translation performance by using COMET, a neural metric that has a high correlation with human evaluation. However, MBR decoding requires quadratic time since it computes the expected score between a translation hypothesis and all reference translations. We propose centroid-based MBR (CBMBR) decoding to improve the speed of MBR decoding. Our method clusters the reference translations in the feature space, and then calculates the score using the centroids of each cluster. The experimental results show that our CBMBR not only improved the decoding speed of the expected score calculation 5.7 times, but also outperformed vanilla MBR decoding in translation quality by up to 0.5 COMET in the WMT'22 En$\leftrightarrow$Ja, En$\leftrightarrow$De, En$\leftrightarrow$Zh, and WMT'23 En$\leftrightarrow$Ja translation tasks.

Shintaro Ozaki, Yuta Kato, Siyuan Feng et al.

Retrieval Augmented Generation (RAG) complements the knowledge of Large Language Models (LLMs) by leveraging external information to enhance response accuracy for queries. This approach is widely applied in several fields by taking its advantage of injecting the most up-to-date information, and researchers are focusing on understanding and improving this aspect to unlock the full potential of RAG in such high-stakes applications. However, despite the potential of RAG to address these needs, the mechanisms behind the confidence levels of its outputs remain underexplored. Our study focuses on the impact of RAG, specifically examining whether RAG improves the confidence of LLM outputs in the medical domain. We conduct this analysis across various configurations and models. We evaluate confidence by treating the model's predicted probability as its output and calculating several evaluation metrics which include calibration error method, entropy, the best probability, and accuracy. Experimental results across multiple datasets confirmed that certain models possess the capability to judge for themselves whether an inserted document relates to the correct answer. These results suggest that evaluating models based on their output probabilities determine whether they function as generators in the RAG framework. Our approach allows us to evaluate whether the models handle retrieved documents.

Yuto Nishida, Naoki Shikoda, Yosuke Kishinami et al.

Understanding what kinds of factual knowledge large language models (LLMs) memorize is essential for evaluating their reliability and limitations. Entity-based QA is a common framework for analyzing non-verbatim memorization, but typical evaluations query each entity using a single canonical surface form, making it difficult to disentangle fact memorization from access through a particular name. We introduce RedirectQA, an entity-based QA dataset that uses Wikipedia redirect information to associate Wikidata factual triples with categorized surface forms for each entity, including alternative names, abbreviations, spelling variants, and common erroneous forms. Across 13 LLMs, we examine surface-conditioned factual memorization and find that prediction outcomes often change when only the entity surface form changes. This inconsistency is category-dependent: models are more robust to minor orthographic variations than to larger lexical variations such as aliases and abbreviations. Frequency analyses further suggest that both entity- and surface-level frequencies are associated with accuracy, and that entity frequency often contributes beyond surface frequency. Overall, factual memorization appears neither purely surface-specific nor fully surface-invariant, highlighting the importance of surface-form diversity in evaluating non-verbatim memorization.

Aru Maekawa, Tsutomu Hirao, Hidetaka Kamigaito et al.

Recently, decoder-only pre-trained large language models (LLMs), with several tens of billion parameters, have significantly impacted a wide range of natural language processing (NLP) tasks. While encoder-only or encoder-decoder pre-trained language models have already proved to be effective in discourse parsing, the extent to which LLMs can perform this task remains an open research question. Therefore, this paper explores how beneficial such LLMs are for Rhetorical Structure Theory (RST) discourse parsing. Here, the parsing process for both fundamental top-down and bottom-up strategies is converted into prompts, which LLMs can work with. We employ Llama 2 and fine-tune it with QLoRA, which has fewer parameters that can be tuned. Experimental results on three benchmark datasets, RST-DT, Instr-DT, and the GUM corpus, demonstrate that Llama 2 with 70 billion parameters in the bottom-up strategy obtained state-of-the-art (SOTA) results with significant differences. Furthermore, our parsers demonstrated generalizability when evaluated on RST-DT, showing that, in spite of being trained with the GUM corpus, it obtained similar performances to those of existing parsers trained with RST-DT.

Seiji Gobara, Hidetaka Kamigaito, Taro Watanabe

Education that suits the individual learning level is necessary to improve students' understanding. The first step in achieving this purpose by using large language models (LLMs) is to adjust the textual difficulty of the response to students. This work analyzes how LLMs can implicitly adjust text difficulty between user input and its generated text. To conduct the experiments, we created a new dataset from Stack-Overflow to explore the performance of question-answering-based conversation. Experimental results on the Stack-Overflow dataset and the TSCC dataset, including multi-turn conversation show that LLMs can implicitly handle text difficulty between user input and its generated response. We also observed that some LLMs can surpass humans in handling text difficulty and the importance of instruction-tuning.

Huy Hien Vu, Hidetaka Kamigaito, Taro Watanabe

Despite significant improvements in enhancing the quality of translation, context-aware machine translation (MT) models underperform in many cases. One of the main reasons is that they fail to utilize the correct features from context when the context is too long or their models are overly complex. This can lead to the explain-away effect, wherein the models only consider features easier to explain predictions, resulting in inaccurate translations. To address this issue, we propose a model that explains the decisions made for translation by predicting coreference features in the input. We construct a model for input coreference by exploiting contextual features from both the input and translation output representations on top of an existing MT model. We evaluate and analyze our method in the WMT document-level translation task of English-German dataset, the English-Russian dataset, and the multilingual TED talk dataset, demonstrating an improvement of over 1.0 BLEU score when compared with other context-aware models.

Jesse Atuhurra, Hiroyuki Shindo, Hidetaka Kamigaito et al.

Many attempts have been made in multilingual NLP to ensure that pre-trained language models, such as mBERT or GPT2 get better and become applicable to low-resource languages. To achieve multilingualism for pre-trained language models (PLMs), we need techniques to create word embeddings that capture the linguistic characteristics of any language. Tokenization is one such technique because it allows for the words to be split based on characters or subwords, creating word embeddings that best represent the structure of the language. Creating such word embeddings is essential to applying PLMs to other languages where the model was not trained, enabling multilingual NLP. However, most PLMs use generic tokenization methods like BPE, wordpiece, or unigram which may not suit specific languages. We hypothesize that tokenization based on syllables within the input text, which we call syllable tokenization, should facilitate the development of syllable-aware language models. The syllable-aware language models make it possible to apply PLMs to languages that are rich in syllables, for instance, Swahili. Previous works introduced subword tokenization. Our work extends such efforts. Notably, we propose a syllable tokenizer and adopt an experiment-centric approach to validate the proposed tokenizer based on the Swahili language. We conducted text-generation experiments with GPT2 to evaluate the effectiveness of the syllable tokenizer. Our results show that the proposed syllable tokenizer generates syllable embeddings that effectively represent the Swahili language.

Shintaro Ozaki, Kazuki Hayashi, Miyu Oba et al.

A large part of human communication relies on nonverbal cues such as facial expressions, eye contact, and body language. Unlike language or sign language, such nonverbal communication lacks formal rules, requiring complex reasoning based on commonsense understanding. Enabling current Video Large Language Models (VideoLLMs) to accurately interpret body language is a crucial challenge, as human unconscious actions can easily cause the model to misinterpret their intent. To address this, we propose a dataset, BQA, a body language question answering dataset, to validate whether the model can correctly interpret emotions from short clips of body language comprising 26 emotion labels of videos of body language. We evaluated various VideoLLMs on BQA and revealed that understanding body language is challenging, and our analyses of the wrong answers by VideoLLMs show that certain VideoLLMs made significantly biased answers depending on the age group and ethnicity of the individuals in the video. The dataset is available.

Kazuki Hayashi, Hidetaka Kamigaito, Shinya Kouda et al.

Retrieval-Augmented Generation (RAG) has emerged as a way to complement the in-context knowledge of Large Language Models (LLMs) by integrating external documents. However, real-world applications demand not only accuracy but also interpretability. While dense retrieval methods provide high accuracy, they lack interpretability; conversely, sparse retrieval methods offer transparency but often fail to capture the full intent of queries due to their reliance on keyword matching. To address these issues, we introduce IterKey, an LLM-driven iterative keyword generation framework that enhances RAG via sparse retrieval. IterKey consists of three LLM-driven stages: generating keywords for retrieval, generating answers based on retrieved documents, and validating the answers. If validation fails, the process iteratively repeats with refined keywords. Across four QA tasks, experimental results show that IterKey achieves 5% to 20% accuracy improvements over BM25-based RAG and simple baselines. Its performance is comparable to dense retrieval-based RAG and prior iterative query refinement methods using dense models. In summary, IterKey is a novel BM25-based approach leveraging LLMs to iteratively refine RAG, effectively balancing accuracy with interpretability.