Papers by Daisuke Kawahara

A Benchmark Suite of Japanese Natural Questions (2024.starsem-1)

Challenge:	Large-scale pre-training corpora are essential for large language models, but if such content remains unfiltered, there is a risk that LLMs may memorize it and leak it through their outputs.
Approach:	They construct a Japanese text corpora dataset and train machine learning models to detect SCPI in text.
Outcome:	The proposed classifier can detect information related to SCPI in Japanese text.

Copied to clipboard

Takuya Uematsu, Hao Wang, Daisuke Kawahara, Tomohide Shibata

Question Answering

Challenge:	Existing studies to solve QA tasks in an integrated manner are not available in other languages because of the lack of QA datasets.
Approach:	They build a Japanese version of Natural Questions using natural questions from query logs of a search engine and crowdsource it using crowdsourcing.
Outcome:	The proposed datasets are based on natural questions from Japanese search engines and crowdsourced.

Demystifying Mixed Outcomes of Self-Training: Pre-training Analyses on Non-Toy LLMs (2026.findings-eacl)

Copied to clipboard

Yusuke Nakamura, Hirokazu Kiyomaru, Chaoran Liu, Shuhei Kurita, Daisuke Kawahara

Semantic Text Processing

Challenge:	Recent studies on self-training report seemingly contradictory outcomes.
Approach:	They use OLMo-2 models as non-toy LLMs and perform multiple rounds of continual pre-training using self-generated text with different prompting strategies and data filtering.
Outcome:	The proposed model collapse is inherent to the training procedure itself, while self-improvement is likely owes its success to human-designed, strategic synthetic pipelines that inject external intelligence.

Grounding in social media: An approach to building a chit-chat dialogue model (2022.naacl-srw)

Copied to clipboard

Ritvik Choudhary, Daisuke Kawahara

Shrinking Japanese Morphological Analyzers With Neural Networks and Semi-supervised Learning (N19-1)

Challenge:	Existing open-domain dialogue models fail to capture and utilize external knowledge, leading to repetitive or generic responses to unseen utterances.
Approach:	They propose to use social media comments to improve the raw conversation ability of open-domain dialogue systems.
Outcome:	The proposed model improves the raw conversation ability of open-domain dialogue systems by mimicking human responses through casual interactions found on social media.

Copied to clipboard

Arseny Tolmachev, Daisuke Kawahara, Sadao Kurohashi

Syntactic Text Processing

Challenge:	Modern neural morphological analyzers consume gigabytes of memory.
Approach:	They propose a method which uses unigram character embeddings to train a model on labels produced by a state-of-the-art analyzer.
Outcome:	The proposed model outperforms dictionary-based methods in Japanese and Chinese . it uses less than 15 megabytes of space and is much smaller than the dictionary- based one .

Multi-Task Learning of Generation and Classification for Emotion-Aware Dialogue Response Generation (2021.naacl-srw)

Copied to clipboard

Tatsuya Ide, Daisuke Kawahara

Scaling Data-Constrained Language Models with Synthetic Data (2026.findings-eacl)

Challenge:	Existing models for human-like interaction with humans are not expected to improve the accuracy of emotion recognition, but instead focus on generating emotion-aware responses.
Approach:	They propose a neural response generation model with multi-task learning of generation and classification, focusing on emotion.
Outcome:	The proposed model makes generated responses more emotionally aware.

Copied to clipboard

Hirokazu Kiyomaru, Yusuke Oda, Takashi Kodama, Chaoran Liu, Daisuke Kawahara

Diversity-aware Event Prediction based on a Conditional Variational Autoencoder with Reconstruction (D19-60)

Challenge:	Large language models (LLMs) improve with more training data, but practical limitations on data collection constrain further scaling.
Approach:	They compare three strategies to generate Japanese text, repeat the limited Japanese Web text, and use English Web text to fill the data shortfall.
Outcome:	The proposed model outperforms baselines and achieves the performance achieved when the entire token budget is filled with additional organic Japanese Web text.

Copied to clipboard

Hirokazu Kiyomaru, Kazumasa Omura, Yugo Murawaki, Daisuke Kawahara, Sadao Kurohashi

Theoretical Linguistics Rivals Embeddings in Language Clustering for Multilingual Named Entity Recognition (2023.acl-srw)

Challenge:	Typical event sequences are important class of commonsense knowledge . previous work in event prediction uses sequence-to-sequence models . however, what can happen after a given event is usually diverse .
Approach:	They propose to incorporate a conditional variational autoencoder into seq2seq for its ability to represent diverse next events as a probabilistic distribution.
Outcome:	The proposed model outperforms deterministic models in terms of precision and recall . the proposed model is based on a conditional variational autoencoder .

Copied to clipboard

Sakura Imai, Daisuke Kawahara, Naho Orita, Hiromune Oda

Information Extraction & Text Mining

Challenge:	Existing studies have used descriptive typological features and a coarse language family classification as baselines for language clustering.
Approach:	They propose two types of language groupings based on morpho-syntactic features in a nominal domain and one based upon a head parameter.
Outcome:	The proposed methods outperform state-of-the-art embedding-based models in multilingual named entity recognition (NER) . their results suggest that theoretical linguistics plays a significant role in multi-lingual learning tasks.

Time-aware COMET: A Commonsense Knowledge Model with Temporal Knowledge (2024.lrec-main)

Copied to clipboard

Eiki Murata, Daisuke Kawahara

Commonsense Reasoning

Challenge:	Existing commonsense knowledge models do not consider granularity or time axes, and can't handle commonsensical knowledge, which is tacit.
Approach:	They propose to use ChatGPT to create a time-aware commonsense knowledge model, TaCOMET, and use it to continually fine tune existing models.
Outcome:	The proposed model outperforms existing models on a robotic decision-making task when proper times are input.

Knowledge-Enriched Two-Layered Attention Network for Sentiment Analysis (N18-2)

Copied to clipboard

Abhishek Kumar, Daisuke Kawahara, Sadao Kurohashi

Sentiment Analysis

Challenge:	Existing sentiment analysis systems are prone to word shortening, exaggeration, lack of grammar and appropriate punctuation.
Approach:	They propose a two-layered attention network based on Bidirectional Long Short-Term Memory for sentiment analysis using the Knowledge Graph Embedding generated using the WordNet.
Outcome:	The proposed model outperforms the state-of-the-art system on the benchmark dataset of SemEval 2017 Task 5 by 1.7 and 3.7 points respectively.

Reinforcement Learning for Edit-Based Non-Autoregressive Neural Machine Translation (2024.naacl-srw)

Copied to clipboard

Hao Wang, Tetsuro Morimura, Ukyo Honda, Daisuke Kawahara

Machine Translation

Challenge:	Non-autoregressive (NAR) language models have a performance gap due to the large decoding space and difficulty in capturing dependency between target words accurately.
Approach:	They propose to use reinforcement learning to enhance the performance of edit-based NAR models by using stepwise reward maximization and episodic reward maximisation.
Outcome:	The proposed model outperforms autoregressive models in the evaluation of an edit-based model.

Developing Japanese CLIP Models Leveraging an Open-weight LLM for Large-scale Dataset Translation (2025.naacl-srw)

Copied to clipboard

Issa Sugiura, Shuhei Kurita, Yusuke Oda, Daisuke Kawahara, Naoaki Okazaki

JGLUE: Japanese General Language Understanding Evaluation (2022.lrec-1)

Challenge:	lack of large-scale open Japanese image-text pairs poses a significant barrier to the development of vision-language models.
Approach:	They construct large-scale Japanese image-text pairs using machine translation and pre-trained CLIP models on a Japanese dataset.
Outcome:	The results show that pre-trained models achieve competitive average scores on Japanese culture tasks compared to models of similar size.

Copied to clipboard

Kentaro Kurihara, Daisuke Kawahara, Tomohide Shibata

Challenge:	There is no benchmark for Japanese to evaluate and analyze NLU ability from different perspectives.
Approach:	They build a Japanese NLU benchmark from scratch without translation to measure general NLU ability in Japanese.
Outcome:	a Japanese NLU benchmark is built from scratch without translation to measure general NLU ability in Japanese.

Improving Crowdsourcing-Based Annotation of Japanese Discourse Relations (L18-1)

Copied to clipboard

Yudai Kishimoto, Shinnosuke Sawada, Yugo Murawaki, Daisuke Kawahara, Sadao Kurohashi

Discourse & Pragmatics

Challenge:	Discourse parsing is an important task in natural language processing, but few languages have corpora annotated with discourse relations . crowdsourcing-based annotations are of poor quality and require expensive and time-consuming . et al. (2009) evaluated the quality of annotations using expert annotations.
Approach:	They construct a Japanese corpus with discourse annotations through crowdsourcing . they propose improvement techniques based on language tests .
Outcome:	The proposed methods improve the quality of the annotations, and will make them publicly available.

Constructing a Japanese Rap Lyric Generation Model with GRPO (2026.acl-srw)

Copied to clipboard

Hayato Ogawa, Daisuke Kawahara

Challenge:	Rap is a vocal style rooted in Hip-Hop culture, characterized by producing rhymes in synchrony with a rhythmic beat.
Approach:	They propose a method for generating Japanese rap lyrics with a large language model . the model's rhyming behavior is improved by using existing Japanese rhapsodysts as training data.
Outcome:	The proposed method improves outputs that receive moderate or high human ratings on rhyme-related criteria.

Building a Japanese Typo Dataset from Wikipedia’s Revision History (2020.acl-srw)

Copied to clipboard

Yu Tanaka, Yugo Murawaki, Daisuke Kawahara, Sadao Kurohashi

Challenge:	Typographical errors (typos) also occur in user generated content (UGC).
Approach:	They extract over half a million Japanese typo–correction pairs from Wikipedia’s revision history and combine character-based extraction rules, morphological analyzers to guess readings, and various filtering methods to address these challenges.
Outcome:	The proposed dataset extracts over half a million typo–correction pairs from Wikipedia’s revision history.

Tree-structured Decoding for Solving Math Word Problems (D19-1)

Copied to clipboard

Qianying Liu, Wenyv Guan, Sujian Li, Daisuke Kawahara

Numerical Reasoning

Challenge:	Existing approaches to solve math word problems do not consider an abstract syntax tree.
Approach:	They propose a tree-structured decoding method that generates an abstract syntax tree of an equation in a top-down manner and can stop during decoding without a redundant stop token.
Outcome:	The proposed method achieves state-of-the-art performance on the largest dataset on this task.

Kanbun-LM: Reading and Translating Classical Chinese in Japanese Methods by Language Models (2023.findings-acl)

Copied to clipboard

Hao Wang, Hirofumi Shimizu, Daisuke Kawahara

Cross-lingual Knowledge Projection Using Machine Translation and Target-side Knowledge Base Completion (C18-1)

Challenge:	Classical Chinese was introduced to Japan approximately 2,000 years ago . it was gradually adapted to a Japanese form called Kanbun-Kundoku (Kanbun) in Japanese reading and translating methods .
Approach:	They construct a dataset that compares Classical Chinese and Kanbun in Japan using character reordering and machine translation tasks.
Outcome:	The proposed dataset compares the current language models with human scores and compared them with human-level models.

Copied to clipboard

Naoki Otani, Hirokazu Kiyomaru, Daisuke Kawahara, Sadao Kurohashi

JDCFC: A Japanese Dialogue Corpus with Feature Changes (L18-1)

Challenge:	Existing efforts to build commonsense knowledge bases are expensive and lack quantity and quality between languages.
Approach:	They propose to project English commonsense knowledge into Japanese and Chinese with high precision.
Outcome:	The proposed method achieves top-10 accuracy on the crowdsourced English–Japanese benchmark and 18,747 facts of accurate Japanese commonsense within a very short period.

Copied to clipboard

Tetsuaki Nakamura, Daisuke Kawahara

Neural Adversarial Training for Semi-supervised Japanese Predicate-argument Structure Analysis (P18-1)

Challenge:	Existing corpora focus on emotional expressions in conversations, but there are no large-scale corpors focusing on the relationships between emotions and utterances.
Approach:	They propose a Japanese Feature Change Knowledge Base (JFCKB) that focuses on emotional expressions in conversations.
Outcome:	The proposed corpus can recognize reasonableness of a given conversation.

Copied to clipboard

Shuhei Kurita, Daisuke Kawahara, Sadao Kurohashi

Responsible & Trustworthy NLP

Challenge:	Japanese predicate-argument structure analysis involves zero anaphora resolution . state-of-the-art models for PAS analysis achieve an accuracy of around 50% for zero pronouns .
Approach:	They propose a Japanese PAS analysis model based on semi-supervised adversarial training with a raw corpus.
Outcome:	The proposed model outperforms existing models for Japanese PAS analysis . the model is based on semi-supervised adversarial training with a raw corpus .

Juman++: A Morphological Analysis Toolkit for Scriptio Continua (D18-2)

Copied to clipboard

Arseny Tolmachev, Daisuke Kawahara, Sadao Kurohashi

Morphology

Challenge:	a morphological analyzer is useful for languages without natural word boundaries, but it is difficult to improve it without creating costly annotations.
Approach:	They propose a toolkit for developing morphological analyzers for languages without natural word boundaries using lattices and neural nets.
Outcome:	The proposed morphological analyzer of Japanese achieves new SOTA on Jumandic-based corpora while being 250 times faster than the previous one.

Evaluating the Impact of SAE-based Language Steering on LLM Performance (2026.eacl-srw)

Copied to clipboard

Sebastian Zwirner, Wentao Hu, Koshiro Aoki, Daisuke Kawahara

Minimize Exposure Bias of Seq2Seq Models in Joint Entity and Relation Extraction (2020.findings-emnlp)

Challenge:	Recent advances in Sparse Autoencoders (SAEs) have revealed interpretable features within large language models (LLMs) however, the impact of SAE-based language steering on output quality and task performance remains unclear.
Approach:	They apply language-specific SAE feature steering to three LLMs from two model families and evaluate it on a translation task and a multilingual question-answering task.
Outcome:	The proposed approach outperforms prompting and language neuron-based steering on translation and multilingual question-answering tasks.

Copied to clipboard

Ranran Haoran Zhang, Qianying Liu, Aysa Xuemo Fan, Heng Ji, Daojian Zeng, Fei Cheng, Daisuke Kawahara, Sadao Kurohashi

KWJA: A Unified Japanese Analyzer Based on Foundation Models (2023.acl-demo)

Challenge:	Existing methods to extract relation triplets from plain text introduce exposure bias . prior work has focused on pipeline methods that ignore intrinsic interactions between subtasks and propagate classification errors through the tasks.
Approach:	They propose a model that reduces the decoding length to three within a triplet and removes the order among triplets.
Outcome:	The proposed model overfits to both datasets while showing better generalization.

Copied to clipboard

Nobuhiro Ueda, Kazumasa Omura, Takashi Kodama, Hirokazu Kiyomaru, Yugo Murawaki, Daisuke Kawahara, Sadao Kurohashi

Challenge:	KWJA supports a wide range of tasks including typo correction, word segmentation, word normalization, named entity recognition, dependency parsing, PAS analysis, bridging reference resolution, coreference resolution, and discourse relation analysis.
Approach:	They propose to build a Japanese text analyzer based on foundation models that performs a wide range of tasks.
Outcome:	The proposed model performs better in a multi-task manner than other analyzers with specialized models.

Machine Comprehension Improves Domain-Specific Japanese Predicate-Argument Structure Analysis (D19-58)

Copied to clipboard

Norio Takahashi, Tomohide Shibata, Daisuke Kawahara, Sadao Kurohashi

Natural Language Interfaces

Challenge:	a lack of gold datasets and knowledge about PAS analysis makes it difficult to create accurate PAS analyses.
Approach:	They construct a Japanese blog-QA dataset and a reading comprehension QA dataset using crowdsourcing.
Outcome:	The proposed method is most effective, pre-training model to acquire domain knowledge and fine-tuning model based on PAS-QA dataset.

Generate, Evaluate, and Select: A Dialogue System with a Response Evaluator for Diversity-Aware Response Generation (2022.naacl-srw)

Copied to clipboard

Ryoma Sakaeda, Daisuke Kawahara

Acquiring Social Knowledge about Personality and Driving-related Behavior (2020.lrec-1)

Challenge:	Existing dialogue systems that generate fluent responses are difficult to evaluate due to the one-to-many nature of dialogue, which means the existence of multiple appropriate responses is not appropriate.
Approach:	They propose a generator-evaluator model that evaluates multiple responses generated by a response generator and selects the best response by an evaluators.
Outcome:	The proposed model is compared with a baseline system and its outputs were judged to be better than the baseline system.

Copied to clipboard

Ritsuko Iwai, Daisuke Kawahara, Takatsune Kumada, Sadao Kurohashi

Sentiment Analysis

Challenge:	Using crowdsourcing, we acquire human-specific knowledge about personality and driving.
Approach:	They propose a psychological approach to collect human-specific social knowledge from a text corpus using NLP techniques.
Outcome:	The proposed approach collects human-specific social knowledge from a text corpus, and then implements it into a system.

Building a Personalized Dialogue System with Prompt-Tuning (2022.naacl-srw)

Copied to clipboard

Tomohito Kasahara, Daisuke Kawahara, Nguyen Tung, Shengzhe Li, Kenta Shinzato, Toshinori Sato

Building Japanese Creativity Benchmarks and Applying them to Enhance LLM Creativity (2025.acl-srw)

Challenge:	We build a dialogue system that can respond based on a given character setting (persona) this method is not suitable because the more persona information is added, the longer the input text becomes.
Approach:	They propose to use prompt-tuning to build a dialogue system that responds based on a persona . they conduct automatic and manual evaluations on English and Japanese .
Outcome:	The proposed method can build a dialogue system with more natural responses with less computational resources than fine-tuning.

Copied to clipboard

So Fukuda, Hayato Ogawa, Kaito Horio, Daisuke Kawahara, Tomohide Shibata

Semantic Text Processing

Challenge:	a recent study evaluated the creativity of large language models (LLMs) in Japanese based on a Torrance test of creative thinking . previous research on LLM creativity focused on English, but differences exist in how it manifests and is evaluated across languages and cultures.
Approach:	They construct three benchmarks to evaluate LLM creativity in Japanese . they use Japanese Creativity Questions (JCQ), Divergent Association Task (DAT) and Story Alteration Task (SAT)
Outcome:	The benchmarks evaluate the creativity of large language models (LLMs) in Japanese . the benchmarks are Japanese Creativity Questions (JCQ), Divergent Association Task (DAT), and Story Alteration Task (SAT).

JFCKB: Japanese Feature Change Knowledge Base (L18-1)

Copied to clipboard

Tetsuaki Nakamura, Daisuke Kawahara

Information Extraction & Text Mining

Challenge:	constructing commonsense knowledge including connotational meanings is challenging . a recent study focused on denotation and connotations, but few studies focused on connotating meanings .
Approach:	They propose to construct a Japanese knowledge base where arguments in event sentences are associated with feature changes caused by events.
Outcome:	The proposed knowledge base is able to generate anaphora resolution tasks in Japanese . it is useful for computers to understand texts, but it is difficult to acquire it .

BERT-based Cohesion Analysis of Japanese Texts (2020.coling-main)

Copied to clipboard

Nobuhiro Ueda, Daisuke Kawahara, Sadao Kurohashi

Semantic Text Processing

Challenge:	Recent advances in neural networks have significantly improved natural language processing tasks . they include self training-based language models such as BERT .
Approach:	They tackle a systematic analysis of cohesion in Japanese texts using BERT models . they find that coreference resolution is different in nature from other tasks .
Outcome:	The proposed analysis outperforms existing studies on cohesion in Japanese texts.

A Method for Building a Commonsense Inference Dataset based on Basic Events (2020.emnlp-main)

Copied to clipboard

Kazumasa Omura, Daisuke Kawahara, Sadao Kurohashi

Commonsense Reasoning

Challenge:	Existing approaches to acquire commonsense are limited by the general-purpose language models.
Approach:	They propose a method for building a commonsense inference dataset using crowdsourcing and automatic extraction from a corpus.
Outcome:	The proposed method can solve 104k commonsense inference problems in a Japanese corpus with high accuracy, but low bias.

Leveraging High-Resource English Corpora for Cross-lingual Domain Adaptation in Low-Resource Japanese Medicine via Continued Pre-training (2025.findings-emnlp)

Copied to clipboard

Kazuma Kobayashi, Zhen Wan, Fei Cheng, Yuma Tsuta, Xin Zhao, Junfeng Jiang, Jiahao Huang, Zhiyi Huang, Yusuke Oda, Rio Yokota, Yuki Arase, Daisuke Kawahara, Akiko Aizawa, Sadao Kurohashi

Constructing a Japanese Verdict Prediction Dataset for Fact-Checking of LLM-Generated Texts (2026.acl-srw)

Challenge:	low-resource language corpora in professional domains like medicine hinder cross-lingual domain adaptation of pre-trained large language models.
Approach:	They examine how linguistic features affect performance on a Japanese–English medical knowledge benchmark.
Outcome:	The proposed model can leverage English-language resources in medical domains while ensuring sufficient coverage of language-specific expressions in a target language.

Copied to clipboard

Miwa Masano, Hirokazu Kiyomaru, Atsushi Keyaki, Kaito Horio, Rei Minamoto, Ribeka Keyaki, Kouta Nakayama, Hideyuki Tachibana, Daisuke Kawahara

Responsible & Trustworthy NLP

Challenge:	Text generated by Large Language Models (LLMs) may contain plausible but incorrect information known as hallucinations.
Approach:	They extend the label set for verdict prediction to capture claim-evidence relationships humans would commonly interpret as supported or refuted.
Outcome:	The proposed system improves F1 by 4 percentage points compared to baseline.

Building a Dialogue Corpus Annotated with Expressed and Experienced Emotions (2022.acl-srw)

Copied to clipboard

Tatsuya Ide, Daisuke Kawahara

Investigating Web Corpus Filtering Methods for Language Model Development in Japanese (2024.naacl-srw)

Challenge:	a human would recognize the emotion of an interlocutor and respond with an appropriate emotion, such as empathy and comfort.
Approach:	They propose to build a dialogue corpus annotated with two kinds of emotions . they collect tweets and annotate them with the emotion they put into the utterance .
Outcome:	The proposed method shows that it is difficult to recognize experienced emotions and multitask learning is effective.

Copied to clipboard

Rintaro Enomoto, Arseny Tolmachev, Takuro Niitsuma, Shuhei Kurita, Daisuke Kawahara