吉永 直樹 / Naoki Yoshinaga, PhD

Associate Professor (Tenured) and The University of Tokyo Excellent Young Researcher

Institute of Industrial Science, The University of Tokyo
Department of Information and Communication Engineering, The University of Tokyo

Ee-503, 4-6-1 Komaba, Meguro-Ku, Tokyo, 153-8505, Japan
ynaga (at) iis [guess from URL]

Curriculum Vitae

Full CV

Research Interests

We're studying various aspects on natural language processing (NLP) and computational linguistics (CL), especially

• Mechanistic interpretability and Evaluation: [ACL-20 SRW, BlackBoxNLP-21, AACL-23 SRW, EACL-24, EMNLP-24, ACL-25, EACL-26]
• Multimodal and Multilingual Models: [CoNLL-15, ACL-17, NAACL-19, CoNLL-19, EMNLP-20, NAACL-21, EACL-24, NAACL-24 SRW, ACL-25, AACL-25]
• Domain Adaptation and Personalization; beyond RAG: [IJCNLP-13, ACL-13, ACL-17 SRW, NAACL-19, EMNLP-20, EACL-23, ACL-26]
• Extremely efficient language modeling: [EMNLP-09, ACL-10, COLING-10, COLING-14, ACL-23, CoNLL-24, SIGIR-26]
• Knowledge acquisition from social big data: [EMNLP-12, IJCAI-16, IJCAI-19, F. EMNLP-21, ACL-23]
• Past: information visualization for NLP [PacificVis-11, IUI-16, PacificVis-18]

I like to design important NLP tasks [EMNLP-12, ACL-13, COLING-14, IJCAI-16, IJCAI-19, ACL-23, ACL-25, AACL-25, ACL-26], rather than solving classic tasks on worn-out datasets.

Research Grants

• Google Research Grant (2025-2026): 5,000,000 JPY (Principal Investigator)
• IIS Tenkai Kenkyuu (2024-2025): 8,000,000 JPY (Principal Investigator)
• IIS Sentei Kenkyuu (2023-2024): 4,000,000 JPY (Principal Investigator)
• Grant-in-Aid for Scientific Research (B) (21H03494; 2021-2024): 17,420,000 JPY (Principal Investigator)
• NII CRIS Collaborative Research operated by NII CRIS and LINE Corporation (2020-2023): 24,500,000 JPY (Principal Investigator)
• NII CRIS Contract Research 2019 (2019-2020): 2,499,000 JPY (Principal Investigator)
• Grant-in-Aid for Scientific Research (B) (16H02905; 2016-2019): 17,810,000 JPY (as a collaborative researcher; Principal Investigator: Masashi Toyoda)
• U-Tokyo Excellent Young Researchers Start-up (2016-2018): 6,000,000 JPY (Principal Investigator)
• Grant-in-Aid for Young Scientists (B) (16K16109; 2016-2018): 3,900,000 JPY (Principal Investigator)
• IIS Sentei Kenkyuu (2016-2017): 2,000,000 JPY (Principal Investigator)
• Grant-in-Aid for JSPS Fellows (2005-2008): 3,400,000 JPY (Principal Investigator)
• Grant-in-Aid for JSPS Fellows (2002-2005): 3,000,000 JPY (Principal Investigator)

Softwares for NLP/CL

• Data Structures:
  • cedar: An efficient, updatable trie implementation based on double array.
• Algorithms:
  • opal: A scalable kernel-based online learner (polynomial kernel is supported).
  • pecco: An efficient classifier for a model pre-trained with conjunctive features (or polynomial kernel).
  • yakmo: A robust, efficient alternative k-means clustering.
• Systems:
  • Jagger: Super fast morphological analyzer for Japanese.
  • J.DepP: Very fast dependency parsers with the state-of-the-art accuracy for Japanese.
  • RenTAL (no longer maintained): A grammar compiler from Lexicalized TAG to HPSG-style grammar.

Note that some of the above softwares are no longer experimental codes; they are substantially elaborated from the original codes to have a better performance (2-5x speed-up, or +1% in accuracy etc.; track History section of each software). Those who want to reproduce the experimental results may want to use the oldest release of the softwares.

Selected Publications (with links to presentations, codes, and datasets)

• Recasting Web-Scale Query Suggestion as dense retrieval: Efficient, Up-to-Date, and Context-Aware Suggestions
  SIGIR-26 (full, acceptance rate: 18.4%). Joint work with S. Nishikawa, N. Kaji
  TL;DR: We reformulate query suggestion (QS) as retrieval (CADE-QS) instead of generation to meet realistic QS requirements.
• Query-Focused Individual Simulation with Progressive Persona Completion
  ACL-26 (Findings (long), acceptance rate 37%). Joint work with W. Su, M. Toyoda
  TL;DR: We address a cold-start problem in user simulation by progressively predicting and retrieving missing but relevant persona.
• Tracing Multilingual Knowledge Acquisition Dynamics in Domain Adaptation: A Case Study of Biomedical Adaptation --- slide / code
  EACL-26 (long, acceptance rate: 20.1%). Joint work with X. Zhao, Y. Tsuta, and A. Aizawa
  TL;DR: We propose AdaXEval, which generates evaluation data from the training data, to study knowledge acquisition dynamics.
• CLICKER: Cross-Lingual Knowledge Editing via In-Context Learning with Adaptive Stepwise Reasoning --- poster / code / data
  EACL-26 (Findings (long), acceptance rate: 36.7%). Joint work with Z. Jiang, X. Zhao, and Y. Kumadaki
  TL;DR: We propose CLICKER, an adaptive reasoning for cross-lingual knowledge editing.
• Is He Extroverted? Identifying Missing Relevant Personas for Faithful User Simulation --- poster / data
  EACL-26 SRW. Joint work with W. Su and Y. Zhou, Z. Wang, and M. Toyoda
  TL;DR: We set up a task of identifying missing persona for individual simulation and build a benchmark, PICQ.
• Commentary Generation from Multimodal Game Data for Esports Moments in Multiplayer Strategy Games --- data
  AACL-25 (Findings (long), acceptance rate: 31%). Joint work with Z. Wang
  TL;DR: We set up a task (datasets, LoL19-trimodal) of generating moment commentaries from eSports data record and screenshot.
• Neuron Empirical Gradient: Discovering and Quantifying Neurons' Global Linear Controllability --- poster / code / data
  ACL-25 (long, acceptance rate: 20.3%). Joint work with X. Zhao and Z. Jiang
  TL;DR: We reveal a linear link between neuron activations and probabilities, introducing Neuron Empirical Gradient.
• A-TASC: Asian TED-Based Automatic Subtitling Corpus --- poster / data
  ACL-25 (long, acceptance rate: 20.3%). Joint work with Y. Zhou
  TL;DR: We develop a large-scale corpus, A-TASC, and metrics for automatic subtitling in Asian languages.
• Further Compressing Distilled Language Models via Frequency-aware Partial Sparse Coding of Embeddings --- poster
  CoNLL-24. Joint work with K. Tamura and M. Neishi
  TL;DR: We compress rare token embeddings using their nearest neighbor embeddings for common tokens as basis embeddings.
• What Matters in Memorizing and Recalling Facts? Multifaceted Benchmarks for Knowledge Probing in Language Models --- slide / poster / data
  EMNLP-24 (Findings (long), acceptance rate: 37.7%). Joint work with X. Zhao and D. Oba
  TL;DR: We present BELIEF, a comprehensive knowledge probing benchmark for LLMs, using a newly-built diverse multi-prompt dataset, MyriadLAMA.
• Commentary Generation from Data Records of Multiplayer Strategy Esports Game (journal ver.) --- poster / data
  NAACL-24 SRW. Joint work with Z. Wang
  TL;DR: We set up a task (datasets, LoL19) of generating commentaries from eSports data records.
• Tracing the Roots of Facts in Multilingual Language Models: Independent, Shared, and Transferred Knowledge (arXiv ver.) --- slide / code
  EACL-24 (long, acceptance rate 17.8%). Joint work with X. Zhao and Daisuke Oba
  TL;DR: We reveal knowledge representations in multilingual language models.
• Rethinking Response Evaluation from Interlocutor's Eye for Open-Domain Dialogue Systems --- data
  IJCNLP-AACL-23 SRW. Joint work with Y. Tsuta, S. Sato, and M. Toyoda
  TL;DR: We propose automatic personalized evaluation for dialogue systems in terms of engagement.
• Back to Patterns: Efficient Japanese Morphological Analysis with Feature-Sequence Trie --- slide / poster / code
  ACL-23 (short, acceptance rate 16.53%). single-authored work!
  TL;DR: You know, patterns are just great -> Jagger.
• Early Discovery of Disappearing Entities in Microblogs --- slide / poster / data
  ACL-23 (long, acceptance rate 23.50%). Joint work with S. Akasaki and M. Toyoda
  TL;DR: We set up and tackle a task of detecting disappearing entities from Twitter timelines.
• A Unified Generative Approach to Product Attribute-Value Identification
  ACL-23 (Findings (long), acceptance rate:). Joint work with Keiji Shinzato, Yandi Xia and Wei-Te Chen
  TL;DR: You should use generation to solve product attribute-value identification, rather than extraction and classification.
• Self-Adaptive Named Entity Recognition by Retrieving Unstructured Knowledge (journal ver.) --- poster
  EACL-23 (long, acceptance rate 28.3%). Joint work with Kousuke Nishida and Kyousuke Nishida
  TL;DR: We propose an NER model that autonomically search for knowledge required to recognize and type inconfident entities.
• Entity Embedding Completion for Wide-Coverage Entity Disambiguation
  EMNLP-22 (Findings (long), acceptance rate:). Joint work with D. Oba, I. Yamada and M. Toyoda
  TL;DR: We develop a lightweight method for adapting entity disambigation models to a new domain with out-of-vocabulary entities.
• Simple and Effective Knowledge-Driven Query Expansion for QA-Based Product Attribute Extraction --- slide / data
  ACL-22 (short, acceptance rate 20.75%). Joint work with K. Shinzato, Y. Xia and W.-T. Chen
  TL;DR: We show that internal knowledge in the training data improves the performance of QA-based attribute value extraction.
• Exploratory Model Analysis Using Data-Driven Neuron Representations --- poster
  BlackBoxNLP Workshop. Joint work with D. Oba and M. Toyoda
  TL;DR: We establish a methodology for exploratory, hypothesis-free analysis of neural NLP models.
• Fine-grained Typing of Emerging Entities in Microblogs --- code & data
  EMNLP-21 (Findings (long), acceptance rate 34.9%). Joint work with S. Akasaki and M. Toyoda
  TL;DR: We set up and tackle a task of typing emerging entities in microblogs.
• Speculative Sampling in Variational Autoencoders for Dialogue Response Generation --- slide / code & data
  EMNLP-21 (Findings (short), acceptance rate 34.9%). Joint work with S. Sato, M. Toyoda, and M. Kitsuregawa
  TL;DR: We redundantly pick latent variables in the training of variational models to obtain better latent space.
• Context-aware Decoder for Neural Machine Translation using a Target-side Document-Level Language Model --- slide / poster / code & data
  NAACL-21 (long, acceptance rate 28%). Joint work with A. Sugiyama
  TL;DR: We utilize PMI computed by a document-level LM to perform context-aware decoding with a sentence-level NMT model.
• Robust Backed-off Estimation of Out-of-Vocabulary Embeddings --- code
  EMNLP-20 (Findings (long), acceptance rate 37%). Joint work with N. Fukuda and M. Kitsuregawa
  TL;DR: Inspired by two processes of creating words, we propose a simple word-based method to estimate OOV embeddings.
• Vocabulary Adaptation for Domain Adaptation in Neural Machine Translation --- code
  EMNLP-20 (Findings (long), acceptance rate 37%). Joint work with S. Sato, J. Sakuma, M. Toyoda, M. Kitsuregawa
  TL;DR: We transplant target-domain vocabularies to source-domain NMT model for effective fine tuning.
• uBLEU: Uncertainty-Aware Automatic Evaluation Method for Open-Domain Dialogue Systems (journal ver.) --- slide / code
  ACL-20 SRW (acceptance rate 36%). Joint work with Y. Tsuta and M. Toyoda
  TL;DR: Our υBLEU performs retrieval-augmented evaluation for open-domain dialogues allowing diverse responses.
• Data augmentation using back-translation for context-aware neural machine translation --- code
  DiscoMT-19. Joint work with A. Sugiyama
  TL;DR: We investigated the impact of data augmentation on context-aware neural MT.
• Multilingual model using cross-task embedding projection --- slide / code
  CoNLL-19 (oral, acceptance rate: 22%). Joint work with J. Sakuma
  TL;DR: Our locally-linear mapping optimizes multilingual models based on cross-lingual word embeddings to any tasks.
• On the Relation between Position Information and Sentence Length in Neural Machine Translation --- poster / code
  CoNLL-19 (acceptance rate: 22%). Joint work with M. Neishi
  TL;DR: We revealed that Transformer is bad at handling inputs of unseen lengths, and fixed it using RNN as relative position encoder.
• Early Discovery of Emerging Entities in Microblogs --- slide / poster / data
  IJCAI-19 (acceptance rate: 13.7%). Joint work with S. Akasaki and M. Toyoda
  TL;DR: We set up and tackle a task of detecting emerging entities from Twitter timelines using timely distant supervision.
• Learning to Describe Unknown Phrases with Local and Global Contexts --- slide / code & data
  NAACL-19 (long, oral; acceptance rate 26%). Joint work with S. Ishiwatari, H. Hayashi, G. Neubig, S. Sato, M. Toyoda, and M. Kitsuregawa
  TL;DR: Our LOG-CaD can explain unknown phrases; evaluated on new Wikipedia datasets.
• Modeling Personal Biases in Language Use by Inducing Personalized Word Embeddings (journal ver.) --- slide
  NAACL-19 (short, oral; acceptance rate 21%). Joint work with D. Oba, S. Sato, S, Akasaki, and M. Toyoda
  TL;DR: We enable to model personalized usage of words, accompanied with analysis and applications.
• Information Integrated Visualization System for Heavy Rainfall Risk Analysis
  PacificViz-18 (poster). Joint work with M. Itoh, T. Sagara, U. Suzuki, K. Umemoto, M. Toyoda, K. Zettsu, and Y. Kidawara
  TL;DR: We integrate and visualize social sensor data to analyze heavy rainfall risks.
• A Bag of Useful Tricks for Practical Neural Machine Translation: Embedding Layer Initialization and Large Batch Size (journal ver.) --- poster / code
  WAT-17 (oral). Joint work with M. Neishi, J. Sakuma, S. Tohda, S. Ishiwatari and M. Toyoda
  TL;DR: We confirmed the effectiveness of CBoW-based embedding layer initialization and Large Batch Size in NMT training.
• Modeling Situations in Neural Chat Bots --- poster
  ACL-17 SRW (acceptance rate 36%). Joint work with S. Sato, M. Toyoda and M. Kitsuregawa
  TL;DR: We enable to model and incorporate user profiles and time in open-domain dialogue systems.
• Chunk-based Decoder for Neural Machine Translation --- poster
  ACL-17 (long; acceptance rate 26%). Joint work with S. Ishiwatari, J. Yao, S. Liu, M. Li, M. Zhou, M. Kitsuregawa and W. Jia
  Inspired by phrase-based SMT, we propose chunk-based decoding in NMT.
• Ordering Concepts Based on Common Attribute Intensity --- slide / poster / code & data
  IJCAI-16 (acceptance rate <25%). Joint work with T. Iwanari, N. Kaji, T. Nishina, M. Toyoda and M. Kitsuregawa; a follow-up paper at COLING-16 (Demo) --- poster / software & data
  TL;DR: We set up and tackle a task of inducing your sense of values from your writings.
• Spatio-temporal Event Visualization from a Geo-parsed Microblog Stream --- poster / demo by Prof. Itoh
  IUI-16 (poster). Joint work with M. Itoh and M. Toyoda
  TL;DR: We visualize spatio-temporal tweets on word-clouds in the sky.
• Accurate Cross-lingual Projection between Count-based Word Vectors by Exploiting Translatable Context Pairs --- poster
  CoNLL-15 (short, acceptance rate <30%). Joint work with S. Ishiwatari, N. Kaji, M. Toyoda and M. Kitsuregawa
  TL;DR: We incorporate surface-based dimensional correspondences into count-based word vectors.
• A Self-adaptive Classifier for Efficient Text-stream Processing --- poster / software
  COLING-14 (acceptance rate 32%). Joint work with M. Kitsuregawa
  TL;DR: We propose a method of accelerating NLP classifiers when the processed text becomes redundant.
• Modeling User Leniency and Product Popularity for Sentiment Classification (journal ver.) --- poster
  IJCNLP-13. Joint work with W. Gao, N. Kaji, and M. Kitsuregawa
  TL;DR: We enable to model annotation and selection biases in sentiment analysis for unseen users and targets.
• Predicting and Eliciting Addressee's Emotion in Online Dialogue --- poster
  ACL-13 (long; acceptance rate 26%). Joint work with T. Hasegawa, N. Kaji, and M. Toyoda
  TL;DR: We enable to model emotions in corpus-based open-dialogue systems.
• Identifying Constant and Unique Relations by using Time-Series Text --- slide
  EMNLP-12 (oral; acceptance rate 17%). Joint work with Y. Takaku, N. Kaji, and M. Toyoda
  Use of massive Web in knowledge acquisition introduces enormous contradictions; solution provided.
• Analysis and Visualization of Temporal Changes in Bloggers' Activities and Interests --- demo / demo by Prof. Itoh
  PacificVis-12 (acceptance rate 34%). Joint work with M. Itoh, M. Toyoda, and M. Kitsuregawa
  TL;DR: We visualize dependencies in weblogs to understand the state of the world.
• Kernel Slicing: Scalable Online Training with Conjunctive Features --- slide / software & data
  COLING-10 (oral; acceptance rate 19%). Joint work with M. Kitsuregawa
  The kernel slicing generalizes kernel splitting (ACL 2008) to pack computations in online learning with polynomial kernel.
• Efficient Staggered Decoding for Sequence Labeling --- software
  ACL-10 (long; acceptance rate 25%). Joint work with N. Kaji, Y. Fujiwara, and M. Kitsuregawa
  TL;DR: We made structured prediction scalable to the number of classes.
• Polynomial to Linear: Efficient Classification with Conjunctive Features (journal ver.) --- poster / software & data
  EMNLP-09 (acceptance rate 34%). Joint work with M. Kitsuregawa
  TL;DR: We speed up testing of non-linear classifier with polynomial kernel.
• Boosting Precision and Recall of Hyponymy Relation Acquisition from Hierarchical Layouts in Wikipedia --- software
  LREC-08. Joint work with A. Sumida and K. Torisawa
  TL;DR: We developed a method of quickly obtaining a large-scale hyponymy relations from Wikipedia.
• Open-Domain Attribute-Value Acquisition from Semi-Structured Texts --- demo
  ISWC-07 workshop, OntoLex. Joint work with K. Torisawa; a follow-up paper on writing-support environment
  TL;DR: We developed a method of extracting attributes and their values for a given object using unsupervised wrapper induction.
• Improving the Accuracy of Subcategorizations Acquired from Corpora
  ACL-04 SRW (acceptance rate 28%)
• A Debug Tool for Practical Grammar Development
  ACL-03 (poster; acceptance rate 43%). Joint work with A. Yakushiji, Y. Tateisi, Y. Miyao, and J. Tsujii
• Comparison between CFG Filtering Techniques for LTAG and HPSG (journal ver.)
  ACL-03 (poster; acceptance rate 43%). Joint work with Y. Miyao, K. Torisawa, and J. Tsujii
• Grammar Conversion from LTAG to HPSG (journal ver.) --- software
  ESSLLI 2011 Student Session (oral; acceptance rate 26%). Joint work with Y. Miyao.

More publications

Awards