| 電子情報通信学会 研究会発表申込システム 講演論文 詳細 |
技報閲覧サービス [ログイン] 技報アーカイブ |
| トップに戻る | 前のページに戻る | [Japanese] / [English] |
| 講演抄録／キーワード | ||
| 講演名 | 2026-02-19 10:30 A Rationale-Guided and Rule-Based LLM Framework for Analyzing Psychological Counseling Dialogues ○Zhihao Shao（UTokyo）・Ryo Sekizaki（WelcomeToTalk）・Shengzhou Yi・Toshihiko Yamasaki（UTokyo）  ITS2025-42 IE2025-57 |
|
| 抄録 | （和） | This study focuses on the automatic severity assessment of student counseling dialogues, where the input consists of counseling records written by students and the output is a psychological risk label~(low risk or high risk) together with the rationale for the prediction. Existing large language model–based approaches often rely on surface-level emotional cues, exhibit unstable reasoning, and lack transparent links between evidence and decisions. To address these issues, we propose a rationale-guided and rule-based framework that integrates rationale generation with rule-driven self-improvement. Before prediction, the system generates and contrasts paired rationales for both possible labels to clarify evidence usage and decision boundaries. After prediction, misclassified samples are analyzed to automatically extract interpretable rules and representative error examples, which are then incorporated as guidance and constraints for subsequent predictions. Experiments are conducted on 6,481 real counseling records collected from the Welcome to Talk online psychological counseling platform, annotated by trained volunteers and stratified into training, validation, and test sets under a binary risk scheme. Results show that our method achieves the best performance among the evaluated methods~(0.911) and MCC~(0.541), outperforming multiple strong baselines. The performance improvement arises from the synergy between rationale guidance and rule-based reflection, enhancing evidence alignment, boundary discrimination, and error suppression. |
| （英） | This study focuses on the automatic severity assessment of student counseling dialogues, where the input consists of counseling records written by students and the output is a psychological risk label~(low risk or high risk) together with the rationale for the prediction. Existing large language model–based approaches often rely on surface-level emotional cues, exhibit unstable reasoning, and lack transparent links between evidence and decisions. To address these issues, we propose a rationale-guided and rule-based framework that integrates rationale generation with rule-driven self-improvement. Before prediction, the system generates and contrasts paired rationales for both possible labels to clarify evidence usage and decision boundaries. After prediction, misclassified samples are analyzed to automatically extract interpretable rules and representative error examples, which are then incorporated as guidance and constraints for subsequent predictions. Experiments are conducted on 6,481 real counseling records collected from the Welcome to Talk online psychological counseling platform, annotated by trained volunteers and stratified into training, validation, and test sets under a binary risk scheme. Results show that our method achieves the best performance among the evaluated methods~(0.911) and MCC~(0.541), outperforming multiple strong baselines. The performance improvement arises from the synergy between rationale guidance and rule-based reflection, enhancing evidence alignment, boundary discrimination, and error suppression. | |
| キーワード | （和） | Mental Health Classification / Large Language Models / Self-Reflection / Rationale-Guided Reasoning / / / / |
| （英） | Mental Health Classification / Large Language Models / Self-Reflection / Rationale-Guided Reasoning / / / / | |
| 文献情報 | 信学技報, vol. 125, no. 356, IE2025-57, pp. 13-18, 2026年2月. | |
| 資料番号 | IE2025-57 | |
| 発行日 | 2026-02-12 (ITS, IE) | |
| ISSN | Online edition: ISSN 2432-6380 | |
| 著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します．(許諾番号：10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) | |
| PDFダウンロード | ITS2025-42 IE2025-57 |
|
| 研究会情報 | |
| 研究会 | IE ITS ITE-MMS ITE-ME ITE-AIT ITE-SIP |
| 開催期間 | 2026-02-19 - 2026-02-20 |
| 開催地（和） | 北海道大学 |
| 開催地（英） | |
| テーマ（和） | 画像処理、一般 |
| テーマ（英） | |
| 講演論文情報の詳細 | |
| 申込み研究会 | IE |
| 会議コード | 2026-02-IE-ITS-MMS-ME-AIT-SIP |
| 本文の言語 | 英語 |
| タイトル（和） | |
| サブタイトル（和） | |
| タイトル（英） | A Rationale-Guided and Rule-Based LLM Framework for Analyzing Psychological Counseling Dialogues |
| サブタイトル（英） | |
| キーワード(1)（和/英） | Mental Health Classification / Mental Health Classification |
| キーワード(2)（和/英） | Large Language Models / Large Language Models |
| キーワード(3)（和/英） | Self-Reflection / Self-Reflection |
| キーワード(4)（和/英） | Rationale-Guided Reasoning / Rationale-Guided Reasoning |
| キーワード(5)（和/英） | / |
| キーワード(6)（和/英） | / |
| キーワード(7)（和/英） | / |
| キーワード(8)（和/英） | / |
| 第1著者 氏名（和/英/ヨミ） | 邵 之昊 / Zhihao Shao / シャオ チーハオ |
| 第1著者 所属（和/英） | 東京大学 (略称： 東大) The University of Tokyo (略称： UTokyo) |
| 第2著者 氏名（和/英/ヨミ） | 関﨑 亮 / Ryo Sekizaki / セキザキ リョウ |
| 第2著者 所属（和/英） | 株式会社Welcome to talk (略称： WelcomeToTalk) Welcome to talk Co. Ltd. (略称： WelcomeToTalk) |
| 第3著者 氏名（和/英/ヨミ） | 易 聖舟 / Shengzhou Yi / イー ショウシュウ |
| 第3著者 所属（和/英） | 東京大学 (略称： 東大) The University of Tokyo (略称： UTokyo) |
| 第4著者 氏名（和/英/ヨミ） | 山崎 俊彦 / Toshihiko Yamasaki / ヤマサキ トシヒコ |
| 第4著者 所属（和/英） | 東京大学 (略称： 東大) The University of Tokyo (略称： UTokyo) |
| 講演者 | 第1著者 |
| 発表日時 | 2026-02-19 10:30:00 |
| 発表時間 | 15分 |
| 申込先研究会 | IE |
| 資料番号 | ITS2025-42, IE2025-57 |
| 巻番号（vol） | vol.125 |
| 号番号（no） | no.355(ITS), no.356(IE) |
| ページ範囲 | pp.13-18 |
| ページ数 | 6 |
| 発行日 | 2026-02-12 (ITS, IE) |
[研究会発表申込システムのトップページに戻る] (日本語ページの行間係数: 1)