Skip navigation

DSpace JSPUI

DSpace preserves and enables easy and open access to all types of digital content including text, images, moving images, mpegs and data sets

Learn More
DSpace logo
English
中文
  • Browse
    • Communities
      & Collections
    • Publication Year
    • Author
    • Title
    • Subject
    • Advisor
  • Search TDR
  • Rights Q&A
    • My Page
    • Receive email
      updates
    • Edit Profile
  1. NTU Theses and Dissertations Repository
  2. 電機資訊學院
  3. 資訊工程學系
Please use this identifier to cite or link to this item: http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/100182
Title: 基於人與人對話進行未知使用者意圖探勘、建構及命名方法
Unseen User Intent Discovery and Induction from Human-Human Conversations
Authors: 林宗聖
Tzung-Sheng Lin
Advisor: 陳縕儂
Yun-Nung Chen
Keyword: 新意圖探勘,對比學習,大型語言模型,意圖命名,對話系統,
New Intent Discovery,Contrastive Learning,Large Language Models,Intent Naming,Dialogue Systems,
Publication Year : 2025
Degree: 碩士
Abstract: 在開放世界的對話系統中,使用者輸入的語句可能包含訓練階段未曾出現的意圖,因此「新意圖探勘」成為一項關鍵任務。
本研究提出一套創新的新意圖探勘與命名架構,結合對比學習與凍結的大型語言模型嵌入向量,在保留預訓練模型泛化能力的同時,學得具辨識性的語意表示。我們設計的模組化流程將嵌入生成、語意表示學習與聚類三個階段明確分離,具備高度彈性,可因應未來大型語言模型嵌入向量或聚類技術的演進而彈性替換或擴充。為提升實務應用中的可解釋性,我們進一步引入自動命名機制,為新探勘出的意圖指派具可讀性的語意標籤。此一面向在過去文獻中常被忽略,或受限於固定命名模板,然而對於實際部署至關重要。
此外,我們亦提出一套基於命名的評估方法,量化模型所產生之意圖名稱與人工標註名稱的一致性,直接反映意圖探勘結果的語意品質。我們在多項標準資料集及一套具挑戰性的真實多輪對話語音轉文字之資料上進行實驗,驗證本方法在效能、泛化能力與穩健性上的優勢。
Open-world dialogue systems must address utterances with evolving intents not seen during training, making new intent discovery (NID) essential.
This paper proposes a novel NID framework that combines contrastive learning with frozen large language model (LLM) embeddings to generate discriminative representations without compromising pretrained generalization. Our modular pipeline explicitly separates embedding generation, representation learning, and clustering, enabling flexible integration of future advances in LLMs and clustering techniques. To improve practical interpretability, we introduce an automated naming mechanism that assigns human-readable labels to newly discovered intents, which is often an overlooked but critical component in real-world deployment.
We further present a naming-based evaluation framework to directly assess the quality of intent discovery by measuring alignment between system-generated and human-annotated intent names. Experiments on benchmarks and a challenging in-house dataset
of multi-turn ASR transcripts demonstrate the effectiveness, generalizability, and robustness of our approach.
URI: http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/100182
DOI: 10.6342/NTU202504217
Fulltext Rights: 同意授權(限校園內公開)
metadata.dc.date.embargo-lift: 2028-12-31
Appears in Collections:資訊工程學系

Files in This Item:
File SizeFormat 
ntu-113-2.pdf
  Restricted Access
3.28 MBAdobe PDFView/Open
Show full item record


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

社群連結
聯絡資訊
10617臺北市大安區羅斯福路四段1號
No.1 Sec.4, Roosevelt Rd., Taipei, Taiwan, R.O.C. 106
Tel: (02)33662353
Email: ntuetds@ntu.edu.tw
意見箱
相關連結
館藏目錄
國內圖書館整合查詢 MetaCat
臺大學術典藏 NTU Scholars
臺大圖書館數位典藏館
本站聲明
© NTU Library All Rights Reserved