中文文本蘊含辨識

Textual Entailment Recognition in Chinese

施政瑋

指導教授 ： 許聞廉

國立清華大學/電機資訊學院/資訊系統與應用研究所/博士(2014年)

摘要

文本蘊含辨識為辨識兩文句之間蘊含關係（inference）的技術。對許多自然語言處理應用來說，文本蘊含辨識的技術可以提供許多的幫助。本論文針對中文文本蘊含辨識進行深入的研究與探討，首先回顧世界各國目前對文本蘊含辨識研究的現況，包括相關的活動及競賽、常用資源與工具、以及主流的文本蘊含辨識方法。其次、本研究使用常用的語言特徵實做一個基礎文本蘊含辨識系統以便於實際觀察文本蘊含辨識的癥結點以及中文文本蘊含辨識特有的現象。於本研究中亦提出在驗證基礎系統中所觀察到的詞彙的相似度及句子間的矛盾問題以及改進的方法並將及其整合於進一個新的中文文本蘊含辨識。相較於基礎系統的表現，本研究所實作改進的系統在兩次NTCIR RITE國際競賽的資料集中不管是辨識各種蘊含關係的正確率(Accuracy)，或是系統的強固性(Robustness)皆取得顯著的提升，根據其驗證結果本研究也同時提出中文文本蘊含辨識未來可能的發展方向。

關鍵字

Textual Entailment Recognition ； Chinese

並列摘要

Recognizing Textual Entailment is a task that recognizes pairs of natural language expressions. It is useful in a wide range of applications in NLP such as text summarization, question answering, and machine translation. In this thesis we focus on textual entailment recognition in Chinese. We first conduct an extensive survey of the vast literature on textual entailment recognition and related events, resources, tools, and common-used approaches. Besides, our experience of constructing a baseline Chinese textual entailment recognition system and the evaluation results are demonstrated. Based on the observation of the evaluation results, a sense decomposition-based word similarity measurement and a novel approach of detecting contradictions in text are proposed to tackle the issues of word similarity measurement and contradiction detection observed in baseline evaluations. The following evaluations show that our new integrated textual entailment recognition system outperforms the baseline system in almost every aspects in the dataset of NTCIR-10 RITE. Finally, the potential research directions of Chinese textual entailment recognition are listed in the end of the thesis.

參考文獻

[109] Yaoyun Zhang, Jun Xu, Chenlong Liu, Xiaolong Wang, Ruifeng Xu, Qingcai

[100] Marta Tatu and Dan Moldovan. Cogex at rte3. In Proceedings of the ACLPASCAL

[60] Cheng-Wei Lee, Chuan-Jie Lin, Hideki Shima, and Wen-Lian Hsu. Evaluating

Kanayama, Cheng-Wei Lee, Chuan-Jie Lin, Shuming Shi, Teruko Mitamura,

[39] Pamela Forner, Anselmo Pe~nas, Eneko Agirre, I~naki Alegria, Corina Forascu,