基于模式学习的形式化答案抽取技术与置信度评价方法  

Formalized Answer Extraction Based on Pattern Learning and Confidence Estimation

作　　者：李鹏[1] 乔佩利[1] 王晓龙[2] 王宝勋[2] 

机构地区：[1]哈尔滨理工大学计算机科学与技术学院,黑龙江哈尔滨150001 [2]哈尔滨工业大学计算机科学与技术学院,黑龙江哈尔滨150001

出　　处：《电子学报》2008年第12期2339-2343,共5页Acta Electronica Sinica

基　　金：国家自然科学基金重点项目(No.60435020);国家863高技术研究发展计划重点项目(No.2006AA01Z197)

摘　　要：本文提出一种基于模式学习的形式化答案抽取方法,区别于传统基于特征的答案抽取方法,通过问题模式和答案模式的自动匹配,直接获取问题答案.本文通过机器学习的方法自动生成用于答案抽取的形式化模板,克服了人工方法费时、费力以及覆盖率低等问题.本文创造性地采用逻辑回归的方法对所学习到的模式进行置信度评价.对比实验表明,本文的方法取得了比较好的答案抽取效果.本文方法实际应用于国际TREC QA评测,评测结果证明本文的方法与传统基于特征的答案抽取方法具有很好的互补性.This paper presents a method of formalized answer extraction based on pattern learning. Our method extracts answers directly by matching the question patterns and the answer patterns automatically, which is different from the other traditional feature-based ones. The method gains formalized patterns automatically by using a machine learning strategy. Therefore, huge work load and low coverage of manually methods are avoided. Creatively, we apply logistic regression to estimate the confidence of the learned patterns. The contrast experiments give our method encouraging results, and it achieves fair and objective test in TREC QA track, the result of which shows that the method is a supplementary to the traditional feature based answer-extracting approaches.

关 键 词：模式学习 问答系统 答案抽取 置信度 逻辑回归 

分 类 号：TP391.2[自动化与计算机技术—计算机应用技术]