基于Markov逻辑网的问答答案排序方法研究

发布时间：2018-03-02 06:29

本文关键词： 问答系统排序学习 Markov逻辑网受限域多特征　出处：《昆明理工大学》2012年硕士论文　论文类型：学位论文

【摘要】：与传统搜索引擎相比,问答系统更能符合用户对所需信息的要求。答案排序是问答系统的一个重要部分,排序结果的好坏直接决定了系统的性能。影响答案排序的主要因素是答案候选文本中的特征及排序学习方法,因此,本文针对特定领域问答特点,就答案排序中的特征选取及融合多特征的学习排序方法开展研究,主要完成的工作如下： (1)定义了问句和答案间相关性的三方面特征：上下文相关特征,基于密度方法的特征和在线知识库特征。其中上下文相关特征包含问句和候选答案间在词法上的相似性以及问句和候选答案在深层语义上的相似程度等。 (2)构建了基于Markov逻辑网的实体类答案排序模型。针对领域问答系统中的事实类和列表类答案的特点,构建基于Markov逻辑网的领域类答案排序模型。以谓词公式来描述问句和候选答案间及答案和知识库间的相关特征,并将其融入到Markov逻辑网中,采用判别式训练的学习算法学习特征参数的权值,并用MC-SAT算法进行推理得到问句和答案的相关度,实现答案排序。基于Markov逻辑网的答案排序方法能够根据特征的相关性不同赋予其不同的权重,实验结果表明,答案准确率和召回率相对其它方法有较大幅度的提升。 (3)针对云南旅游领域,设计实现了基于Markov逻辑网的排序学习领域问答原型系统。
[Abstract]:Compared with the traditional search engine, question answering system can conform to the user of the information needed. Answer ranking is an important part of question answering system, results directly determine the performance of the system. The main factors affecting the answer ranking is characteristic and ranking learning method, candidate answer in the text so, according to the specific question characteristics, research learning feature ranking method in answer ranking selection and multi feature fusion, the main work is as follows:
(1) defines three characteristics of questions and answers: the correlation between contextual features, feature density method and online knowledge base. Based on the features of the context sensitive features include questions and candidate answers between the lexical similarity between questions and candidate answers and the deep degree of semantic similarity.
(2) to construct the Markov logic network entity answer ranking model based on domain question answering system. According to the characteristics of factoid and list answers, constructing Markov logic network domain class answer ranking model based on. To describe the relevant characteristics of questions and candidate answers and answer and knowledge base to predicate formulas, and it will be integrated into Markov logic network, using discriminative training learning algorithm learning feature weights, and using MC-SAT algorithm of relevance reasoning questions and answers, the answer ranking. Markov logic network answer ranking method can give different weights according to the different characteristics of the correlation based on the experimental results show that the answer, accuracy rate and recall rate relative to other methods have greatly improved.
(3) aiming at the field of tourism in Yunnan, a QA prototype system based on Markov logic network is designed and realized.

【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.3

【参考文献】