用户交互式问答系统中问题推荐机制的研究

发布时间：2018-05-22 20:46

本文选题：用户交互式问答系统 + 问题推荐机制　；参考：《中国科学技术大学》2012年博士论文

【摘要】：在计算机网络技术迅速发展的今天,互联网应用得到迅速普及。用户交互式问答系统作为Web2.0时代的典型应用已经成为现今最流行的社交网络应用之一,它为互联网用户提供了一个搜索信息和共享知识的平台。相较于从搜索引擎获取信息的方式,用户在交互式问答系统中通过简单的提问和回答方式快速准确的获取所需信息,而不是从搜索引擎返回的大量相关文档中繁琐地去查找信息。用户交互式问答系统虽然为人们提供了获取信息的便捷服务,但是依然存在着各种各样的问题,用户等待答案时间长和答案质量差是其中最显著的两个问题。在用户交互式问答系统中,提问者有时候需要等待几个小时甚至是几天的时间来等待其他用户提供答案。此外,一些用户为了获取交互式问答系统中的积分等级,提供很多不相关答案甚至是垃圾答案,这些问题都大大降低了用户获取所需信息的效率。为了提高交用户互式问答系统的性能,解决系统中存在大量零回答问题和低质量答案的问题,本文提出了在用户交互式问答系统中建立问题推荐的机制。将系统中尚未被人回答的问题,通过推荐机制将其推送给合适的专家用户去回答,以提高回答效率和答案质量。本文首先对于用户交互式问答系统中的问题推荐机制给出了定义并详细描述了问题推荐的模型。在此基础上,本文随后提出了两种不同的问题推荐策略,分别是将未解决的问题推荐给领域问答专家和将问题推荐给特定的问答专家来回答。在第一种问题推荐策略中,同一类别下的问题将会被推荐给该类别领域中的问答专家用户。本文分别提出了语义链接分析方法和语义语言模型方法来发现领域问答专家。在语义链接分析方法中,我们首先根据在用户交互式问答系统中各个用户之间的问答关系构建用户问答关系图。在这个关系图中,每一个结点代表一个用户,结点之间的每一条连接边代表用户之间的问答关系。其次,我们从用户所参与问题会话的交互行为和问答内容中抽取出不同类型的语义信息,并将这些语义信息结合到传统链接分析方法中,衍生出新的语义链接分析方法。在该新方法中,用户之间的问答链接关系融入了诸如问题难度、答案相关性、答案质量、隐性链接等语义信息,从而产生出具有不同权重的链接关系。最后,我们在带有语义信息的用户问答关系图上执行一个名为繁殖计算的链接分析方法,来计算每一个用户的专家程度值,用户获得较高计算值的将会被认为更加专家,获得最高值的前1%用户将会被认为是领域问答专家。在语义语言模型方法中,我们在传统语言模型的方法中融入抽取出的各种语义信息,将其作为权重因素考虑到传统语言模型中。通过计算用户在某一问题类别下是否为问答专家的条件概率来查找出领域问答专家。通过在用户交互式问答系统Yahoo! Answers中获取的数据上进行的实验,验证了我们提出的语义链接分析方法和语义语言模型方法在领域问题专家发现问题上较传统方法有了显著的提高。此外,实验也同样验证了抽取出的语义信息对于提高领域专家发现方法性能的有效性。相较于第一种问题推荐策略的粗粒度性,第二种问题推荐策略旨在发现能回答某一未解决问题的特定专家用户,并将该问题推荐给特定问答专家回答。在此问题推荐策略中,我们首先根据用户回答过的问题信息建立用户问答档案文件,在此基础上建立起一个基于主题的用户兴趣模型,在该模型中用户问答档案被认为是在不同主题上的一个分布,通过吉布斯抽样的方法,我们可以有效的获得用户兴趣在主题上的准确分布。然后,根据已经获取的用户兴趣主题模型,我们可以有效地计算出用户是否为某一问题的特定问答专家的概率值。概率值计算结果越高的用户将会被认为在回答该问题上专家程度越高。通过在用户交互式问答系统Yahoo! Answers中获取的数据上进行的实验,验证了我们所提出的两种不同问题推荐机制的高效性。进一步地,根据实验结果我们对比了两种不同问题推荐策略的性能。从实验结果中,我们可以发现第一种问题推荐策略较优于第二种策略。出现该现象的原因可能是,在第一种发现领域专家的问题推荐策略中覆盖了绝大部分的问答专家用户,而第二种发现特定问题问答专家方法只能查找出部分问答专家,这导致了第二种问题推荐策略在性能上略逊于第一种策略。该对比实验结果为用户交互式问答系统中问题推荐机制的策略选择提供了重要的参考。
[Abstract]:With the rapid development of computer network technology, the Internet application has been popularized rapidly. The user interactive question answering system, as a typical application of the Web2.0 era, has become one of the most popular social network applications. It provides a platform for Internet users to search information and share knowledge. In the way of information, users can get the information quickly and accurately through simple questions and answers in interactive question answering system, instead of searching for information from a large number of relevant documents returned by the search engine. The user interactive question answering system still has a convenient service for people to obtain information. In a variety of questions, the user waiting for a long answer and poor answer is the two most significant problem. In the user interactive question answering system, the questioner sometimes has to wait a few hours or even a few days to wait for the other users to provide answers. In addition, some users have to get the points in the interactive Q & a system. Level, which provides many irrelevant answers or even garbage answers, greatly reduces the efficiency of users getting the information they need.
In order to improve the performance of interuser reciprocal question answering system and solve the problem of zero answer and low quality answer in the system, this paper proposes a mechanism to establish a problem recommendation in the user interactive question answering system. The problem that has not been answered in the system is pushed to the appropriate expert user to answer by the recommendation mechanism. In order to improve the answer efficiency and answer quality, this paper first defines the problem recommendation mechanism in the user interactive question answering system and describes the recommended model in detail. On the basis of this, this paper proposes two different recommendation strategies, which are to recommend the unresolved question to the domain question and answer expert and the question. The question is recommended to a particular question and answer expert to answer.
In the first problem recommendation strategy, the problem under the same category will be recommended to the question and answer expert users in the category domain. The semantic link analysis method and the semantic language model method are proposed to find the domain question and answer expert respectively. In the semantic link analysis, we first based on the user interactive question answering system. In this relationship diagram, each node represents a user, and each connection between nodes represents a question and answer relationship between the users. Secondly, we extract different types of semantic information from the interactive behavior and question and answer content of the user's participation in the question session. And combining these semantic information into the traditional link analysis method, a new method of semantic link analysis is derived. In this new method, the question answer link relationship among users is integrated into the semantic information such as problem difficulty, answer relevance, answer quality, recessive link and so on. We execute a link analysis method named reproduction calculation on the user question answering graph with semantic information to calculate the expert degree value of each user. The top 1% users who obtain higher values will be considered to be more experts. The top 1% users will be considered as domain question and answer experts. In the traditional language model, we incorporate the various semantic information extracted from the traditional language model, and consider it as the weight factor in the traditional language model. By calculating the conditional probability of a question answering expert under a certain problem category, we find out the domain question and answer expert. Through the user interactive question answering system Yahoo! Answers The experiments carried out on the obtained data verify that the semantic link analysis and the semantic language model method proposed by us have improved significantly on the problem of expert discovery in domain problems. In addition, the experiment also validates the effectiveness of the extraction of semantic information for improving the performance of domain expert discovery methods.
Compared to the coarse granularity of the first problem recommendation strategy, the second problem recommendation strategy aims to find a specific expert user who can answer a certain unsolved problem and recommend it to a specific question answer expert. In this recommendation strategy, we first establish a user question and answer file based on the question information returned by the user. On this basis, a user interest model based on the theme is established. In this model, the user question and answer file is considered to be a distribution on different topics. Through the Gibbs sampling method, we can effectively obtain the accurate distribution of the user interest on the subject. Then, according to the subject model of the user interest that has been obtained, I We can effectively calculate the probability of whether a user is a particular question answering expert. The higher the probability of a user will be considered to be more expert in answering the question. The two kinds of experiments we have proposed through the experiment on the data obtained in the user interactive question answering system Yahoo! Answers According to the experimental results, we can find that the first problem recommendation strategy is better than the second strategy. The reason for the emergence of this phenomenon may be the problem recommendation strategy in the first field of discovery experts. It covers most of the question answer expert users, and the second kinds of question answering expert method can only find out some questions and answer experts. This leads to the second problem recommendation strategy less than the first strategy in performance. The comparison experiment results provide the strategy selection of the problem recommendation mechanism in the user interactive question answering system. An important reference.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2012
【分类号】：TP391.3

【引证文献】