信息检索模型风险及其评价方法研究

发布时间：2018-10-13 11:35

【摘要】：随着信息检索技术的不断发展,信息检索模型不同阶段出现的风险问题,如相关性估计中的风险,文档排序中的风险,查询扩展中的风险逐渐地被关注。研究这些风险问题的突破点是设计一种可以同时评价性能均值和模型风险的方法,量化这些风险的大小,进而再找出降低风险的策略。本文的研究重点有两个,其一是对信息检索风险评价指标的研究,具体工作是将基于偏差方差分解的风险评价指标从平均准确率(AP)一般化到其他评价指标,并将该指标中的目标模型设置得更加公平无偏。另一个研究重点是如何降低信息检索模型中存在的查询扩展失败的风险。针对该问题,本文提出了一种基于知识图谱的查询扩展方法来降低检索模型的风险,即增加了查询扩展模型中与查询相关的信息来降低风险,具体做法是从知识图谱中抽取与查询相关的若干实体及实体属性作为扩展词来重构查询,更好地表达用户的信息需求;且在计算扩展词权重时,参考了投资组合理论中的收益-风险分析方法,最大化扩展词和原查询的相关性收益,同时也最小化扩展词可能带来的查询漂移的风险,更进一步控制了查询扩展中的风险。为检验本文所提出的基于偏差方差分解的风险评价方法的合理性,实验部分首先利用该评价方法重新评价了TREC Ad Hoc(1993-1999)和Web Track(2010-2013)两个任务上提交的模型检索结果,说明了所提评价方法对衡量模型整体性能的合理性,并利用偏差和方差对模型的有效性和稳定性之间存在的折中现象做了量化分析。接着针对本文所提基于知识图谱来降低查询扩展风险的策略,实验部分在两个网页数据集上验证了该策略的有效性,并和基于伪相关反馈的查询扩展模型(RM3)做了对比分析,实验结果表明该本文所提扩展模型在有效性和稳定性上都优于RM3。
[Abstract]:With the development of information retrieval technology, the risk problems in different stages of information retrieval model, such as the risk in correlation estimation, the risk in document sorting and the risk in query expansion, have been paid more and more attention. The breakthrough point of studying these risk problems is to design a method that can evaluate both the performance mean and model risk, quantify the size of these risks, and then find out the risk reduction strategy. There are two emphases in this paper. One is to study the risk evaluation index of information retrieval. The specific work is to generalize the risk evaluation index based on deviation variance decomposition from average accuracy (AP) to other evaluation index. And the target model in this index is set more fairly and unbiased. Another research focus is how to reduce the risk of query expansion failure in the information retrieval model. To solve this problem, this paper proposes a query extension method based on knowledge atlas to reduce the risk of retrieval model, that is, to reduce the risk by adding information related to query expansion model. The specific method is to extract a number of entities and entity attributes related to the query from the knowledge map as extension words to reconstruct the query, to better express the information needs of users, and to calculate the weight of the extended words. Referring to the profit-risk analysis method in portfolio theory, the paper maximizes the correlation benefit between the extension word and the original query, and minimizes the risk of query drift caused by the extended word, which further controls the risk in query expansion. In order to test the rationality of the risk assessment method based on deviation variance decomposition, the model retrieval results submitted by TREC Ad Hoc (1993-1999 and Web Track (2010-2013) were reevaluated in the experimental part. The rationality of the proposed evaluation method for measuring the overall performance of the model is explained, and the tradeoff between the validity and stability of the model is analyzed quantitatively by using deviation and variance. Then, aiming at the strategy of reducing the risk of query expansion based on knowledge atlas, the experimental results show that the strategy is effective in two web data sets, and is compared with the query extension Model (RM3) based on pseudo-correlation feedback. The experimental results show that the extended model proposed in this paper is more effective and stable than RM3..
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】