融合多特征的专家列表排序学习方法研究
发布时间：2018-12-06 19:06
【摘要】：传统搜索引擎能通过关键词组合方式检索召回查询相关页面,但还必须经过人工方式选择与查询主题相关的信息。专家检索是当前垂直信息检索研究的热门领域,是针对专家特征而开展的更精确的信息检索方式,其能够提供多种形式主题相关查询,且直接返回与查询主题最相关的专家列表或主页,是当前最有效的专家信息获取手段。专家排序模型是专家搜索的核心,专家排序的效果决定了整个专家检索系统的性能。因此,构造高效的专家排序模型成为关键。本文对专家排序方法作了一定的探讨,致力于如何融合专家证据文档、专家关系及专家元数据等特征信息构建基于列表的专家排序模型,进而提高专家排序效果。主要在以下几个方面展开深入研究,取得了一定的成果： (1)分析了影响专家排序的因素,定义了用于专家排序的三大种类特征。针对专家排序任务,研究查询与专家页面及证据文档之间的相关性,分析专家证据文档、专家关系网、专家元数据等因素对专家检索排序影响,提取相似度特征、BM25评分、专家页面内容特征、专家关联关系特征。后续的实验表明,融入上述特征有效地提高了专家排序的效果。 (2)提出了基于ListNet的多特征融合的专家排序方法。该方法首先对专家的页面特点进行分析,选取查询和专家候选页面相关性特征、专家页面内容及专家页面间关联关系特征,然后,将特征融合到ListNet排序模型中,通过梯度下降法学习参数,构建基于列表的融合多特征的专家排序模型,最后,利用训练好的模型进行专家排序对比实验,实验结果表明提出方法有较好的效果,相比基于数据对的专家排序方法NDCG@1值提升14.2%,基于列表的融合多特征方法能够提高专家排序的效果。 (3)提出了基于关联特征的专家列表学习排序方法。该方法首先构建基于专家证据文档的相关性模型、构建基于专家关系网的相关性模型、构建基于专家元数据的相关性模型,在获得以上三个基于关联特征的相关性模型基础上,我们提出Expert-ListNet算法,然后训练得到基于关联特征的专家列表学习排序模型。通过实验证明了提出方法的有效性和优越性。 (4)利用上述研究成果,设计实现了融合多特征的专家列表排序学习原型系统。
[Abstract]:The traditional search engine can retrieve the relevant pages of recall query by keyword combination, but it must select the information related to the query subject manually. Expert retrieval is a hot field in the research of vertical information retrieval at present. It is a more accurate information retrieval method based on the characteristics of experts, and it can provide various forms of topic related queries. It is the most effective method to obtain expert information by directly returning the list of experts or the home page which is most relevant to the query topic. The expert sorting model is the core of expert search, and the effect of expert sorting determines the performance of the whole expert retrieval system. Therefore, the construction of efficient expert sorting model becomes the key. In this paper, the method of expert sorting is discussed, and how to combine the characteristic information of expert evidence document, expert relation and expert metadata to construct the expert sort model based on list is discussed in order to improve the result of expert sort. The main results are as follows: (1) the factors influencing the expert ranking are analyzed, and the three kinds of characteristics used in the expert ranking are defined. Aiming at the task of expert sorting, this paper studies the correlation between query and expert page and evidence document, analyzes the influence of expert evidence document, expert relation network, expert metadata and other factors on expert retrieval sorting, extracting similarity feature, BM25 score, etc. Expert page content feature, expert association relationship feature. The subsequent experiments show that the integration of the above features can effectively improve the effect of expert ranking. (2) an expert sorting method for multi-feature fusion based on ListNet is proposed. This method firstly analyzes the characteristics of the experts' pages, selects the correlation features of query and expert candidate pages, and then integrates the features into the ListNet sorting model, including the content of the expert pages and the associated features between the expert pages. Through gradient descent method to learn the parameters, the expert ranking model based on list fusion and multiple features is constructed. Finally, the comparison experiment of expert ranking is carried out by using the trained model. The experimental results show that the proposed method has good results. Compared with the expert sorting method based on data pair, the NDCG@1 value is increased by 14.2. the multi-feature method based on list can improve the effect of expert sorting. (3) an expert list learning ranking method based on association feature is proposed. Firstly, this method constructs the correlation model based on expert evidence document, the correlation model based on expert relation network, and the correlation model based on expert metadata. We propose Expert-ListNet algorithm, and then train to get the ranking model of expert list learning based on association feature. The effectiveness and superiority of the proposed method are proved by experiments. (4) based on the above research results, an expert list ranking learning prototype system is designed and implemented.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP18
[Abstract]:The traditional search engine can retrieve the relevant pages of recall query by keyword combination, but it must select the information related to the query subject manually. Expert retrieval is a hot field in the research of vertical information retrieval at present. It is a more accurate information retrieval method based on the characteristics of experts, and it can provide various forms of topic related queries. It is the most effective method to obtain expert information by directly returning the list of experts or the home page which is most relevant to the query topic. The expert sorting model is the core of expert search, and the effect of expert sorting determines the performance of the whole expert retrieval system. Therefore, the construction of efficient expert sorting model becomes the key. In this paper, the method of expert sorting is discussed, and how to combine the characteristic information of expert evidence document, expert relation and expert metadata to construct the expert sort model based on list is discussed in order to improve the result of expert sort. The main results are as follows: (1) the factors influencing the expert ranking are analyzed, and the three kinds of characteristics used in the expert ranking are defined. Aiming at the task of expert sorting, this paper studies the correlation between query and expert page and evidence document, analyzes the influence of expert evidence document, expert relation network, expert metadata and other factors on expert retrieval sorting, extracting similarity feature, BM25 score, etc. Expert page content feature, expert association relationship feature. The subsequent experiments show that the integration of the above features can effectively improve the effect of expert ranking. (2) an expert sorting method for multi-feature fusion based on ListNet is proposed. This method firstly analyzes the characteristics of the experts' pages, selects the correlation features of query and expert candidate pages, and then integrates the features into the ListNet sorting model, including the content of the expert pages and the associated features between the expert pages. Through gradient descent method to learn the parameters, the expert ranking model based on list fusion and multiple features is constructed. Finally, the comparison experiment of expert ranking is carried out by using the trained model. The experimental results show that the proposed method has good results. Compared with the expert sorting method based on data pair, the NDCG@1 value is increased by 14.2. the multi-feature method based on list can improve the effect of expert sorting. (3) an expert list learning ranking method based on association feature is proposed. Firstly, this method constructs the correlation model based on expert evidence document, the correlation model based on expert relation network, and the correlation model based on expert metadata. We propose Expert-ListNet algorithm, and then train to get the ranking model of expert list learning based on association feature. The effectiveness and superiority of the proposed method are proved by experiments. (4) based on the above research results, an expert list ranking learning prototype system is designed and implemented.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP18
【参考文献】
相关期刊论文 前1条
1 陈，
本文编号：2366512
本文链接：https://www.wllwen.com/kejilunwen/sousuoyinqinglunwen/2366512.html
