Automatically Identifying Latent User Goals to Improve Web Search

• Main Authors: Kuan-yu He, 何寬禹
• Other Authors: Wen-hsiang Lu
• Format: Others
• Language: en_US
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/43727438899071435912

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 95 === With the quick expansion of the Web, users often suffer from the problem of information overloading. Therefore, it is eagerly expected for users that search engines could quickly respond exact results what they want. However, it seems difficult for most existing search engines to understand user needs exactly behind diverse short queries with limited information. There is much evolution in the history of developing search mechanisms after the advent of search engines in the world. The search mechanisms, such as keyword matching which is the classical technique in informational retrieval, link-structure algorithms which consider the link structure between Web pages, and the related search techniques which employ user’s past click-through data, all can bring a wave of improvement to search engines in the times. As the resources in Web are getting more and more heterogeneous, the needs of users are also getting more and more various. As the role of a window to facilitate users to quickly obtain the resources in Web, search engines are required more heavily by Web users day by day. In the current wave of improving Web search, search mechanisms based on user goals become a main stream. Just accurately understanding what the user needs are, that is, why users do Web search, can bring an evolution of improvement to the Web. However, nowadays, current researches on user goals simply discussed the characteristics of user goals, or proposed an automatic approach to judge the category of the possible user goal behind an issued query, not mentioned that they can identify possible user goals to further improve Web search. In this thesis, we propose an enhanced approach to utilizing search-result snippets to identify latent user goals for improving our previous method which employs syntactic structures (verb-object pairs) to discover a variety of latent user goals. Our new approach employs supervised-learning and boostrapping techniques to learn hint verbs, and considers URL information and title information to classify snippets into three major categories, which are resource-seeking, informational, and navigational. Also, we propose three different methods to identify three different categories of diverse latent user goals from classified snippets. In addition, we propose a unified user goal model to unify three categories of user goals and finally employ our proposed search model utilizing unified user goals to re-rank search results. The most valuable contribution in this thesis is that our method can identify heterogeneous resource-seeking goals, which are very difficult to identify, in the form of verb-object pair. The experimental results show that the performance of our new method of resource-seeking goal identification is better than our previous method and we also identify two other categories (informational and navigational) of latent user goals ignored by our previous work. The experimental results also show that the re-ranked search results based on user goals fitting to users can more satisfy users’ search need.