话题检测与跟踪算法改进分析-计算机应用技术专业论文.docxVIP

华中科技大学硕士学位论文 华 中 科 技 大 学 硕 士 学 位 论 文 II II Abstract With the rapid development of the Internet ,the Internet information shows explosive growth .Effective organization and management of these information has become more difficult and it often appears the phenomenon of information overload .For effectively organizing and managing these information ,topic detection and tracking technology came into being .For a variety of news report of information flow ,detecting new topics and tracking the follow-up reports of known topics are the main purpose . According to the characteristics of the topic detection, hierarchical clustering is used to detect topics where it doesn’t set the number of categories when clustering. Hierarchical clustering can well adapt to needs of topic detection, and on this basis named entity has high-sensitive characteristics in the topic report. Increasing the weight value of the named entity in the calculation can improve overall system performance during similarity calculation. The experiments in the existing corpus and experimental data show that the improvement on the similarity calculation improve the rate of correct of topic detection and reduce system spending. When commonly used traditional K-nearest neighbor algorithm is applied to track the topics ,it requires balance between the reported number of topic and this shortcoming will cause the topic to offset to a certain extent. Support vector machine algorithm is used in the training phase of K-nearest neighbor algorithm to determine support vector to replace the value of K and it can eliminate the dependence of value of K. It decreases topic offset problems because of uneven between the reported number of topic. Experimental results show that to a certain extent these methods improve the performance of topic tracking and it tests and verifies that the correct rate of improved K-nearest neighbor algorithm isn’t affected by parameter K. Key words: topic detection, topic tracking, hierarchica