Adaptive Topic Tracking Technology Based On Incremental Learning Research

Topic tracking is one of the subtask of topic detection and tracking, the objective of it ismonitoring news and recognizing the news which are related to the given topic, which is anindispensable part of public opinion monitoring. Aimed at the problems in the traditional topictracking, this paper improved it and proposed an adaptive topic tracking method based onfeedback stories, and applied to monitoring system. The main research contents include,1.Research into the key techniques of topic tracking, and summarize the difficulties existin it.The two main techniques of topic tracking are text representation and text categorization.Through analyzing characteristics of topics, we find the core of a topic is changing with timelapse, and the related stories of one topic which clustered from topic detection are few, whichproduces a big problem to topic tracking. So it is necessary to amend topic modelsdynamically.2.Improve Vector Space Model based on named entities to make the topic model expressa topic betterFor the reason that, the named entities could differentiate different topics more effectivethan nouns, verbs and other words, so increasing the weight of named entity when build topicmodels. We identify the named entities by the part-of-speech tagging of word segmentationsystem, and increase the weight in the process of feature extraction. Experimental results showthat, the improved topic models could express a topic better.3.Propose a feature expanding method of topic models based on feedback storiesAim at the characteristic of a topic that it evolutes dynamically, utilizing collectingfeedback stories, extract items which have high weight to extend topic model, and adjust theweight of existed feature items. This method could amend topic models dynamically, andimprove the problem of topic tracking accuracy descending brought by topic shifting.4.Reduce the noise data by feedback stories which collected by dynamic thresholdsBecause of the importance of the accuracy of feedback stories, which would influence theefficiency of topic tracking, so it is important to avoid noise data in feedback stories. In thispaper, a dynamic threshold is utilized to collect feedback stories, the setting of the thresholdrelated to several parameters, the similarity between feedback stories and original topic modelis taken as base number, and the proportion of related stories and total stories is used to amendthe threshold to avoid that the threshold is too high to cause a high missing rate. In addition, acoefficient is added to raise the threshold to reduce noise data.5. According to the previous research, propose an adaptive topic tracking method basedon feedback storiesThe adaptive topic tracking method proposed in this paper, which based on the improvedVector Space Model based on named entities, feature expanding of topic models and the dynamic threshold setting. The method could improve traditional topic tracking well, andreduce the problem of topic shifting. Experimental results show that it could increase theefficiency of topic tracking obviously.6.Apply the adaptive topic tracking method proposed in this paper to public opinionmonitoring system, which improve the efficiency of the hot topic tracking module.Design and complete the public sentiment monitoring system, accomplish the three mainmodules, information collection, public sentiment analysis and public sentiment present. Andapply the adaptive topic tracking method to the hot topic tracking module, to improve thewhole performance of the system.