Research On Short Text Topic Model And Topic Evolution

Topic discovery and evolution has always been an important research topic in the natural language processing area.Topic discovery is to extract the keyword group from large corpus,which can succinctly represent the summary information of the cor-pus.And topic evolution is to further analyze the evolution process of topic content or strength over time of the extracted topics,helping researchers understand the process of the emergence and development of topics.Currently,research work on topic evolution mostly focus on long texts such as web pages,news and blogs.However,with the rise of short text data streams such as Weibo and WeChat,the traditional method of long text topic evolution is no longer suitable for sparse short texts.Therefore,there is an urgent need for new solutions for short text data stream processing.This paper focuses on the study of the topic evolution process in short text data streams.The main work of this paper contains:1)For the "short" feature of short text data streams,this paper improves on the traditional short text topic model BTM and proposes a new short text topic modeling method RIBS-TM.RIBS-TM uses recurrent neural network RNN to train the relation-ship between word pairs,and introduces the inverse document frequency IDF value to attenuate the influence of high-frequency words,thereby improving topic mining effect;2)This paper proposes a complete topic evolution model based on RIBS-TM and topic association.The model uses the improved RIBS-TM algorithm to model the top-ics of each time window's document set,and uses the topic association method to per-form association analysis of the topics between adjacent time windows.Besides,the topic association filtering rules are added to filter out invalid topic associations;3)This paper explores the different types of topic content evolution,and verifies the existence of these topic evolution phenomena on the user complaint dataset and the Sogou news title dataset.The development and changes of people's attention in different time periods are reflected in the different types of topic evolution,making the topic modeling results more practical and meaningful.