Topic Extraction From Short Texts On Social Media

The prosperity of social media results in the appearance of a large number of short and noisy texts.Topic extraction from short texts aims to reveal thematic information of the underlying collection,which can be used in summarization,hashtag recommendation,response generation and so on.Now,the methods of topic extraction from short texts on social media can be classified into aggregation strategy based,prior knowledge based and combining text content and network structure.However,they all ignore the shared context information across events and the dynamic user behaviours on the social media.Based on the two points,we further explore topic extraction from short texts on social media.1.The existing researches use word embeddings as prior knowledge to guide modeling or integrate conversation structures to enrich context,which can be employed to infer the topics of a single event on social media.However,the shared context across a large number of events is ignored,which can be used as prior knowledge to reinforce coherent topic generation for each event.Thus,this paper proposes a Reinforced Knowledge LDA(RKLDA)for discovering topics of each event.2.The existing topic models only consider text information or simultaneously model the posts and the static characteristics of social network.They ignore the dynamic user behaviors in social media,which may result in the poor consistency of generated topics.However,one discusses diverse topics when dynamically interacting with different people.Moreover,people who talk about the same topic have various effects on the topic.These phenomena provide us the potential and useful clues.Inspired by network embedding and social user behavior researches,this paper proposes an Interaction Aware Topic Model(IATM)based on dynamic interactions and user attention mechanism.Experiments on Sina Weibo datasets show the effectiveness of the proposed models by comparing with the previous methods.