Interaction-Aware Topic Model for Microblog Conversations through Network Embedding and User Attention

Proceedings of the 27th International Conference on Computational Linguistics, pages 1398–1409 Santa Fe, New Mexico, USA, August 20-26, 2018. 1398 Interaction-Aware Topic Model for Microblog Conversations through Network Embedding and User Attention Ruifang He1,2,∗, Xuefei Zhang1,2,∗, Di Jin1,2,†, Longbiao Wang1,2, Jianwu Dang1,2,3, Xiangang Li4 1School of Computer Science and Technology, Tianjin University, Tianjin, China 2Tianjin Key Laboratory of Cognitive Computing and Application, Tianjin, China 3Japan Advanced Institute of Science and Technology, Ishikawa, Japan 4AI Labs, Didi Chuxing, Beijing, China {rfhe,jindi,longbiao wang}@tju.edu.cn,zhangxuefei29@163.com jdang@jaist.ac.jp,lixiangang@didichuxing.com Abstract Traditional topic models are insufficient for topic extraction in social media. The existing meth- ods only consider text information or simultaneously model the posts and the static characteristics of social media. They ignore that one discusses diverse topics when dynamically interacting with different people. Moreover, people who talk about the same topic have different effects on the topic. In this paper, we propose an Interaction-Aware Topic Model (IATM) for microblog con- versations by integrating network embedding and user attention. A conversation network linking users based on reposting and replying relationship is constructed to mine the dynamic user be- haviours. We model dynamic interactions and user attention so as to learn interaction-aware edge embeddings with social context. Then they are incorporated into neural variational inference for generating the more consistent topics. The experiments on three real-world datasets show that our proposed model is effective. 1 Introduction The prosperity of microblog platforms, such as Twitter1 and Sina Weibo2 brings the large scale, noisy and user-generated short posts. Automatically extracting topics in social media aims to reveal thematic information of the underlying collection, which can be used in summarization (Zhuang et al., 2016), hashtag recommendation (Li et al., 2016b), response generation (Xing et al., 2017) and so on. The conventional topic models, like Latent Dirichlet Allocation (LDA) (Blei et al., 2003), infer the hidden topics based on word co-occurrences in documents. They could not be directly transferred in social media due to the data sparsity and the noise of short texts. The existing relevant researches can be roughly categorized into: (1) Methods exploit the content of short texts by aggregation strategy (Mehrotra et al., 2013) or modeling word-pair co-occurrence (Yan et al., 2013). (2) Models incorporating with word embedding try to deeply understand the posts by repre- sentation learning (Sridhar, 2015; Hu and Tsujii, 2016), yet they ignore the social context of microblog messages. Essentially, social media content and network structures influence each other (Bi et al., 2014), the only content analysis is insufficient. (3) (Li et al., 2016a; Chen and Ren, 2017) take into account the content and static characteristics of network structures to deduce topics. However, they ignore dynamic user behaviours. Actually, a user may talk about various topics when interacting with diverse neighbours in a social network. For instance, Fig. 1(a) shows two conversation trees, where [U0] communicates with [U1] and [U2] about the topic of United Kingdom European Union membership referendum, while [U0] talks a game with [U3]. Hence, we need to infer topics according to …

Content truncated to 500 words