Automatic Labeling of Topic Models Using Text Summaries

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 2297–2305, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Automatic Labeling of Topic Models Using Text Summaries Xiaojun Wan and Tianming Wang Institute of Computer Science and Technology, The MOE Key Laboratory of Computational Linguistics, Peking University, Beijing 100871, China {wanxiaojun, wangtm}@pku.edu.cn Abstract Labeling topics learned by topic models is a challenging problem. Previous studies have used words, phrases and images to label topics. In this paper, we propose to use text summaries for topic labeling. Several sentences are extracted from the most related documents to form the sum- mary for each topic. In order to obtain summaries with both high relevance, cov- erage and discrimination for all the topics, we propose an algorithm based on sub- modular optimization. Both automatic and manual analysis have been conducted on two real document collections, and we find 1) the summaries extracted by our proposed algorithm are superior over the summaries extracted by existing popular summarization methods; 2) the use of summaries as labels has obvious ad- vantages over the use of words and phrases. 1 Introduction Statistical topic modelling plays very important roles in many research areas, such as text mining, natural language processing and information re- trieval. Popular topic modeling techniques in- clude Latent Dirichlet Allocation (LDA) (Blei et al., 2003) and Probabilistic Latent Semantic Anal- ysis (pLSA) (Hofmann, 1999). These techniques can automatically discover the abstract “topics” that occur in a collection of documents. They model the documents as a mixture of topics, and each topic is modeled as a probability distribution over words. Although the discovered topics’ word distribu- tions are sometimes intuitively meaningful, a ma- jor challenge shared by all such topic models is to accurately interpret the meaning of each topic (Mei et al., 2007). The interpretation of each topic is very important when people want to browse, understand and leverage the topic. However, it is usually very hard for a user to understand the dis- covered topics based only on the multinomial dis- tribution of words. For example, here are the top terms for a discovered topic: {fire miles area north southern people coast homes south damage northern river state friday central water rain high california weather}. It is not easy for a user to fully understand this topic if the user is not very familiar with the document collection. The situa- tion may become worse when the user faces with a number of discovered topics and the sets of top terms of the topics are often overlapping with each other on many practical document collections. In order to address the above challenge, a few previous studies have proposed to use phrases, concepts and even images for labeling the discov- ered topics (Mei et al., 2007; Lau et al., 2011; Hulpus et al., 2013; Aletras and Stevenson, 2013). For example, we may automatically extract the phrase “southern california” to represent the ex- ample topic mentioned earlier. These topic labels can help …

Content truncated to 500 words