Topic Modeling Using Community Detection on a Word Association Graph

Proceedings of Recent Advances in Natural Language Processing, pages 908–917 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_098 908 Topic Modeling Using Community Detection on a Word Association Graph Mahfuzur Rahman Chowdhury Brac University Dhaka, Bangladesh mahfuzur.rahman@bracu.ac.bd Intesur Ahmed Brac University Dhaka, Bangladesh intesur.ahmed@bracu.ac.bd Farig Sadeque Brac University Dhaka, Bangladesh farig.sadeque@bracu.ac.bd Muhammad Nur Yanhaona Brac University Dhaka, Bangladesh nur.yanhaona@bracu.ac.bd Abstract Topic modeling of a text corpus is one of the most well-studied areas of information re- trieval and knowledge discovery. Despite sev- eral decades of research in the area that begets an array of modeling tools, some common prob- lems still obstruct automated topic modeling from matching users’ expectations. In partic- ular, existing topic modeling solutions suffer when the distribution of words among the un- derlying topics is uneven or the topics are over- lapped. Furthermore, many solutions ask the user to provide a topic count estimate as in- put, which limits their usefulness in modeling a corpus where such information is unavailable. We propose a new topic modeling approach that overcomes these shortcomings by formu- lating the topic modeling problem as a commu- nity detection problem in a word association graph/network that we generate from the text corpus. Experimental evaluation using mul- tiple data sets of three different types of text corpora shows that our approach is superior to prominent topic modeling alternatives in most cases. This paper describes our approach and discusses the experimental findings. 1 Introduction The goal of topic modeling is to find the underly- ing semantic structure in a corpus that succinctly describes the documents and the text forming the corpus without compromising the corpus’s statisti- cal characteristics. It is one of the oldest and most researched problems in the field of information re- trieval and has numerous direct and downstream applications such as document grouping, classifi- cation, retrieval, and summarization. For a long time, the most prominent solutions to topic mod- eling are Latent Dirichlet Allocation (LDA) (Blei et al., 2003) and its variants. LDA works under the principle of ‘interchangeability’ of documents and describes a document as a random mixture of a fixed number of latent ‘topics’ drawn from a Dirichlet distribution where each topic is a distri- bution of words. LDA is a bag of words model as it disregards the position of words in a document. LDA’s assumption of a latent Dirichlet distribu- tion for the topics and interchangeability of doc- uments under the bag of words model is both for the mathematical tractability of the problem, as its authors admit (Blei et al., 2003), as opposed to its conformance with any empirical rule describing real texts (e.g., the Zipf’s law (Zipf, 1935)). How- ever, LDA provides the first principled approach to group both the documents and the words of a cor- pus and remains widely applicable. Consequently, most subsequent works on topic modeling focused on improving the LDA model which led to many LDA variants. For example, Biterm (Yan et al., 2013) adapts LDA for short texts, HDP (Teh et al., 2004) eliminates …

Content truncated to 500 words