Automatic Labelling of Topics with Neural Embeddings

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 953–963, Osaka, Japan, December 11-17 2016. Automatic Labelling of Topics with Neural Embeddings Shraey Bhatia1,2, Jey Han Lau1,2 and Timothy Baldwin2 1 IBM Research 2 Dept of Computing and Information Systems, The University of Melbourne shraeybhatia@gmail.com, jeyhan.lau@gmail.com, tb@ldwin.net Abstract Topics generated by topic models are typically represented as list of terms. To reduce the cognitive overhead of interpreting these topics for end-users, we propose labelling a topic with a succinct phrase that summarises its theme or idea. Using Wikipedia document titles as label candidates, we compute neural embeddings for documents and words to select the most relevant labels for topics. Compared to a state-of-the-art topic labelling system, our methodology is simpler, more efficient, and finds better topic labels. 1 Introduction Topic models are a popular approach to detecting trends and traits in document collections, e.g. in tracing the evolution of a scientific field through its publications (Hall et al., 2008), enabling visual navigation over search results (Newman et al., 2010a), interactively labelling document collections (Poursabzi-Sangdeh et al., 2016), or detecting trends in text streams (Wang and McCallum, 2006; AlSumait et al., 2008). They are typically unsupervised, and generate “topics” ti in the form of multinominal distributions over the terms wj of the document collection (Pr(wj|ti)), and topic distributions for each document dk in the collection, in the form of a multinomial distribution over topics (Pr(ti|dk)). Traditionally, this has been carried out based on latent Dirichlet allocation (LDA: Blei et al. (2003)) or extensions thereof, but more recently there has been interest in deep learning approaches to topic modelling (Cao et al., 2015; Larochelle and Lauly, 2012). In contexts where the output of the topic model is presented to a human user, a fundamental concern is the best way of presenting the rich information generated by the topic model, in particular, the topics themselves, which provide the primary insights into the document collection. The de facto topic representation has been a simple term list, in the form of the top-10 terms in a given topic, ranked in descending order of Pr(wj|ti). The cognitive overhead in interpreting the topic presented as a list of terms can be high, and has led to interest in the task of generating labels for topics, e.g. in the form of textual descriptions (Mei et al., 2007; Lau et al., 2011; Kou et al., 2015), visual representations of the topic words (Smith et al., to appear), or images (Aletras and Stevenson, 2013). In the former case, for example, rather than the top-10 terms of ⟨school, student, university, college, teacher, class, education, learn, high, program ⟩, a possible textual label could be simply EDUCATION. Recent work has shown that, in the context of a timed information retrieval (IR) task, automatically-generated textual labels are easier for humans to interpret than the top-10 terms, and lead to equivalent-quality relevance judgements (Aletras et al., 2014). Despite this, the accuracy of state-of-the-art topic generation methods is far from …

Content truncated to 500 words