More Efficient Topic Modelling Through a Noun Only Approach

More Efficient Topic Modelling Through a Noun Only Approach Fiona Martin Department of Computing, Macquarie University, NSW, 2109, Australia fiona.martin@students.mq.edu.au Mark Johnson Department of Computing, Macquarie University, NSW, 2109, Australia mark.johnson@mq.edu.au Abstract This study compared three topic models trained on three versions of a news cor- pus. The first model was generated from the raw news corpus, the second was gen- erated from the lemmatised version of the news corpus, and the third model was generated from the lemmatised news cor- pus reduced to nouns only. We found that the removing all words except nouns improved the topics’ semantic coherence. Using the measures developed by Lau et al (2014), the average observed topic coher- ence improved 6% and the average word intrusion detection improved 8% for the noun only corpus, compared to modelling the raw corpus. Similar improvements on these measures were obtained by simply lemmatising the news corpus, however, the model training times are faster when reducing the articles to the nouns only. 1 Introduction A challenge when analysing a large collection of text documents is to efficiently summarise the multitude of themes within that collection, and to identify and organise the documents into particular themes. Document collections such as a newspa- per corpus contain a wide variety of themes or top- ics, with each individual article referencing only a very small subset of those topics. Such topics may be broad and coarse grained, such as politics, fi- nance or sport. Alternatively, topics may be more specific, such as articles related to earthquakes in southern California, or to Napa Valley wineries. Topic modelling is one way to examine the themes in large document collections. Topic mod- elling considers documents to be a mixture of la- tent topics. A more formal definition of topics, as provided by Blei (2012), is that a topic is a multi- nomial distribution over a fixed vocabulary. One of the most prominent algorithms for topic mod- elling is the Latent Dirichlet Allocation (LDA) al- gorithm, developed by Blei et al. (2003). Typ- ically the most frequent function words are ex- cluded prior to topic modelling with LDA (termed stop word removal). The topics then generated by LDA can be a mixture of nouns, verbs, adjectives, adverbs and any function words not previously ex- cluded. The LDA algorithm treats all word tokens as having equal importance. It is common to examine the most frequent words associated with the topic, to determine if these words together suggest a particular theme. For example, a topic with the most frequent words {water plant tree garden flower fruit val- ley drought} suggests a possible label of “garden”, whereas a topic of {art good house room style work fashion draw} seems to combine multiple themes, and is harder to label. Manually assign- ing a meaning to a topic (e.g. “gardening”) is eas- ier for a reviewer if the most frequent words in the topic are semantically coherent. One issue iden- tified with topic modelling is that it can generate ‘junk’ topics …

Content truncated to 500 words