Topic Modeling Using Contextual Cues

Abstract

This paper proposes a new solution for topic modeling using contextual cues by applying Automatic Term Recognition (ATR) to extract domain-specific terms in the text preprocessing step. The vocabularies used for topic modeling are constructed using linguistic patterns to determine the inner structure of each document analyzed and, by only taking into account the document's domain relevant terms and ignoring other words which do not bring any kind of informational gain, the proposed method manages to extract more coherent and human readable topics with a higher grade of separability as shown in our experiments. To evaluate our method, we use publicly available data sets containing scientific and news articles which are already labeled. These classes are used as ground truth for the clustering evaluation methods and as an evaluation technique for human readability. Moreover, the coherence of topics is evaluated by using the CV method proposed in the Palmetto framework.