A Semantic Similarity Based Topic Evaluation for Enhancing Information Filtering

Topic Modelling has been applied in many successful applications in data mining, text mining, machine learning and information filtering. The limitation is that the quality of topics generated from modelled corpus are not always good because many topics contain intrusive and ambiguous words. This negative drawback would affect the performance of text based application systems based on topic models. Hence, topic evaluation to assess and to rank the topics is really important for the good quality topics before applying those topics to text based applications. In this study, we proposed an ontology-based topic evaluation method for enhancing information filtering, named as STRbTCM. This new model assesses the quality of topics by matching topic models with headings in Library Congress Subject Heading (LCSH) ontology. To evaluate the effectiveness of our proposed model, we compare the model with two existing topic evaluation methods applied to information filtering system. In addition, we also compare our proposed model to term-based model BM25 and two other models based on topics: TNG and LDA_words. Through extensive experiments, we find that our proposed model performed better than other baseline models according to four main evaluating measures.