Exploring a new approach by combining topic modeling and text summarization to improve the quality of users ' reviews summaries

Abstract

Many organizations want to have a clearer insight of the content of their users ’ reviews. There are many purposes to study users ’ reviews. One of them is to obtain a summary of the content. Many approaches exist to achieve the task. One of them is to use Topic model to detect hidden topics and extract sentences that represent most each topic to form the summary. Another is simply to perform text summarization techniques to obtain a text summary with the main ideas of the reviews. In this work, we argue that combining the two can improve the quality of the produced summary instead of applying them on their own. Our proposed approach was performed on the Opinosis dataset which contains topics with several users ’ reviews sentences. Using ROUGE metrics to evaluate the quality of the produced summaries, the results show that on one hand our proposed approach can perform better than forming a summary by only extracting the most representative sentences of each topic found by topic modeling. However, it can also underperform depending on the chosen text summarization techniques. On the other hand, it never outperforms simply applying text summarization techniques. We argue that the cause could come from the topic model and its clusters of sentences. They degrade slightly the information in the original text for the application of text summarization techniques. Therefore, we conclude that text summarization techniques created for the specific purpose of summarizing texts are still the best alternatives. However, in this work only Latent Dirichlet Allocation was implemented for topic modeling and only extractive graph-based methods were implemented for text summarization. Thus, we suggest that further research can be done to study the reliability of the results of this work with other techniques.