TSSRD: A Topic Sentiment Summarization Framework Based on Reaching Definition

Exposure to massive information in daily lives makes it necessary for people to obtain major points efficiently, promoting the development of text summarization technology. However, existing sentiment-based text summarization methods only pay attention to the sentiment polarity of either a single sentence or a whole document, ignoring changes of sentiments along with sentences or sentiment flow across the whole document. To incorporate the above two aspects into the summarization process to generate high-quality summaries, we propose a topic sentiment summarization framework based on reaching definition (TSSRD). In the framework, we first use topic models to model documents and calculate topic sentiment embeddings. Then, we analyze document structures from different perspectives to design data flow diagrams, in which improved reaching definition is used to analyze sentiment changes and sentiment flow. Finally, topic sentiment summaries are generated based on sentiments in steady states of the reaching definition. To evaluate our summarization framework, we introduce an extrinsic evaluation method. In this method, a sentiment classifier is trained by the topic sentiment summaries, and accuracy of the sentiment classification is used as a quality score. Experimental results demonstrate that our summarization framework is at least 2.32% better than baselines on IMDb and Amazon datasets.

Automated summary

The exposure to massive information in daily lives has made it necessary for people to efficiently obtain major points. This article proposes a topic sentiment summarization framework based on reaching definition (TSSRD) to generate high-quality summaries by incorporating sentiment changes and flow. Experimental results demonstrate the effectiveness of the framework.