A Comparative Study of Sentiment-Based Graphs of Text Summaries

摘要

Sentiment included in a sentence can indicate whether a sentence may have positive, negative or neutral polarity. Polarity of the sentences is deemed important in text summarization, especially when summarizing narrative texts. This paper proposes to discover the patterns and sentiment scores of the summaries generated by established summarization methods: Luhn, Latent Semantic Analysis (LSA) and LexRank. This is done by conducting a study and comparison on the generated sentiment-based graphs of the summaries. A comparative study is conducted on the sentiment-based graph of the generated summaries with two different sentiment lexicons, namely SentiWordNet and VADER. The analysis is conducted by comparing the patterns of the sentiment-based graph and their sentiment scores as well. In the experiments conducted, there is an obvious pattern for the two sentiment lexicons. This implies that sentiment-based graph's pattern and score are helpful in generating a compact summary. The analysis will alleviate future research on sentiment-based summarization and motivates a new method which can be considered as a graph-based summarization to extract a summary based on its sentiment score.