Monotone Submodularity in Opinion Summaries

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 169–178, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Monotone Submodularity in Opinion Summaries Jayanth Jayanth IIT Bombay jayanthjaiswal10 @cse.iitb.ac.in Jayaprakash Sundararaj IIT Bombay osjayaprakash @gmail.com Pushpak Bhattacharyya IIT Bombay pb @cse.iitb.ac.in Abstract Opinion summarization is the task of pro- ducing the summary of a text, such that the summary also preserves the sentiment of the text. Opinion Summarization is thus a trade-off between summarization and sen- timent analysis. The demand of com- pression may drop sentiment bearing sen- tences, and the demand of sentiment de- tection may bring in redundant sentences. We harness the power of submodularity to strike a balance between two conflict- ing requirements. We investigate an in- cipient class of submodular functions for the problem, and a partial enumeration based greedy algorithm that has perfor- mance guarantee of 63%. Our functions generate summaries such that there is good correlation between document sentiment and summary sentiment along with good ROUGE score, which outperforms the- state-of-the-art algorithms. 1 Introduction Sentiment Analysis is often addressed as a classi- fication task, which aims at determining the sen- timent of a word, sentence, paragraph or a docu- ment as a whole into positive, negative or neutral classes (Pang et al., 2002). Summarization, on the other hand is the task of aggregating and represent- ing information content from a single document or multiple documents in a brief and fluent manner. Due to the explosive growth of data, fine grained sentiment analysis as well as summarization on the whole chunk of data can be a very time-consuming task. Sentiment Analysis also requires filtering of text portions as either objective (factual informa- tion) or subjective (expressing some sentiment or opinion) during pre-processing and then, classify- ing the subjective extracts as positive or negative. Subjective extracts can also be provided to users as a summary of the sentiment-oriented content of the reviews in search engines. In this paper, we address the problem of generic extractive sum- marization of reviews, a task commonly known as Opinion Summarization (Liu, 2012). The goals of opinion summarization are: 1. Present a short summary that conveys the essence as well as the sentiment of the review 2. Provide a short subjective extract to NLP pipeline for faster execution (e.g. sentiment analysis, review clustering etc.). In this paper, we use movie reviews for opinion summarization task as they often have the follow- ing parts: 1. Plot - Description of the story, which is fac- tual in nature 2. Critique - Opinion about the movie, which is sentiment bearing Clearly, opinion summary to be generated will have a trade-off between the two opposing parts - subjective critique and objective plot. Our goal is to strike a balance through linear combination of suitable submodular functions in our paper. Joint models of relevance and subjectivity have a great benefit in that they have a large degree of freedom as far as controlling redundancy goes. In con- trast, conventional two-stage …

Content truncated to 500 words