Summarize What You Are Interested In: An Optimization Framework for Interactive Personalized Summarization

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1342–1351, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Summarize What You Are Interested In: An Optimization Framework for Interactive Personalized Summarization Rui Yan Department of Computer Science and Technology, Peking University, Beijing 100871, China r.yan@pku.edu.cn Jian-Yun Nie D´epartement d´informatique et de recherche op´erationnelle, Universit´e de Montr´eal, Montr´eal, H3C 3J7 Qu´ebec, Canada nie@iro.umontreal.ca Xiaoming Li Department of Computer Science and Technology, Peking University, Beijing 100871, China lxm@pku.edu.cn Abstract Most traditional summarization methods treat their outputs as static and plain texts, which fail to capture user interests during summa- rization because the generated summaries are the same for different users. However, users have individual preferences on a particular source document collection and obviously a universal summary for all users might not al- ways be satisfactory. Hence we investigate an important and challenging problem in sum- mary generation, i.e., Interactive Personalized Summarization (IPS), which generates sum- maries in an interactive and personalized man- ner. Given the source documents, IPS captures user interests by enabling interactive clicks and incorporates personalization by model- ing captured reader preference. We develop experimental systems to compare 5 rival al- gorithms on 4 instinctively different datasets which amount to 5197 documents. Evalua- tion results in ROUGE metrics indicate the comparable performance between IPS and the best competing system but IPS produces sum- maries with much more user satisfaction ac- cording to evaluator ratings. Besides, low ROUGE consistency among these user pre- ferred summaries indicates the existence of personalization. 1 Introduction In the era of information explosion, people need new information to update their knowledge whilst infor- mation on Web is updating extremely fast. Multi- document summarization has been proposed to ad- dress such dilemma by producing a summary de- livering the majority of information content from a document set, and hence is a necessity. Traditional summarization methods play an im- portant role with the exponential document growth on the Web. However, for the readers, the impact of human interests has seldom been considered. Tra- ditional summarization utilizes the same methodol- ogy to generate the same summary no matter who is reading. However, users may have bias on what they prefer to read due to their potential interests: they need personalization. Therefore, traditional summa- rization methods are to some extent insufficient. Topic biased summarization tries for personaliza- tion by pre-defining human interests as several gen- eral categories, such as health or science. Readers are required to select their possible interests before summary generation so that the chosen topic has priority during summarization. Unfortunately, such topic biased summarization is not sufficient for two reasons: (1) interests cannot usually be accurately pre-defined by ambiguous topic categories and (2) user interests cannot always be foreknown. Often users do not really know what general ideas or detail information they are interested in until they read the summaries. Therefore, more flexible interactions are required to establish personalization. Due to all the insufficiencies of existed sum- marization approaches, we …

Content truncated to 500 words