Résumé: pt Automatically summarizing a group of short texts that mainly share one topic is a fundamental task in many applications, e.g., summarizing the main symptoms for a disease based on a group of medical texts that are usually short, i.e., tens of words. Conventional unsupervised short text summarization techniques tend to find the most representative short text document. However, they may cause privacy issues, e.g., personal information in the medical texts may be exposed. Moreover, compared with the complete short text where some unimportant words may exist, a summary consisting of only a few keywords is more preferable by the user due to its clear and concise form. Due to the above reasons, in this article, we aim to solve the problem of unsupervised derivation of keyword summary for short texts. Existing keyword extraction methods such as Latent Dirichlet Allocation cannot be applied to solve this problem, since (1) the ordering relations among the extracted keywords are ignored, which causes troubles for people to capture the main idea of the event, and (2) short texts contain limited context, which makes it hard to find the optimal words for semantic coverage. Hence, we propose a simple but yet effective method named Frequent ClosedWordsets Ranking (FCWRank) to derive the keyword summary from a short text cluster. FCWRank is an unsupervisedmethod that builds on the idea of frequent closed itemsetmining in transaction database. FCWRank firstmines all frequent closed wordsets from a cluster of short texts and then selects the most important wordset based on an importance model where the similarity between closed wordsets and the relation between the closed wordset and the short text document are considered simultaneously. To make the keywords within the wordset more understandable, FCWRank further unfolds the semantics behind them by sorting them. Experiments on real-world short text collections show that FCWRank outperforms the state-of-the-art baselines in terms of Recall-Oriented Understudy for Gisting Evaluation-Longest common subsequence F1, precision and recall scores.

Langue: Anglais