Keeping keywords fresh: a BM25 variation for personalized keyword extraction

Abstract

Keyword extraction from web pages is essential to various text mining tasks including contextual advertising, recommendation selection, user profiling and personalization. For example, extracted keywords in contextual advertising are used to match advertisements with the web page currently browsed by a user. Most of the keyword extraction methods mainly rely on the content of a single web page, ignoring the browsing history of a user, and hence, potentially leading to the same advertisements or recommendations. In this work we propose a new feature scoring algorithm for web page terms extraction that, assuming a recent browsing history per user, takes into account the freshness of keywords in the current page as means of shifting users interests. We propose BM25H, a variant of BM25 scoring function, implemented on the client-side, that takes into account the user browsing history and suggests keywords relevant to the currently browsed page, but also fresh with respect to the user's recent browsing history. In this way, for each web page we obtain a set of keywords, representing the time shifting interests of the user. BM25H avoids repetitions of keywords which may be simply domain specific stop-words, or may result in matching the same ads or similar recommendations. Our experimental results show that BM25H achieves more than 70% in precision at 20 extracted keywords (based on human blind evaluation) and outperforms our baselines (TF and BM25 scoring functions), while it succeeds in keeping extracted keywords fresh compared to recent user history.