Multilingual User Modeling for Personalized Re-ranking of Multilingual Web Search Results M. Rami Ghorab, Dong Zhou, Seamus Lawless, Vincent Wade Centre for Next Generation Localisation, Knowledge & Data Engineering Group, School of Computer Science & Statistics, Trinity College Dublin, Ireland {ghorabm, dong.zhou, seamus.lawless, vincent.wade}@scss.tcd.ie Abstract. This paper proposes a novel method to represent user models in a multilingual manner which caters for multilingual Web search users. Further- more, an evaluation is presented which examines a result re-ranking algorithm that is based upon that model. Keywords: User Modeling, Personalized Multilingual Information Retrieval. 1 Introduction Global multilinguality is becoming an important feature of Web users’ daily interac- tion with information. The process of searching for information on the Web has two dimensions: the multilingual Web and a multilingual user. The first dimension has been subject to much research in the literature; many studies have investigated im- proving various aspects of the accessibility and search-ability of multilingual content, such as machine translation quality [1] or cross-lingual / multilingual 1 search effec- tiveness [2]. On the other hand, the multilingual user dimension has not been the sub- ject of the same amount of research. Modeling users’ interests for personalized search [3] has been studied, but only on a monolingual level. This research study argues that the user’s searches may be influenced by language. For example, a multilingual user may prefer to use his/her native language when searching for certain types of content (e.g. news), yet choose to use English when searching for other types (e.g. technical content). Furthermore, in multilingual search, the user may choose to click on documents coming from certain languages depending on the type of information sought. This kind of behavior may suggest that a multilin- gual user has multiple behavioral personas in Web search. This paper proposes a nov- el method for modeling the user’s search interests in a multilingual fashion. Instead of traditional methods which store interest terms (keywords) in a single language, this method stores the terms in multiple languages. Such user model representation aims at capturing the multiple personas of a multilingual user. Furthermore, an algorithm for re-ranking and merging multilingual search results is proposed and evaluated. 1 Multilingual search involves performing cross-language search in multiple target languages and then merging and/or translating the search results before presenting them to the user.