Topic-Aware Learning-to-Rank for Cross-Lingual Information Retrieval

Cross-lingual information retrieval (CLIR) enables search across languages, 
avoiding storage of a cross-lingual index by either translating the search 
query or the document repository in order to apply monolingual retrieval. 
The primary goal of cross-lingual retrieval is to find relevant documents, 
however, optimizing translations to preserve meaning and structure in 
bitexts while being agnostic of the search task does not necessarily yield 
optimal search results. In addition, the meaning of a query is highly 
dependent on contextual information such as topic, category, or authorship. 
We propose a multilingual CLIR system that does not rely on machine 
translation but directly models relevance between queries and documents 
across languages. Our approach avoids the "lost in translation" problem 
often found in cross-lingual pipeline models. The system is trained and 
evaluated on multilingual data from Wikipedia, but can easily be adapted 
to other domains where structured document information is available such as 
patent prior art search.