Dear all, This week, Alessandro Sordoni from RALI will tell us about how a hierarchical recurrent neural net can be used to suggest a information-retrieval query based on a history of previous queries/returns. Hope to see many of you there! - K === Speaker: Alessandro Sordoni Date/Time: 27 March @14.00 Place: AA3195 Title: Generating Context-Aware Query Suggestions with Hierarchical Recurrent Neural Networks Abstract: Query suggestion consists in presenting related queries to users after they issue one or several queries to a web search engine. Typically, suggestions are computed on the basis of the most recent query only. Taking into account the full history of previous queries as important context for suggestion is crucial but challenging. We propose a hierarchical generative model of query sequences that allows to generate synthetic suggestions given an unbounded history of previous queries as context. Our approach is based on a novel neural network approach that generates two levels of sequences, each with its own recurrent state: sequences of words, i.e. user queries, and sequences of queries, i.e. user sessions. Our model outperforms several context-aware approaches in next query prediction setting. Additionally, the model is able to generate fluent recommendations even for long-tail queries. Our findings are corroborated by a user study. Short bio: Alessandro Sordoni is a PhD student at RALI, under the direction of Jian-Yun Nie.