Dear all,Alessandro kindly shared his slides with us. The slides are attached.Best,- KOn Wed, Mar 25, 2015 at 9:49 AM, Kyung Hyun Cho <cho.k.hyun@gmail.com> wrote: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 SordoniDate/Time: 27 March @14.00Place: AA3195Title: Generating Context-Aware Query Suggestions with Hierarchical Recurrent Neural Networks
Abstract:Query suggestion consists in presenting related queries to users afterthey 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 importantcontext for suggestion is crucial but challenging. We propose ahierarchical generative model of query sequences that allows togenerate synthetic suggestions given an unbounded history of previousqueries as context. Our approach is based on a novel neural networkapproach that generates two levels of sequences, each with its ownrecurrent state: sequences of words, i.e. user queries, and sequencesof queries, i.e. user sessions. Our model outperforms severalcontext-aware approaches in next query prediction setting.Additionally, the model is able to generate fluent recommendationseven for long-tail queries. Our findings are corroborated by a userstudy.Short bio:Alessandro Sordoni is a PhD student at RALI, under the direction ofJian-Yun Nie.