Query Splitting for Context-Driven Federated Recommendations

Context-driven query extraction for content-based recommender systems faces the challenge of dealing with queries of multiple topics. In contrast to manually entered queries, for automatically generated queries this is a more frequent problem. For instances if the information need is inferred indirectly viathe user's current context. Especially for federated search systems were connected knowledge sources might react vastly differently on such queries, an algorithmic way how to deal with such queries is of high importance. One such method is to split mixed queries into their individual subtopics. To gain insight how a multi topic query can be split into its subtopics we conducted an evaluation where we compared a naive approach against amore complex approaches based on word embedding techniques: One created using Word2Vec and one created using GloVe. To evaluate these two approaches we used the Webis-QSeC-10 query set, consisting of about 5,000 multi term queries. Queries of this set were concatenated and passed through the algorithms with the goal to split those queries again. Hence the naive approach is splitting the queries into several groups, according to the amount of joined queries, assuming the topics are of equal query term count. In the case of the Word2Vec and GloVe based approaches we relied on the already pre-trained datasets. The Google News model and a model trained with a Wikipedia dump and theEnglish Gigaword newswire text archive. The out of this datasets resulting query term vectors were grouped into subtopics usinga k-Means clustering. We show that a clustering approach based on word vectors achieves better results in particular when the query is not in topical order. Furthermore we could demonstrate the importance of the underlying dataset.