Enforcing Topic Diversity in a Document Recommender for Conversations

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 588–599, Dublin, Ireland, August 23-29 2014. Enforcing Topic Diversity in a Document Recommender for Conversations Maryam Habibi Idiap Research Institute and EPFL Rue Marconi 19, CP 592 1920 Martigny, Switzerland maryam.habibi@idiap.ch Andrei Popescu-Belis Idiap Research Institute Rue Marconi 19, CP 592 1920 Martigny, Switzerland andrei.popescu-belis@idiap.ch Abstract This paper addresses the problem of building concise, diverse and relevant lists of documents, which can be recommended to the participants of a conversation to fulfill their information needs without distracting them. These lists are retrieved periodically by submitting multiple implicit queries derived from the pronounced words. Each query is related to one of the topics identified in the conversation fragment preceding the recommendation, and is submitted to a search engine over the English Wikipedia. We propose in this paper an algorithm for diverse merging of these lists, using a submodular reward function that rewards the topical similarity of documents to the conversation words as well as their diversity. We evaluate the proposed method through crowdsourcing. The results show the superiority of the diverse merging technique over several others which not enforce the diversity of topics. 1 Introduction We present a diverse retrieval technique for ranking documents that are spontaneously retrieved and recommended to people during a conversation. These documents represent potentially useful information for the conversation participants. The information needs of the participants are represented by implicit queries which are built in the background based on their current speech, specifically from keywords obtained from the conversation transcripts. Since people usually mention several topics even during a short conversation span, such keyword sets are made of content words related to different topics. When juxtaposed in an implicit query, these topics may have noisy effects on the retrieval results (Bhogal et al., 2007; Carpineto and Romano, 2012). The purpose of this paper is to present a method for merging lists of documents retrieved through multiple implicit queries prepared for short conversations spans. Several topically-separated queries are constructed from keywords, and generate several lists of documents. The goal of the method proposed here is to generate a unique and concise list of documents that can be recommended in real time to the conversation participants. The list should cover the maximum number of implicit queries and therefore topics. To merge the lists of documents according to these criteria, we use inspiration from extractive text summarization (Lin and Bilmes, 2011; Li et al., 2012) and from our own previous work on diverse keyword extraction (Habibi and Popescu-Belis, 2013). The method proposed here rewards at the same time topic similarity – to select the most relevant documents to the conversation fragment – and topic diversity – to cover the maximum number of implicit queries and therefore topics in a concise and relevant list of recommendations, if more than one topic is discussed in the conversation fragment. Several studies have been previously carried out on merging lists of results in information retrieval. Despite the superficial similarity, …

Content truncated to 500 words