Entity Retrieval via Entity Factoid Hierarchy

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 514–523, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Entity Retrieval via Entity Factoid Hierarchy∗ Chunliang Lu, Wai Lam, Yi Liao Key Laboratory of High Confidence Software Technologies Ministry of Education (CUHK Sub-Lab) Department of Systems Engineering and Engineering Management The Chinese University of Hong Kong {cllu,wlam,yliao}@se.cuhk.edu.hk Abstract We propose that entity queries are gener- ated via a two-step process: users first se- lect entity facts that can distinguish tar- get entities from the others; and then choose words to describe each selected fact. Based on this query generation paradigm, we propose a new entity rep- resentation model named as entity fac- toid hierarchy. An entity factoid hierar- chy is a tree structure composed of fac- toid nodes. A factoid node describes one or more facts about the entity in different information granularities. The entity fac- toid hierarchy is constructed via a factor graph model, and the inference on the fac- tor graph is achieved by a modified variant of Multiple-try Metropolis algorithm. En- tity retrieval is performed by decompos- ing entity queries and computing the query likelihood on the entity factoid hierarchy. Using an array of benchmark datasets, we demonstrate that our proposed framework significantly improves the retrieval perfor- mance over existing models. 1 Introduction Entity retrieval, which aims at returning specific entities to directly answer a user’s query, has drawn much attention these years. Various entity retrieval tasks have been proposed, such as TREC Entity (Balog et al., 2012; Wang et al., 2011) and INEX-LD (Wang et al., 2012; Wang and Kang, 2012). Many existing entity retrieval models fol- low the document retrieval assumption: when is- suing queries, users choose the words that may ∗ The work described in this paper is substantially sup- ported by grants from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project Codes: 413510 and 14203414) and the Direct Grant of the Faculty of Engineering, CUHK (Project Code: 4055034). appear in the “entity pseudo-document”. Based on the assumption, these models construct in- ternal entity representations by combining vari- ous entity descriptions, and use these representa- tions to compute the rank of the candidate enti- ties for a given entity query. These models in- clude fielded versions of BM25 and Mixture of Language Models (Neumayer et al., 2012), Entity Language Model (Raghavan et al., 2004), Hierar- chical Expert Model (Petkova and Croft, 2006), Structured Positional Entity Language Model (Lu et al., 2013). However, a closer examination of entity queries reveals that most of them are not simple uniform word samples from the “entity pseudo-document”. Instead, they can be decomposed into multiple parts, where each part describes a fact about target entities. For example, the query “National capitals situated on islands” describes two facts regarding a target entity: it is a national capital; it is lo- cated on an island. Compared to the assumption in document …

Content truncated to 500 words