Improving Related Entity Finding via Incorporating Homepages and Recognizing Fine-grained Entities

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 174–182, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Improving Related Entity Finding via Incorporating Homepages and Recognizing Fine-grained Entities Youzheng Wu Chiori Hori Hisashi Kawai Hideki Kashioka Spoken Language Communication Group,MASTAR Project National Institute of Information and Communications Technology (NiCT) 2-2-2 Hikaridai, Keihanna Science City, Kyoto 619-0288, Japan {youzheng.wu, chiori.hori, hisashi.kawai, hideki.kashioka}@nict.go.jp Abstract This paper describes experiments on the TREC entity track that studies retrieval of homepages representing entities rele- vant to a query. Many studies have fo- cused on extracting entities that match the given coarse-grained types such as or- ganizations, persons, locations by using a named entity recognizer, and employ- ing language model techniques to calcu- late similarities between query and sup- porting snippets of entities from which entities are extracted to rank the entities. This paper proposes three improvements over baseline, i.e., 1) incorporating home- pages of entities to supplement support- ing snippets, 2) recognizing fine-grained named entities to filter out or negatively reward extracted entities that do not match the specified fine-grained types of entities such as a university, airline, author, and 3) adopting a dependency tree-based sim- ilarity method to improve language model techniques. Our experiments demonstrate that the proposed approaches can signifi- cantly improve performance, for instance, the absolute improvements of nDCG@R and P@1 scores are 8.4%, and 27.5%. 1 Introduction Many user information needs would be better an- swered by presenting a ranked list of entities di- rectly, instead of just a list of relevant documents. Based on this assumption, increasing attention has been devoted to related entity finding tasks that aimed at finding documents representing entities of a correct type that are relevant to a query. The TREC expert finding track (Nick, 2005), for exam- ple, focused on creating an ordered list of experts who have skills and experiments on a given topic. The INEX entity ranking task (Vries, 2007) stud- ied at ranking Wikipedia entities given a query, in which target entity types are shifted from a single type of entity (person) to any Wikipedia category. The TREC related entity finding (REF) track (Ba- log, 2010) started in 2009, is defined as: Given an input entity, by its name and homepage, the type of the target entity1, as well as the nature of their relation, described in free text, find related enti- ties that are of a target type, standing in the re- quired relation to the input entity. The REF task is also similar to a combination of the TREC list QA (Voorhees, 2003) and homepage finding (Hawk- ing, 2001) tasks. In short, all these entity finding tasks generally aim at performing entity-oriented search tasks on the Web. This paper is concerned with the TREC REF track. Figure 1 shows an ex- ample of this. <query> <num>7</num> <entity_name>Boeing 747</entity_name> <entity_URL>clueweb09-en0005-75-02292</entity_URL> <target_entity>organization</target_entity> <narrative>Airlines that currently use Boeing 747 planes.</narrative> </query> Figure 1: Test query in TREC 2009 entity track. The key challenge in the REF task …

Content truncated to 500 words