Target Type Identification for Entity-Bearing Queries

Date: August 8, 2017
Venue: Tokyo, Japan. The 40th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '17)
Corresponding article: https://arxiv.org/abs/1705.06056

Please cite the paper, and link to or credit this presentation when using it or part of it in your work.

#InformationRetrieval #IR #EntityRanking #EntityRetrieval #ER #EntityTypes #EntityOrientedSearch #KnowledgeBases #SemanticSearch

Transcript

1. CIM M A X (t,R) CIM A t t R (t,R) CIM A Vt (t,R) TC B M 15 (t,R) E N TITIE C(t) E C B M 15,1(( (t,R) E C B M 15,5( (t,R) CIB UIN t C(t) E C B M 15,1( (t,R) C I IU; R E N (t) I; BCJM (t) I; BA Vt (t) ON O JN C(t,R) E C B M 15,1( (t,R) OTE R M C 1 (t,R) E C B M 15,5 (t,R) ; E PTI (t) E C UM ,1(( (t,R) UE N t TI (t) E C UM ,5( (t,R) TC UM (t,R) E C UM ,1( (t,R) E C UM ,1( (t,R) E C UM ,5 (t,R) OTE R M C 1 (t,R) FeDtures 0.0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 Gini sFore 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 1DCG@5 Gini sFore 1DCG@5 TARGET TYPE IDENTIFICATION FOR ENTITY-BEARING QUERIES Darío Garigliotti Faegheh Hasibi Krisztian Balog University of Stavanger Norwegian University of University of Stavanger bit.ly/sigir2017-querytypes We thank SIGIR for the Student Travel Grant Thing Agent Organization Company … Mean of Transportation Automobile … … … … Type taxonomy Query finland car industry manufacturer saab sisu Ranked target types Company 0.57 Automobile 0.42 Ranked relevant entities Sisu Auto Automotive industry by country Saab Automobile Saab V8 Patria (company) . . . We propose a supervised learning approach for automatically identifying the target types of a query. It substantially outperforms existing methods, when evaluated with a purpose-built test collection. Science and Technology References: 
 [1] Krisztian Balog and Robert Neumayer. Hierarchical Target Type
 Identification for Entity-oriented Queries. In Proc. of CIKM’12. Target type identification performance ,1EX_LD LLst6eDrch 4ALD2 6em6eDrch_E6 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 1DCG@5 EC, L0 7C, L0 L75 RESULTS • Baseline methods: entity-centric (EC) and type-centric (TC) [1], each using BM25 or LM • Our approach significantly and substantially outperforms all baselines - Relative improvement over 43% on any metric, with p < 0.001 - Robust across different query categories Method EC, LM TC, LM LTR NDCG@1 0.1417 0.2341 0.4842 NDCG@5 0.3161 0.3780 0.6355 Semantic similarities Baseline signals Knowledge base features Type label statistics Lexical similarities CONTRIBUTIONS • A Learning-to-Rank approach with a rich set of features • A purpose-built test collection: 485 queries annotated with DBpedia types via Crowdsourcing - Job: Select the single most specific type, if any, that can cover all results the query asks for FEATURE ANALYSIS • Query-type semantic similarities, enriched with distributional representations, are the most effective features Performance break-down by query category