360 likes | 455 Views
Natural Language Questions for the Web of Data. Mohamed Yahya, Klaus Berberich, Gerhard Weikum Max Planck Institute for Informatics, Germany Shady Elbassuoni Qatar Computing Research Institute Maya Ramanath Dept. of CSE, IIT-Delhi, India Volker Tresp
E N D
Natural Language Questionsfor the Web of Data Mohamed Yahya, Klaus Berberich, Gerhard Weikum Max Planck Institute for Informatics, Germany Shady Elbassuoni Qatar Computing Research Institute Maya Ramanath Dept. of CSE, IIT-Delhi, India Volker Tresp Siemens AG, Corporate Technology, Munich, Germany EMNLP 2012
Introduction • Natural language question qNL • “Which female actor played in Casablanca and is • married to a writer who was born in Rome?”. • SPARQL • ?x hasGender female ?x isa actor • ?x actedIn Casablanca (film) ?x marriedTo ?w • ?w isa writer ?w bornIn Rome Problem : SPARQL is way too difficlut. Target : Convert SPARQL to qNL.
Knowledge Base: Yago2 • Yago2is a huge semantic knowledge base, derived from Wikipedia, WordNet and GeoNames.
Framework • DEANNA (DEep Answers for maNy Naturally Asked questions)
Framework Phrase Detection Phrase Mapping Q-Unit Generation Disambiguation of Phrase Mappings Query Generation
Phrase Detection • A detected phrase p is a pair < Toks, l > where Toks is a phrase and l is a label, l∈ {concept, relation}. • Such as • qNL : “Which female actor played inCasablanca and is • married toa writer who was born in Rome?”. • Concept phrase : < a writer, concept> • < Casablanca, concept> • Relation phrase : < played in, relation > • < is married to, relation > Framework
Phrase Detection Conceptdetection • Using the Yago2 knowledge base. Framework
Phrase Detection Relation detection • Using the ReVerb (Fader et al., 2011) which is a relation detector. qNL : “Which female actor played in Casablanca and is married to a writer who was born in Rome?”. Framework
Phrase Detection Framework
Phrase Mapping • The mapping of concept phrases also relies on the phrase-concept dictionary. : Using Yago2 knowledge base. • The mapping of relation phrases relies on a corpus of • textual patterns to relation mappings of the form. Framework
Phrase Mapping Framework
Dependency Parsing & Q-Unit Generation Dependency Parsing • Dependency parsing identifies triples of tokens, • or triploids, <trel, targ1, targ2>, where • trel, targ1, targ2 ∈ qNL are seeds for phrases. Framework
Dependency Parsing & Q-Unit Generation • qNL : “Which female actorplayed inCasablanca and is • married to a writer who was born in Rome?”. played / played in actor Casablanca • Triploid : < played in, actor, Casablanca > < played, actor, Casablanca > Framework
Dependency Parsing & Q-Unit Generation Q-Unit Generation • A q-unit is a triple of sets of phrases, • <{prel ∈ Pr}, {parg1 ∈ Pc}, {parg2 ∈ Pc}>, where • trel ∈ prel and similarly for arg1 and arg2. Framework
Dependency Parsing & Q-Unit Generation Framework
Dependency Parsing & Q-Unit Generation Framework
Disambiguation of Phrase Mappings • Disambiguation Graph • Esim⊆ Vp × Vs • Ecoh⊆ Vs × Vs • Eq⊆ Vq×Vp×d, where d ∈ {rel, arg1, arg2} is a q-edge. Framework
Disambiguation of Phrase Mappings Disambiguation Graph(Cohsem) • For Yago2, the characterize an entity e by its inlinks InLinks(e): the set of Yago2 entities whose corresponding Wikipedia pages link to the entity. InLinks(Taipei_zoo): Framework
Disambiguation of Phrase Mappings Disambiguation Graph(Cohsem) • For classc with entities e, its inlinks are defined as follows: InLinks(Taiwan): Framework
Disambiguation of Phrase Mappings Disambiguation Graph(Cohsem) • For classr with entities e, its inlinks are defined as follows: Framework
Disambiguation of Phrase Mappings Disambiguation Graph(Simsem) • For entities • How often a phrase refers to a certain entity in Wikipedia. • For classes • Normalized prior the reflects the Number of members in a class • For relations • The maximumn-gram similarity between the phrase and any of the • relation’s surface forms Framework
Disambiguation of Phrase Mappings • Objective function is : Framework
Disambiguation of Phrase Mappings Definitions: Framework
Disambiguation of Phrase Mappings Definitions: Framework
Disambiguation of Phrase Mappings Constraints: Framework
Disambiguation of Phrase Mappings Constraints: Framework
Disambiguation of Phrase Mappings Constraints: Framework
Query Generation Framework
Evaluation • Experiments are based on two collections of questions: • QALD-1 (27 questions out of 50) • NAGA (44 questions out of 87) • Using 19 questions from the QALD-1 Test set for tuning hyperparameters (α, β, γ) in the ILP objective function
Evaluation • Evaluating the output of DEANNA at three stages in the • processing pipeline: • a) Disambiguation • b) Query Generation • c) Question Answering • At each of the three stages, the output was shown to two human assessors. If the two were in disagreement, then a third person resolved the judgment.
Evaluation • Define coverage and precision as follows:
Evaluation a) Disambiguation
Evaluation b) Query Generation
Evaluation c) Question Answering
Conclusions A method for translating natural-language questions into structured queries.