Data-Driven and Ontological Analysis of FrameNet for Natural - PowerPoint PPT Presentation

Data-Driven and Ontological Analysis of FrameNet for Natural Language Reasoning for Natural Language Reasoning EkaterinaOvchinnikova 1 , Laure Vieu 2,3 , Alessandro Oltramari 2 , Stefano Borgo 2 , Theodore Alexandrov 4 1 University of Osnabrück, 2 LOA-ISTC-CNR Trento, 3 IRIT-CNRS Toulouse, 4 University of Bremen May, 20th - LREC, Valetta

Introduction � Lexical- semantic knowledge for reasoning � WordNet [ Morato et al., 2004 ] • search • information extraction • information extraction • … � FrameNet • question answering [ Shen and Lapata, 2007 ] • recognizing textual entailment [ Burchardt et al., 2009 ] • …

Introduction � Shortcomings of FrameNet with regard to NL reasoning � low coverage [ Shen and Lapata, 2007; Cao et al., 2008 ] � conceptual inconsistency and lack of axiomatization Our focus methodology for improving the conceptual structure of FrameNet for the goals of NL reasoning

Outline FrameNet for Reasoning 1. Proposed Methodology 2. Conceptual Problems 3. Data-Driven Analysis 4. OntologicalAnalysis 5. Case Study 6. Conclusion 7.

Outline FrameNet for Reasoning 1. Proposed Methodology 2. Conceptual Problems 3. Data-Driven Analysis 4. OntologicalAnalysis 5. Case Study 6. Conclusion 7.

FrameNet for reasoning

FrameNet for reasoning (a) [ John ] DONOR [ gave ] Giving [ Mary ] RECIPIENT [ a flower ] THEME (b) [ Mary ] RECIPIENT [ got ] Getting [ a flower ] THEME [ from John ] SOURCE causes Giving Getting DONOR SOURCE RECIPIENT RECIPIENT THEME THEME

Frame relations Inheritance: 441 1. � Vehicle – Arfitact, Motion_directional - Motion Precedence: 55 2. � Being_awake – Fall_asleep Perspective: 43 3. � Buy, Sell – Goods_transfer Buy, Sell – Goods_transfer Causation: 49 4. � Giving - Getting Subframe: 87 5. � Trial, Sentencing – Criminal_process Using: 426 6. � Recovery – Medical_conditions See_also: 669 7. � Scrunity - Seeking

Research goals Axiomatizing frame relations 1. Finding missing frame relations 2. Cleaning up frame relations 3. Applying frame relations to NL reasoning 4.

Outline FrameNet for Reasoning 1. Proposed Methodology 2. Conceptual Problems 3. Data-Driven Analysis 4. OntologicalAnalysis 5. Case Study 6. Conclusion 7.

Proposed improvement methodology Conceptual problems in FrameNet : 1. Frame-Annotated Corpus for Textual Entailment (FATE) Clustering frames 2. Ontological analysis of frames and frame relations Ontological analysis of frames and frame relations 3. axiomatizing frame relations � constraints on frame relations � Evaluation: enriched, axiomatized and cleaned up 4. frame relations in RTE

Outline FrameNet for Reasoning 1. Proposed Methodology 2. Conceptual Problems 3. Data-Driven Analysis 4. OntologicalAnalysis 5. Case Study 6. Conclusion 7.

Frame-AnnotatedCorpus forTextual Entailment FATE [ Burchardt & Pennacchiotti, 2008 ] � 800 T-H entailment pairs annotated with FrameNet frames and roles � we have analized cases when T was known to entail H (400 pairs) applying a frame matching strategy

FATE analysis results � 170 pairs: matching is possible � 131 pairs: this approach does not work � annotation disagreements � different conceptualizations of T and H � different conceptualizations of T and H � 99 pairs: the same facts in T and H are represented by different frames which are related semantically and could be mapped on each other with the help of reasoning � FrameNet enables inferences only for 17 pairs

Discovered problems missing relations 1. (t) … X [ survived ] Surviving Sars … (h) … X [ recovered ] Recovery from Sars … problems in the relational structure 2. …[ parts ] Part_whole [ of Aceh province ] WHOLE … Part_whole → → Part_piece , WHOLE → → SUBSTANCE → → → → missing axiomatization of the relations 3. (t) … X [ recovered ] Recovery from Sars… (h ) … X [ was ill ] Medical_conditions … Recovery uses Medical_conditions

Outline FrameNet for Reasoning 1. Proposed Methodology 2. Conceptual Problems 3. Data-DrivenAnalysis 4. OntologicalAnalysis 5. Case Study 6. Conclusion 7.

Clustering frames For every two frames f1 and f2 we apply similarity measures based on [ Pennacchiotti & Wirdth, 2009 ] : overlapping frame elements in f1 and f2 overlapping frame elements in f1 and f2 1. 1. co-occurrence of lexemes evoking f1 and f2 in corpora 2. (pmi)

Clustering results Clusters based on overlapping frame elements 1. 228 clusters in total � 1497 relations not contained in FrameNet � 73 clusters from 100 random contain semantically related � frames frames (2 experts, agreement 0.85) (2 experts, agreement 0.85) Clusters based on co-occurence of lexemes evoking 2. frames � 113 clusters in total 1149 relations not contained in FrameNet � 65 clusters from 100 random contain semantically related � frames (2 experts, agreement 0.85)

Frame clusters: visualization ( Pajek tool )

Outline FrameNet for Reasoning 1. Proposed Methodology 2. Conceptual Problems 3. Data-Driven Analysis 4. Ontological Analysis 5. Case Study 6. Conclusion 7.

Frames and situations What do frames describe? � Frames abstract from natural language expressions (predicates with their arguments) (predicates with their arguments) � Natural language expressions describe situations � Frames can be seen as abstractions from situations

Types of situations From which types of situations do frames abstract? � categories from the DOLCE ontology [ Masolo et al.,2002 ] for describing types of situations Types of situations: Types of situations: „Event“ situation 1. • e.g. Motion ( John is running in the park ) „Object“ situation 2. e.g. People ( A man ) • „Quality“ situation 3. • e.g. Color ( This rose is red ) „Relation“ situation 4. • e.g. Part_whole ( This park is a part of the town )

Situations and time Situations having temporal qualities 1. John is running in the park , a clerk , This rose is red , John is � next to Mary can participate in temporal relations ( precedence , temporal � inclusion etc.) inclusion etc.) Non-temporal situations 2. A man , The war lasted four years , Einstein‘s birth preceded my � birth � cannot participate in temporal relation

Causation: f 1 is causative of f 2 ∀ s 1 ( f 1 (s 1 ) → ∃ s 2 ( f 2 (s 2 ) ∧ causes (s 1 ,s 2 ))) ∀ s 1 s 2 ( causes (s 1 ,s 2 ) → ¬ starts_before (s 2 , s 1 ))) ∀ s 1 s 2 ( causes (s 1 ,s 2 ) → ¬ starts_before (s 2 , s 1 )))

Subframe: f 1 is subframe of f 2 Subframe of “Events” 1. ∀ s 1 s 2 ( sub_ev (s 1 ,s 2 ) → ( strict_temp_inc (s 2 ,s 1 ) ∧ spatially_inc (s 2 ,s 1 ))) part presupposes whole • ∀ s 1 ( f 1 (s 1 ) → ∃ s 2 ( f 2 (s 2 ) ∧ sub_ev (s 1 ,s 2 ))) whole presupposes part whole presupposes part • • ∀ s 2 ( f 2 (s 2 ) → ∃ s 1 ( f 1 (s 1 ) ∧ sub_ev (s 1 ,s 2 ))) Subframe of “Objects” 2. part presupposes whole • ∀ s 1 en 1 ( f 1 (s 1 ) ∧ FE 1 (s 1 , en 1 ) → ∃ s 2 en 2 ( f 2 (s 2 ) ∧ FE 2 (s 2 , en 2 ) ∧ part_of (en 1 ,en 2 ))) whole presupposes part • ∀ s 2 en 2 ( f 2 (s 2 ) ∧ FE 2 (s 2 , en 2 ) → ∃ s 1 en 1 ( f 1 (s 1 ) ∧ FE 1 (s 1 , en 1 ) ∧ part_of (en 1 ,en 2 )))

Using and See_also � the most frequent relations in FN � sometimes can be represented in terms of other axiomatized relations � otherwise � otherwise ∀ s 1 ( f 1 (s 1 ) → ∃ s 2 ( f 2 (s 2 ) ∧ depends (s 1 ,s 2 ))) � often represent typical rather than necessary dependence (e.g. Medical_professionals - Cure )

Mapping frame elements If f 1 is related to f 2 with a relation in FN then ∀ s 1 s 2 (( f 1 (s 1 ) ∧ f 2 (s 2 )) → ( rel (s 1 ,s 2 ) ↔ ∀ x(FE 1 (x,s 1 ) ↔ FE 2 (x,s 2 ))), ( rel (s 1 ,s 2 ) ↔ ∀ x(FE 1 (x,s 1 ) ↔ FE 2 (x,s 2 ))), where FE 1 in f 1 is mapped to FE 2 in f 2 .

Example ∀ s 1 ( Giving (s 1 ) → ∃ s 2 ( Getting (s 2 ) ∧ causes (s 1 ,s 2 ))) ∀ s 1 s 2 (( Giving (s 1 ) ∧ Getting (s 2 )) → ∀ s 1 s 2 (( Giving (s 1 ) ∧ Getting (s 2 )) → ( causes (s 1 ,s 2 ) ↔ ∀ x( DONOR (x,s 1 ) ↔ SOURCE (x,s 2 )))

Cleaning up constraints Given frames f 1 and f 2 connected with a relation r define the types of situations that instantiate f 1 and f 2 1. if r is a temporal relation, make sure that both f 1 and f 2 2. refer to „temporal“ situations refer to „temporal“ situations define whether r has a typical or a necessary character 3. check whether the frame relation axioms apply to all 4. instantiations of f 1 and f 2

Outline FrameNet for Reasoning 1. Proposed Methodology 2. Conceptual Problems in FrameNet 3. Data-Driven Analysis of FrameNet 4. OntologicalAnalysis of FrameNet 5. Case Study 6. Conclusion 7.

Case Study: „medical cluster“

Enriched and cleaned up „medical“ cluster