Generating & Recognizing Paraphrases Paraphrases < IJCNLP 2005, Oct. 11th, 2005 > � “Alternative ways to convey the same information” (IWP) Middleware for a wide range of application Exploiting Lexical Conceptual Structure � Generation for Paraphrase Generation � Text simplification ! Carroll et al., 1999 "! Inui et al., 2003 " � Pre- and post-editing for MT ! Shirai et al., 1995 " Atsushi FUJITA (1) , Kentaro INUI (2) , Yuji MATSUMOTO (2) � Recognition (1) Kyoto University � QA ! Hermjakob et al., 2002 "! Takahashi et al., 2004 " (2) Nara Institute of Science and Technology � Multi-document summarization ! Barzilay et al., 2003 " 2 Variety Lexically compositional paraphrases (in Japanese) Lexical paraphrase Paraphrasing of light-verb constructions (LVCs) film - NOM him - DAT impression - ACC to give - ACTIVE Emma burst into tears and he tried to comfort her. ( The film made an impression on him. ) Emma cried and he tried to console her. film - NOM him - ACC to be impressed - CAUSATIVE ! Barzilay et al., 2001 " ( The film impressed him. ) Syntactic paraphrase Locative alteration It was his best suit that John wore to the dance last night. - syntactically regular he - NOM wall - DAT paint - ACC to splay - semantically ( He splayed paint on the wall. ) John wore his best suit to the dance last night. he - NOM wall - ACC paint - with to splay compositional ! Dras, 1999 " Lexically compositional paraphrase ( He splayed the wall with paint. ) Category shifting Steven made an attempt to stop playing Hearts. room - NOM already- ADV to warm- Verb - Passive - Perfective Steven attempted to stop playing Hearts. ( The room has already been warmed up. ) room - NOM already- ADV be warm- Adjective - Present ! Dras, 1999 " ( The room is already warm. ) 3 4 Issues Paraphrasing of LVCs Issue: to explore… LVCs single verb phrases paraphrase � what sorts of lexical properties affect � Syntactic and semantic properties of two verbs interact � how existing framework of lexical semantics can be used head of semantics head of syntax to represent them film - NOM him - DAT impression - ACC to give - ACTIVE Our attempt ( The film made an impression on him. ) � Exploit Lexical Conceptual Structure (LCS) film - NOM him - ACC to be impressed - CAUSATIVE ! Jackendoff, 1990 " ( The film impressed him. ) � Examine current theory and implementation of LCS for price - NOM exchange - DAT influence - ACC to give - ACTIVE Japanese ! Kageyama, 1996 "! Takeuchi et al., 2002 " ( The stock price gives an influence to the foreign exchanges. ) price - NOM exchange - DAT to influence - ACTIVE � Develop an LCS-based paraphrase generation model ( The stock price influences the foreign exchanges. ) � Case study on paraphrasing of LVCs in Japanese 5 6
Contents Overview of LCS Issues and goals What’s LCS? ! Jackendoff, 1990 " 1. LCS � A verb classification which reflects 2. several syntactic and semantic properties of verbs Paraphrasing of LVCs in Japanese 3. � Agentivity: “ '(&% ” (to locate) : Non-agentive LCS-based paraphrase generation model 4. “ )* ” (to play) : Agentive Experiments 5. � Focus of statement: “ +,% ” (to give) : Agent “ -$% ” (to receive) : Goal Conclusion 6. � Link between syntax and semantics (Linking): “ ./&% ” (to transit) : ( NOM , ACC ) = (Theme, Goal) “ #$% ” (to deliver) : ( NOM , ACC , DAT ) = (Agent, Theme, Goal) 7 8 Example Electronic resource and application “ #$% ” (to deliver), “ ! & ” (to translate) English LCS Verb Lexicon � Agentivity: Agentive � 4,163 verbs / 468 LCS types [ x CONTROL [BECOME [ y BE AT z ]]] � Focus: Agent � MT ! Dorr, 1997 "! Habash et al., 2003 " , NLG ! Traum et al., 2000 " ! Takeuchi et al., 2002 " � Linking: Takeuchi’s Japanese LCS dictionary CONTROL (Agent, Theme, Goal) � 1,165 verbs / 16 LCS types Agent = ( NOM , ACC , DAT ) � Compound noun analysis ! Takeuchi et al., 2002 " ! shopper " x ! shopper " x BECOME � Etc. Further projects are running (for Japanese) NOM ! Kato et al., 2005 "! Takeuchi et al., 2005 " BE AT shopper - NOM customer - DAT Theme Goal product - ACC to deliver - ACTIVE ! product " y ! product " y ! customer " z ! customer " z ( The shopper delivers the product to the customer. ) ACC DAT 9 10 Contents Intriguing points Issues and goals 1. DAT is not necessarily changed to ACC “give” != CAUSATIVE LCS 2. film - NOM him - DAT impression - ACC to give - ACTIVE ( The film made an impression on him. ) Paraphrasing of LVCs in Japanese 3. film - NOM him - ACC to be impressed - CAUSATIVE LCS-based paraphrase generation model ( The film impressed him. ) 4. TV - NOM kids - DAT stimulation - ACC to give - ACTIVE Experiments 5. ( TV gives kids stimulation. ) Conclusion 6. TV - NOM kids - ACC to stimulate - ACTIVE ( TV stimulates kids. ) price - NOM exchange - DAT influence - ACC to give - ACTIVE ( The stock price gives an influence to the foreign exchanges. ) price - NOM exchange - DAT to influence - ACTIVE ( The stock price influences the foreign exchanges. ) 11 12
How LCS is expected useful? Contents To determine voice and syntactic cases Issues and goals 1. film - NOM him - ACC to be impressed - CAUSATIVE LCS 2. Syntactic cases Voice Paraphrasing of LVCs in Japanese 3. � Voice: how the event is described ! Muraki, 1991 " LCS-based paraphrase generation model 4. � Who causes the event Agentivity, Focus Experiments 5. � Who is influenced by the event Conclusion 6. � Syntactic cases: � Which marker should be assigned for each nominal element Linking 13 14 LCS as lexical constraints / a tool for transfer Step 1: Semantic analysis BECOME BECOME Surface 01 LCS (light-verb) BE WITH BE WITH BECOME ACT ON ! Ken " z ! Ken " z MOVE FROM TO 2. Semantic transfer NOM ! film " x ! Ken " y BE WITH ! inspiration " y ! film " x ! Ken " z (2) NOM ACC ACC DAT 1. Semantic analysis 3. Surface generation ! Ken " z ! Ken " z MOVE FROM TO Syntactic transfer NOM to receive to inspire to inspire -??????? -PASSIVE ! inspiration " y ! film " x ! inspiration " y ! film " x ! Ken " z ! Ken " z Ken- NOM inspiration- ACC Ken- NOM Ken-??? ACC DAT film- DAT film-??? film- DAT Ken - NOM film - DAT (to receive) (1) inspiration - ACC to receive - ACTIVE Ken - NOM film - DAT inspiration - ACC to receive - ACTIVE Ken - NOM film - DAT to inspire - PASSIVE Ken - ??? film - ??? to inspire - ??????? ( Ken received an inspiration from the film. ) LCS dic. 15 16 Step 2: Semantic transfer Step 3: Surface generation LCS (light-verb) 01 LCS (deverbal noun) BECOME LCS (deverbal noun) 01 Surface Focus of “ ukeru ” BECOME (3) BECOME BE WITH Ken - NOM film - DAT to inspire - PASSIVE Ken - NOM film - DAT to inspire - PASSIVE Compatible predicate classes BE WITH BE WITH ( Ken was inspired by the film. ) MOVE FROM TO 21 ACT ON Alteration rule: ! Ken " z if (Focus != Agent) (2) ! Ken " z MOVE FROM TO ACT ON ! Ken " z then passivize (1) film - NOM Ken - ACC to inspire - ACTIVE ACT ON NOM ! inspiration " y ! film " x ! Ken " z ! film " x ! film " x ! Ken " y ! Ken " y ( The film inspired him. ) NOM ACC ACC DAT ! film " x ! Ken " y (2) NOM ACC (1) (to inspire) (0) Argument matching rules LCS dic. 17 18
Recommend
More recommend
