Automatically Extracting new Search Patterns for VerbOcean Kuzman - PowerPoint PPT Presentation

Automatically Extracting new Search Patterns for VerbOcean Kuzman Ganchev 1

Overview • VerbOcean refresher • How patterns are used • What makes a good pattern • How we look for patterns • Results 2

VerbOcean Refresher: relations Find relations between verbs Relation Examples Synonymous Stronger persuade → advise overhaul → amend Similar absolve → vindicate allot → earmark Opposite absolve → vindicate allot → earmark Causes centralize → overhaul privatize → improve Happens Before advise → persuade survive → avenge 3

VerbOcean Refresher: patterns Search google for verb pairs in “patterns”. Relation Pattern Examples Synonymous “X ie Y” “Xed ie Yed” Stronger “X even Y” “Yed or at least Xed” Similar “Xed and Yed” “X and Y” Opposite “either Xed or Yed” “to X * but Y” Causes “Xed by Ying the” “Xed by Ying or” Happens Before “to X and then Y” “Xed and later Yed” “Xed and later Yed” + “buy → sell” = “bought and later sold” Pattern Verb Pair 4

How Patterns are Used � S p ( V 1 , V 2 ) > C p where: hits ( V 1 ,p,V 2 ) N S p ( V 1 , V 2 ) = hits est ( p ) × hits (” toV 1 ”) × C v × hits (” toV 2 ”) × C v N N N for asymmetric relations and hits ( V 1 ,p,V 2 ) hits ( V 2 ,p,V 1 ) N N S p ( V 1 , V 2 ) = 2 × hits est ( p ) × hits (” toV 1 ”) × C v × hits (” toV 2 ”) × C v N N N for symmetric relations. 5

What Makes a Good Pattern For asymmetric relations: S p ( V 1 , V 2 ) = hits ( V 1 , p, V 2 ) arg max hits est ( p ) p For symmetric relations: S p ( V 1 , V 2 ) = hits ( V 1 , p, V 2 ) × hits ( V 2 , p, V 1 ) arg max . hits est ( p ) p 6

How we look for patterns • List of verb pairs in known relations • Wildcard Search: “buy * * sell”, “bought * * sold” • Extract patterns: “X and then Y”, “Xed and later Yed” • Rank patterns. 7

Ranking Patterns: ideas • find arg max p � V 1 ,V 2 S p ( V 1 , v 2 ) P V 1 ,V 2 ∈ R hits ( V 1 ,p,V 2 ) • use the odds ratio: arg max p P V 1 ,V 2 �∈ R hits ( V 1 ,p,V 2 ) 8

Training data 100 verb-pairs (used in testing VerbOcean ). Relation Occurances Examples Synonymous 0 persuade → advise overhaul → amend Stronger 19 analyse → monitor devastate → batter doom → complicate erode → impact Similar 26 absolve → vindicate allot → earmark avenge → survive bellow → scream bobble → throw commute → reverse Opposite 6 absolve → vindicate allot → earmark avenge → survive bellow → scream bobble → throw commute → reverse Causes 2 centralize → overhaul privatize → improve Happens Before 12 advise → persuade survive → avenge batter → devastate grab → carry commence → suspend decree → reconsider 9

Results 10

Comments • Need lots of verb pairs for this to be viable. • Restricting contribution of each verb pair. • Getting wildcard patterns like “Xed * by Ying the” 11