Anthony Fader, Stephen Soderland, and Oren Etzioni
Transcription
Anthony Fader, Stephen Soderland, and Oren Etzioni
Identifying Relations for Open Information Extraction Anthony Fader, Stephen Soderland, and Oren Etzioni 2011 TextRunner TR2 TR3 WOE ReVerb OLLIE 2007 2008 2009 2010 2011 2012 Agenda • • • Solution overview Solution details Evaluation Why? WOE and TextRunner extractions are incoherent, uninformative Why? What? WOE and TextRunner extractions are incoherent, uninformative Improve the quality Why? What? WOE and TextRunner extractions are incoherent, uninformative Improve the quality How? Add constraints both syntactic and lexical Why? WOE and TextRunner extractions are incoherent, uninformative Samples Incoherentextraction phrase has no meaningful interpretation Why? WOE and TextRunner extractions are incoherent, uninformative Samples Incoherentextraction phrase has no meaningful interpretation { The Extractor makes a decision about each word separately } up to 13-30% of output are incoherent Why? WOE and TextRunner extractions are incoherent, uninformative How? Syntactic constraint: multiword relation 1. begins with a verb; 2. end with a preposition; 3. is a contiguous sequence of words in a sentence; {… made a deal with …} Samples Incoherentextraction phrase has no meaningful interpretation { The Extractor makes a decision about each word separately } up to 13-30% of output are incoherent Why? WOE and TextRunner extractions are incoherent, uninformative Samples Uninformative extraction that omit critical information Why? WOE and TextRunner extractions are incoherent, uninformative Samples Uninformative extraction that omit critical information { The Extractor handles improperly verb-noun relation phrases (LVC) } 4-7% are uninformative Ex. Faust made a deal with the devil. (Faust, made, a deal)-(Faust, made a deal with, the devil) Why? WOE and TextRunner extractions are incoherent, uninformative How? Samples Uninformative extraction that omit critical information { The Extractor handles improperly verb-noun relation phrases (LVC) } 4-7% are uninformative Syntactic constraint: multiword relation 1. begins with a verb; 2. end with a preposition; 3. is a contiguous sequence of words in a sentence; {… made a deal with …} Ex. Faust made a deal with the devil. (Faust, made, a deal)-(Faust, made a deal with, the devil) Demo time Are we perfect now? No, overly specific relation phrases Example The Obama administration is offering only modest greenhouse gas reduction targets at the conference No, overly specific relation phrases Example [The Obama administration] is offering only modest greenhouse gas reduction targets at [the conference] No, overly specific relation phrases Example [The Obama administration] {is offering only modest greenhouse gas reduction targets at } [the conference] How: what is inside? Constraints Syntactic How: what is inside? Constraints Syntactic How: what is inside? Constraints Syntactic Lexical How: what is inside? Constraints Syntactic Lexical Valid relation phrase should take many distinct arguments in a large corpus How: what is inside? Constraints [ arg1 - rel - arg2 ] common relation for OIE Syntactic Lexical Valid relation phrase should take many distinct arguments in a large corpus How: what is inside? Constraints [ arg1 - rel - arg2 ] common relation for OIE Syntactic Lexical Valid relation phrase should take many distinct arguments in a large corpus Extendicare agreed to buy Arbor Health Care for about US $432 million in cash and assumed debt. How: what is inside? Constraints [ arg1 - rel - arg2 ] common relation for OIE Syntactic Lexical Valid relation phrase should take many distinct arguments in a large corpus “Extendicare agreed to buy Arbor Health Care for about US $432 million in cash and assumed debt.” TextRunner output: (Arbor Health Care, for assumed, debt). First evaluation: how much do we lose? Loosing recall First test set - Random web pages 300 sentences 327 verb relation phrases First evaluation: how much do we lose? Loosing recall First test set - Random web pages 300 sentences 327 verb relation phrases ? First evaluation: how much do we lose? Loosing recall First test set - Random web pages 300 sentences 327 verb relation phrases dependency parsers are still slow on web-scale ReVerb Sentence POSed (X1; R1; Y1) (X2; R2; Y2) … (Xn; Rn; Yn) ReVerb 1. Relation extraction Sentence POSed 2. Argument extraction (X1; R1; Y1) (X2; R2; Y2) … (Xn; Rn; Yn) ReVerb Sentence POSed 1. Relation extraction (find a verb-> expand it satisfying constraints) 2. Argument extraction (find the nearest on the left, find the nearest on the right) (X1; R1; Y1) (X2; R2; Y2) … (Xn; Rn; Yn) ReVerb Sentence POSed 1. Relation extraction (find a verb-> expand it satisfying constraints) 2. Argument extraction (find the nearest on the left, find the nearest on the right) (X1; R1; Y1) (X2; R2; Y2) … (Xn; Rn; Yn) We talk about Open Information Extraction PRP VBP IN NNP NNP NNP B-NP B-VP B-PP B-NP I-NP I-NP ReVerb Sentence POSed 1. Relation extraction (find a verb-> expand it satisfying constraints) 2. Argument extraction (find the nearest on the left, find the nearest on the right) (X1; R1; Y1) (X2; R2; Y2) … (Xn; Rn; Yn) We [talk] about Open Information Extraction PRP [VBP ] IN NNP NNP NNP B-NP [B-VP] B-PP B-NP I-NP I-NP ReVerb Sentence POSed 1. Relation extraction (find a verb-> expand it satisfying constraints) 2. Argument extraction (find the nearest on the left, find the nearest on the right) (X1; R1; Y1) (X2; R2; Y2) … (Xn; Rn; Yn) We [talk about] Open Information Extraction PRP [VBP IN ] NNP NNP NNP B-NP [B-VP B-PP ] B-NP I-NP I-NP ReVerb Sentence POSed 1. Relation extraction (find a verb-> expand it satisfying constraints) 2. Argument extraction (find the nearest on the left, find the nearest on the right) (X1; R1; Y1) (X2; R2; Y2) … (Xn; Rn; Yn) {We } [talk about] Open Information Extraction {PRP } [VBP IN ] NNP NNP NNP {B-NP} [B-VP B-PP ] B-NP I-NP I-NP ReVerb Sentence POSed 1. Relation extraction (find a verb-> expand it satisfying constraints) 2. Argument extraction (find the nearest on the left, find the nearest on the right) (X1; R1; Y1) (X2; R2; Y2) … (Xn; Rn; Yn) {We } [talk about] {Open Information Extraction} {PRP } [VBP IN ] {NNP NNP NNP } {B-NP} [B-VP B-PP ] {B-NP I-NP I-NP } ReVerb Sentence POSed 1. Relation extraction (find a verb-> expand it satisfying constraints) 2. Argument extraction (find the nearest on the left, find the nearest on the right) (X1; R1; Y1) (X2; R2; Y2) … (Xn; Rn; Yn) {We } [talk about] {Open Information Extraction} {PRP } [VBP IN ] {NNP NNP NNP } {B-NP} [B-VP B-PP ] {B-NP I-NP I-NP } (‘we’; ’talk about’; ‘open information extraction’) But! Still low precision even though recall is comparatively high But! Still low precision even though recall is comparatively high ? Confidence function train conf. function based on Web+Wiki, all extractions from 1000 sent. Confidence function train conf. function based on Web+Wiki, all extractions from 1000 sent. Confidence function train conf. function based on Web+Wiki, all extractions from 1000 sent. Evaluation results 500 sent. from Web 2 judges, 0.68 agr. Evaluation results 500 sent. from Web 2 judges, 0.68 agr. boost over lex flow TR & WOE 1. 2. 3. Auto Labeling the sentence An extractor is learned using sequence labeling graphical model extraction: arguments, label the relation between args as part of relation phrase ReVerb 1. 2. Relation extraction Argument extraction Evaluation results Achievements Achievements Incoherent extractions elimination if much better Achievements Incoherent extractions elimination if much better Outperforming precision-recall Achievements Incoherent extractions elimination if much better Outperforming precision-recall Faster ReVerb Error analysis Possible improvements Precision ReVerb Error analysis Possible improvements Recall Thank you Open Language Learning for Information Extraction Mausam, Michael Schmitz, Robert Bart, Stephen Soderland, and Oren Etzioni TextRunner TR2 TR3 WOE ReVerb OLLIE 2007 2008 2009 2010 2011 2012 2 Agenda Introduction Relation Extraction Context Analysis Evaluation 3 why? Reverb and WOE V | V P | V W*P Only for verbs OLLIE Nouns, adjectives, and more 4 why? Reverb and WOE V | V P | V W*P Only for verbs No context is taken into account OLLIE Nouns, adjectives, and more Including contextual information 5 why? Reverb and WOE V | V P | V W*P Only for verbs No context is taken into account OLLIE Nouns, adjectives, and more Including contextual information XXXX XXXX Not factual extractions 6 why? Reverb and WOE V | V P | V W*P Only for verbs No context is taken into account The last version (Nov 2015) doesn’t give these results OLLIE Nouns, adjectives, and more Including contextual information XXXX XXXX Not factual extractions 7 Introduction + conf.function 8 Bootstrapping set Goal Create large training set Hypothesis Every relation can be expressed in Reverb style Sentences express original tuple 110th seed tuples from ReVerb from ClueWeb (Students,build, bootstrap set) Extract all sentences with the same content words … (Bootstrap set is built by students) (while working on OIE, students built the set) 9 Bootstrapping set Goal Create large training set Hypothesis Every relation can be expressed in Reverb style Sentences express original tuple 110th seed tuples from ReVerb from ClueWeb (Students,build, bootstrap set) Extract all sentences with the same content words Bootstrapping error reduction (Bootstrap set is built by students) (while working on OIE, students built the set) (students worked on a set of tasks, workers built a new cafe on the campus) linear path size<5 10 Open Pattern Learning Goal Learn general patterns that encode diff types of relations 11 Open Pattern Learning Goal Learn general patterns that encode diff types of relations Open pattern templates dep path Sample for 2.: open extraction We do interesting things. 12 Open Pattern Learning Goal Learn general patterns that encode diff types of relations Open pattern templates dep path open extraction But what are syntactic and semantic patterns? 13 Open Pattern Learning Purely syntactic patterns There are no slot nodes in the path Relation node in bw/ (arg1, arg2) The prep edge in the pattern matches the prep in relation Path has no nn and amor edges 14 Open Pattern Learning Purely syntactic patterns Semantic/lexical patterns There are no slot nodes in the path together with words/types the pattern is used with Relation node in bw/ (arg1, arg2) (generalize words into types The prep edge in the pattern matches if it’s possible) the prep in relation Path has no nn and amor edges 15 Open Pattern Learning Purely syntactic patterns Semantic/lexical patterns There are no slot nodes in the path together with words/types the pattern is used with Relation node in bw/ (arg1, arg2) (generalize words into types The prep edge in the pattern matches if it’s possible) the prep in relation Path has no nn and amor edges 16 Semantic/lexical patterns Example 17 Context analysis Conditional truth Attribution ClausalModifier AttributeTo advcl dep edge only if we meet {if, then, although} ccomp dep edge match context verb with list of comm/cogn verbs from VerbNet 18 Comparison 19 20 21 22 Thank you 23