Semantic Roles & Labeling LING 571 — Deep Processing in NLP November 18, 2019 Shane Steinert-Threlkeld 1
Announcements ● HW7: 89.4 average ● Only common mistake: similarity vs. distance sim(u, v) = 1 - distance(u, v) 2
Questions on HW #8 ● For the mc_similarity portion c 1 , c 2 [ sim resnik ( c 1 , c 2 ) ] ● You should use From Resnik (1999), eq. 2 wsim ( w 1 , w 2 ) = max ● The numbers in the example_output are random. No meaning to them being < 1! ● For the WSD algorithm ( mea culpa ): ● The pseudocode is confusing so: 3
Alternative Resnik WSD Pseudocode Given: input word w 0 and probe words {p 1, …,p n } for p i in {p 1, …,p n } : supported_sense = null most_information = 0.0 for sense w in S ENSES ( w 0 ): for sense p in S ENSES ( p i ): lcs synset = L OWEST C OMMON S UBSUMER ( sense w , sense p ) lcs info = I NFORMATION C ONTENT ( lcs synset ) if lcs info > most_information : most_information = lcs info supported_sense = sense w increment support[ supported_sense ] by most_information 4
Semantic Roles 5
Semantic Analysis ● Full, deep compositional semantics ● Creates full logical form ● Links sentence meaning representation to logical world model representation ● Powerful, expressive, AI-complete ● Domain-specific slot-filling: ● Common in dialog systems, IE tasks ● Narrowly targeted to domain/task ● e.g. ORIGIN_LOC, DESTINATION_LOC, AIRLINE, … ● Often pattern-matching ● Low cost, but lacks generality, richness, etc 6
Semantic Role Labeling ● Typically want to know ● Who did what to whom ● … where , when , and how ● Intermediate level: ● Shallower than full deep composition ● Abstracts away (somewhat) from surface form ● Captures general predicate-argument structure info ● Balance generality and specificity 7
Examples Yesterday Tom chased Jerry Yesterday Jerry was chased by Tom Tom chased Jerry yesterday Jerry was chased yesterday by Tom ● Semantic roles: ● Chaser : Tom ● ChasedThing : Jerry ● TimeOfChasing : yesterday ● Same across all sentence forms 8
Full Event Semantics ● Neo-Davidsonian Style: ● ∃ e Chasing ( e ) ∧ Chaser ( e , Tom ) ∧ ChasedThing ( e , Jerry ) ∧ TimeOfChasing ( e , Yesterday ) ● Same across all examples ● Roles: Chaser , ChasedThing , TimeOfChasing ● Specific to verb “chase” ● a.k.a. “Deep roles” 9
Main Idea ● Extract the semantic roles without doing full semantic parsing ● Easier problem, but still useful for many tasks ● More data ● Better models 10
Issues & Challenges ● How many roles for a language? ● Arbitrary! ● Each verb’s event structure determines sets of roles 11
Issues & Challenges ● How can we acquire these roles? ● Manual construction? ● Some progress on automatic learning ● Mostly successful on limited domains (ATIS, GeoQuery) 12
Issues & Challenges ● Can we capture generalities across verbs/events? ● Not really, each event/role is specific 13
Thematic Roles ● Solution to instantiating a specific role for every verb ● Attempt to capture commonality between roles 14
Thematic Roles ● Describe common semantic roles of verbal arguments ● e.g. subject of break is A GENT ● A GENT : volitional cause ● T HEME : things affected by action ● Enables generalization over surface order of arguments ● John A GENT broke the window T HEME ● The rock I NSTRUMENT broke the window T HEME ● The window T HEME was broken by John A GENT 15
Thematic Roles ● Verbs take different roles ● The break verb could be formed as: ● A GENT /Subject, T HEME /Object ( John broke the window ) ● A GENT /Subject, T HEME /Object, I NSTRUMENT /PP with ( John broke the window with a rock ) ● I NSTRUMENT /Subject, T HEME /Object ( The rock broke the window ) ● T HEME /Subject ( The window was broken ) 16
Thematic Roles ● Thematic grid, Θ -grid, case frame ● Set of thematic role arguments of verb ● subject: A GENT ; Object: T HEME , or ● subject: I NSTR ; Object:T HEME ● Verb/Diathesis Alternations ● Verbs allow different surface realizations of roles ● Doris A GENT gave the book T HEME to Carv G OAL ● Doris A GENT gave Carv G OAL the book T HEME 17
Canonical Roles Thematic Role Example A GENT The waiter spilled the soup E XPERIENCER John has a headache F ORCE The wind blows debris from the mall into our yards. T HEME Only after Benjamin Franklin broke the ice … R ESULT The French government has built a regulation-size baseball diamond… C ONTENT Mona asked “You met Mary Ann at a supermarket?” I NSTRUMENT He turned to poaching catfish, stunning them with a shocking device … B ENEFICIARY Whenever Ann Callahan makes hotel reservations for her boss… S OURCE I flew in from Boston. G OAL I drove to Portland . 18
Thematic Role Issues ● Hard to produce From Levin and Rappaport Hovav 2005: ● Standard set of roles a. John broke the window with a rock. b. The rock broke the window. ● Fragmentation: Often need to make more specific a. Swabha ate the banana with a fork. ● e.g. I NSTRUMENT s can be subject or not b. * The fork ate the banana. ● Standard definition of roles ● Most A GENT s: animate, volitional, sentient, causal ● But not all… e.g.? [Google] Agent found the answer. 19
Thematic Role Issues ● Strategies: ● Generalized semantic roles: P ROTO -A GENT /P ROTO -P ATIENT ● Defined heuristically: PropBank ● Define roles specific to verbs/nouns: FrameNet 20
PropBank ● Sentences annotated with semantic roles ● Penn and Chinese Treebank ● Roles specific to verb sense ● Numbered: Arg 0 , Arg 1 , Arg 2 , … ● Arg 0 : P ROTO -A GENT ; Arg 1 : P ROTO -P ATIENT , etc 21
PropBank ● Arguments >1 are Verb-specific ● e.g. agree.01 ● Arg 0 : Agreer ● Arg 1 : Proposition ● Arg 2 : Other entity agreeing ● Ex1: [ Arg0 The group] agreed [ Arg1 it wouldn’t make an offer] 22
PropBank ● Resources: ● Annotated sentences ● Started w/Penn Treebank ● Now: Google answerbank, SMS, webtext, etc ● Framesets: ● Per-sense inventories of roles, examples ● Span verbs, adjectives, nouns (e.g. event nouns) 23
PropBank ● propbank.github.io ● Recent status: ● 5940 verbs w/8121 framesets ● 1880 adjectives w/2210 framesets ● Continued into OntoNotes ● [CoNLL 2005 and 2012 shared tasks] 24
AMR ● “Abstract Meaning Representation” ● Sentence-level semantic representation ● Nodes: Concepts ● English words; PropBank: predicates; or keywords (‘person’) ● Edges: Relations ● PropBank thematic roles (ARG0-ARG5) ● Others including ‘location,’ ‘name,’ ‘time,’ etc… ● ~100 in total 25
AMR 2 From Liu et. al (2015) ● AMR Bank: (now) ~40K annotated see-01 sentences ARG0 ARG1 ● JAMR parser: 63% F-measure (2015) i dog poss ARG0-of ● Alignments between word spans & graph fragments person run-02 ● Example: “I saw Joe’s dog, which was name location running in the garden.” name garden op1 “Joe” 26
AMR 3 ● Towards full semantic parsing ● “Deeper” than base PropBank, but: ● No real quantification ● No articles ● No real vs. hypothetical events (e.g. “wants to go”) 27
FrameNet (Fillmore et al) ● Key insight: ● Commonalities not just across different sentences w/same verb but across different verbs (and nouns and adjectives) ● PropBank ● [ Arg0 Big Fruit Co.] increased [ Arg1 the price of bananas]. ● [ Arg1 The price of bananas] was increased by [ Arg0 BFCo]. ● [ Arg1 The price of bananas] increased [ Arg2 5%]. ● FrameNet ● [ A TTRIBUTE The price] of [ I TEM bananas] increased [ D IFF 5%]. ● [ A TTRIBUTE The price] of [ I TEM bananas] rose [ D IFF 5%]. ● There has been a [ D IFF 5%] rise in [ A TTRIBUTE the price] of [ I TEM bananas]. 28
FrameNet ● Semantic roles specific to frame ● Frame: script-like structure, roles (frame elements) ● e.g. C HANGE _P OSITION _ ON _S CALE : increase, rise ● A TTRIBUTE ; I NITIAL _V ALUE ; F INAL _V ALUE ● Core, non-core roles ● Relationships between frames, frame elements ● Add causative: C AUSE _C HANGE _P OSITION _ ON _S CALE 29
Change of position on scale dwindle move soar escalation shift VERBS: advance edge mushroom swell explosion tumble climb explode plummet swing fall decline fall reach triple fluctuation ADVERBS: decrease fluctuate rise tumble gain increasingly diminish gain rocket growth dip grow shift hike NOUNS: double increase skyrocket decline increase drop jump slide decrease rise 30
Recommend
More recommend
