Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did...

44
Semantic Roles & Labeling LING 571 — Deep Processing in NLP November 18, 2019 Shane Steinert-Threlkeld 1

Transcript of Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did...

Page 1: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Semantic Roles & LabelingLING 571 — Deep Processing in NLP

November 18, 2019 Shane Steinert-Threlkeld

1

Page 2: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Announcements● HW7: 89.4 average

● Only common mistake: similarity vs. distance

2

sim(u, v) = 1 - distance(u, v)

Page 3: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Questions on HW #8● For the mc_similarity portion● You should use

● 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

wsim(w1, w2) = maxc1,c2

[simresnik (c1, c2)] From Resnik (1999), eq. 2

https://arxiv.org/pdf/1105.5444.pdf
Page 4: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Alternative Resnik WSD PseudocodeGiven: input word w0 and probe words {p1,…,pn}

for pi in {p1,…,pn}: supported_sense = null most_information = 0.0 for sensew in SENSES(w0): for sensep in SENSES(pi):

lcssynset = LOWESTCOMMONSUBSUMER(sensew, sensep)lcsinfo = INFORMATIONCONTENT(lcssynset)if lcsinfo > most_information:

most_information = lcsinfosupported_sense = sensew

increment support[supported_sense] by most_information

4

Page 5: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Semantic Roles

5

Page 6: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

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

Page 7: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

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

Page 8: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

ExamplesYesterday Tom chased JerryYesterday Jerry was chased by TomTom chased Jerry yesterdayJerry was chased yesterday by Tom

● Semantic roles:● Chaser: Tom ● ChasedThing: Jerry ● TimeOfChasing: yesterday

● Same across all sentence forms

8

Page 9: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

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

Page 10: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Main Idea● Extract the semantic roles without doing full semantic parsing

● Easier problem, but still useful for many tasks● More data● Better models

10

Page 11: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Issues & Challenges● How many roles for a language?● Arbitrary!● Each verb’s event structure determines sets of roles

11

Page 12: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Issues & Challenges● How can we acquire these roles?● Manual construction?● Some progress on automatic learning● Mostly successful on limited domains (ATIS, GeoQuery)

12

Page 13: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Issues & Challenges● Can we capture generalities across verbs/events?● Not really, each event/role is specific

13

Page 14: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Thematic Roles● Solution to instantiating a specific role for every verb

● Attempt to capture commonality between roles

14

Page 15: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Thematic Roles● Describe common semantic roles of verbal arguments● e.g. subject of break is AGENT

● AGENT: volitional cause● THEME: things affected by action

● Enables generalization over surface order of arguments● JohnAGENT broke the windowTHEME

● The rockINSTRUMENT broke the windowTHEME

● The windowTHEME was broken by JohnAGENT

15

Page 16: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Thematic Roles● Verbs take different roles

● The break verb could be formed as:● AGENT/Subject, THEME/Object (John broke the window)

● AGENT/Subject, THEME/Object, INSTRUMENT/PPwith (John broke the window with a rock)

● INSTRUMENT/Subject, THEME/Object (The rock broke the window)

● THEME/Subject (The window was broken)

16

Page 17: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Thematic Roles● Thematic grid, Θ-grid, case frame● Set of thematic role arguments of verb● subject: AGENT; Object: THEME, or● subject: INSTR; Object:THEME

● Verb/Diathesis Alternations● Verbs allow different surface realizations of roles● DorisAGENT gave the bookTHEME to CarvGOAL

● DorisAGENT gave CarvGOAL the bookTHEME

17

Page 18: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Canonical Roles

18

Thematic Role ExampleAGENT The waiter spilled the soupEXPERIENCER John has a headacheFORCE The wind blows debris from the mall into our yards.THEME Only after Benjamin Franklin broke the ice…RESULT The French government has built a regulation-size baseball diamond…CONTENT Mona asked “You met Mary Ann at a supermarket?”INSTRUMENT He turned to poaching catfish, stunning them with a shocking device…BENEFICIARY Whenever Ann Callahan makes hotel reservations for her boss…SOURCE I flew in from Boston.GOAL I drove to Portland.

Page 19: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Thematic Role Issues● Hard to produce

● Standard set of roles● Fragmentation: Often need to make more specific● e.g. INSTRUMENTs can be subject or not

● Standard definition of roles● Most AGENTs: animate, volitional, sentient, causal● But not all… e.g.?

19

From Levin and Rappaport Hovav 2005:

a. John broke the window with a rock.b. The rock broke the window.

a. Swabha ate the banana with a fork.b. * The fork ate the banana.

[Google]Agent found the answer.

Page 20: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Thematic Role Issues● Strategies:● Generalized semantic roles: PROTO-AGENT/PROTO-PATIENT

● Defined heuristically: PropBank

● Define roles specific to verbs/nouns: FrameNet

20

https://framenet.icsi.berkeley.edu/fndrupal/
Page 21: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

PropBank● Sentences annotated with semantic roles● Penn and Chinese Treebank● Roles specific to verb sense● Numbered: Arg0, Arg1, Arg2, …● Arg0: PROTO-AGENT; Arg1: PROTO-PATIENT, etc

21

Page 22: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

PropBank● Arguments >1 are Verb-specific● e.g. agree.01● Arg0: Agreer● Arg1: Proposition● Arg2: Other entity agreeing● Ex1: [Arg0 The group] agreed [Arg1 it wouldn’t make an offer]

22

Page 23: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

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

Page 24: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

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

http://propbank.github.io/
https://catalog.ldc.upenn.edu/LDC2013T19
Page 25: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

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

Page 26: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

AMR 2

26

see-01

i dog

person run-02

name

“Joe”

garden

ARG0 ARG1

poss

name

op1

location

ARG0-of

● AMR Bank: (now) ~40K annotated sentences

● JAMR parser: 63% F-measure (2015)● Alignments between word spans & graph

fragments

● Example: “I saw Joe’s dog, which was running in the garden.”

From Liu et. al (2015)

https://amr.isi.edu/
http://www.aclweb.org/anthology/N15-1114
Page 27: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

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

Page 28: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

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● [ATTRIBUTE The price] of [ITEM bananas] increased [DIFF 5%].● [ATTRIBUTE The price] of [ITEM bananas] rose [DIFF 5%].● There has been a [DIFF 5%] rise in [ATTRIBUTE the price] of [ITEM bananas].

28

https://www.aclweb.org/anthology/L02-1140/
Page 29: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

FrameNet● Semantic roles specific to frame● Frame: script-like structure, roles (frame elements)● e.g. CHANGE_POSITION_ON_SCALE: increase, rise● ATTRIBUTE; INITIAL_VALUE; FINAL_VALUE

● Core, non-core roles● Relationships between frames, frame elements● Add causative: CAUSE_CHANGE_POSITION_ON_SCALE

29

Page 30: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Change of position on scale

30

VERBS: dwindle move soar escalation shiftadvance edge mushroom swell explosion tumbleclimb explode plummet swing falldecline fall reach triple fluctuation ADVERBS:decrease fluctuate rise tumble gain increasinglydiminish gain rocket growthdip grow shift NOUNS: hike

double increase skyrocket decline increasedrop jump slide decrease rise

Page 31: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Core Roles

31

Core RolesATTRIBUTE The ATTRIBUTE is a scalar property that the ITEM possesses.DIFFERENCE The distance by which an ITEM changes its position on the scale.FINAL_STATE A description that presents the ITEM’s state after the change in the ATTRIBUTE’s value as an independent predication.FINAL_VALUE The position on the scale where the ITEM ends up.INITIAL_STATE A description that presents the ITEM’s state before the change in the ATTRIBUTE’s value as an independent predication.INITIAL_VALUE The initial position on the scale from which the ITEM moves away.ITEM The entity that has a position on the scale.VALUE_RANGE A portion of the scale, typically identified by its end points, along which the values of the ATTRIBUTE fluctuate.

Some Non-Core RolesDURATION The length of time over which the change takes place.SPEED The rate of change of the VALUE.GROUP The GROUP in which an ITEM changes the value of an ATTRIBUTE in a specified way.

Page 32: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

FrameNet● Current status:● 1224 frames● 13669 lexical units (mostly verbs, nouns)● 10749 frame element relations● Annotations over:● Newswire (WSJ, AQUAINT)● American National Corpus

● Under active development

● Still relatively limited coverage

32

https://framenet.icsi.berkeley.edu/fndrupal/current_status
Page 33: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Semantic Role Labeling

33

Page 34: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Semantic Role Labeling● Task of automatically assigning semantic roles for each argument

34

Page 35: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Typical Strategy● Assign Parse to Input String

● Traverse parse to find all predicates

● For each predicate, examine each node and decide semantic role (if any)

35

Page 36: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Typical Strategyfunction SEMANTICROLELABEL(words) returns labeled treeparse←PARSE(words)

for each predicate in parse dofor each node in parse dofeaturevector←EXTRACTFEATURES(node, predicate, parse)CLASSIFYNODE(node, featurevector, parse)

36

J&M 3rd ed, ch 20.6

https://web.stanford.edu/~jurafsky/slp3/20.pdf#section.20.6
Page 37: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Semantic Role Labeling Features● Governing predicate

● Phrase Type (NP, VP, etc)

● Headword of constituent

● Headword POS

● PATH from current node to predicate (NP↑S↓VP↓VBD)

● …

37

Page 38: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Typical Strategy

38

7

24�7&.%6+�

(8

8LI

224

7ER

224

*VERGMWGS

224

)\EQMRIV

:4

:&(8%6+)8

MWWYIH

24%6+�

(8

E

..

WTIGMEP

22

IHMXMSR

44�814%6+1814

-2

EVSYRH 24

22

RSSR

24�814

]IWXIVHE]<latexit sha1_base64="Vzpp1dZk5gNKdoPwuUbDXkZKAKI=">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</latexit><latexit sha1_base64="Vzpp1dZk5gNKdoPwuUbDXkZKAKI=">AAALO3ictVbNbttGEGactlHZpo2bYy+LOAGKwnIkp7Gjg4HEVuTUiFKllpSgomAslyN5oeWS2V0GVgk+Wh+hD9Bzb0WvvXf4oyaUHYcqYEogBzPfzs5+M5gdNxRcm0bjj2tr1z/59LMbtc/tL768+dXXt9a/GeogUgwGLBCBeu1SDYJLGBhuBLwOFVDfFfDKnR2k9ldvQWkeyL6ZhzD26VTyCWfUoOpk/cbvjgbjUy4ngTTxIReCHFOpE/seeTLskuP+oNOxHfrW14EyGcbhJrEdDya4ZRZAzEQSP+t3nyfx9sHuQauzZHb/M+/sPnjw6OGymS3MD1vpL7HtPCZzuhTTSFE5hb0ovB9PFcBs8zB7CzyL3HyevWXkJ+MPrCf7gfAWTtzJ/3Wz8MDNB9ena98/InjTXFqctNU5aLWXidCRq5nioSlBO52d5k5rCSrLGKS9fY72MmTBrePClMt4EijQJiH4JQbD34up4FO5x0AaUIk9IjE5TjZJGvneu/g3zSlnM9QXAaSwF7368f5R7Fx4CHL3JCZPfj5sJHeTau7a/Wo4LEPSP4VLwDmWkHH6zyJ90avuG1N9Zb47WEGMaxZc2Q5Pz6iPlaCqblDsMay4xXC/fVnC+5jxp/3KKU8D5lpH4K3ER+8jNde8kpqjq8R4dFTdsQ6BcSpWK4nq7sHjactfsSB6vXq/+zGiuymmMtk/rhA0VUEkq1VFXhIVW9YKIcigOmlFO0zZqOx/jn0YlEfnK2Tm/b/tgPQW/fzk1kZjq5E95LzQLIQNq3h6J+trA8cLWORj32eCaj1qNkIzjqkynAnA+yLSEFI2o1OIQ6rwRsTboqQeoSipD3ocF2VxDzVedrWkNyfJtCVH1Nd67ruITK9XvWxLlRfZRpGZPBrHXIaRAcnyjSaRICYg6XBDPK6AGTFHgWKd4hEIO6WKMmQYPbUBj6igi35/CkFRE6jvY4eqKbbKJHbcCSnky6D07B0U5VLshs9+rb9JGaqzwA+pucDsQYgJw/DnZSOOWGXFWc5mSVek2cbprIP0LtmkSXtIeUUkOQs8qGeclieabFQhXbQq/OT2lNBsD8FdRdU8xsFyGuEcyhnOhE5uo2Eo5gUAJw4aCaPLQBsLsblcdueF4fZWs7HVfLm98Xi/KMma9a11x/rOalq71mPrmdWzBhar/VD7pcZqnv2b/af9l/13Dl27Vqy5bZUe+59/Abne9Is=</latexit><latexit sha1_base64="Vzpp1dZk5gNKdoPwuUbDXkZKAKI=">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</latexit><latexit sha1_base64="eWepl42E4iD5l8hRND5CHVRxxJU=">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</latexit><latexit sha1_base64="Jp98yY87SdUU3yFDABSmZVpNVas=">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</latexit><latexit sha1_base64="Jp98yY87SdUU3yFDABSmZVpNVas=">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</latexit><latexit sha1_base64="Z+D20rHgiQGEaUNV8lqA+EhJoJ0=">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</latexit><latexit sha1_base64="Vzpp1dZk5gNKdoPwuUbDXkZKAKI=">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</latexit><latexit sha1_base64="Vzpp1dZk5gNKdoPwuUbDXkZKAKI=">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</latexit><latexit sha1_base64="Vzpp1dZk5gNKdoPwuUbDXkZKAKI=">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</latexit><latexit sha1_base64="Vzpp1dZk5gNKdoPwuUbDXkZKAKI=">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</latexit><latexit sha1_base64="Vzpp1dZk5gNKdoPwuUbDXkZKAKI=">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</latexit><latexit sha1_base64="Vzpp1dZk5gNKdoPwuUbDXkZKAKI=">AAALO3ictVbNbttGEGactlHZpo2bYy+LOAGKwnIkp7Gjg4HEVuTUiFKllpSgomAslyN5oeWS2V0GVgk+Wh+hD9Bzb0WvvXf4oyaUHYcqYEogBzPfzs5+M5gdNxRcm0bjj2tr1z/59LMbtc/tL768+dXXt9a/GeogUgwGLBCBeu1SDYJLGBhuBLwOFVDfFfDKnR2k9ldvQWkeyL6ZhzD26VTyCWfUoOpk/cbvjgbjUy4ngTTxIReCHFOpE/seeTLskuP+oNOxHfrW14EyGcbhJrEdDya4ZRZAzEQSP+t3nyfx9sHuQauzZHb/M+/sPnjw6OGymS3MD1vpL7HtPCZzuhTTSFE5hb0ovB9PFcBs8zB7CzyL3HyevWXkJ+MPrCf7gfAWTtzJ/3Wz8MDNB9ena98/InjTXFqctNU5aLWXidCRq5nioSlBO52d5k5rCSrLGKS9fY72MmTBrePClMt4EijQJiH4JQbD34up4FO5x0AaUIk9IjE5TjZJGvneu/g3zSlnM9QXAaSwF7368f5R7Fx4CHL3JCZPfj5sJHeTau7a/Wo4LEPSP4VLwDmWkHH6zyJ90avuG1N9Zb47WEGMaxZc2Q5Pz6iPlaCqblDsMay4xXC/fVnC+5jxp/3KKU8D5lpH4K3ER+8jNde8kpqjq8R4dFTdsQ6BcSpWK4nq7sHjactfsSB6vXq/+zGiuymmMtk/rhA0VUEkq1VFXhIVW9YKIcigOmlFO0zZqOx/jn0YlEfnK2Tm/b/tgPQW/fzk1kZjq5E95LzQLIQNq3h6J+trA8cLWORj32eCaj1qNkIzjqkynAnA+yLSEFI2o1OIQ6rwRsTboqQeoSipD3ocF2VxDzVedrWkNyfJtCVH1Nd67ruITK9XvWxLlRfZRpGZPBrHXIaRAcnyjSaRICYg6XBDPK6AGTFHgWKd4hEIO6WKMmQYPbUBj6igi35/CkFRE6jvY4eqKbbKJHbcCSnky6D07B0U5VLshs9+rb9JGaqzwA+pucDsQYgJw/DnZSOOWGXFWc5mSVek2cbprIP0LtmkSXtIeUUkOQs8qGeclieabFQhXbQq/OT2lNBsD8FdRdU8xsFyGuEcyhnOhE5uo2Eo5gUAJw4aCaPLQBsLsblcdueF4fZWs7HVfLm98Xi/KMma9a11x/rOalq71mPrmdWzBhar/VD7pcZqnv2b/af9l/13Dl27Vqy5bZUe+59/Abne9Is=</latexit>

NP↑S ↓VP↓VBDPATH(NP-SBJ)

Page 39: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Some Semantic Role Labeling Applications● Question answering:● Who did what to whom?

● Machine translation● Maintain agents/thematic roles through translation

● Dialogue systems

39

Page 40: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Scaling up SRL

40

Page 41: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

Neural SRL

41

He et al 2017

Can be global or contextual [contextual tends to improve]

No “detour” through syntactic parse

https://www.aclweb.org/anthology/P17-1044/
Page 42: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

QA-SRL

42

the paper

Editorial:should’ve been /casserole/

https://www.aclweb.org/anthology/D15-1076/
Page 43: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

QA-SRL vs. PropBank

43

Page 44: Semantic Roles & Labeling · 2020. 9. 24. · Semantic Role Labeling Typically want to know Who did what to whom …where, when, and how Intermediate level: Shallower than full deep

QA-SRL● Much more info, including live data explorer:● http://qasrl.org/

● AI2 NLP Highlights podcast most recent episode ft. Luke Zettlemoyer: ● https://soundcloud.com/nlp-highlights/96-question-answering-as-an-annotation-

format-with-luke-zettlemoyer

44

http://qasrl.org/
https://soundcloud.com/nlp-highlights/96-question-answering-as-an-annotation-format-with-luke-zettlemoyer
https://soundcloud.com/nlp-highlights/96-question-answering-as-an-annotation-format-with-luke-zettlemoyer

High-performance Multilingual Semantic Role Labeling

High-performance Multilingual Semantic Role Labeling

Semantic Role Labeling. Introduction Semantic Role Labeling AgentThemePredicateLocation.

Semantic Role Labeling. Introduction Semantic Role Labeling AgentThemePredicateLocation.

On Labeling Schemes for the Semantic Web

On Labeling Schemes for the Semantic Web

LowResource++ Semantic+Role+Labeling+mgormley/papers/gormley+al.acl.2014.slides-clsp… · LowResource++ Semantic+Role+Labeling+ Matthew’R.Gormley’ Margaret’Mitchell’ BenjaminVan

LowResource++ Semantic+Role+Labeling+mgormley/papers/gormley+al.acl.2014.slides-clsp… · LowResource++ Semantic+Role+Labeling+ Matthew’R.Gormley’ Margaret’Mitchell’ BenjaminVan

Using Semantic Role Labeling to Combat Adversarial SNLIzhangmf/CS224u/NLU_final_report.pdf · Using Semantic Role Labeling to Combat Adversarial SNLI Brett Szalapski brettski@stanford.edu

Using Semantic Role Labeling to Combat Adversarial SNLIzhangmf/CS224u/NLU_final_report.pdf · Using Semantic Role Labeling to Combat Adversarial SNLI Brett Szalapski [email protected]

CHAPTER 22 Semantic Role Labeling

CHAPTER 22 Semantic Role Labeling

High-Resolution Semantic Labeling with Convolutional Neural Networksgcharpia/Maggiori_HighResLabeling... · 2018. 9. 24. · to semantic pixel labeling (i.e., assigning a class to

High-Resolution Semantic Labeling with Convolutional Neural Networksgcharpia/Maggiori_HighResLabeling... · 2018. 9. 24. · to semantic pixel labeling (i.e., assigning a class to

Case Grammar Semantic Role Labeling - Syracuse Universityclasses.ischool.syr.edu/ist664/NLPSpring2010/SemanticRoleLabeling… · Semantic Role Labeling. Semantics of events in sentences

Case Grammar Semantic Role Labeling - Syracuse Universityclasses.ischool.syr.edu/ist664/NLPSpring2010/SemanticRoleLabeling… · Semantic Role Labeling. Semantics of events in sentences

Neural Semantic Role Labeling: What works and what’s nextluheng.github.io/files/deep_srl_slides_long.pdf · Neural Semantic Role Labeling: What works and what’s next or: What

Neural Semantic Role Labeling: What works and what’s nextluheng.github.io/files/deep_srl_slides_long.pdf · Neural Semantic Role Labeling: What works and what’s next or: What

Automatic Semantic Role Labeling - Stanford University

Automatic Semantic Role Labeling - Stanford University

Recursive Context Propagation Network for Semantic Scene Labeling

Recursive Context Propagation Network for Semantic Scene Labeling

Automatic Labeling of Semantic Roles

Automatic Labeling of Semantic Roles

Semantic Labeling of Places

Semantic Labeling of Places

Semantic Role Labeling Tutorial - NAACL HLT 2013naacl2013.naacl.org/Documents/semantic-role-labeling-part-1-naacl... · Semantic Role Labeling Tutorial NAACL, June 9, 2013 Part 1:

Semantic Role Labeling Tutorial - NAACL HLT 2013naacl2013.naacl.org/Documents/semantic-role-labeling-part-1-naacl... · Semantic Role Labeling Tutorial NAACL, June 9, 2013 Part 1:

Towards Semantic Role Labeling of Hand-drawn Sketches · Labeling of natural languages - as described in the current literature, - describe Sketch Semantic Role Labeling, and then

Towards Semantic Role Labeling of Hand-drawn Sketches · Labeling of natural languages - as described in the current literature, - describe Sketch Semantic Role Labeling, and then

Semantic labeling

Semantic labeling

Semantic Labeling of 3D Point Clouds for Indoor Scenespapers.nips.cc/paper/4226-semantic-labeling-of-3d-point...Semantic Labeling of 3D Point Clouds for Indoor Scenes Hema Swetha Koppula,

Semantic Labeling of 3D Point Clouds for Indoor Scenespapers.nips.cc/paper/4226-semantic-labeling-of-3d-point...Semantic Labeling of 3D Point Clouds for Indoor Scenes Hema Swetha Koppula,

Semantic Role Labeling

Semantic Role Labeling

SEMANTIC ROLE LABELING USING LEXICALIZED TREE ADJOINING ...anoop/students/yudong_liu/thes-yudong.pdf · SEMANTIC ROLE LABELING USING LEXICALIZED TREE ADJOINING GRAMMARS by ... Semantic

SEMANTIC ROLE LABELING USING LEXICALIZED TREE ADJOINING ...anoop/students/yudong_liu/thes-yudong.pdf · SEMANTIC ROLE LABELING USING LEXICALIZED TREE ADJOINING GRAMMARS by ... Semantic

Combination Strategies for Semantic Role Labeling

Combination Strategies for Semantic Role Labeling