Foundations of Natural Language Processing

Lecture 16
Semantic Role Labelling and Argument Structure
Alex Lascarides
(Slides based on those of Schneider, Koehn, Lascarides)

13 March 2020

Alex Lascarides FNLP Lecture 16 13 March 2020

 Language is Flexible

• Often we want to know who did what to whom (when, where, how and why)

• But the same event and its participants can have different syntactic realizations.

Sandy broke the glass. vs. The glass was broken by Sandy.
She gave the boy a book. vs. She gave a book to the boy.

• Instead of focusing on syntax, consider the semantic roles (also called

thematic roles) defined by each event.

Alex Lascarides FNLP Lecture 16 1

 Argument Structure and Alternations

• Mary opened the door

The door opened

• John slices bread with a knife

This bread slices easily
The knife slices cleanly

• Mary loaded the truck with hay

Mary loaded hay onto the the truck
The truck was loaded with hay (by Mary)
The hay was loaded onto the truck (by Mary)

• John gave a present to Mary

John gave Mary a present

Alex Lascarides FNLP Lecture 16 2

 Stanford Dependencies

cf Mary ate the sandwich with Kim!

Alex Lascarides FNLP Lecture 16 3
 Syntax-Semantics Relationship

Alex Lascarides FNLP Lecture 16 4

 Outline

• syntax 6= semantics

• The semantic roles played by different participants in the sentence are not
trivially inferable from syntactical relations

• . . . though there are patterns!

• The idea of semantic roles can be combined with other aspects of meaning
(beyond this course).

Alex Lascarides FNLP Lecture 16 5

 Commonly used thematic roles

Role Example
Agent The boy kicked his toy
Theme The boy kicked his toy
Experiencer The boy felt sad
Result The girl built a shelf with power tools
Instrument The girl built a shelf with power tools
Source She came from home
... ...

• J&M give definitions and additional roles

Alex Lascarides FNLP Lecture 16 6

 Issues with thematic roles

• No universally agreed-upon set of roles

• Items with the “same” role (e.g., Instrument) may not behave quite the same
Sandy opened the door with a key The key opened the door
Sandy ate the salad with a fork *The fork ate the salad

• The two main NLP resources for thematic roles avoid these problems by
defining very fine-grained roles:
– Specific to individual verbs only (PropBank)
– Specific to small groups of verbs (FrameNet)

Alex Lascarides FNLP Lecture 16 7

 Semantic role labelling

• The NLP task of identifying which words/phrases play which roles in an event.

• Supervised classification:
– Resource data is PropBank: Repository of frame files for each verb
(more shortly) plus annotations on constituents in Penn treebank with their
semantic roles (wrt the relevant frame file).
– Features are mostly related to syntactic structure and the particular words
involved
(i.e., assumes pipeline architecture)

• Current research focuses on reducing the need for training data (e.g., to work
on non-English languages)

Alex Lascarides FNLP Lecture 16 8

 Example Frame Roles: load
Mary loaded the truck with hay at the depot on Friday

• load: load.01 ‘cause to be burdened’

Arg0-PAG: loader,agent
Arg1-GOL: beast of burden
Frame roles:
Arg2-PPT: cargo
Arg3-MNR: instrument

• load up: load.02 ‘phrasal cause to be burdened’

Frame roles are the same as load.01

• load: load.03 ‘fix, set up to cheat’

Arg0-PAG: cheater
Frame roles: Arg1-GOL: thing loaded (dice, the deck, etc)
Arg2-PPT: with what

• All sentences can have temporal, spatial adjuncts (AM-TMP, AM-LOC). . .

Alex Lascarides FNLP Lecture 16 9

 PropBank
Penn treebank annotated with Arg0, Arg1 etc, and verb with its sense;
so specific semantic role recoverable.

Alex Lascarides FNLP Lecture 16 10

 Semantic Role Labelling
Traditional pipeline:

1. Either assume or compute syntactic parse and predicate senses

2. Argument identification (deterministic): select the predicate’s argument

phrases (by parsing the parse tree)

3. Argument classification: select a role for each argument (wrt to the frame
role for the predicate’s sense).
• Useful feature: predicate-to-argument path in the tree (e.g., NP-S-VP-V).

Alex Lascarides FNLP Lecture 16 11

 Problems

• Numbered roles are predicate specific:

– load.01.ARG1: beast of burden
put.01.ARG1: thing.put
put.01.ARG2: beast of burden.

• FrameNet tries to generalise via verb classes;

but less treebank data.

Alex Lascarides FNLP Lecture 16 12

 Paraphrase
James snapped a photo of me with Sheila.
Sheila and I had our picture taken by James

Alex Lascarides FNLP Lecture 16 13

 Paraphrase
James snapped a photo of me with Sheila.
Sheila and I had our picture taken by James

Alex Lascarides FNLP Lecture 16 14

 Photo Labelled. . .
photograph vs. take picture vs. snap picture. . .

Alex Lascarides FNLP Lecture 16 15

 Idealised Stanford Dependencies

• James snapped a photo of me with Sheila

nsubj(snap, James)
dobj(snap, photo)
prep of(photo,me)
prep with(me,Sheila)
det(photo,a)

• Sheila and I had our picture taken with James

nsubjpass(taken,Sheila)
nsubjpass(taken,I)
conj and(Sheila,I)
aux(taken,had)
dobj(taken,picture)
poss(picture,our)
prep with(picture,James)

Here, agent is the complement introduced with by in a passive construction. . .

Alex Lascarides FNLP Lecture 16 16

 FrameNet: Meanings are reletavised to scenes!

• Tries to capture relationships among word and phrase meanings by assigning

them the same frame (and so captures paraphrases).

• ≈ 1000 frames represent scenarios.

– Most are associated with lexical units (predicates);
but some are phrases

• Frames are explained with textual descriptions and linguistic examples.

Alex Lascarides FNLP Lecture 16 17

 Example: Create physical artwork
Definition:
A Creator creates an artefact that is typically an iconic Representation of an
actual or imagined entity or event. The Representation may also be evocative of
an idea while not based on resemblance.

• Diagrams must be clearly drawn on construction paper.

I took his picture and told him it came out well.

Frame Elements:

Core: creator, representation

Non-Core manner, location of representation . . .

Alex Lascarides FNLP Lecture 16 18

 FrameNet Resources

• FrameNets for several languages

• Some (limited!) data annotated with Frame elements from FrameNet

• SEMAFOR is a frame-semantic parser

– Ongoing research at CMU, Google, Edinburgh. . .

Alex Lascarides FNLP Lecture 16 19

 Summary

• Grammatical relations on their own don’t determine who did what to whom

• You need to (also) know about word and phrase meanings and how they relate
to grammatical roles

• There is flexibility in how a verb realises its participants syntactically (connected

with the kind of event that the verb denotes)

• One must exploit those patterns to obtain NL understanding (e.g., predict

entailments, paraphrases etc).

Alex Lascarides FNLP Lecture 16 20