LING 406 Intro to Computational Linguistics Features and Unification

1
LING 406Intro to Computational
LinguisticsFeatures and Unification

• Richard Sproat
• URL http//catarina.ai.uiuc.edu/L406_08/

2
This Lecture

• Features/Unification
• What are feature structures used for?
• Unification
• Integrating unification with parsing

3
Problems with pure CFGs

• Agreement
• This num sg person num sg likes num sg Spam
• Those num pl people num pl likes num sg
  Spam
• This num sg person num sg like num pl Spam
• Those num pl people num pl like num pl Spam
• Those num pl person num sg likes num sg
  Spam
• Subcategorization
• I gave Bill some Spam (V NP NP)
• I gave Spam to Bill (V NP PP)
• I think that Bill likes Spam (V that-S)
• I told Mary that Bill likes Spam (V NP that-S)
• I gave Spam that Bill likes other Hormel products
• I thought Spam to Mary
• I told that Bill likes Spam

4
Problems with pure CFGs

• Rules like the following are not restrictive
  enough they overgenerate.
• S ? NP VP
• VP ? V NP
• VP ? V NP PP
• VP ? V that-S
• VP ? V NP that-S
• NP ? Det Nom

5
Expanding the CFG

• This gets messy fast

6
Solution feature structures

• A schematic feature structure

7
Some features
8
Feature values can also be feature structures
9
Directed acyclic graph (DAG) representation
10
Feature paths

• Feature path is a sequence of features from the
  root to a value. For instance

11
Reentrancy
12
Unification

• Unification allows you to reconstruct a complete
  description from partial descriptions.

13
Unification of feature structures
14
Unification with reentrancy
15
No reentrancy here
16
Further issues

• Look over the discussion in JM (Ch 16) for
• Further complex examples involving reentrancy
• The notion of subsumption. But briefly a more
  general feature structure subsumes a more
  specific one.
• Also see Section 16.4 for a discussion of how
  unification is implemented.

17
Adding feature structures to rules
18
Head features
19
Subcategorization
(cf. I never used to believe Bill that John liked
Spam)
believe
believe but
20
Subcategorization
21
Parsing with unification augmented Earleys
algorithm
22
Unification with Earleys algorithm

• New test for Completer operation only advance
  those states if feature structures unify.
• New test for whether to enter state in chart. You
  might have the same dotted rule, but with
  different DAGs
• Dont add states that have DAGs that are more
  specific than states in chart i.e., is new state
  subsumed by existing states?

23
Tree of Lexical-Functional-Grammar f-structures
24
How f-structures combine
my mothers f-structure
unifies with my daughters f-structure
25
Tree of Lexical-Functional-Grammar f-structures
26
Subcategorization non-configurational languages
27
Subcategorization non-configurational languages
28
Cf.
29
Non-configurational encoding in English

• John saw Mary in the park at sundown with her
  dog.
• John saw Mary in the park with her dog at
  sundown.
• John saw Mary with her dog in the park at
  sundown.
• John saw Mary with her dog at sundown in the
  park.
• John saw Mary at sundown with her dog in the
  park.
• John saw Mary at sundown in the park with her
  dog.