Features and Augmented Grammars

Reference: Chapter 4 of Allen, James: Natural Language Understanding, 2nd ed., Benjamin Cummings, 1995.

Aim:

To describe feature systems, principally syntactic ones, and how grammars may be augmented using featural restrictions. We also enhance a basic chart parser to handle features.

Keywords: agreement, augmented grammar, complement, feature, head feature, head subconstituent, subcat, subcategorization, vform

Plan:

agreement and features, notation, feature examples (vform, subcat, agr) augmenting grammars with features, notation, head features parsing with features - using augmented rules and a chart parser

Features

• A feature is a piece of information associated with a word or phrase. The info can be syntactic or semantic.

• An example feature is agreement: a man is OK but a men is not.

• Agreement can be enforced by attaching to nouns and articles a feature that indicates whether the word is singular or plural, and then reading the usual grammar rule NP → DET N as requiring that the features of DET and N agree.

• Features have a name and a value.

• In this case the name is agr and the value might be s or p. (In practice, 3s or 3p for third-person singular/plural are more usual. Or 1s, 1p, 2s, 2p.)

Binary and Multiple-valued Features

• Two-valued features ("binary features") are very common, and the values are often expressed as + or –.

• For example, the inv feature is a binary feature that indicates whether or not an S structure has an inverted subject (as in a yes/no question.)

• The S structure for Jack laughed would have inv –, while that for Did Jack laugh? would have inv +.

• Often the value appears as a prefix: +inv, –inv.

Binary and Multiple-valued Features 2

• Features may be multiple valued (e.g. "you" - agr {2s, 2p} ).

• Most nouns (NPs) have agr 3s - the dog, agr 3p - the dogs or agr {3s,3p} - the sheep.

• Examples for verbs: loves has {3s}, love has {1s,2s,1p,2p,3p}.

Testing for Agreement

• When checking agreement with multiple-valued features, two features unify through intersection, and agree if the intersection is non-empty.

• For example, the sheep love the grass is OK because {3s,3p} intersection {1s,2s,1p,2p,3p} = {3p} is non-empty , and the dog love the grass is not OK because {3s} intersection {1s,2s,1p,2p,3p} is empty.

The vform Feature

• Verb form (vform): this feature indicates the form of the verb, best explained by describing them all:

The subcat Feature

• Verb complements: complement structure of a verb (i.e. what can/must follow that verb by way of NPs, etc.) can be encoded using features, too;

• Complement structure is often referred to as the subcategorisation of the verb.

• Attach a subcat feature to each verb to record this.

• Some verbs have several allowable complements.

The subcat Feature 2

• Some possibilities:

  subcat value	Example Verb	Example
  none	laugh	Jack laughed _
  np	find	Jack found a key
  np_np	give	Jack gave Sue the paper
  vp:inf	want	Jack wants to fly
  np_vp:inf	tell	Jack told the man to go
  vp:ing	keep	Jack keeps hoping for the best
  np_vp:ing	catch	Jack caught Sam cheating
  np_vp:base	watch	Jack watched Sam buy the drinks

  See Figures 4.2 and 4.4 in Allen for more examples.

subcat Features for Adjectives

• In be-as-copula sentences (Adrian was sad to be rejected by his girlfriend) the complement is determined by the adjective (sad). Thus adjectives can have subcat features too.

4.4 A Simple Grammar Using Features

• In this grammar, subcat and vform features for verbs are used extensively.

• VP(vform(inf)) will be abbreviated VP[inf].

• V(subcat(np_vp:inf)) as V[np_vp:inf].

Grammar 4.7 of Allen

	Rule	Example
1	S[–inv] → NP(agr(?a)) VP[{pres past}](agr(?a))
2	NP → DET(agr(?a)) N(agr(?a))	a man
3	NP → PRO	he
4	VP → V[none]	[he] cries
5	VP → V[np] NP	[he] sees her
6	VP → V[vp:inf] VP[inf]	[he] wants to see her
7	VP → V[np_vp:inf] NP VP[inf]	[he] wants her to see him
8	VP → V[adjp] ADJP	[he] is happy to help
9	VP[inf] → TO VP[base]	to help
10	ADJP → ADJ	happy
11	ADJP → ADJ[vp:inf] VP[inf]	happy to help

Head_features(S, VP) = vform, agr; Head_features(NP) = agr