Word Grammar in Theory Dick Hudson Cardiff, May 2013.

Word Grammar in Theory

Dick Hudson

Cardiff, May 2013

History

• 1963: PhD on Beja grammar at SOAS• Halliday or Chomsky?

– Halliday was more convincing

• 1964-71: worked with Halliday– 64-67 with Huddleston on a corpus study– 66-67 met Terry Winograd (AI)– read about Lamb's network theory– wrote first generative Systemic Grammar book

Systemic Grammar

Dependency theory

• 1976: why no word-word dependencies?

• Systemic + Dependency = Daughter-Dependency Grammar

Psychological reality

• AI (e.g. Winograd, Schank, Anderson)

• Mental networks (Lamb, Halliday)

• Linguists (Bresnan, Chomsky)

• Sociolinguistics (e.g. Labov, Gumperz)

Word Grammar 1984

• Why recognise phrases?

• Daughter-Dependency Grammar + psychological reality – phrases = WG

What am I?

• a lapsed systemicist• a dependency

grammarian• a cognitive linguist• a sociolinguist• a descriptive linguist• a theoretical linguist• an educational linguist

What do I believe?

• Theory matters and can be improved

• Truth matters and can be aimed at.

• Minds matter and can be modelled

• Texts matter and can be explained

• Education matters and needs linguistics

• Neighbouring disciplines matter, and can be learned from.

A nice quote

Wray 2009

One has to ask … whether a particular claim or assumption in one domain can really be true, given what we know about another domain.

What don't I believe in?

• macrofunctions in structure

• system networks

• the rank scale

• constructions as signs

Macrofunctions in structure

• Many artefacts have multiple functions

• But the functions are all 'realized' by the same structure.– e.g. chair:

• comfortable for sitting

• strong, durable

• good-looking

• affordable

For instance

Hearer is part of the meaning of

'question', but not otherwise.

Why 'interpersonal'?

What does 'theme' mean?

System networksWhy is this the

first choice?

How might we organise

knowledge so tidily? And why?

The rank scale

sentence

clause

group

word

Hurry!

hurry

hurry

hurry

unary branching

Headedness

Classification of mother is always predictable from its head daughter

objections

Constructions

• CxG: a language consists of nothing but constructions.

• A construction is a form-meaning pair.– So: every 'form' has a meaning

• But some forms have no meaning, e.g.– extraposition– (arguably) all of morphology, e.g. {z}, {ing}

What does WG actually claim?

• Language is just ordinary knowledge applied to words. So:

• Language is a mental network.

• The network nodes (concepts) are atoms.

• Concepts are defined solely by their links to other concepts.

• Concepts are organised in 'isa' hierarchies.

For instance …

around

… messing around… looked around

… all around were …

… the soil around the base of the plant …

roundagroundon

{on} {a} {round}

around1 around2

labels are NOT part of the analysis

isa hierarchy

Tokens and types

• Tokens are created ad hoc, types are stored.• The type-token ratio = types/tokens

– e.g. for The cat sat on the mat TTR = 5/6

• Tokens have distinct properties, so they must be distinct concepts.

• Each token isa at least one type. – e.g. He wandered around

• Tokens show modifying effects of context.

isa around1

An utterance as part of the grammar

The cat sat on the mat.

THE CAT SIT past ONMAT

noun verb preposition

word

SIT, past

Tokens and syntax

• Each word token is distinct from its type:

• It shows the effects of its syntactic context.– the/1 means 'the cat'– on means 'on the mat'– sat means 'sat on the mat'– but also 'the cat sat on the mat'

• So phrases aren't needed.

Tokens of tokens …

The1 cat sat on the2 mat.

'mat'

'the mat'

'on the mat'

'sat on the mat'

'the cat sat on the mat'

'cat'

'the cat'

sat1 on1 the21the11

sat2

Usage

• The network is 'usage based'– built out of episodic memories

• It contains 'ex-tokens': words plus context– so formulae– and socially constrained

• So any properties may be generalised– 'linguistic' – relations to other words– 'non-linguistic' – relations to social constructs

Processing

• Node creation: one node per token

• Default inheritance to enrich token nodes

• Binding tokens to existing nodes

• Spreading activation– activation guides inheritance and binding– super-active tokens become permanent types

An utterance as part of the grammar

THE CAT SIT past ONMAT

noun verb preposition

word

SIT, past

?

The cat sat on the mat

Summary

• Mentalist – just cognitive structures• Just a network of atoms

– Declarative network– Organised round an isa hierarchy

• Procedures:– node creation– node enrichment (inheritance, binding)– spreading activation