Systemic Functional Grammar

unknownlippsAI and Robotics

Oct 16, 2013 (4 years and 2 months ago)

81 views

A Probabilistic Representation of
Systemic Functional Grammar

Robert Munro

Department of Linguistics,

SOAS,

University of London

2

Outline


Introduction


Functions in the nominal group


Machine learning


Testing framework


Classification vs unmarked function


Gradational realization


Delicacy


Conclusions

3

Introduction


An exploration of the ability of machine
learning to learn and represent functional
categories as fundamentally probabilistic


Gauged in terms of the ability to:


computationally learn functions from labeled
examples and apply to new texts.


represent functions probabilistically: a gradation
of potential realization between categories.


explore finer layers of delicacy.

4

Functions in the nominal group


Functions:


Deictic, Ordinative, Quantitative, Epithet,
Classifier, Thing (Halliday 1994)





Gradations:


Here, ‘red’ functions also functions as an Epithet.


The uptake of such marked classifiers will not be
uniform.


Overlap does not necessarily limit significance.

Deictic


The

Ordin.


first

Quant.


three

Epith.


tasty

Class.


red

Thing


wines

5

Machine Learning


Machine learning:


computational inference from specific examples.


A learner named
Seneschal
was developed
for the task here:


probabilistic


seeks sub
-
categories (improves both
classification and analysis)


allows categories to overlap


not too dependent on the size of the data set

6

Machine Learning

y

x

?

?

?

?

?

?

?


The task here:


Given categories & with known values for
x

and
y
, infer a probabilistic model (potentially with
sub
-
categories) that can classify new examples:


7

Machine Learning


It is important that attributes (
x,y,z
...) :


represent features that distinguish functions


can be discovered automatically (for large scales)


are meaningful for analysis…?


Compared to manually constructed parsers:


greater scales than are practical


more features/dimensions than are possible
(100’s are common)


8

Testing Framework


The model was learned from 10,000 labeled
words from Reuters sports newswires from
1996


23 features:


part
-
of
-
speech and its context


punctuation


group / phrase contexts


collocational tendencies


probability of repetition

9

Testing Framework


Accuracy:


The ability to correctly identify the dominant
function in 4 test corpora (1,000 words each):

1.
Reuters Sports Newswires (1996)

2.
Reuters Sports Newswires (2003)

3.
Bio
-
informatics abstracts

4.
Extract from Virginia Woolf’s ‘The Voyage Out’





10

Testing Framework


Gradational model of realization:


calculated as the probability of a word realizing
other functions, averaged between all clusters.


Finer layers of delicacy:


Manual analysis of clusters found within a
function.

11

Unmarked function


Unmarked function:

function defined by only
part
-
of
-
speech (POS) and word order.


eg: adjective = Epithet, non
-
final noun = Classifier


Previous functional parsers have assumed that
most instances are unmarked:


POS taggers are almost 100% accurate


word order is trivial


…so the problem is solved?

12

Unmarked function


This is a false assumption.


Across the corpora:


< 40% of non
-
final adj’s realized Epithets


< 50% of Classifiers were nouns


44% of Classifiers were ‘marked’!


13

Unmarked function


This task halved the classification error:

14

Gradational
Realization

Deictic


The

Ordin.


first

Quant.


three

Epith.


tasty

Class.


red

Thing


wines

Deictic

Ordin.

Quant.

Epith.

Class.

Thing


Nominal functions are typically represented
deterministically:



Although described as probabilistic,



With relationships existing between all
functions

15

Delicacy

Deictic

Numerative

Epithet

Expansive

Thing

Demonstrative

Possessive

Ordinative

Quantitative

Hyponymic

Classifier

First Name

Intermediary

Last Name

non
-
Nom.

Stated

Described

Discursive

Nominative

Named Entity

Group
-
Releasing

Nominal

Tabular

16

Delicacy


More delicate functions for Classifiers
(Matthiessen 1995) :


Hyponymic:
describing a taxonomy or general
‘type
-
of’ relationship eg: ‘
red

wine’, ‘
gold
medal

,

neural network

architecture'


Expansive:
expands the description of the
Head. eg: ‘
knee

surgery’, ‘
optimization

problems',

sprint

champion’,



17

Delicacy

18

Delicacy


More delicate descriptions can be found:


more features


more instances / registers


other algorithms / parameters


Methodology can be applied to:


other parts of a grammar


other languages


19

Conclusions


Gradational modeling of functional realization
is desirable


Sophisticated methods are necessary for
computationally modeling functions:


Markedness is common


Machine learning is a useful tool and
participant in linguistic analysis.


20

Thank you


Acknowledgments:


Geoff Williams


Sanjay Chawla


The slides and extended paper will be
published at:


www.robertmunro.com/research/