Reference and Working Memory: What Discourse Can Tell us about Cognition

1
Reference and Working MemoryWhat Discourse
Can Tell us about Cognition

• Andrej A. Kibrik (kibrik_at_chat.ru)
• (Institute of Linguistics, Moscow,
• and MPI for Evolutionary Anthropology, Leipzig)

2
INTRODUCTIONThe phenomenon Referential choice
in discourse

• When people speak or write, they constantly
  mention various referents (persons, animals,
  objects, abstract notions, etc.)
• Basic referential choice
• full noun phrase (full NP)
• pronoun
• zero form

3
An example (from the web page of the city of
Dresden)

• 1. Johann Friedrich Böttger
• Ø Alchemist and inventor, Ø born 4.2.1682 in
  Schleiz, Ø died 13.3.1719 in Dresden. Böttger was
  imprisoned as an alchemist in Königstein Fortress
  in 1703. In 1707 his laboratory was transferred
  to the Jungfernbastei, a bastion of the Dresden
  City fortifications. It was here, a year later,
  that he discovered the formula for the first
  European porcelain and the world's first hard
  porcelain. Böttger also achieved certain results
  as a botanist in Dresden, Ø setting up a
  greenhouse with over 400 rare plants. In 1710 he
  was ordered to Meissen as administrator of the
  royal porcelain manufactory.

zero
Full NP
pronoun
4
Summary of the talk

• Part I A linguistic study of referential choice
  in natural discourse
• Part II Consequences of that study for the
  broader field of working memory research

5
PART I THE LINGUISTIC STUDY The problem

• How does the speaker make the referential choice
  between full noun phrases and reduced noun
  phrases, such as pronouns?
• Note This problem is really fundamental at
  least every third word in natural discourse
  depends on referential choice.

6
Prior studies huge literature, including

• Linguistics, e.g.
• Fox, Barbara. 1987. Discourse structure and
  anaphora in written and conversational English.
  Cambridge CUP
• Psycholingustics, e.g.
• Gernsbacher, Morton Ann. 1990. Language
  comprehension as structure building. Hillsdale,
  NJ Erlbaum.
• Cognitive psychology and neuroscience, e.g.
• Streb, Judith Roesler, Frank Henninghausen,
  Erwin. 1999. Event-related responses to pronoun
  and proper name anaphors in parallel and
  nonparallel discourse structures. Brain and
  Language 70 273-286.

7
Different terminologies

• Coreference
• Anaphora
• Reference tracking
• Reference maintenance
• Management of reference
• Referential choice

8
Important terms
coreferential

• Johni was sitting at the table. Hei was
  daydreaming about the weekend

antecedent
Referential device/expression
9
Goal

• To construct a model of referential choice in
  discourse

10
Properties of the present model

• speaker-oriented (rather than addressee-oriented
  or text-centered)
• sample-based (rather than based on a
  heterogeneous set of examples)
• general (rather than tolerant to exceptions)
• predictive and finite (rather than post-hoc
  repairing to account for individual cases)
• explanatory (rather than based on the black box
  ideology)
• cognitively based (rather than relying on
  home-made quasi-cognitive concepts)
• multi-factorial (rather than assuming one
  omnipotent factor)
• testable and calculative (rather than declarative)

11
The cognitive assumptions

• The primary cognitive determiner of referential
  choice is activation of the referent in question
  in the speakers working memory (WM).
• Referents activation score (AS) varies within a
  certain range (e.g. between 0 and 1).
• If the current activation score is above a
  certain threshold, then a semantically reduced
  (pronoun or zero) reference is possible, and if
  not, a full NP is used.

12
This model continues the lines of

• Cognitively minded linguistic research, such as
• Chafe, Wallace. 1994. Discourse, consciousness,
  and time. The flow and displacement of conscious
  experience in speaking and writing. Chicago
  University of Chicago Press.
• Tomlin, Russell Pu, Ming-Ming. 1991. The
  management of reference in Mandarin discourse.
  Cognitive Linguistics 2 6593
• Kibrik, Andrej A. 1991. Maintenance of reference
  in sentence and discourse. In Lehmann, Winfred
  P. Hewitt, Helen-Jo J. (eds.) Language
  typology. Amsterdam Benjamins, 57-84.

13
And attempts to be compatible with

• Cognitive-psychological and neuropsychological
  work on working memory
• Baddeley, Alan. 1990. Human Memory Theory and
  Practice. Needham Heights, Mass Allyn Bacon.
• Cowan, Nelson. 1995. Attention and Memory An
  Integrated Framework. New York Oxford Oxford
  University Press.
• Smith, E.E. Jonides J. 1997. Working memory A
  view from neuroimaging. Cognitive Psychology 33,
  542.

14
The cognitive multifactorial model of reference
in discourse production

Discourse context
Referents activation score (AS)
Referential choice
Filters
Properties of the referent
Activation factors
15
The original study

• Referential choice in Russian narrative discourse
  (Kibrik 1996)
• Main results
• About seven to ten significant activation factors
• Numerical model of factor interaction
• Complete prediction of the data in corpus
• Almost complete prediction in the test corpus

16
A study of referential choice in English
narrative discourse

• The Maggie B. by Irene Haas
• Discourse type
• written narrative
• simple, clear style
• basic event types physical events, interaction
  of people, human reflections
• Number of discourse units 117
• Number of referents 76
• Number of referent mentions 225
• Number of important referents 14
• Number of protagonist referents mentions
  Margaret 72, James 28, the ship 12
• Number of relevant referential devices full
  NPs 39, activation-based pronouns 40

17
Stages of model construction

• I will explain the heuristics of model
  construction in terms of five consecutive stages,
  A through E.

18
Stage A Identify alterable vs. unalterable
referential devices

• 1502 A storm was coming!1503 Margaret must
  make the boat ready at once.1601 She
  (Margaret) took in the sail1602 and tied it
  tight.
• 1603 She (Margaret) dropped the anchor1604
  and stowed all the gear lt...gt
• Alterable and unalterable devices correspond to
  different activation levels

19
Attribution of referent mentions to potential
referential form categories
20
Stage B Identify the significant activation
factors, as opposed to insignificant
21
Stage C Specify the list of the significant
activation factors

• with the indication of
• the distinction between primary and secondary
  factors
• logical structure of each factor
• values of each factor
• corresponding numerical weights of each value

22
Primary activation factors (variables), their
values, and numerical activation weights
23
An example of a rhetorical graph (in accordance
with the Rhetorical Structure Theory of Mann and
Thompson)
24
Example of RhD ? LinD(RhD is low and LinD is
high)

• 1201 After juice-and-cookie time, she gave James
  his counting lesson,
• 1202 and this is how she did it.
• 1203 One, two, three, four, five, once I caught
  a fish alive,
• 1204 six, seven, eight, nine, ten, but I let
  him go again.
• 1205 Why did you let him go?
• 1206 because he bit my finger so.
• 1207 Which finger did he bite?
• 1208 This little one upon the right.
• 1209 And she gave James' little finger a nibble

RhD1
LinD7
25
Reference and discourse structure

• Referential choice is fundamentally conditioned
  by discourse structure
• The strongest activation factor is the rhetorical
  (hierarchical) distance to the antecedent
• Reduced NPs are more likely to occur in coherent
  contexts

26
Primary activation factors...(continuation)
27
Reference and the properties of
antecedent/referent

• Antecedent role is the second strongest
  activation factor subjects make very good
  antecedents
• More permanent referent properties
  (protagonisthood, animacy) play the role of
  correction/compensation factors

28
Activation weights

• Present model weights found through a trial-and
  error procedure, by hand
• Ideal model weights found through a
  computational procedure, automatically

29
Stage D Identify the mechanism of factor
interaction

• Present model addition of all relevant
  activation weights the resulting AS varies
  within the limits from 0 to a bit over 1.
• Ideal model multiplication or more complex
  interaction of the factors activation weights

30
Stage E Identify referential strategies, or
mappings AS ? referential choice
31
A probabilistic reinterpretation of referential
strategies 4 thresholds
Pronoun only
Pronoun OK
Full NP OK
Full NP only
32
An example of calculating a referents current AS
33
PART II Consequences for working memory studies

• Some classical issues in WM research
• (1) WM capacity how much information can WM hold
  at one time?
• (2) Control of WM through what mechanism does
  information enter WM?
• (3) Forgetting through what mechanism does
  information quit WM?

34
Issue 1 Capacity

• The procedure of calculating the referents ASs
  does not depend on whether a given referent is
  actually mentioned at the present point
• For any referent, its AS can be identified at any
  time
• Therefore, summary (grand) activation of all
  referents can be calculated for any moment of
  discourse

35
The dynamics of two protagonist referents
activation and of grand activation in an excerpt
of English narrative
36
Generalizations about WM capacity

• Grand activation is an estimate of the
  specific-referent portion of WM
• The maximal values of grand activation are
  between 3 and 4 (cf. an identical estimate in
  Cowan 2000)
• Grand activation varies much less than activation
  of individual referents
• In the course of coherent stretches of discourse
  (paragraphs) grand activation gradually builds up
• At the points of incoherence (paragraph
  boundaries) WM is reset or updated.

37
Issue 2 Control of WM

• WM is controlled by the attentional system of the
  brain (Baddeley 1990, Cowan 1995, Posner
  Raichle 1994 173).
• Focal attention is linguistically rendered by the
  syntactic status of subject (Tomlin 1995)
• Subjects are the best antecedents, both
  discourse- and sentence-wide

38
Cognitive and linguistic interplay between
attention and WM

• Attention feeds WM What is attended at moment tn
  becomes activated in WM at moment tn1
• Linguistic moments are discourse units
• Focally attended referents (moment tn) are coded
  by subjects
• Activated referents (moment tn1) are coded by
  reduced NPs (pronouns)

39
Cognitive and linguistic interplay between
attention and WM Summary
40
Issue 3 Forgetting

• Trace decay theory Forgetting is a function of
  time
• Interference theory Forgetting is a result of
  displacement by new incoming information

41
Linguistic data are compatible with the trace
decay theory

• Activation decreases as distance to the
  antecedent becomes greater
• The limit on the number of concurrently activated
  referents can be explained by WM capacity
  limitations
• The balanced system of activation factors
  activates and deactivates referents in accordance
  with WM capacity limitations

42
Conclusions

• Capacity of WM for referents is severely limited
  (3 to 4 times maximal activation of a single
  referent)
• Referents enter WM through the mechanism of
  attentional control
• Referents can be forgotten from WM through the
  mechanism of decay

43
Really final conclusions

• Linguistic discourse analysis can indeed
  contribute to explorations of the human cognitive
  system
• It is the time for a close cooperation between
  linguistics and psychology in the study of
  cognition