CCCT05-austin - Derek J. Smith

blabbedharborAI and Robotics

Feb 23, 2014 (3 years and 5 months ago)

87 views

The Problem of Context in
Sentence Production


Surely A Case to Re
-
Convene the
Data Base Task Group?


Derek J. SMITH

Centre for Psychology

University of Wales Institute, Cardiff

smithsrisca@btinternet.com

http://www.smithsrisca.co.uk


As presented to the

3
rd

International Conference

on Computing, Communications, and
Control Technologies

Austin, TX, Wednesday 27
th

July 2005



A BIT MORE ABOUT THE AUTHOR


1980s
-

specialized in the design and operation of very large DBTG
databases.



Since 1991 taught cognitive science and neuropsychology to Speech
and Language Pathologists.



Hence interdisciplinary in database, cognitive neuropsychology, and
psycholinguistics.

THE PROBLEM


Information systems love duplicating your data. E.g. duplicate
postings to a transaction file, duplicate entries in a master file, entire
duplicated files (try moving house, and see how long it takes for the
old address to stop being used!)



To help cure this problem,
by the late 1950s steps had been taken to
specify organizational data more accurately using “data dictionaries“
and “data models".



This accumulation of "metadata"
-

data about data


was then used to
specify a central shared
-
access data “base”, and the software
products which managed the whole process became known as
"database management systems", or "DBMSs".



And yet those not directly involved with DBMSs know little about their
technical construction, evolution, or how to design or operate them
effectively (Haigh, 2004).

THE PLAN OF ATTACK


This paper concerns itself with the network database.



It reminds us of the history, not just of the database itself, but of the
whole idea of associative networks …..



….. and then considers the trans
-
disciplinary relevance of the
underlying concepts and mechanisms to the science of
psycholinguistics,
because they might just help solve some of that
science's long
-
standing problems.



So history ..... networks ..... words in the mind .....



and always in the nomansland between pure IT and pure psychology



and if nothing else, you will at least learn the difference between
speech and language!



NETWORK DATABASE HISTORY (1)


The story of the network database begins in the early 1960s at the
General Electric Corporation's laboratories in New York, where
Charles W. Bachman

had been given the job of building GE a DBMS.



The resulting system was the "Integrated Data Store" (IDS), and was
built around a clever combination of two highly innovative design
features, namely (1) a “direct access” facility similar to IBM’s
acclaimed
RAMAC
, and (2) Bachman’s own
"data structure diagram"

(soon to become famous as the "Bachman Diagram")
.



This is how the direct access part of the equation is implemented …..

NETWORK DATABASE HISTORY (2)


Bachman Diagrams prepare
your data for maximum
usability by analyzing it on a
set owner/set member basis.



Owner records are stored using
the direct access facility, and
their related members are
identified using chain pointer
addressing.



"Via
-
clustering" is often used to
keep member records
physically close to their
owners.
(This cannot always be
done, but is very efficient in
disc accesses when it can be.)

NETWORK DATABASE HISTORY (3)


And here, from
Maurer and Scherbakov (2005, online)

is a typical owner
-
member set (left) and the corresponding Bachman Diagram (right) …..

NETWORK DATABASE HISTORY (4)


Bachman had the IDS prototype running early 1963, and by 1964 it was
managing GE's own stock levels.



Initial user feedback was so positive that the Bachman
-
GE approach
soon came to the attention of CODASYL, the committee set up by the
Pentagon in May 1959 to produce a general purpose programming
language.



However, CODASYL had published the specifications for COBOL in
January 1960, so it predated IDS by four years and had accordingly
not been designed to support the particular processing requirements
of DBMSs …..

NETWORK DATABASE HISTORY (5)


In fact, it was so difficult for COBOL to implement IDS's chain pointer
sets (or "lists") of records, that in October 1965 CODASYL established
a List Processing Task Force (LPTF) to look into possible
improvements to the specification.



The LPTF meetings immediately became so dominated by database
issues in general that they renamed themselves the
Data Base Task
Group

(DBTG).



We may thus refer to IDS as a “DBTG database”, a “CODASYL
database”, or a “network database”. All these terms are synonymous
and used inter
-
changeably in the literature.

NETWORK DATABASE HISTORY (6)


A curious turn of events then saw IDS development taken over by one
of GE's early customers, the B.F. Goodrich Chemical Corporation.
They had been highly impressed with IDS, but wanted greater
functionality, so they bought the rights to develop an IBM version.



By 1969, Goodrich were able to market their improved system in its
own right, badging it as the "Integrated Database Management
System" (IDMS). The new product was heavily deployed in the 1980s,
and survives to this day as Computer Associates' CA
-
IDMS, in which
incarnation it continues to power many of the world's heaviest duty on
line systems.



Bachman was awarded the 1973 A.C.M. Turing Award for his
achievements …..

NETWORK DATABASE HISTORY (7)


….. however, it’s just possible that Bachman had actually been
cleverer with his network memory technology than even he or the
Turing Awards panel realized …..



….. because
in historical terms the idea of associative memory
underlies much of psychology as well. Indeed, it dates all the way
back to the classical Greek philosophers.



So now for some history from a totally different discipline …..

THE BIRTH OF COGNITIVE SCIENCE (1)


Aristotle had suggested back in around 350 BCE that memory was
based on the "incidental association" of one stored concept with
another.



That same general orientation went on to give its name to the entire
"Associationist" tradition of philosophy, culminating in Freud's
"association of ideas" technique of psychoanalysis and the modern
connectionist net and semantic network industries.



Data networks, in other words, are nothing new to students of the
mind, and in this paper we are going to select one application in
particular for attention …..

THE BIRTH OF COGNITIVE SCIENCE (2)


….. namely early attempts at machine translation (MT).



It was one of these early "computational linguists"
-

Cambridge
University's
Richard H. Richens
, plant geneticist by profession but
self
-
taught database designer into the bargain
-

who first coined the
popular modern term "semantic net" (Richens, 1956, p23).



Richens was actually half of an important research partnership in the
history of cognitive science.
Just after the war he had toyed with
Hollerith technology to help him analyze his genetics research data,
and had ended up with a rudimentary punched
-
card database.



This experience convinced him that with the right arrangement of data
and greater processing power it ought to be possible to automate
most anything, including natural language translation. So he began
designing the card layouts for a bilingual machine dictionary


making
him arguably the first database designer?


THE BIRTH OF COGNITIVE SCIENCE (3)


Richens then discovered that his enthusiasm for MT was shared by a
University of London crystallographer named
Andrew D. Booth
,
himself something of an expert in computers. During WW2, Booth had
been a “boffin” in the rubber industry, x
-
raying slices of rubber from
destroyed enemy aircraft and vehicles. And because
X
-
ray
crystallography generates a lot of numbers, B
ooth had built
calculating machinery to assist him. He had continued this work when
he got a peacetime lectureship at Birkbeck College.



This research had then brought him to the attention of the US National
Defense Research Committee's
Warren Weaver
. Weaver duly met with
Booth on 8th March 1947 while the latter was on a fact
-
finding visit to
the University of Pennsylvania's Moore School of Engineering.



Weaver was so enthusiastic about what Booth had to say that he used
his influence to put him forward for a study scholarship under John
von Neumann at Princeton's Institute of Advanced Studies.

THE BIRTH OF COGNITIVE SCIENCE (4)


Booth was at Princeton from March to September 1947, and upon his
return to Britain proceeded to build a small relay [i.e. electro
-
mechanical] computer, complete with one of the first magnetic drum
memories (10 years before IBM’s RAMAC).



A chance meeting of minds then changed the world. It took place
between Richens and Booth on
11th November 1947

(Hutchins, 1997),
and focused on the pair's shared interest in MT.



They concluded that Booth's magnetic drum might provide the sort of
random access technology needed to host Richens' proposed lexical
database
-

giving them 15 years prior claim to Bachman's basic IDS
architecture.

THE BIRTH OF COGNITIVE SCIENCE (5)


T
here followed a decade of collaborative research during which this
-

and eventually many other teams


found out just how complicated
natural language really was!



To start with, MT took the scientific world by storm, with the first MT
conference being organized at MIT by Yehoshua Bar
-
Hillel (an MT
skeptic). This took place 17
-
20th June 1952.



Centers of academic excellence soon emerged at MIT (Victor Yngve),
Washington (Erwin Reifler), and Berkeley (Sydney Lamb). Britain's
effort was concentrated at the Cambridge Language Research Unit,
under Margaret Masterman, where the researchers included Richens
himself, Frederick Parker
-
Rhodes, Yorick Wilks, Michael Halliday, and
Karen Spärck Jones.

THE BIRTH OF COGNITIVE SCIENCE (6)


In short, this was interdisciplinary science at its best, and its target


language
-

lay at the very heart of cognition.
We therefore date the
birth of cognitive science to that foggy November 1947 meeting
between Richens and Booth.



Semantic networks are now a major research area within AI (for an
excellent review, see Lehmann, 1992).

OUR PROBLEM AND OUR PLAN


However, as an IDMS designer
-
programmer turned cognitive scientist,
our personal complaint is that network researchers typically ignore
the explanatory and practical potential of the network database.



To help restore the balance, the present paper will explore how IDMS
concepts might help with one of cognitive science’s most troublesome
problems, that of context in speech production.



So let us move away from all the history and look at some modern
psycholinguistics.
Specifically, we need to look at the staged cognitive
processing which takes place during speech production.



WARNING
: “Language” and “speech” are
-

crucially
-

NOT THE SAME
THING, as we shall shortly be seeing.


SPEECH PRODUCTION STAGES (1)


The notion that voluntary speech production involves a succession of
hierarchically organized processing stages may be seen in a number
of influential 19th century models of cognition, but the subject was
largely ignored until UCLA's
Victoria A. Fromkin

reawakened interest
in it in the early 1970s (Fromkin, 1971).



Fromkin proposed six processing stages.
The first three stages
constitute the
language

part of the speech and language equation,
while the latter three provide the
speech

to go with it.



Reassuringly, there is virtual unanimity amongst authors ancient and
modern as to where in the overall scheme of things to place the bulk
of the semantic network …..



….. you simply attach the semantic network to the command and
control module at the top of the cognitive hierarchy, to serve as that
module's resident knowledge base.

SPEECH PRODUCTION STAGES (2)


The result is a mental
champagne
-
cascade …..



….. with ideas pouring down
from the top …..



….. words being added on the
way down …..



..... the ideas plus the words give
you your
language

.....



….. sounds being added below
that .....



..... and “linear”
speech

emerging
at the bottom.

SPEECH PRODUCTION STAGES (3)


This diagram is from Ellis
(1982) and shows how
psycholinguists typically
summarize the flow of
information between cognitive
modules.



Click here

to see full sized
diagram and
here

for a
detailed explanatory
commentary.

SPEECH
PRODUCTION


STAGES (4)


Here we see the speech
production (lower left) leg of
Ellis (1982) in close
-
up.



Note the three successive
modules. Fromkin’s six stages
map roughly two each onto
these hierarchically separated
processing levels …..

SPEECH PRODUCTION STAGES (5)

STAGE #1
-

PURE IDEATION


Stage #1
-

Propositional Thought:

This is the selective activation of
propositions within the semantic network, as part of the broader
phenomenon of reasoning,
and it is vitally important to students of the
mind because it e
stablishes the semantic “context” for whatever
happens next,
and especially the use and interpretation of words
.



This stage is known by Associationist epistemologists as
"ratiocinative" thought.

SPEECH PRODUCTION STAGES (6)

STAGE #2
-

SPEECH ACTS


Stage #2
-

Speech Act Volition:

This is where a carefully selected
subset of the aforementioned stream of propositions is converted into
a "speech act" of some sort.



Speech acts are preverbal linguistic manipulations of the social
environment, each calculated to achieve some discrete behavioral
effect.



Fully functioning adult humans have a repertoire of around 1000
different speech acts to choose from (see Bach and Harnish, 1979, for
a fuller list).



I
n the Chomskyan sense, speech acts give us much of our "deep"
sentence structure.



This structure is what gets passed down to Stage #3, thus interfacing
the original thought with the spoken word.

SPEECH PRODUCTION STAGES (7)

THE POINT ABOUT SPEECH ACTS


Because it is the final outcome which matters, speech acts are free to
generate sentences which use words ironically or figuratively.



E.g. such everyday phrases as



"when you have a moment" (i.e. now)



and




"if you don't mind" (i.e. whether you do or not)
.

SPEECH PRODUCTION STAGES (8)

THE POINT ABOUT ENCODING


Note very carefully that all the mental content we have talked about so
far has been NONVERBAL.



In fact, you should think of it as encoded in “images”, “icons”,
“sprites”, “ideograms”, etc.,
both concrete and abstract
.



This is very awkward in practice, because you usually end up having
to describe in words something whose very essence is that the words
haven't yet been selected.

SPEECH PRODUCTION STAGES (9)

STAGE #3


“LEXICALIZATION”

(REPLACING IDEAS WITH WORDS)


Stage #3


Word Finding:

The deep structure produced by Stage #2 is
now passed block by block (grammarians call them "phrases") down
the motor hierarchy.



Stage #3 determines the surface words to be used and how they will
need to be combined syntactically. Identifying the “agent” of a
sentence is particularly vital. For example, consider the ideation
“<IDEO = Fido> <IDEO = bite> <IDEO = Derek> <SPEECH ACT =
warn>

.



If you get the agent
-
object relationship confused, then the sentence
“Derek bites Fido”

will be just as likely to occur as
“Fido bites Derek”
.

CONTEXT IN SPEECH PRODUCTION (1)

THE PROBLEM OF PRONOUNS


There is an even bigger problem with pronouns, thus .....


“Fido is going to bite Derek”


“Fido is going to bite
him




He

is going to bite Derek”



He

is going to bite
him




Context allows the most appropriate NOUN
-
PRONOUN option to be
selected,
hence the process is
highly sensitive to the prior state of the
concept network, IN BOTH SPEAKER AND LISTENER
.



Indeed, it is fair to say that it is the mind’s context maintenance
mechanisms


whatever they are


which allow everyday conversation
to rely so heavily on what is NOT being said!


CONTEXT IN SPEECH PRODUCTION (2)

THE PROBLEM OF DEIXIS


The use of language to point in some way at a thing referred to is
known as
"deixis"
. Here are some examples of its subtypes .....



Example 1
:

"It is bad enough when
it

might have been mentioned
many words beforehand, but you also get
“forward deixis”
, where
the
referent

is still to come.



Example 2
:


They

have particular problems with pronoun deixis,
MT
programmers
, because
they

have to work out
-

occasionally from
phrases not yet spoken

-

what they are supposed to be translating.”



Example 3
:

"You also get
non
-
explicit deixis
, where the referent is left
to establish itself without specific mention, as in
'
They

are out to get
me' ”.


A CROSS
-
DISCIPLINARY EXPERIMENT


So what would happen if we used IDMS
-

a network architecture by
design
-

to implement the knowledge network at the top of the speech
motor hierarchy?



Would its systems internals be able to cope (where rival systems have
not) with the combined load of philosophical, psychological,
psycholinguistic, and linguistic problems?



Specifically, might it help machines master
language

as well as
speech
?



Well it is going to take a sustained research effort to answer these
questions fully, but the DBTG metaphor certainly promises much in
three important areas, as follows .....

DBTG PROMISE #1

HASH RANDOM ADDRESSING


The IDMS hash random
facility would be ideal for
storing noun concepts such
as
<IDEO = Fido> and <IDEO
= Derek> …..



….. giving us cumulatively
our personal knowledge
base.

DBTG PROMISE #2

CHAIN POINTER ADDRESSING


The IDMS chain pointer
facility is already ideal for
implementing Bachman’s
logical data sets, weaving
the individual data
fragments into a complex
yet "navigable" lattice.



Chain pointers thus give
more than two millennia's
worth of philosophers their
associative network.

DBTG PROMISE #3

SET CURRENCY ADDRESSING


Perhaps more importantly
still, the IDMS device known
as the "set currency" does
for the DBTG database what
calcium
-
modulated synaptic
sensitization appears to be
doing in biological memory
systems (Smith, 1997).



Biological set currencies
allow specific memories to
be sustained up to an hour
after first activation. E.g the
pronoun
“him”

in the earlier
example can point to one
noun in particular out of
potentially many tens of
thousands.


DBTG PROMISE #3

A BIOLOGICAL SET CURRENCY?


Readers may simulate the phenomenon of memory sensitization right
now by trying to recall the year of Bar
-
Hillel's MIT conference. You
have ten seconds …..

DBTG PROMISE #3

A BIOLOGICAL SET CURRENCY?


The year in question


1952
-

is perhaps ten minutes of listening time
ago,
but its “engram”


its memory trace
-

is nonetheless still in a
raised state of excitation

…..

DBTG PROMISE #3

A BIOLOGICAL SET CURRENCY?


….. It is long
-
term memory
left "glowing" in some way
by the original activation.



This is possibly the
mechanism of maintaining
referential context and
supporting deixis over time
-
extensive thought or
conversation.



Click here
for a more
detailed introduction to the
biochemistry of memory.

CONCLUSION (1)


We have been considering the trans
-
disciplinary relevance of the
concepts and mechanisms underlying the DBTG database to the
science of psycholinguistics.



Our central complaint was that despite a long tradition of semantic
network simulations in computational linguistics none of the
established research technologies really implements a network data
model as a network physical form. Instead, they prefer to keep the
physical storage relatively simple, typically in a "flat file" format.



By contrast, the only architecture which has ever been able to cope
with volatile data
in bulk

is the DBTG architecture.
This is because it is
largely self
-
indexing, often via
-

clustered, and uses pre
-
allocated
expansion space.
(
This is precisely why CA
-
IDMS is still supporting
the heavy end of the world's OLTP industry, despite repeated attempts
to dislodge it.)

CONCLUSION (2)


Our humble
(and not entirely tongue
-
in
-
cheek)

proposal is therefore that the DBTG
-

having
delivered on behalf of the volatile data industry in
the 1960s
-

now needs to be reconstituted in the
interests of a better understanding of the mind
-

the
ultimate database.



We are ourselves currently researching the nature of
the interdisciplinary collaboration which such an
exercise would involve.


REFERENCES


Bach, K. and Harnish, R.M. (1979).
Linguistic Communication and Speech Acts.

Cambridge, MA: MIT Press.


Fromkin, V.A. (1971). The non
-
anomalous nature of anomalous utterances.
Language,

Vol. 47, pp. 27
-
52.


Haigh, T. (2004). A veritable bucket of facts. In M. E. Bowden and B. Rayward (Eds.),
The History and Heritage of
Scientific and Technical Information System
, Medford, NJ: Information Today.


Hutchins, W.J. (1997). From first conception to first demonstration.
Machine Translation
, Vol. 12, No. 3, pp. 195
-
252.


Lehmann, F. (Ed.) (1992).
Semantic Networks in Artificial Intelligence
. Oxford: Pergamon. [Being a special issue of
the journal
Computers and Mathematics with Applications
, 23(2
-
9).]


Maurer, H. and Scherbakov, N. (2005, online).
Network (CODASYL) Data Model
. [Electronic document retrieved
17th July 2005 from
http://coronet.iicm.edu/wbtmaster/allcoursescontent/netlib/ndm1.htm
)


Richens, R.H. (1956). Preprogramming for mechanical translation.
Mechanical Translation
, Vol. 3, No. 1, pp. 20
-
25.


Smith, D.J. (1997).
The IDMS Set Currency and Biological Memory
. Cardiff: UWIC. [ISBN: 1900666057]
[Workbook to support poster presented 10th March 1997 at the Interdisciplinary Workshop on Robotics, Biology,
and Psychology, Department of Artificial Intelligence, University of Edinburgh.]