PUS, A Paradigmatic Language Acquisition System: An Overview
Proceedings of the Sixth International Joint Conference on Artificial Intelligence. .
© 1979 International Joint Conferences on Artificial Intelligence
Abstract
This paper informally describes the capabilities of a teachable analogr-baaed languageindependent natural language acquisition system known aa PW. Ita language comprehension ia intended to resemble that of pre-school children in certain significant aspects. It can be taught, through examplea, to underatand and acquire the situational aspects (i.e., physical objects, agents, tbeir aensori-110tor properties and apatial relations) described in a text. It has no built-in knowledge of English or the domain of discourse. Neither it infers any formal grammar of English.
- Sembugamoorthy, V. 1 A Paradigmatic I.enguage Acquisition System', Ph.D.Thesis, Indian Institute of Tech., Bombay, India, Ma7, 1978.
- Narasimhan, R. 'Modelling lenguage Behaviour', Tech.Report, Computer group, TIFR, April,l973.
- Jordan, R.S. 'A Natural lenguage Understand- er Based on a Freely Associated Learned Memory Net', Int.J.Comput.Inf.Sci.6:1 (1977) 10-25.
- Reeker,L.H. 'The Coarputational Study of Language Acquisition' in Advances in Computers,
- Sikloss,.,L. 'Natural lenguage learning by ~omputer' in Represenlation and ~aning: I!P.!riments with Information Process~teiiiiS, Prentice-Hall, NJ, 1972, pp.288-328. -- (61
- Harris, L.R. 'A System for Primitive Na{ural Language Acquisition', Int. J.Man-Mach. Stud.9 (1977) 153-206.
- Moore, J. and Newell,A. 'How can Merlin Understand?' in Know led&'!._ and -~it ion, Gregg, L. (Ed), Lawrence Erlbaum Assoc. Poto11100, Hd., 1973.
References were extracted automatically from the PDF and may contain errors.
Full text extracted automatically from the PDF; figures, tables and equations may differ. The PDF is the version of record.
1. lHTROOOCTiaf
This paper infonall.7 describe!! the capabilit iea of a primitive natural language acquis it ion syatem known as PIAS. The versien reported here baa been imple·nted asing UCI LlSP on the IEC-lD s,.st .. at the BCSOO'l', 'l'IFR, Bomba:r. Its linguistic co~~prebensien is intended te resemble that ot a pre-school child in certain significant aspects. PIAS can be taaght, through examples, to understand and acquire language utterances representing pb)'S icel objects, agents, their senseri-110tor properties (e.gs colour, beigbt, etc.) and spatial relations (a.gs on the chair, to the lett ot 1 etc.). To this end, P!AS has to acquire and use the knowledge to ana178e and interpret linguistic utterances, anderstand profor.s, obtain the subject ot a santence,and identifJ and associate the sensori-.ator pro pertiea and relations with the subject.
PIAS differs significantl:r troa the reported language acquisition ayftelll8 (3, 4, S, 61 in one or more of tbe following aspectss Pf.As is neither baaed on an:r built-in knowledge of English nor designed to acqdre an:r formal p-a1111ar of English. The language understanding and generation procedures are analogr-based. The primit bee of the internal representation ot language behaviour and world knowledge are either acquired through nataral lanpuage inter action, or built-in and independent ot an:r language 1 task or de·in·
Within ita scope, PLAS provides a solation to tbe iaane of assimilation and acco..odation raiaed in (7).
The .. chiner,r ot PLAS baa been JDOtivated primaril:r bJ oar long-term goal to model child language behaviour. llevertbelaaa,the computational 111achinery should be useful towards constracting practical expansible nataral language interaction syetelll8. Refer to (1, 2] tor the pBJCbological aspects of PLAS, the explicit substantiation of ita features mentioned aOO.e and the detailed discussion ot ita co~tational machiner,r.
2. AN EXAMPlE OF IBTERACTIDN
Initially PIAS does not possess en:r knowledge or English or the domain or discourse. Using a simulated teaching environment called 'Teach names' 1 the teacher first associates a set of names with tbe physical objects and agents that he proposes to deal with in his interaction. Bote that neither all names of an object/agent nor all objects that a word or a phrase can name need be aaaociated using 'Teach-names·. A tJpieal input to 'Teach-names· is:
- CHAm 1, CHAIR2; CRAm; (OBJECT MAlE-OFWOOD).
This means that the language expression 'CHAm· ia a name of the pb.7aical objects 1CHAml' and 'CHAIR2 1 ; 1 (OBJECT MAlE-OF-WOOD)' is called a taglis:t,. It is intended to simulate the structured complex or sensori-motor mechanisms which is responsible tor the perception/mani pulation ot 1 CHAml' and 1CHAIR21 · However, since articulating this structure is itself a challenging open problem, a taglist is present ly assumed to be an unstructured set of tags where each tag relates to a sensori-1110tor mechanism. As we shall see in aection ~' the taglist plsys an i~rtant role in understand ing linguistic utterances. Moreover, it is significant to note that the language expressions used to name tags (e.g: HAnK-OF WOOD) do not tunctio~ as lsnguage expressions as tar as PLAS is concerned. let as assume that ··ing 'Teach-names·, the teacher has taught PLAS the following wordS and phrases:
2. TABlE, THE TABlE; ). WINOOW; 4. IESK; 5. ROOM, THE ROOM; 6. POUN'l'A IN, THE FOUNT A IN; 7. 1m; 8. SRI!: IF.
Rext, the teacher teaches, ·sing another teaching device called 'Teach-aspects·, a set ot words and phrases known as .!!,E!ct to~ as shown below (Aspect tokens are lsnguap ex pressions that denote properties of, and relstiona between, objects and/or agents):
- ~ (FitOWN; (COIOOR)). This me-ans that 1 mOWN· is an aspect token of a New Aspect CategorJ (HAC for short). '(COLOUR)' is -the tagliat associated with I mOWN' at the sensori-1110tor le'Yel.
10. (GIU!'.ER; (COlOUR)) SAC (moWN; (COlOUR)). The aspect tokens 1 GREBN 1 and 1 BROWN' belong to the ~ame A,apect [atepey (SAC) ·
ll. (RED; (COI.OOR)) SAC (GREEN; (COlOUR)). 12. (RED 1R COIOOR; {ColOUR)) S.AS (RED; (COlOUR)).
'RED IN COIDUR' and 'RBD' represent the Same A!Pflct (S.AS). let as aaau11111 that using rTeaoh aspects·, the teacher baa taagbt the following aspect tokens:
- PlASTIC; 14. WOOD; 15. DAR THE POORT.&IR; 16. IN THE MIDDlE OF THE P'OUN'l'AIR; 17. lNSIIB THB IESK; 18. Cl'l THE Im; lCJ. MAil!! OP' PlASTIC; 20. HEAVY 21. LIDHr.
Next, the teacher asea a teaching device called 'Teacb-Proforme' to teach proform tokens as shown below:
22. (THE CHAm NEAR THE TABlE) ASS (THE ONE NEAR THE TABlE). This 11111ana that 1THE CHAm NEAR THE TABlE' is an ASSignment (ASS for abort) of the proform 'THE ONE REAR THE TABIE 1 ·
- (THE CHAm) ASS (IT). Before we proceed turtber, it is worthwhile to point out that (i) in principle, 1Teach-namea 1 , 'Teach-aspects' and 'Teach-proforma' can be employed in an;y order and an;y n12mber of tiDIBBJ (ii) the teaching devices 1 Teacb-aspecta' and
1 Teach-proforme 1 are anrealistic from the Yiew point of 1110delling the learning environment of pre-school children and can be eliminated, as shown in ( 1 ], by incorporating, in PLAS, a question-answering environment.
Next, the teacher teaches a text (call Textl) as shown below:
24. ((A CHAm) REP' 1 CHAml 1 AND (A TABlE) REF 1 T ABIEl i) REP' (CO LIECTIONl AGO ffiml, TABIEl}ARE NEAR THE WlNOOW. This mean a that the tokens · A CHA m· , 'A TABlE' and 1 A TABU: .&RD A CHAm· refer to 1 CHAml·, 1 TABIE1 1 and 1 COILECTION1' respectbe1y; 1COLI.ECTION11 consists of 1CHAml1 and1TABIE11 · The sentence taagbt is 1 A CHAm ARD A TABlE ARE NEAR THE WlNOOW'. REP' and AGO, which stand tor 'reference-of' and 1aggregate-of1 , simulate the extra-linguistic means employed by the lsnguap community to associate a language expression with its referent, and to a_pecity the members of an aggregate of objects/agents respectiY8lJ. Besides interpreting these extra-lingwistic markers, PLAS bas to analyse and interpret 'NEAR THE WlNDOW1 using 15 i.e, step 15; identity 1 A CHAm AND A TABlE' ae the subject token; associate the aspect token 1NEAR THE WlNOOW1 with 1COLIECTION11 end thereby with 1 CHAllll 1 and 1 TABIE1 1 ·
25. THE CHAm lS BROWN IN COlOUR. Using 1 and 2 i.e., asing the knowledge acquired in 1 and 2, PLAS interprets 1THE CHAIR' as 1 CHAml·; acquires 1 BROWN IN OOIDUR' ·Bing 12; identitiea the subject token 'TIE CHAm' J associates 1 BROWN lN COlOUR' with 'CHAllll'.
26. THE TABU: lS RED IN COIDUR. PLAS uses 25 to underetand this sentence since it is similar to the one in 25.
- A Im1C ARD A CIV.:r;R ARE IN THE MIDDlE OF TH! ROOM.
1 A r:H..U'P.1 is interpreted as · f.IIA'!R'2' because 1 Cl!A~l 1 bas been associated with 'NEAR THE WlNOOW 1 in 24. 1 1N THE MIDDlE OP' THE ROOM1 is understood using 16.
28. THE BROWN CHAm IS MAlE OF WOOD. PIAS interprets 1THE BROWN CHAm· as 1CHAml· and not as 1CHAlll2 1 · Note that 1CHAIR11 baa been associated in 25 with · BROWN IN COIDUR'. PLAS ueea 12 to equate 1 BROWN IN COlOUR' and 1 BROWN' , and to correctly interpret 'THE BROWN CHAm· as 1CHAml1 ·
- THE ONE :m THE MIDDlE OF THE ROOM IS MAlE OF PLASTIC. PIAS assigns 1THE CHAIR lN THE MIDDlE OF THE
ROOM' to the proform 'THE ONE lN THE MIDDlE OF THE ROOM' · Note that the teacher has taught, in 22, only the assignment 'THE CHAlR NEAR THE TABlE' · Moreover, PLAS has to use 1!:7 to inter pret 1 THE CRAlR IN THE MIDDlE OF THE ROOM' as 1 CHAlR21 and not as 1 CHAlR1 1 ·
The teacher next tests PLAS b.1 asking it to understand a similar new text. For example,
~ THlS lS A CHAlR. THERE lS A OOX ON THE CHAlR. IT lS HEAVY. THE OOX JNS IIE THE SHElF lS LIGHI'. THE HEAVY OOX lS GREBN lN COIDUR. Note that PLAS can understand Text2 even though the underlying sentential forme and the described situational aspects vary significant !,. from those of Textl. Moreover, PLAS should interpret 'IT' in the third sentence as the box referred to in the second sentence even though the teacher has associated only 'THE CHAlR' with 'IT' in 23. Now, as requested b.1 the teacher, PLAS generates the following sentences using 1CHAlR1 and 1 RED 1 :
THE HEAVY CHAlR lS RED IN COIDUR. THE LIGHI' CHAlR lS RED lN COIDUR. THE RED CHAlR lS MAIE OF WOOD. THE RED CHAlR lS MADE OF PLASTIC. THE CHAlR IS RED lN COIDUR.
Note that none of these sentences appears in Textl or Text2. Moreover, PLAS has not generated any semantically anemoloaa sentence like 'THE RED CHAlR IS GREEN lN COLOUR'.
3. THI!: COMPUTATIONAL MACHJNERY
Interaction procedures simulate the teaching and testing environments discussed in section2. They invoke sensori-motor~f!dureJ! and lan~&! behaviour .J!rocedy..,!! to process the sensori-motor level information (i.e,pqysical objects, taglists, extra-linguistic markers) and the linguistic express ions respectively. In what follows we shall sketch only the language understanding procedure. To under stand a language expression, say 1 0N THE TABL!', PLAS first invokes the !!_!g!!nta_iion pr~cedure to analyse it into a known language schema, sa,. 'on -' ( 1 - 1 stands for a variable) and the filler 'THE TABlE'; 'THE TABlE' itself has to be understood before it is accepted as a filler. The segmented language expression :lJ! then va!!dat_~. For instance, if 'NEAR THE CHAlR' is an already known valid filler for thP second variable of the schema '- lS -', then 'ON THE TABLE' is accepted as a valid filler for the same variable provided the taglists associated with the two aspect tokens match (presently two tagliats match iff they contain .identical tags). Tb! .. !nt!:x:.P:r:.~t_atio~cedure then interprets the validated token. For example, using the knowledge that the referent of 1 J0Ifl 1 S OOOK' :lJ!
aa7 1 000Kl0 1 , 1 JILL 1 S T.ABIE 1 will be interpret ed as, say 1 TABIE5 1 · A successfull7 segmented, validated and interpreted language expression is said to be \lnderstood. If the schema ·On -' is not known already, it will be abstracted out of it by the ~matization pr~ed~ b.1 replacing understandable or already known language expressions in it. In this case also, before 'ON THE TABlE' is declared understood, it has to be validated and interpreted. Note that what underlies the language understanding pro cedure is not any language-specific information but are the capabilities to assimilate the unknown in terms of the known, and to acco111110date the unknown if assimilation fails. These capabilities are analogy-based. To enable these capabilities is the main function of the data spaces of PLAS.
ACKNOWlEDGEMENTS
The valuable guidance of Prof.R.Narasimhan is gratefully acknowledged.
REFERENCFS
- [1) Sembugamoorthy, V. 1 A Paradigmatic I.enguage Acquisition System', Ph.D.Thesis, Indian Institute of Tech., Bombay, India, Ma7, 1978.
- [2) Narasimhan, R. 'Modelling lenguage Behaviour', Tech.Report, Computer group, TIFR, April,l973.
- [3) Jordan, R.S. 'A Natural lenguage Understand er Based on a Freely Associated Learned Memory Net', Int.J.Comput.Inf.Sci.6:1 (1977) 10-25.
- (41 Reeker,L.H. 'The Coarputational Study of Language Acquisition' in Advances in Computers, Rubinoff, H. , and Yov its, M.-C., (Eds) Academic Press, 1976, pp.lBl-235.
- f5) Sikloss,.,L. 'Natural lenguage learning by ~omputer' in Represenlation and ~aning: I!P.!riments with Information Process~teiiiiS, Prentice-Hall, NJ, 1972, pp.288-328. --
- (61 Harris, L.R. 'A System for Primitive Na{ural Language Acquisition', Int. J.Man-Mach. Stud.9 (1977) 153-206.
- Moore, J. and Newell,A. 'How can Merlin Understand?' in Know led&'!._ and -~it ion, Gregg, L. (Ed), Lawrence Erlbaum Assoc. Poto11100, Hd., 1973.
@inproceedings{ijcai1979p173,
title = {PUS, A Paradigmatic Language Acquisition System: An Overview},
author = {Sembugamoorthy, V.},
booktitle = {Proceedings of the Sixth International Joint Conference on
Artificial Intelligence, {IJCAI-79}},
publisher = {International Joint Conferences on Artificial Intelligence},
editor = {Bruce G. Buchanan},
year = {1979},
month = {8},
url = {https://ijcai.org/proceedings/1979/173},
}
TY - CPAPER TI - PUS, A Paradigmatic Language Acquisition System: An Overview AU - Sembugamoorthy, V. T2 - Proceedings of the Sixth International Joint Conference on Artificial Intelligence, IJCAI-79 PB - International Joint Conferences on Artificial Intelligence PY - 1979 UR - https://ijcai.org/proceedings/1979/173 ER -Copied
Machine-readable versions of this paper, for tools and AI agents: