A Computer Model of Child Language Acquisition

ABSTRACT

Children learn d~fferent aspects of language in a characteristic order, . and make cnaracteristic er­ rors during acquisition. This paper focusses on five specific data, and explains these data in terms or two hypotheses regarding the relationship of comprehension to generation and the relation­ ship between meaning and syntax. A computer model, CHILD, is described wnich embodies these hy­ potheses and which manifests the same five data. CHILD's performance suggests that the explanations are plaus~ble.

~ INTRODUCTION

The question of why children's language ac­ quisition has the distinct characteristics it does is an important question for the development of computer models of language learning. This paper considers five specific characteristics of child language learning. It proposes explanations for them based on a computer model of language learn­ ing, CHILD, which displays those characteristics.

These characteristics involve the order of acquisition of various aspects of language, and errors producea during development. First, comprehension of certain aspects of language pre­ cedes the generation of those aspects [1]. Second, generation is thought to develop from shorter utterances to longer utterances [3]. Third, knowledge of language meaning precedes knowledge of language syntax [12]. Fourth, children make comprehension errors when the intended meaning of an utterance differs from the semantical ly most likely meaning. [8,5]. Fifth, children misunder­ stand utterances whose syntax indicates an in­ terpretation different from that which might be obtained using a semantics-only interpretation. In such cases, younger children will respond as if they produced the semantics-only interpretation, while older children will respond correctly using syntactic knowledge [11].

In order to address these issues, this paper explores two hypotheses: that knowledge of language syntax is indexed under knowledge of language meaning, and that learning to generate language is driven by learning to understand language. These hypotheses imply that what chil­ dren learn during language development is the con­ tents of their lexicon, and that the Processes of comprehension, generation, and inference are not

learned.

A COMPUTER MODEL OF CHILD LANGUAGE ACQUISITION

Mallory Selfridge

Department of EE and CS University of Connecticut Storrs, CT. 06268

The first hypothesis, the "semantically­ indexed syntax hypothesis", is focussed around the representation of word meaning as conceptual case frames. The process of understanding is seen as the problem of inferring the filler for a case slot in the meaning or one word from among the meanings of the other words in the utterance. Syn­ tax functions in situations in which semantic in­ formation is insufficient, to supply positional information as to where in the utterance a filler might be found. The process of generation is seen as finding a word to express the top level case frame of a concept, and finding other words to ex­ press case fillers not spanned by the meaning of the first word. The purpose of syntax in genera­ tion is to specifY the appropriate position of the words to express the fillers relative to the word expressing the case frame, and relative to words expressing other fillers. The semantically-indexed syntax hypothesis assumes that comprehension and generation both access the same word meanings and syntactic knowledge. It states that syntactic knowledge is stored associated with a slot in a

case frame which is the meaning of a word.

The second hypothesis, the "comprehension­ driven generation hypothesis", assumes that the principle effect of learning to understand is the addition to the lexicon of the meanings of new words, the refinement of the meanings or words al­ ready in the lexicon, the addition or syntactic knowledge to the meanings or words in the lexicon, and the refinement of that syntactic knowledge. It states that the ability to generate language grows as the lexicon develops during the development of

comprehension.

These two hypotheses are certainly not the only factors in language development, but they may help explain certain aspects of acquisition. I n particular, a computer model of language acquis~­ tion which incorporates them manifests the same five characteristics or child language development describea above.

ll ~ .cli.1J.JL PROGRAM

CHILD is a computer model ot· the development of language comprehension and generation abilities in a child, Joshua, between the ages ot· one ana two years. CHILD begins with the kind of knowledge that Joshua had at age 1 when he understood no language, and learns to understand commands in­ volving action, object, and spatial relation words at Joshua's age 2 level. It also learns to gen­ erate descriptions or the actions of others it "observes" which are appropriate to Joshua's age 2 level. CHILD learns by being given the kind of language experiences Joshua had between the ages of 1 and 2, and making use of rules to a) infer the meaning of utterances, b) attend to words, and c ) learn language meaning and syntax. While CHILD only comprehends imperatives, and only generates declaratives, these are a significant subset of Joshua's language abilities at age 2, and thus are appropriate subjects for computer modelling. Fur ther, CHILD focusses only on modelling how Joshua learned the words for concepts which he al­ ready knew; since at age 1 his conceptual reper­ t oire was large, this appears reasonable.

The first component or a computer model of t he development of Joshua's comprehension is Joshua's knowledge prior to his language learning. Observations of Joshua, reported in [10), suggest that Joshua had considerable knowledge or objects, actions, spatial relations, and gestures at age 1. CHILD contains such knowledge represented in Con­ ceptual Dependency (CD) [9).

CHILD models Joshua's knowledge of word mean­ i ng and syntax by maintaining a "dictionary" con­ taining words and their associated meaning and syntax. Meanings are represented as CD struc­ t ures , while syntactic knowledge is represented using "sequential structure" [10) which specifies t he location of the fillers for slots in CD con­

ce pts in CHILD's short-term memory.

The language understanding process which CHILD uses is based on an adult-level language understanding program called CA [2). CA operates by reading the words of a sentence one by one, lef t to right. When it reads a word, it places the CD concept representing its meaning on the CLIST, an immediate processing memory. Then, the semantic and structural requirements associated wi th the slots in the concepts from previous words are checked to see if the new concept can be used to fill any of them. Likewise, the requirements associated with the new concept are cnecked to see i f they can be filled by any of the concepts al­ ready on the CLIST. In this way, the meaning of the whole sentence is built up. When there are no more words, the resulting combination of concepts is made available to inferential process1ng, which uses more global world knowledge to combine any

uncombined concepts and fill any unfilled slots.

While children say things for many reasons [7), CHILD only models one aspect of Joshua's gen­ eration ability: his ability to describe ongoing actions of others. CHILD's generation component has been developed from an aault-level generator called CGEN [b). CGEN begins when a knowledge structure is placed on its CLIST, its immediate processing memory. When CGEN finds a concept on the front of the CLIST, it searches its dictionary to find a word whose CD meaning matches the con­ cept being generated. When it finds one, it places t his word on the front of the CLIST. When CGEN finds a word ·on the front of the CLIST, CGEN

"says" the word by printing it. If, however, the dictionary contains only a word whose meaning matches the root CD concept of the overall concept being expressed, then CGEN remembers that word on the CLIST, and inserts the CD concepts which filled slots in the root concept into their ap­ propriate places on the CLIST. The word which matched the root concept may have supplied struc­ tural knowledge to determine where in the CLIST the fill-concepts should be inserted. If not, their position is arbitrary. If CGEN fails to find a word whose meaning even partially matches the concept to be expressed, it generates "um". CGEN continues by examining the front of the CLIST, either printing the words it finds, or looking up words for the cocepts, Thus, when CHILD wants to generate a concept, it places that con­ cept on CGEN's CLIST, and CGEN proceeds to express that concept in English as well as it can.

In order to learn, CHILD must infer the mean­ ing of utterances it hears. Since at the age of 1 Joshua had knowledge of the function and proper­ ties of objects, knowledge of the meanings of ges­ tures, and knowledge of actions, CHILD has been given this knowledge, and uses it to infer utter­ ance meaning. If CHILD has learned some of the words in the utterance, it can also use its par­ tial understanding as it infers utterance meaning. Since the utterances it hears are imperative, CHILD can verify the inference by "behaving" and awaiting parental approval or disapproval. I" ord­ er to model Joshua's knowledge of such parental responses, CHILD has been given knowledge of the keywords "yes" and "no", and of such gestures as head nodding, head shaking, and smiling. If none of these are provided after a behavior, CHILD as­ sumes that implicit approval has been given.

For Joshua to learn language at his age 2 level beginning with his age 1 conceptual knowledge required the experiences he had between the ages of 1 and 2. These experiences were pri­ marily ones in which Joshua could infer the mean­ ing oi· much of the language he heard. Since Joshua received this kind of experiences, CHILD has been provided with similar experiences.

CHILD's strategies for learning word meaning and syntax are still under development; the current strategies have proven adequate for the tasks undertaken so far. The basic orientation of these learning strategies corresponds to learning rules proposed by a number of others [4,13).

CHILD takes the meaning of a word to be that component or the utterance meaning which is common to all occurrences of that word. This means that CHILD cannot learn words with multiple meanings. Further, it is assumed that the meanings of words appearing in the same utterance do not overlap. Thus if some words in the utterance are already known, then their meaning can be factored out or the meaning of the utterance as a whole and only the remainder considered as the meaning of the unknown word. Thus the first time a word is heard, it is given a meaning corresponding to the in­ ferred meaning of the entire utterance with mean­ ings oi known words factored out. Each further time a word occurs in an utterance, the meaning of that utterance is then "intersected" with the word's current meaning and the result is the word's new meaning, If the intersection is empty, then the word is considered meaningless and marked as such.

CHILD's syntax learning strategy require that a representation of the meaning of the utterance be obtained, along with a copy of the short-term memory used during the analysis phase, The posi­ tions of the words whose meanings fill slots in the meanings of other words are remembered, and the sequential structure of their position is stored under that slot or the meaning of the word in the word's definition.

lli. .c.lW.D.' .S. PERFOBMANCE

Most of CHILD was originally developed at Yale University, while current ·work proceeds at the University of Connecticut. The excerpts have been taken from a complete run of CHILD, and are presented in the order in which they appeared in that. Each excerpt represents CHILD at a dif­ ferent stage in development as it progresses from Joshua's age 1 abilities to Joshua's age 2 abili­ ties. Messages concerning CHILD's internal state, such as rules being activated, have been largely omitted, This run represents one possible develoP­ mental progression for CHILD. Other initial

.. nowledge or other language experiences would lead to a different progression.

At the onset, CHILD has only the kind of knowledge ascribed to Joshua age 1. In the excerpt below, the "parent" types a lowercase utterance to CHILD, which responds with a message stating its lack of understanding, When the parent provides gestures via simulated visual input, CHILD under­ stands, and prints the CD representation of its response.

IPARENT SAYS: give me the ball I ICHILD STARES BLANKLY AT PARENT ICHILD RETURNS TO PLAY I ICHILD SEES: (PARENT HOLDS OUT HAND) I (PARENT LOOKS AT BALL1) I ICHILD RESPONDS: I(ATRANS ACTOR (CHILD) OBJECT (BALL1) I TO (POSS VAL (PARENT))

Thus at the stage corresponding to Joshua at age 1, CHILD understands no language, though it will respond to gestures that it knows. At this stage its generation abilities are also poor. As the following example shows, CHILD has little to say about an action that it observes,

!CHILD SEES: (PARENT PUTS BALL1 ON TABLE1) I !INTERPRETING GESTURE AS: I(PTRANS ACTOR (PARENT) OBJECT (BALL1) I TO (TOP VAL (TABLE1))) I

ICHILD SAYS: um

Hence CHILD's initial stage is that in which it knows neither word meanings nor syntax. Its transition out of the first stage occurs when it begins to learn word meaning; for example the meaning of "ball." When CHILD knows the meaning of one word, it can demonstrate rudimentary under­ standing. In the excerpt below, the parent tells child to "give me the ball." CHILD understands only the word "ball", and must use its knowledge of possible actions that can involve balls to infer possible responses. The first response is not what the parent intended, so CHILD tries again, successfully.

IPARENT SAYS: give me the ball I ICHILD RESPONDS: I ( PTRANS ACTOR (CHILD) OBJECT (BALL 1) I TO (TOP VAL (TABLE1))) I I IPARENT SAYS: no I ICHILD RESPONDS: I(ATRANS ACTOR (CHILD) OBJECT (BALL1) I TO (POSS VAL (PARENT))) I

IPARENT SAYS: (smiles and nods)

CHILD at this stage can respond to words and generate responses on the basis of its understand­ ing of them, as shown in the following ·

ICHILD SEES: (PARENT PUTS BALL1 ON TABLE1) I !INTERPRETING GESTURE AS: I(PTRANS ACTOR (PARENT) OBJECT (BALL1) I TO (TOP VAL (TABLE1))) I !CHILD SAYS: ball

In the third stage, shown in the excerpts below, CHILD has learned the meaning of several words, and understands some utterances correctly. In this case, it has learned the words "put", "ball", and "box". Notice, however, that although it responds correctly to the first utterance given by the parent, it misunderstands the second. This sort of error is reported in [8]. Not knowing "on", CHILD incorrectly infers that the aP­ propropriate relationship is containment, based on its knowledge of the usual relationships that ob­

tain between balls and boxes.

IPARENT SAYS: put the ball in the box I ICHILD RESPONDS: I(PTRANS ACTOR (CHILD) OBJECT (BALL1) I TO (CONT VAL (BOX1)) I IPARENT SAYS: put the ball on the box I ICH!LD RESPONDS: I(PTRANS ACTOR (CHILD) OBJECT (BALL1) I TO (CONT VAL (BOX1))

CHILD can also use its increased knowledge of word meaning to generate better descriptions of

observed actions, as shown in the following:

!CHILD SEES: (PARENT PUTS BALL1 IN BOX1) I !INTERPRETING GESTURE AS: l(PTRANS ACTOR (PARENT) OBJECT (BALL1) I TO (CONT VAL (BOX1))) I

!CHILD SAYS: ball put CHILD's generative abillty 1ncreases w~tn tne learning of additional word meaning, In this case it has learned the additional object words "slippers", ·"piano", "ball", "block", and "table," CHILD can use these new words to generate better descriptions of parental action,

!CHILD SEES: (PARENT PUTS SLIPPERS1 IN BOX1) I I INTERPRETING GESTURE AS: l(PTRANS ACTOR (PARENT) OBJECT (SLIPPERS1) I TO (CONT VAL (BOX1))) I . !CHILD SAYS: box slippers put

Although CHILD has learned enough words to express certain concepts at the three word stage, it knows no syntax. It is now taught syntax for the word "put", and learns that the fillers of the OBJECT and TO slots are found FOLLOWing "put", and that the TO slot is found FOLLOWing the OBJECT filler. This new syntactic knowledge enables CHILD to gen­ erate the description of the observed action with the words in the correct order:

!CHILD SEES: (PARENT PUTS SLIPPERS1 IN BOX1) I !INTERPRETING GESTURE AS: I ( PTRANS ACTOR (PARENT) OBJECT ( SLIPPERS1) I TO (CONT VAL (BOX1))) I !CHILD SAYS: put slippers box

The transition to the next stage occurs when CHILD learns the meaning of ·on," This allows it to understand the unlikly utterance, "Put the slippers on the piano", whereas at any earlier stage it would not have understood this correctly.

This is shown in the following example.

!PARENT SAYS: put the slippers on the piano I !CHILD RESPONDS: l (PTRANS ACTOR (CHILD) OBJECT (SLIPPERS1)

I TO (TOP VAL (PIAN01)))

Although this stage represents Joshua's age 2 understanding of word meaning, CHILD has not yet learned sufficient syntax to interpret utterances whose meanings contradict the syntactic cues. It interprets the following unlikely utterance in­ correctly in accord with its knowledge of the usu­ al relationships between objects. This sort of er­

ror is similar to that reported in [11].

!PARENT SAYS: put the table on the ball I I CHILD RESPONDS: l (PTRANS ACTOR (CHILD) OBJECT (BALL1) I TO (TOP VAL (TABLE1)))

CHILD's generation at this stage also re­ flects the lack of syntactic knowledge. It knows the syntax for ·put,· but not for "on.· This leads it to generate utterances whose order is not to­ tally correct.

!CHILD SEES: (PARENT PUTS BOX1 ON TABLE1) I !INTERPRETING GESTURE AS: f(PTRANS ACTOR (PARENT) OBJECT (BOX1) I TO (TOP VAL (TABLE1))) I ICHILD SAYS: put box table on

The progression to CHILD's final stage is accom­ plished by learning the syntax of "on". When CHILD learns that the filler of the VAL slot of "on's" meaning is found following "on·, it can then understand utterances whose syntax contradicts the most likely meaning of the utterance. CHILD is asked to "Put the table on the ball," and its correct response to this utterance should be in­ terpreted as modelling a child who understands and who might go over to the table and tug at it a bit while looking at the ball.

IPARENT SAYS: put the table on the ball I I CHILD RESPONDS: f(PTRANS ACTOR (CHILD) OBJECT (TABLE1) I TO (TOP VAL (BALL1))) At this stage, CHILD also finally generates a description of the observed action correctly. !CHILD SEES: (PARENT PUTS SLIPPERS1 ON TABLE1) I !INTERPRETING GESTURE AS: I (PTRANS ACTOR (PARENT) OBJECT (SLIPPERS1) I TO (TOP VAL (TABLE1))) I

ICHILD SAYS: put slippers on table

This progression shows that CHILD's perfor­ mance manifests those language characteristics described earlier. First, its comprehension of language meaning and syntax precedes its genera­ tion or that same meaning and syntax. second, the length of its utterances increases during the pro­ gression. Third, its knowledge of language mean­ ing precedes its knowledge of language syntax, since it comprehends and generates utterances without syntax before it comprehends and generates them with syntax. Fourth, CHILD incorrectly understands utterances on the basis of the semantically most likely meaning when it doesn't know all of the words in the utterance, Fifth, it misunderstands utterances whose syntax indicates an interpretation different from the semantically most likely interpretation before it has learned

the syntax for the words in the utterance.

ll CONCLUSIONS

CHILD's manifestation of the five data sup­ ports both the se.mantically-indexed syntax hy­ pothesis and the comprehenslon-driven generation hypothesis. In particular, it offers explanations for this data. Generation follows comprehension because generation cannot occur until comprehen­ sion learning adds words to the dictionary. Length of utterance increases because the number of words available to express a concept increases during comprension learning. Knowledge of language mean­ ing precedes knowledge of language syntax because syntax is indexed under word meanings and hence cannot be learned before the word meaning. Misunderstanding of utterances whose correct in­ terpretation is not the semantically most probable occurs because children use their knowledge ot" probable meanings to augment gaps in their under­ standing. Finally, misunderstanding of utterances whose syntax suggests an interpretation different from the semantically most likely interpretation occurs because knowledge of syntax is learned after knowledge of meaning. This error is made when enough knowledge of meaning has been acquired to produce an interpretation, but not enough syn­ tax has been learned to produce a correct in­ terpretation.

There are certainly many other factors in child language acquisition which have not been considered here, but this paper offers support for the semantically-indexed syntax hypothesis and the comprehension-driven generation hypothesis as com­ ponents of a complete model of child language ac­ quisition. Further research must address questions posed by the data of more advanced child language acquis1tion, which could support other methods of indexing syntactic knowledge and other possible relationships between comprehension and genera­ tion.

ACKNOWLEDGEMENTS

Dr. Roger Schank's assistance in this work was in­ valuable. Dr. Richard Cullingford has made many valuable suggestions, and Dr. Katherine Nelson has provided many insights into problems of modelling child language learning. Peter Selfridge, Larry Birnbaum, Marie Bienkowski and Jamie Callan have contributed both ideas and criticisms.

REfERENCES

1. Benedict, H. (1976). Language Cgmorebension .!!1 ll 12 ~ H2.n.th.-.2l.l1. Infants. (thesis) Department of Psychology, Yale University, New Haven, Connecticut.
2. Birnbaum, L.,andSelfridge, M. (1981). in Inside Artificial Intelligence: Five Programs plus Miniatures. Schank R. and Riesbeck C.K. (eds.), Lawrence Earlbaum Associates, Hills­ dale, NJ.
3. [ 3] Bloan, L. , 197 3. ,OM. 1l2.c.l! At. .il ~. Mouton, The Hague.
4. [4) Bruner, J.S., Goodnow, J. J., and G. A., ( 1956). A~ 9!.. Thinking. ley an~ Sons, New York Austin, John Wi-
5. Clark, E.V. and Psychology ~ Jovanovich. Clark, Language, H. H., ( 1978). Harcourt Brace
6. Cullingford, R.E., Krueger, M.W., fridge, M. (1981) A Picture and Words: Autanated Explanations as a of a Computer-Aided Design System. ing. and Sel­ a Thousand Component Forth com-
7. Halliday, M.A.K., (1975). Learning ~ 12 .Mull: Explorations ..1n 1h§. Development 9f. Language Edward Arnold, London.
8. Hoodenraad, R., Grieve, R., Baldwin, P., and Campbell, R. (1976). Comprehens1on as an In­ teractive Process. In"R. N. Campell and P. T. Smith (eds.) ~ Adyances .!!1 ~ Psychology £(Language., Plenum Press, New York
9. [9) Schank, R. c., (1973). Identification of Con­ ceptualizations Underlying Natura! Language. In R. C. Schank and K. M. Col by ( eds. ) Computer ~ 9f. Thought .i.lli!. Language w. H. Freeman and Co., San Fransisco.
10. Selfridge, M. (1980). A Process Model of Language Acquisition. Computer Science Techn­ ical Report 172, Yale University, New Haven, Ct.
11. Strohrer, H. and Nelson, K.E., (1974). The Young Child's Development of Sentence Comprehension: Influence of Event Probabili­ ty, Non-verbal Context, Syntactic Form, and Strategies. .CI1il.sl. .nu_., 45:567-576
12. Westone, H. and Friedlander, (1973). The Ef­ fect of Word Order on Young Children's Responses to Simple Questions and Commands. .c.ll.U.Q. .nu.. 4 4 : 73 4-7 40
13. Winston, P. (1975). Learning Structural Descriptions from Examples. In P.H. Winston (ed.) ~ Psychology 9!.. Computer yisiori. McGraw-Hill, New York.