A Computer Model of Child Language Learning

Mallory SelfridgeAAAI 1980

Abstract

This paper presents a computational model of child language acquisition, developed in the late 1970s. Utilizing a "Good Old Fashioned AI" (GOFAI) approach, the research simulates how a child progresses from zero linguistic capability to full comprehension over a twelve-month period. By implementing a system that maps spoken utterances onto pre-existing world knowledge and gestural cues, the model demonstrates that complex language structures can be acquired through general-purpose inference mechanisms rather than innate, language-specific brain structures.

Introduction

One of the most enduring debates in cognitive science is the "nature vs. nurture" conflict regarding language acquisition. In the late 1970s, the prevailing view—championed by Noam Chomsky—was that human language is too complex to be learned through experience alone, necessitating an innate "Universal Grammar." This paper challenges that assumption by proposing a computer model capable of learning language through exposure to situational context and world knowledge.

The model, named "Joshua," is based on longitudinal observations of a real child between the ages of 1 and 2. The goal was to determine if a program starting only with an understanding of objects and actions could learn to comprehend complex commands simply by being "raised" in a simulated environment mirroring a child's experience.

Methodology

System Architecture

The model was implemented in LISP on a DECSYSTEM-20. To represent meaning, the system utilized Conceptual Dependency (CD), a framework developed by Roger Schank. CD breaks down sentences into fundamental primitive actions—such as PTRANS for the physical transfer of an object—allowing the program to store and manipulate concepts independently of specific words.

Learning via Inference Rules

The core of the system relies on three primary inference rules that allow the program to "guess" the meaning of unknown linguistic input:

  1. Gestural Meaning Inference: When a simulated parent speaks while performing a gesture (like pointing), the program assumes the utterance refers to the object or action indicated.
  2. Response Completion Inference: If the program understands a portion of a sentence, it uses the situational context to "subtract" known meanings and isolate the definition of the remaining unknown words.
  3. Structure Learning Rule: This rule tracks word positioning. For example, if a specific object consistently follows the word "on," the program updates its internal dictionary to recognize "on" as a marker for a spatial relationship involving the subsequent word.

Experimental Design and Data

The primary dataset consisted of longitudinal observations of a single child, Joshua. The program was provided with "simulated visual input" (descriptions of objects and gestures) and segmented text utterances (e.g., "give me the ball"). To ensure the model's mechanisms were not specific to English, the author also tested the program on subsets of Japanese, Russian, Chinese, Hebrew, and Spanish, finding the learning rules to be cross-linguistically applicable.

Results: The Four Stages of Development

The program successfully replicated Joshua’s linguistic progression through four distinct stages:

  • Stage 1 (Age 1.0): The model possesses no linguistic knowledge and responds exclusively to physical gestures.
  • Stage 2: The model begins to learn basic nouns and verbs. However, it frequently confuses spatial relations, often treating "on" and "in" as interchangeable concepts.
  • Stage 3: The model understands word meanings well enough to handle "unlikely" or "silly" requests (e.g., "Put your slippers on the piano"). This stage is crucial as it proves the system is processing linguistic structure rather than just guessing based on common sense.
  • Stage 4 (Age 2.0): The model masters the structure of the language, allowing it to interpret highly unusual commands (e.g., "Put the table on the ball") and even identify them as semantically anomalous.

Discussion and Insights

Meaning Before Grammar

A central insight of the model is that children do not learn language in a vacuum. They utilize "World Knowledge" as a bootstrap; because they already understand that cups are for drinking and balls are for throwing, language learning becomes a mapping exercise. Consequently, the model suggests that syntax is acquired only after word meanings are established. Grammar, in this view, is essentially a set of expectations regarding where specific meanings will appear in a sequence.

Psychological Plausibility

The model’s validity is supported by its error patterns. Like real children, the program made over-extension errors (such as using "on" to mean "inside"). These mistakes suggest that the model's learning path mirrors human cognitive development.

Limitations and Future Work

While successful, the model operates under several simplifications. It receives perfectly segmented input, whereas real children must parse continuous speech. It also ignores polysemy (words with multiple meanings) and assumes that a child’s conceptual understanding of the world is static, rather than evolving alongside their language. Future research directions include moving from comprehension to language generation and handling more complex syntactic structures like relative clauses.

Summary updated

BibTeX
@inproceedings{selfridge1980computer,
  author = {Mallory Selfridge},
  title = {{A Computer Model of Child Language Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure