Analogy-Based Acquisition of Utterances Relating to Temporal Aspects

V. Sembugamoorthy

Introduction

In the late 1970s, artificial intelligence research faced a significant challenge in language acquisition: bridging the gap between rigid, hard-coded grammatical rules and "blank slate" systems that struggled with abstract concepts. While early systems like the Paradigmatic Language Acquisition System (PLAS) could learn language through interaction and analogy, they often lacked the internal machinery to represent and reason about the temporal dimension. This paper introduces PLAS.T, an extension designed to ground temporal concepts—such as "yesterday," "last week," or "before"—in a simulated sensorimotor world model.

The PLAS.T Framework

The core philosophy of PLAS.T is that language acquisition should be rooted in sensorimotor experience. The author, V. Sombugamoorthy, proposes a two-stage approach. The first stage focuses on grounding linguistic units in actions and perceptions, while the second stage builds formal logic and complex grammar upon this foundation.

To represent knowledge in a language-independent way, PLAS.T utilizes Roger Schank’s Conceptual Dependency (CD) theory. This allows the system to model actions (e.g., "walking," "coming") and agents without relying on specific linguistic strings. In this view, natural language expressions themselves are treated as the primitives of knowledge representation.

Methodology: Learning by Analogy and "Delta"

The system learns primarily through analogy. When presented with a new utterance, PLAS.T compares it to a known utterance. If the linguistic difference (the "delta") between the two sentences corresponds to a specific difference in the internal world model (e.g., different blocks on a simulated calendar), the system creates a mapping between the new phrase and the temporal state.

Figure 1: Mapping of the utterance 'JOHN CAME TO BOMBAY LAST WEEK' to a specific temporal state ('SECOND WEEK OF AUGUST') via the DESCRIBES marker.

This "Learning by Delta" principle means the system does not require explicit definitions for terms like "last week." Instead, it identifies these terms as the remaining linguistic components that explain the difference between two analogous situations.

Extra-Linguistic Markers

To facilitate teaching, PLAS.T employs several extra-linguistic markers that simulate human communicative intent:

  • DESCRIBES: Links a sentence directly to a specific event in the world model.
  • IS-A-RECOLLECTION-OF: Instructs the system to search its past memory for a matching state.
  • IS-AN-IMAGINATION-OF: Signals the system to model a hypothetical or future state.

These markers allow the system to perform complex tasks such as tense conversion. By utilizing the RECOLLECTION and IMAGINATION markers, PLAS.T can transform a present-tense statement into a past or future context.

Figure 2: A CD representation associated with the temporal marker 'JULY'.

Discussion and Results

The results of PLAS.T are qualitative but demonstrate a significant step toward autonomous language acquisition. The system successfully associates previously unknown temporal units with their sensorimotor counterparts. Furthermore, it generalizes using a "near miss" principle, which allows for the simultaneous acquisition of grammar and vocabulary through teacher interaction.

The author also highlights a Whorfian influence: by learning the language, the system effectively learns how to structure its model of the world. However, the system has notable limitations. The world model is highly simplified, consisting mainly of abstract calendar blocks. Additionally, the system remains heavily dependent on a human teacher to provide corrective states during the learning process.

Conclusion

PLAS.T represents an early attempt to treat natural language expressions as the primitives of knowledge representation. By grounding temporal language in a world model through analogy, the system suggests that learning a language is intrinsically linked to learning the structure of the world itself. Future work points toward a "second-stage" system that would move from this primitive grounding into formal logic and complex grammatical performance.

Summary updated

BibTeX
@inproceedings{sembugamoorthy1981analogy,
  author = {V. Sembugamoorthy},
  title = {{Analogy-Based Acquisition of Utterances Relating to Temporal Aspects}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}
Figure