📄 Constructivist Development of Grounded Construction Grammar

Summary

The paper proposes a computational model of constructivist language acquisition. According to this approach, grammar is not innate: it gradually emerges from concrete constructions that are initially tied to individual verbs. The author criticizes purely observational, statistical learning on two grounds: it runs into the poverty-of-the-stimulus problem, and it treats language as a fixed system. Instead, he proposes three mechanisms based on Peirce's ideas. The speaker invents new constructions, the hearer abductively guesses their meaning, and induction tests these hypotheses against the success of subsequent interactions. To this end, a unification-based formalism for construction grammar was developed, close to Embodied Construction Grammar. Its rules are reversible, so the same rules serve for both parsing and production. The model was tested in language games. Agents observe puppet scenes through cameras, recognize events, and describe them to each other. As agents reuse markers by analogy between events, generalized roles such as agent and patient gradually emerge in the artificial grammars. The work is significant as one of the first concrete implementations of constructivist theory and as a model of language evolving as an adaptive system.

Key points

• The author argues that three processes are necessary and sufficient for language learning: invention, abduction, and induction. The agent subsequently prefers successful constructions and avoids unsuccessful ones when it has alternatives.
• Meaning is represented in predicate-calculus style, with each event argument expressed as a separate predicate (e.g., give-1, give-2, give-3). This makes it easy to add new components to an event description. Syntactic and semantic structures share common units.
• Experimental setup: puppet scenes (push/pull, give/take) are processed by standard computer vision algorithms and a hierarchical event recognition system. In the description game, a turn is successful if the hearer confirms that the described event actually took place.
• Grammatical marking uses particles such as "ba", together with a word-order constraint. When a single word participates in several rules, hierarchical structure emerges (example: "fall red ball ba").
• After 700 games between two agents, individual markers ("va", "zu") cover several predicate-argument relations each, making the grammar more compact. The invention and abduction strategies are currently hand-coded, and the author leaves their description as meta-rules for future work.