Learning form-meaning mappings for language

The proposed thesis research addresses two of the main obstacles to building agents that communicate using natural language: the need for richer representations of linguistic constructions that incorporate aspects of conceptual knowledge, context and goals; and the need for a principled approach to the automatic acquisition of such structures from examples. More generally, it explores the idea that patterns that arise in language are inextricably linked with and motivated by patterns of meaning and experience. This view, along with empirical evidence suggesting that linguistic knowledge at all levels can be characterized as mappings between form and meaning, serves as the basis for a computational model of the acquisition of simple phrasal and clausal constructions. The idea that language involves associations between sounds (form) and relatively richer sets of stimuli (meaning) is neither new nor surprising. Indeed, such an assumption is made without controversy in some recent models of the acquisition of individual words, including those for objects, spatial relations, and actions (Regier 1996; Bailey 1997; Siskind 1997; Roy & Pentland 1998). Thus far, however, models of the learning of larger phrasal and clausal structures have been oriented toward the problem of acquiring symbolic syntactic patterns, usually based around verbal argument structure (e.g., Brent 1994). The meaning of these larger structures is typically assumed to be predictable from the meaning of its constituents. The current work adopts the less strictly compositional framework of Construction Grammar (Goldberg 1994). On this view, although phrasal and clausal constructions require more complex structural description of the relations between their constituents, they can still be described, like words, as mappings between form and meaning. Parameters of form consist primarily of phonological cues, such as features of intonation, words and their inflections, and word order. Parameters of meaning encompass a much larger set of possibilities, including event structure, sensorimotor control, force dynamics, attentional state, perspective-