Children are facile at acquiring the structure and meaning of their native tongues. The ease with which they become language experts belies the complexity of the underlying task – learning from a raging sea of information. Computational approaches to learning aspects of language typically reduce the problem to learning syntax alone, learning a lexicon alone, or learning semantics alone. These simplifications have led to disconnected solutions and some unreasonable assumptions about inputs to their algorithms.

This dissertation addresses the language learning problem for an embodied robot embedded in a rich environment. We present a pair of theoretical contributions and realize them through a suite of algorithmic contributions. Our first theory posits that language learning systems that rely exclusively on syntax can exploit semantics to improve their success. Similarly, systems that learn semantics can exploit syntax to improve their success. We demonstrate this approach through a life-long learning system, and prove a series of theoretical results identifying a class of learnable context-free languages.

When learning syntax, most algorithms expect complete knowledge about the language’s alphabet (or lexicon). When learning a lexicon, most algorithms never consider the usage of the lexical items when acquiring the lexicon. We relax this assumption and utilize this new information, respectively, in our second theoretical contribution: bootstrapping learning tasks. We introduce a bootstrapped approach to learning a lexicon and a grammar for a class of regular languages. Using phonetic transcriptions of natural language, we report on experimental results for this bootstrap. We also extend our approach to learning both syntax and semantics through bootstrapping the two, using limited knowledge about semantics to infer additional knowledge about syntax, and limited knowledge about syntax to infer additional knowledge about semantics. Finally, the ultimate goal of this research is to ground all learning in real-valued sensor data. We present a set of algorithms that point toward new approaches to incorporating real-valued data into our bootstrap learning systems.