Experiments Using Semantics for Learning Language Comprehension and Production
Several questions in natural language learning may be addressed by studying formal language learning models. In this work we hope to contribute to a deeper understanding of the role of semantics in language acquisition. We propose a simple formal model of meaning and denotation using ﬁnite state transducers, and an algorithm that learns a meaning function from examples consisting of a situation and an utterance denoting something in the situation. We describe the results of testing this algorithm in a domain of geometric shapes and their properties and relations in several natural languages: Arabic, English, Greek, Hebrew, Hindi, Mandarin, Russian, Spanish, and Turkish. In addition, we explore how a learner who has learned to comprehend utterances might go about learning to produce them, and present experimental results for this task. One concrete goal of our formal model is to be able to give an account of interactions in which an adult provides a meaning-preserving and grammatically correct expansion of a child’s incomplete utterance.