Using sentence context and implicit contrast to learn sensor-grounded meanings for relational and deictic words: the twig system

摘要

This thesis describes a novel system allows a robot to infer the meanings of new words from their usage in context. TWIG (Transportable Word Intension Generator) can parse simple sentences, determine the reference of any unknown words to objects or people in the environment through sentence context, and can determine over time what the meanings of the new words by building trees imply the word meanings from their structure. The system was originally built to learn pronouns, a word category has previously been unmodeled in the robotic word learning literature, but is general enough to learn some other word categories, including prepositions and transitive verbs. The system was implemented on a physical robot equipped with face detectors, simple vision systems, and a sensor network for object localization. TWIG succeeded in learning I and you refer to the speaker and addressee; he must refer to a person is neither of these; this and that must refer to proximal or distal non-person objects; above and below prepositions refer to relative height; and am and are refer to the identity relation. The system can be used for sentence production as well as comprehension, and was found to produce more correct sentences and fewer incorrect sentences about its environment than similar systems lacked the system's extension inference and definition tree capabilities. The work contains several new approaches in the area of robotic word learning, and can also be interpreted as a computational model of how human infants use contrast to learn word meaning.