Abstract
Speech-controlled user interfaces facilitate the operation of devices andhousehold functions to laymen. State-of-the-art language technology scans theacoustically analyzed speech signal for relevant keywords that are subsequentlyinserted into semantic slots to interpret the user's intent. In order todevelop proper cognitive information and communication technologies, simpleslot-filling should be replaced by utterance meaning transducers (UMT) that arebased on semantic parsers and a \emph{mental lexicon}, comprising syntactic,phonetic and semantic features of the language under consideration. Thislexicon must be acquired by a cognitive agent during interaction with itsusers. We outline a reinforcement learning algorithm for the acquisition of thesyntactic morphology and arithmetic semantics of English numerals, based onminimalist grammar (MG), a recent computational implementation of generativelinguistics. Number words are presented to the agent by a teacher in form ofutterance meaning pairs (UMP) where the meanings are encoded as arithmeticterms from a suitable term algebra. Since MG encodes universal linguisticcompetence through inference rules, thereby separating innate linguisticknowledge from the contingently acquired lexicon, our approach unifiesgenerative grammar and reinforcement learning, hence potentially resolving thestill pending Chomsky-Skinner controversy.