English
Language use is both productive and conventional. Both aspects can be used to explain a language users ability to read, understand or pronounce new utterances or texts he or she has never heard or seen before. The role of convention has been underestimated in (computational) linguistics, mainly because no elegant solutions have been found for representing the massive amount of conventions in language in the economical formalisms that have ruled the field for decades. A new branch of interdisciplinary research on the application of artificial learning systems to language tasks, of which the present proposal is an exponent, is investigating how much of language uses when seen as mappings from one language level to another, can be explained by models that make unlimited use of memory to hold as much observed instances of natural language use as possible, and generalise from these memorised instances to new instances. These memory models of language emulate purely conventional language processing; all new performance is based on stored representations of old experiences. Language processing by convention in this approach is not unproductive repetition of patterns; it is a combination of creative analogical reasoning on the basis of non-trivial similarity functions on the one hand, and adequate representations of language use that capture all context necessary to make task-specific analogical mappings on the other hand. We propose to investigate two language processing tasks for which data is, or is becoming, available: fluent text pronunciation, and dialogue act classification. In the former, the tested mapping is between written text and speech annotated at a level that can be output to a speech synthesizer to produce adequately fluent artificial speech; in the latter task, the tested mapping is between a transcribed dialogue and the underlying intentions behind each of the utterances in the dialogue. For both tasks, traditional approaches suggest a decomposition into subtasks, which could together be combined into a modular software system. Following the Minimal Description Length principle, decompositions are suggested that yield isolated subtasks that can be described with small, understandable models (e.g. principles, rule systems). Often, splitting up a complex task into subtasks ignores the fact that the split subtasks need each others solution; splitting them up can cause information to be blocked, causing the overall system to produce bad solutions. The thesis research of the submittor shows that this is the case for English word pronunciation; leaving this task fully composed yields the best generalization accuracies on new words when a memory model is used. This suggests that it is equally relevant to aim at a minimal number of description levels to arrive at compact models of language processing that describe that process adequately; this is in fact an incarnation of the Mirtimal Description Length principle, but an innovative, largely unexplored one in language modeling.
