Constructing Meaning, Piece by Piece: A Computational Cognitive Model of Human Sentence Comprehension

AI systems with language for robots don’t try to model human processing. Psycholinguistic models of human language processing don’t have operational computational models. To solve these problems, this thesis contributes to progress in answering two interlocking scientific questions: how does the human mind do sentence comprehension, and how can we enable artificial agents to use natural language to collaborate with humans. We do this with a system called Lucia, which is a computational cognitive model of human sentence comprehension that works by constructing the meaning of a sentence piece by piece. The Lucia model is designed according to five overriding qualitative principles of human language comprehension. To show that its results are useful, it does embodied, end-to-end comprehension (E3C) within an artificial agent called Rosie. To model key characteristics of human comprehension, it draws on research in cognitive linguistics, psycholinguistics, artificial intelligence, and robotics to: represent composable knowledge of the meaning of linguistic forms (CKM), do incremental, immediate interpretation processing (I3P), and do it using general cognitive mechanisms (GCM). The model leads to a theory of language acquisition from experience (LAE), some parts of which have been implemented experimentally. To conform to these principles, the Lucia model is implemented in a robotic agent called Rosie to do E3C. It uses Embodied Construction Grammar (ECG) as its method of representing composable knowledge of meaning (CKM), and demonstrates that this knowledge can be processed incrementally (I3P) using a novel comprehension algorithm that relies on the general cognitive mechanisms (GCM) of the Soar cognitive architecture to produce embodied, end-to-end comprehension (E3C). Lucia makes several contributions to answering the broader scientific questions. It provides a novel theory for incremental processing (I3P) based on a three-phase construction cycle. It provides a theory of how memories interact during comprehension. It demonstrates grounded comprehension in an embodied robotic agent. Finally, it provides a detailed, functional model of cognitive E3C processing that can serve as a basis for further research in modeling human language processing in the brain and in designing larger-scale language models for artificial agents.PHDComputer Science & EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/172668/1/plindes_1.pd