An On-Line Computational Model of Human Sentence Interpretation: A Theory of the Representation and Use of Linguistic Knowledge

Abstract

This dissertation presents a model of the human sentence interpretation process, which attempts to meet criteria of adequacy imposed by the different paradigms of sentence interpretation. These include the need to produce a high-level interpretation, to embed a linguistically motivated grammar, and to be compatible with psycholinguistic results on sentence processing.The model includes a theory of grammar called Construction-Based Interpretive Grammar (CIG) and an interpreter which uses the grammar to build an interpretation for single sentences. An implementation of the interpreter has been built called Sal.Sal is an on-line interpreter, reading words one at a time and updating a partial interpretation of the sentence after each constituent. This constituent-by-constituent interpretation is more finegrained and hence more on-line than most previous models. Sal is strongly interactionist in using both bottom-up and top-down knowledge in an evidential manner to access a set of constructions to build interpretations. It uses a coherence-based selection mechanism to choose among these candidate interpretations, and allows temporary limited parallelism to handle local ambiguities. Sal's architecture is consistent with a large number of psycholinguistic results.The interpreter embodies a number of strong claims about sentence processing. One claim is uniformity, with respect to both representation and process. In the grammar, a single kind of knowledge structure, the grammatical construction, is used to represent lexical, syntactic, idiomatic, and semantic knowledge. CIG thus does not distinguish between the lexicon, the idiom dictionary, the syntactic rule base, and the semantic rule base. Uniformity in processing means that there is no distinction between the lexical analyzer, the parser, and the semantic interpreter. Because these kinds of knowledge are represented uniformly, they can be accessed, integrated, and disambiguated by a single mechanism.A second claim the interpreter embodies is that sentence processing is fundamentally knowledge-intensive and expectation-based. The representation and integration of constructions uses many diverse types of linguistic knowledge. Similarly, the access of constructions is sensitive to top-down and bottom-up, syntactic and semantic knowledge, and the selection of constructions is based on coherence with grammatical knowledge and the interpretation.iii Acknowledgements This dissertation would not have been possible without the support and advice of many people. First, my advisor Robert Wilensky, who has given me guidance and generous support throughout my dissertation work. He carefully read and critiqued my work, and our discussions resulted in many of the ideas presented here. The other members of my committee, Charles Fillmore, who was also my undergraduate advisor, Jerry Feldman, and Paul Kay, were extremely supportive and helpful throughout.I owe an especial debt of gratitude to Nigel Ward, Graeme Hirst, and Jane Edwards, who read every chapter, and whose comme...

A Probabilistic Model of Lexical and Syntactic Access and Disambiguation

“…Second, we argue that these results demonstrate the need for probabilistic augmentation to grammatical theories if they are to be able to deal with a broad range of phenomena as a processing model. In this vein our model is part of a larger effort involving computational implementations of linguistic theory (Jurafsky, 1992) and large-scale implementations of probabilistic augmentations to linguistic models for speech recognition .…”

Trajectories through Knowledge Space

“…Jurafsky (1992) defines an "on-line interpretation" as one in which a continually updated partial interpretation of each sentence is maintained at all times. 3 Recognizing the on-line nature of interpretation however, is not sufficient by itself to form a correct interpretation.…”

“…The advantage of such schemes is that it becomes possible to pursue in parallel, efficiently and within a uniform framework, many distinct inference paths, in contrast to more computationally expensive linguistic approaches (e.g., Jurafsky, 1992 andHobbs et al, 1993). (See also Dolan, 1989;Lange and Dyer, 1989;Sumida and Dyer, 1989;and Sumida, 1991 for some connectionist approaches that are parallel at the knowledge level.)…”

“…In contrast, symbolic computational language models (e.g., Charniak and Goldman, 1989;Jurafsky, 1992;Norvig, 1989) provide separate mechanisms to deal with this problem. For example, Charniak and Goldman construct an interpretation of a story by accessing a knowledge base in order to construct, on the fly, a Bayesian network that represents that particular story.…”

6. Degeneracy: The maintenance of constructional networks

“…The term constructicon is said to be originally coined byJurafsky (1992).Degeneracy: The maintenance of constructional networks…”