This research is supported by NIH under NRSA 1F32DC00459-01 and previously by Texas Higher Education Coordinating Board under grant ARP-003658-444-1995. Most of our projects are described below; for more details and for other projects, see publications in Natural Language Processing.
The sound structure of language changes over time due to the interaction of socio-cultural and communicative factors. We hypothesize that the socio-cultural factors induce change, whereas the communicative factors define the direction of change. In this research, we develop generative models of language in order to examine the relative effects of these factors on the rate of change and to predict different paths of sound change. The goal is to understand the process by which languages differentiate through time.
The SARDSRN model demonstrates how the performance of a Simple Recurrent Network (SRN) parser can be enhanced significantly by the addition of a SARDNET module. SARDNET is an extension to the Self-Organizing Map architecture (SOM) which allows self-organization of sequences. The SARDNET module in the SARDSRN parser explicitly represents the input sequence in a self-organizing map. The distributed SRN component leads to good generalization and robust cognitive properties, whereas the SARDNET map provides exact representations of the sentence constituents. This combination allows SARDSRN to learn to parse sentences with more complicated structure than can the SRN alone.
Syllable systems across languages share a number of common patterns. A particularly compelling explanation for these patterns is that they orginate from constraints provided by the perceptual and articulatory systems of language users. In this research, we use genetic algorithms to examine how a few experimentally defined perceptual and articulatory constraints on syllables interact to produce different relative distributions of syllable types in evolved vocabularies. The goal is to show how both language regularity and variation arise from optimizing the sound system under these constraints.
This research is concerned with the use of a neural network model to form meaningful representations of the content of documents available electronically. The goal is to show that these representations are well-suited for text management tasks such as categorization, evaluation, retrieval, and question answering.
The integrated processing-decoding network model of St. John and McClelland (1990) was revised to allow extracting the predicate content of complex sentences directly from an incoming stream of word tokens. The input stream was presented to the network without any syntactic markup such as bracketization, and the extraction was done without any explicit emulation of stacking, segmentation, or other such operations that are ordinarily associated with parsing a sentence. The lack of such explicit syntactic operations allowed a simulated neural network of minor complexity to be trained to the task under a simple regimen.
DISLEX was originally developed as the lexicon component for the DISCERN system. It was later extended into a more complete model of the mental lexicon, including semantic, orthographic, and phonological lexical modalities. The organization of each modality and the mappings between them are learned based on co-occurrences of symbols and their meanings. DISLEX shows how dyslexic and category-specific aphasic impairments can arise based on the physical organization of the lexical system.
Semantic Disambiguation in Sentence Processing
Subsymbolic processing of sentences is based on associations of words with past context. As words come in, their possible contexts are combined into the interpretation of the sentence. This model demonstrates how context frequency drives the process of disambiguating word meanings.
SPEC is based on the idea that sentence understanding is a controlled process. In SPEC, subsymbolic networks for parsing, memory, and control are integrated into a large modular system that learns to understand sentences with complex relative clause structure. SPEC shows how productive and systematic linguistic performance can be achieved in an entirely subsymbolic system, while still accounting for subsymbolic cognitive phenomena.
Much of our NLP work originates from DISCERN, a large-scale natural language processing system implemented entirely at the subsymbolic level. In DISCERN, distributed neural network models of parsing, generating, reasoning, lexical processing, and episodic memory are integrated into a single system that learns to read, paraphrase, and answer questions about stereotypical narratives. In this approach, subsymbolic networks are not only plausible models of isolated cognitive phenomena, but also serve as building blocks for large-scale artificial intelligence systems.
With FGREP, distributed representations for words are developed as part of the task. The representations reflect how the words are used in the task, and in this sense, also stand for the meanings of the words: words that are used the same way have similar representations. FGREP representations lead to good generalization and robustness under noise, and the method frees the system designer from having to encode the representations by hand. FGREP is used for example in DISCERN, SPEC, DISLEX, and the semantic disambiguation projects described on this page.