No label defined (Q4023686)

The article of this volume will not be reviewed individually. The book contains a collection of 17 articles on connectionist natural language processing first published in the journal ``Connection Science'' between 1989 and 1991. The articles focus on connectionist parsing, text understanding, speech recognition and synthesis. Three applications of connectionist script-based story understanding systems are given. One recognizes an input story as an instance of a particular script; the recognition is based on a hierarchy of feature maps and the system extracts the hierarchical taxonomy of scripts, tracks and role bindings from experience in an unsupervised self-organization process; the taxonomy can also be used to recognize and fill incomplete instantiations of the scripts (\textit{Miikkulainen}). The second system learns distributed semantic representations of concepts and events in everyday scriptal experiences by recirculating the hidden layer in two auto-associative recurrent PDP networks, performs script-based causal chain completion inferences according to the acquired sequential knowledge, and performs script role association and retrieval during script application (\textit{Lee}, \textit{Flowers} and \textit{Dyer}). The third is a hybrid model of script generation that allows the modeling of scriptal behavior in the computer domain, viz. routine computing and Unix command composition (Mannes and Doane). Another contribution describes a connectionist network which uses surface features and minimal semantic information to identify topic entities within the framework of a simplified representation of the nominal expressions in written narrative discourse; topic identification as well as pronoun resolution is seen as a function of attentional state (\textit{Karen}). One contribution explores the close relationship between connectionism and cognitive linguistics, considering a connectionist model for the diverse meanings of the preposition ``over'' (\textit{Harris}). A general architectural scheme for connectionist language processing situated in a microworld context, the Connectionist Languages Users (CLUES) model, is described (\textit{Allen}); neural networks which act as state machines are applied to a variety of natural language generation (descriptions, planning etc.) and recognition tasks (question answering, pronoun reference resolution, action identification, analogy completion). Problems of syntax receive a lot of attention in many contributions. Two papers consider transformational issues of linguistic theory in a connectionist framework; the first proposes a connectionist model of movement in government-binding theory (\textit{Rager} and \textit{Berg}), while Pollack's Recursive Auto-Associative Memory (RAAM) model is used in the second paper to construct distributed representations of syntactically structured sentences on which holistic structure-sensitive operations (syntactic transformations on sentences, in this case active-passive transformations) are then carried out (\textit{Chalmers}). A single-layer higher-order dynamic topology neural network is shown to achieve remarkable performance results on a context-free grammar recognition task (\textit{Wyard} and \textit{Nightingale}). A cognitive architecture for syntactic aspects of human sentence processing which applies a unification-based tree formation model and principles from simulated annealing to the search for the best configuration of isolated subtree in the final parse tree is shown to account for well-known psycholinguistic parsing effects and a grammatic language performance (\textit{Kempen} and \textit{Vosse}). Hybrid architectures merging connectionist with symbolic computation principles are considered in two contributions, one introducing a hybrid model for the understanding and disambiguation of noun phrases in technical sublanguages which takes advantage of the virtues of distributed and localist connectionist and symbolic representation levels in order to model semantic memory as well as semantic and syntactic constraints (\textit{Wermter} and \textit{Lehnert}), the other discussing the implementation of a connectionist deterministic parser based on the rules from a deterministic grammar (\textit{Kwasny} and \textit{Faisal}). Learning issues are treated in a paper that deals with the grammatical inference of strictly-hierarchical context-free grammars and their mapping onto a connectionist parser in order to perform various pattern recognition tasks; the feasibility of the approach is demonstrated for hand-written character recognition (\textit{Lucas} and \textit{Damper}). The final four papers show the application of connectionist methods to phonetics and phonology. The first one treats the role of similarity in Hungarian vowel harmony (\textit{Hare}), the second paper deals with three approaches to the representation and recognition of temporal patterns as applied to speech recognition, i.e. sound as events in time (\textit{Port}), the third one introduces networks that learn about phonological feature persistence, viz. harmony constraints within morphemes and feature persistence across morpheme boundaries (\textit{Gasser} and \textit{Lee}), and the fourth paper treats single-layer and multi-layer preceptrons as required for different pronunciation tasks for digit sequences in text-to-speech systems (\textit{Ainsworth} and \textit{Warren}).