SCRIPT-BASED INFERENCE AND MEMORY RETRIEVAL IN SUBSYMBOLIC STORY PROCESSING

DISCERN is an integrated natural language processing system built enti rely from distributed neural networks. It reads short narratives about stereotypical event sequences, stores them in episodic memory, genera tes fully expanded paraphrases of the narratives, and answers question s about them. Processing in DISCERN is based on hierarchically-organiz ed backpropagation modules, communicating through a central lexicon of word representations. The lexicon is a double feature map system that transforms each orthographic word symbol into its semantic representa tion and vice versa. The episodic memory is a hierarchy of feature map s, where memories are stored ''one-shot'' at different locations. Seve ral high-level phenomena emerge automatically from the special propert ies of distributed neural networks in this model. DISCERN learns to in fer unmentioned events and unspecified role fillers, generates expecta tions and defaults, and exhibits plausible lexical access errors and m emory interference behavior. Word semantics, memory organization, and appropriate script inferences are automatically extracted from example s. DISCERN shows that high-level natural language processing is feasib le through integrated subsymbolic systems. Subsymbolic control of high -level behavior and representing and learning abstractions are the two main challenges in scaling up the approach to more open-ended tasks.