Donald Bren School of Information and Computer Sciences

Utilizing recurrent network topologies to produce case/role meaning representations for single sentences has become common practice in connectionist natural language processing systems. Typically, these systems train with the complete sentence meaning as the target output for the entire period that the sentence is being processed; i.e., the complete meaning is available starting with the first word of the sentence. Thus, the context feedback provided by these systems is non-incremental in that they use information about the sentence that has not yet been encountered in order to aid in the processing and learning tasks. SAIL1 is a connectionist natural language processing system which builds the sentence meaning representation incrementally, incorporating into the meaning only the information derived from words already processed.