Goal/plan analysis via distributed semantic representations in a connectionist system

摘要

In this paper we describe DYNASTY, a multi-module distributed connectionist system designed to perform a very high-level symbolic task, namely, comprehension of goal/plan-based stories. DYNASTY has two phases of operation: learning and performance. During learning, each DYNASTY module acquires both the knowledge and skill to perform its specified subtask, through backpropagation learning on a data set of propositions. In addition to modifying their connection weights, DYNASTY modules automatically form distributed semantic representations (DSRs) of the lexical and conceptual symbols used in training the modules. Each DSR encodes, as an activation vector, both structural and sequential information inherent in the training data. During performance, DRSs are passed among various connectionist modules, thus supporting communication and modularity. In addition, DSRs of words with similar meanings end up having similar DSRs. This feature gives DYNASTY the ability to generalize, e.g., generate appropriate inferences when given novel yet similar inputs.