A Production Rule System for Message Summarization

Book One

Volume

Proceedings of the AAAI Conference on Artificial Intelligence, 4

Natural Language

In summarizing a message, it is necessary to access knowledge about linguistic relations, subject matter knowledge about the domain of discourse, and knowledge about the user’s goals for the summary. This paper investigates the feasibility of integrating these knowledge sources by using computational linguistic and expert system techniques to generate one-line summaries from the narrative content of a class of Navy messages. For deriving a knowledge representation of the narrative content, we have adapted an approach developed by Sager et al. at New York University. This approach, called information formatting, uses an explicit grammar of English and a classification of the semantic relationships within the domain to derive a tabular representation of the information in a message narrative. A production system, written in OPS5, then interprets the information in the table and automatically generates a summary line. The use of a production rule system provides insight into the mechanisms of summarization. A comparison of computer-generated summaries with those obtained manually showed good agreement, indicating that it is possible to automatically process message narrative and generate appropriate, and ultimately useful, summaries.