Automated Assertion Generation from Natural Language Specifications

Abstract

We explore contemporary natural language processing (NLP) techniques for converting NL specifications found in design documents directly to an temporal logic-like intermediate representation (IR). Generally, attempts to use NLP for assertion generation have relied on restrictive sentence formats and grammars as well as being difficult to handle new sentence formats. We tackle these issues by first implementing a system that uses commonsense mappings to process input sentences into a normalized form. Then we use frame semantics to convert the normalized sentences into an IR based on the information and context contained in the Frames. Through this we are able to handle a large number of sentences from real datasheets allowing for complex formats using temporal conditions, property statements, and compound statements; all order agnostic. Our system can also be easy extended by modifying an external, rather than internal, commonsense knowledge-base to handle new sentence formats without requiring code changes or intimate knowledge of the algorithms used.