GLAsT: Learning Formal Grammars to Translate Natural Language Specifications into Hardware Assertions

The purpose of functional verification is to ensure that a design conforms to its specification. However, large written specifications can contain hundreds of statements describing correct operation which an engineer must use to create sets of correctness properties. This laborious manual process increases both verification time and cost. In this work we present GLAsT, a new learning algorithm which accepts a small set of sentences describing correctness properties and corresponding SystemVerilog Assertions (SVAs). GLAsT creates a custom formal grammar which captures the writing style and sentence structure of a specification and facilitates the automatic translation of English specification sentences into formal SystemVerilog Assertions. We evaluate GLAsT on English sentences from two ARM AMBA bus protocols. Results show that a translation system using the formal grammar generated by GLAsT automatically generates correctly formed SVAs from the targeted AMBA specification as well as from a second, different AMBA bus specification.