nlrpBENCH: A Benchmark for Natural Language Requirements Processing

Abstract

We present nlrpBENCH: a new platform and framework to improve soft- ware engineering research as well as teaching with focus on requirements engineering during the software engineering process. It is available on http://nlrp.ipd. kit.edu. Recent advances in natural language processing have made it possible to process textual software requirements automatically, for example checking them for flaws or translating them into software artifacts. This development is particularly fortunate, as the majority of requirements is written in unrestricted natural language. However, many of the tools in in this young area of research have been evaluated only on limited sets of examples, because there is no accepted benchmark that could be used to assess and compare these tools. To improve comparability and thereby accelerate progress, we have begun to assemble nlrpBENCH, a collection of requirements specifications meant both as a challenge for tools and a yardstick for comparison. We have gathered over 50 requirement texts of varying length and difficulty and organized them in benchmark sets. At present, there are two task types: model extrac- tion (e.g., generating UML models) and text correction (e.g., eliminating ambiguities). Each text is accompanied by the expected result and metrics for scoring results. This paper describes the composition of the benchmark and the sources. Due to the brevity of this paper, we omit example tools comparisons which are also available.