Over 50 papers present natural language processing tools for improving the quality of requirements. However, few of these are adopted by industry. Even worse, most of them are no longer publicly available or supported by their creators. The few available and actively maintained tools exhibit some outstanding features, but also include sub-optimal functionalities. In this paper, we compare the performance of 3 existing tools on how well they automatically detect ambiguity and atomicity defects and deviations in 4 real-world natural language requirements sets. Next, we show how to design a superior tool by combining the best performing approaches of these three. Finally, we introduce a research roadmap toward automatically generating NLP RE tool mashups through the assembly of modular components taken from existing tools.
