Evaluate Similarity of Requirements with Multilingual Natural Language Processing

Finding redundant requirements or semantically similar ones in previous projects is a very time-consuming task in engineering design, especially with multilingual data. Due to modern NLP it is possible to automate such tasks. In this paper we compared different multilingual embeddings models to see which of them is the most suitable to find similar requirements in English and German. The comparison was done for both in-domain data (requirements pairs) and out-of-domain data (general sentence pairs). The most suitable model were sentence embeddings learnt with knowledge distillation.