RE Data Challenge: Requirements Identification with Word2Vec and TensorFlow

Since their introduction over a year ago, Google's TensorFlow package for learning with multilayer neural networks and their Word2Vec representation of words have both gained a high degree of notoriety. This paper considers the application of TensorFlow-guided learning and Word2Vec-based representations to the problems of classification in requirements documents. In this paper, we compare three categories of machine learning techniques for requirements identification for the SecReq and NFR datasets. The first category is the baseline method used in prior work: Naïve Bayes over word count and TF-IDF representations of requirements. The remaining two categories of techniques are the training of TensorFlow's convolutional neural networks on random and pre-trained Word2Vec embeddings of the words found in the requirements. This paper reports on the experiments we conducted and the accuracy results we achieved.