Affiliations: [a] Deparment of Software Engineering, School of Software, Yunnan University, Kunming, China
| [b] Deparment of Network Engineering, School of Software, Yunnan University, Kunming, China
Correspondence:
[*]
Corresponding author. Yan Kang, Deparment of Software Engineering, School of Software, Yunnan University, 650991, Kunming, China. E-mail: kangyan@ynu.edu.cn.
Abstract: In this paper, we present a novel method for data-mining large informal product descriptions rather than extracting requirement features from proprietary project repositories. Our algorithm hybridizes deep-learning algorithms such as word2vec and recurrent neural network (RNN) with classical techniques to improve the performance of text analysis. Given the inaccuracy and incompleteness of the software requirement descriptions on websites, the instance-transfer learning method is utilized to construct a robust classifier and predict domain feature knowledge based on domain knowledge similar to the target domain. The bagging clustering algorithm is utilized with multiple clustering algorithms to help select transfer instances. [Author to confirm changes.]The RNN-based algorithm is utilized as a useful alternative to predict missing features by studying the requirement descriptions of a related software system, while word2vec is utilized to extract sensible feature keywords for the specific software domain. [Author to confirm changes.]Our RNN model for every subclass is based on the clustering result, and we construct subclass classifiers to recommend requirement keywords. Requirement features recommended by our algorithm potentially increase opportunities for requirement classification, promote software requirement quality, and deliver more reliable software products. We explain the details of implementation and perform experimental work on real requirement descriptions to establish its worth.
Keywords: Word2vec, RNN, transfer learning, feature model, software requirement
