基于神经网络集成学习和语义增强特征的需求跟踪方法

A Requirements Traceability Approach Based on Neural Network Ensemble Learning and Semantic-Enhanced Features

现有的基于机器学习的需求跟踪方法存在稳定性差的问题。为了缓解上述问题,提出了一种基于神经网络集成学习和语义增强特征的需求跟踪方法(Ensemble Learning Trace Approach,EMTrace)。该方法将需求跟踪问题转化为分类问题,集成了多个机器学习分类器进行预测,并对这些预测结果进行加权生成跟踪链接。为了自动获取各个基模型的权重,构建了一种基于神经网络的元学习器并利用每个基模型的预测结果进行训练。为了更准确地表达制品之间的跟踪链接,EMTrace方法使用多个词嵌入和句子嵌入模型提取软件制品的语义信息来增强跟踪链接特征的语义表示。实验结果表明,EMTrace方法能够有效提高需求跟踪的稳定性和性能,相比最优的基线方法,EMTrace方法在F1上提升了0.162。

The existing machine learning-based requirements traceability approaches suffer from instability.This paper proposes an Ensemble Learning Trace Approach(EMTrace)based on neural network ensemble learning and semantic-enhanced features,to alleviate these issues.The approach transforms the requirements traceability problem into a classification task by leveraging multiple machine learning classifiers for prediction.These classifiers are integrated,and weights are assigned to their predictions to establish trace links.To automatically obtain the weights of each base model,this paper constructs a neural network-based meta-learner and trains it using the predictions of each base model.Additionally,to accurately represent the trace links between artifacts,the EMTrace approach uses multiple word embedding and sentence embedding models to extract the semantic information of software artifacts and enhance the semantic representation of trace link features.Experimental results demonstrate that the EMTrace approach effectively improves the stability and performance of requirements traceability,achieving a 0.162 improvement in F1 compared to the optimal baseline method.