Adapting Word Embeddings to Traceability Recovery

Maintaining the traceability links of a software is tedious, error-prone task, but an essential requirement. Information retrieval has been approached to help to generate traceability links. Traceability links are usually determined by the similarity between two artifacts. However, methods are put forward mainly based on vector space model, topic model etc. which ignored the word semantic. According to that, this paper adapts the popular word embedding technique to traceability recovery tasks, and handle the out-of-vocabulary words at test time. In the end, a machine learning method is used (learning to rank) to improve our final result. Several contrast experiments are conducted on five public datasets, and the baseline methods are outperformed under the same condition.