| Abstract: |
Requirements traceability provides critical support in helping manage software
system evolvement. Establishing and maintaining trace links are often arduous
problems which require intensive human effort, if traces need to be evaluated
manually. Automatic retrieval tools can help maintain traceability links by
dynamically identifying traces between artifacts. In order to effectively reduce
the effort involved in manual links discrimination, such automatic tools must
achieve high retrieval performance. This paper presents results of experimental
studies to analyze the performance of a dynamic trace retrieval approach
implementing a probabilistic information retrieval network. An implementation of
the retrieval approach described in the paper involves the definition of a small
training set for which traces must be known. This can be accomplished by either
using past knowledge or by manually evaluating the traces in the training set. A
study explores the effect of different size training sets on the retrieval
performance of the automatic tool. A second study analyzes methods for defining
confidence values, which are attached to each (un)retrieved trace to indicate
how confident we are that the dynamically retrieved trace represents a true
link, or vice versa that the not retrieved link is a false link. The results of
this research are beneficial for enhancing the utility and performance of
dynamic tracing tools. |