On the use of positional proximity in IR-based feature location

Abstract

As software systems continue to grow and evolve, locating code for software maintenance tasks becomes increasingly difficult. Recently proposed approaches to bug localization and feature location have suggested using the positional proximity of words in the source code files and the bug reports to determine the relevance of a file to a query. Two different types of approaches have emerged for incorporating word proximity and order in retrieval: those based on ad-hoc considerations and those based on Markov Random Field (MRF) modeling. In this paper, we explore using both these types of approaches to identify over 200 features in five open source Java systems. In addition, we use positional proximity of query words within natural language (NL) phrases in order to capture the NL semantics of positional proximity. As expected, our results indicate that the power of these approaches varies from one dataset to another. However, the variations are larger for the ad-hoc positional-proximity based approaches than with the approach based on MRF. In other words, the feature location results are more consistent across the datasets with MRF based modeling of the features.

Exploiting spatial code proximity and order for improved source code retrieval for bug localization

“…We should also mention a recent study by Hill et al that investigates the effect of incorporating term‐term proximity and order on feature location. This work shows that MRF‐based approaches to leveraging positional proximity leads to more consistent retrieval accuracies in comparison to NLP‐based approaches .…”

Exploring the use of deep learning for feature location

“…Examples include the use of verb and direct object pairs [17] and statement level markov random fields [5]. Deep learning approaches allow the inclusion of broader context than these previous approaches.…”

“…In particular, we investigate the efficacy of document vectors [4] (DVs). DVs capture the influence of the surrounding context, and its order, on each term, which can improve the ranking of results retrieved for a developer query [5]. For example, in the statement diagram.redraw() the word diagram is relevant to the word redraw and this relationship is captured by DVs.…”