Access Type

Open Access Dissertation

Date of Award

January 2013

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science

First Advisor

Andrian Marcus

Abstract

The text found in software artifacts captures important information. Text Retrieval (TR) techniques have been successfully used to leverage this information. Despite their advantages, the success of TR techniques strongly depends on the textual queries given as input. When poorly chosen queries are used, developers can waste time investigating irrelevant results.

The quality of a query indicates the relevance of the results returned by TR in response to the query and can give an indication if the results are worth investigating or a reformulation of the query should be sought instead. Knowing the quality of the query could lead to time saved when irrelevant results are returned. However, the only way to determine if a query led to the wanted artifacts is by manually inspecting the list of results. This dissertation introduces novel approaches to measure and predict the quality of queries automatically in the context of SE tasks, based on a set of statistical properties of the queries. The approaches are evaluated for the task of concept location in source code. The results reveal that the proposed approaches are able to accurately capture and predict the quality of queries for SE tasks supported by TR.

When a query has low quality, the developer can reformulate it and improve it. However, this is just as hard as formulating the query in the first place. This dissertation presents two approaches for partial and complete automation of the query reformulation process. The semi-automatic approach relies on developer feedback about the relevance of TR results and uses this information to automatically reformulate the query. The automatic approach learns and applies the best reformulation approach for a query and relies on a set of training queries and their statistical properties to achieve this. Both approaches are evaluated for concept location and the results show that the techniques are able to improve the results of the original queries in the majority of the cases.

We expect that on the long run the proposed approaches will contribute directly to the reduction of developer effort and implicitly the reduction of software evolution costs.

Download

Share

COinS