Abstract

Our research centers around exploring methodologies for developing reusable software, and developing methods and tools for building inter-enterprise information systems with reusable components. In this paper, we focus on an experiment in which different component indexing and retrieval methods were tested. The results are surprising. Earlier work had often shown that controlled vocabulary indexing and retrieval performed better than full-text indexing and retrieval [IEEE Trans. Software Engng (1994) 1, IEEE Trans. Software Engng 17 (1991) 800], but the differences in performance were often so small that some questioned whether those differences were worth the much greater cost of controlled vocabulary indexing and retrieval [Commun. Assoc. Comput. Mach. 28 (1985) 289, Commun. Assoc. Comput. Mach. 29 (1986) 648]. In our experiment, we found that full-text indexing and retrieval of software components provided comparable precision but much better recall than controlled vocabulary indexing and retrieval of components. There are a number of explanations for this somewhat counter-intuitive result, including the nature of software artifacts, and the notion of relevance that was used in our experiment. We bring to the fore some fundamental questions related to reuse repositories.

Author Keywords: Software reuse; Multi-faceted classification; Boolean retrieval; Plain-text retrieval; Retrieval evaluation; Approximate retrieval

Article Outline

1. Introduction

1.1. Component retrieval: do we still care?

1.2. The component retrieval problem

2. ClassServer: an experimental component repository

2.1. Overview

2.2. A multi-faceted classification of components

2.3. Software component retrieval in ClassServer

2.3.1. Free text indexing and search

2.3.2. Multi-faceted controlled-vocabulary retrieval

2.4. The component library

3. Constructing domain vocabulary

3.1. Extracting a set of concepts

3.2. Constructing a hierarchy of important domain concepts

3.2.1. Principles

3.2.2. A case-study: the Genbank experiment

3.2.3. Constructing the graph based on OSEs on-line documentation

4. Automatic indexing from controlled vocabulary

4.1. The algorithm

4.2. Results

5. Retrieval experiments

5.1. Experimental design

5.2. Queries

5.3. Component relevance: a performance-based evaluation

5.4. Performance results

6. Conclusion and directions

Acknowledgements

References