Supporting Program Indexing and Querying in Source Code Digital Libraries

1
Supporting Program Indexing and Querying in
Source Code Digital Libraries

• Yuhanis Yusof and Omer F. Rana
• Cardiff University, Wales, UK

2
Presentation Outline

• Background
• Software reuse
• Software repositories
• Literature review
• Agent-based architecture in s/code DLs
• Agent Interaction
• Program retrieval
• Results
• Conclusions Future Work

3
Background

• Software reuse creating software systems from
  existing software
• Software artifacts such as source code, design
  specifications, software components
• Software repositories
• contain a wealth of valuable information for
  empirical studies in software engineering
• source control systems store changes to the
  source code as development progresses
• defect tracking systems follow the resolution of
  software defects
• as the reuse of existing software artifacts
  becomes more important - store open-source
  applications in a DL format
• Requires suitable managing operations such as
  retrieval and querying

4
Background

• Current open source site sourceforge.net,
  freshmeat.net
• Keyword search based on initial description of
  the project
• Retrieve the (whole) project developers need to
  manually browse the content to locate required
  component
• Component-based reuse
• Formal language specifications Mili
  et.al(1994), Schumann and Fischer(1997),
  Penix(1999), Nakkrasae and Sophatsathit(2002)
• Faceted index - Prieto-Diaz(1985)
• Semantic net - Sugumaran and Storey(2003)

5
Our Approach

• Target users developers with intention to
  develop reusable software
• Reusable software offers code design
  scavenging, software component reuse and concept
  reuse
• Agents to index and retrieve source code programs
  based on
• Java program structure
• classes, comments, identifiers, packages and
  import statements
• Design patterns
• patterns of classes and communicating objects
  that solve specific design problems
• offers design solutions as reused entities
  concept reuse

6
Our Approach

• Higher level of abstraction to represent source
  code programs

Reused entity is more abstract and is designed to
be configured and adapted for a range of
situations
Concept Reuse
Developers are inevitably constrained by design
decisions specific reusable entities
Component Reuse
abstraction
Object/Function Reuse
Requires detail knowledge of the classes and
function modules
7
Agent-based Architecture

• Examples of query
• Natural language (keyword, phrase)
• Singleton registry class
• Program template

public class Registry private Registry
() public static Registry getRegistry()
synchronized(classlock) if (registry
null) registry new Registry()
return registry
8
Agent-based Architecture
9
Agent-based Architecture

• Portions of agents description for program index
  building is given as follows

10
Agent Interaction Program retrieval -
11
Agent Interaction Program retrieval -

• Agents roles
• Decompose queries
• Identify relevant information to be extracted
  from queries
• Identify collaborating classes (design patterns)
• The Index Creation Agent (ICA) consist of 3
  agents
• Keyword agent (KEMA)
• Terms in query are analyzed separately as an
  individual token
• Java Template agent (TEMA).
• class name(s)
• file name together with its full path
• method name and signatures
• superclass,
• abstract class
• interface class.
• Design Pattern agent (DEPA)
• Determines the existence of design patterns
  (Singleton, Composite Observer)

12
Results

• Similarities measurements Precision and Recall
  
• Precision (fraction of the retrieved programs
  which is relevant)
• retrieve relevant programs
• retrieve programs
• Recall (fraction of the relevant programs which
  has been retrieved)
• retrieve relevant programs
• relevant programs

• Experiment - 7 Java applications constituting of
  477 files
• Retrieval based on the first 50 documents

13
Results

• Both Composite and Observer has higher precision
  at lower recall values the top ranked retrieved
  programs are relevant, however, fail to retrieve
  all relevant programs
• Singleton is superior at higher recall values
  able to retrieve all relevant programs but the
  programs were ranked at the bottom

14
Discussions

• Result analysis - compensate between precision
  and recall. Higher recall produces less
  precision.
• Singleton Not many developers implement
  Singleton pattern (2 out of 477 files) due to the
  fact that it controls the creation of class
  instances.
• Composite Both precision and recall can be
  increase if the task agent (design pattern)
  manage to identify more interface class.
• Observer - balance trade-off between precision
  and recall as both analysis reached more than 80
  of value at the end of the retrieval
• The system is capable of achieving such result
  due to the cooperation between task agents,
  information agents and interface agents.
• Based on the search query and stored programs,
  appropriate information extraction is undertaken
  by the agents

15
Conclusion Future Work

• Program structure and design patterns can be use
  as retrieval methods in source code DLs
• Agents technology has proven to facilitate the
  process of indexing, querying and retrieving
  program source code
• decomposing problem into smaller chunks
• identifying the appropriate information
  extraction and collaborating classes
• Agent negotiation strategies need to be
  identified in generating optimum search results