icpc 2012 - mining source code descriptions
DESCRIPTION
Very often, source code lacks comments that adequately describe its behavior. In such situations developers need to infer knowledge from the source code itself, or to search for source code descriptions in external artifacts. We argue that messages exchanged among contributors/developers, in the form of bug reports and emails, are a useful source of information to help understanding source code. However, such communications are unstructured and usually not explicitly meant to describe specific parts of the source code. Developers searching for code descriptions within communications face the challenge of filtering large amount of data to extract what pieces of information are important to them. We propose an approach to automatically extract method descriptions from communications in bug tracking systems and mailing lists. We have evaluated the approach on bug reports and mailing lists from two open source systems (Lucene and Eclipse). The results indicate that mailing lists and bug reports contain relevant descriptions of about 36% of the methods from Lucene and 7% from Eclipse, and that the proposed approach is able to extract such descriptions with a precision up to 79% for Eclipse and 87% for Lucene. The extracted method descriptions can help developers in understanding the code and could also be used as a starting point for source code re-documentation.TRANSCRIPT
Mining Source Code Descriptionsfrom Developer Communications
Sebastiano Jairo Massimiliano Andrian GerardoPanichella Aponte Di Penta Marcus Canfora
Context: Software Project
Documentation
Source Code
Developer
understanding
describes
Class diagram
Sequencediagram Program
Comprehension
understandingDifficult
Maintenance Tasks
Source Code
Developer
Documentation
Class diagram
understanding
describes
Sequencediagram
Coming back to the reality...
Context: Software Project
Program Comprehension
Maintenance Tasks
understandingDifficult
We argue that messages exchanged among contributors/developersare a useful source of information to help understanding source code.
Idea
In such situations developers need to infer knowledge from,
the source code itselfsource code descriptions in externalartifacts.
Developer
..................................................
When call the method IndexSplitter.split(File
destDir, String[] segs) from the Lucene cotrib
directory(contrib/misc/src/java/org/apache/luc
ene/index) it creates an index with segments
descriptor file with wrong data. Namely wrong
is the number representing the name of segment
that would be created next in this index.
..................................................
CLASS: IndexSplitter METHOD: split
A Five Step-Approach for Mining Method Descriptions
Developer
Step 1: Downloading emails/bugs reports and tracing them onto classes
Two heuristics
The discussion contains a fully-qualified class name (e.g., org.apache.lucene.analysis.MappingCharFilter); or the email contains a file name (e.g., MappingCharFilter.java)
For bug reports, we complement the above heuristic by matching the bug ID of each closed bug to the commit notes, therefore tracing the bug report to the files changed in that commit
Developer Discussion
When call the method .split(File destDir, String[] segs) from the
Lucene cotrib directory (contrib/misc/src/java/org/apache/lucene/index) it creates
an index with segments descriptor file with wrong data. Namely wrong is the number
representing the name of segment that would be created next in this index.
public void split(File destDir, String[] segs) throws IOException {
destDir.mkdirs();
FSDirectory destFSDir = FSDirectory.open(destDir);
SegmentInfos destInfos = new SegmentInfos }
If some of the segments of the index already has this name this results either to
impossibility to create new segment or in crating of an corrupted segment.
CLASS: IndexSplitter
IndexSplitter
Step 2: Extracting paragraphs
Two heuristics
We consider as paragraphs, text section separated by one or more white lines
We prune out paragraph description from source code fragments and/or stack Traces "by using an approach inspired by the work of Bacchelli et al.
Developer Discussion
When call the method IndexSplitter.split(File destDir, String[] segs) from the
Lucene cotrib directory (contrib/misc/src/java/org/apache/lucene/index) it creates
an index with segments descriptor file with wrong data. Namely wrong is the number
representing the name of segment that would be created next in this index.
public void split(File destDir, String[] segs) throws IOException {
destDir.mkdirs();
FSDirectory destFSDir = FSDirectory.open(destDir);
SegmentInfos destInfos = new SegmentInfos }
If some of the segments of the index already has this name this results either to
impossibility to create new segment or in crating of an corrupted segment.
PAR2
PAR3
PAR1
When call the method IndexSplitter.split(File destDir, String[] segs)
from the Lucene cotrib directory it creates an index with segments
descriptor file with wrong data. Namely wrong is the number
representing the name of segment that would be created next in this
index.
......................................................................................
......................................................................................
......................................................................................
......................................................................................
Step 3: Tracing paragraphs onto methods
These paragraphs must
respect the following
two conditions:
A) A valid paragraph must contain the keyword “method”
B) and the method name must be followed by a open parenthesis—i.e., we match “foo(”
Developer Discussion
PAR1
CLASS: IndexSplitter
METHOD: split(
A) B)
Step 4: Heuristic based Filtering
We defined a set of heuristics to further filter the paragraphs associated with
methods that assign each paragraph a score:
a) Method parameters:% of parameters
s1= mentioned in the paragraphs. Value between 0 and 1
1 if the methoddoes not
have parameters
..........................
Problem seems to come from
MainMethodeSearchEngine in
org.eclipse.jdt.internal.ui.launcher
The Method
searchMainMethods
,there's
a call to addSubTypes(List,
IProgressMonitor, IJavaSearchScope)
Method if includesSubtypes flag is
ON. This method add all types sub-
types as soon as the given scope
encloses them without testing if
sub-types have a main! After return
IType[] before the excecution
..........................
CLASS: MainMethodSearchEngine
(IProgressMonitor,
IJavaSearchScope, boolean)
METHOD: serachMainMethods% parameter = 100% -> s1= 1
SCORE
a) Method parameters:% of parameters
s1= mentioned in the paragraphs. Value between 0 and 1
b) Syntactic descriptions (mentioning return values):check whether the paragraph contains the
s2= keyword “return”. If YES Value equal 1, 0 otherwise
1 if the methoddoes not
have parameters
Equal to one if the method is
void.
..........................
Problem seems to come from
MainMethodeSearchEngine in
org.eclipse.jdt.internal.ui.launcher
The Method
searchMainMethods
,there's
a call to addSubTypes(List,
IProgressMonitor, IJavaSearchScope)
Method if includesSubtypes flag is
ON. This method add all types sub-
types as soon as the given scope
encloses them without testing if
sub-types have a main! After
IType[] before the excecution
..........................
CLASS: MainMethodSearchEngine
METHOD: serachMainMethods
SCORE
(IProgressMonitor,
IJavaSearchScope, boolean)
return
1+
% parameter = 100% -> s1= 1
=
Step 4: Heuristic based Filtering
We defined a set of heuristics to further filter the paragraphs associated with
methods that assign each paragraph a score:
a) Method parameters:% of parameters
s1= mentioned in the paragraphs. Value between 0 and 1
b) Syntactic descriptions (mentioning return values):check whether the paragraph contains the
s2= keyword “return”. If YES Value equal 1, 0 otherwise
1 if the methoddoes not
have parameters
Equal to one if the method is
void.
c) Overriding/Overloading:1 if any of the “overload” or
s3=“override” keywords appears in the paragraph, 0 otherwise
d) Method invocations:1 if any of the “call” or
s4=“excecute” keywords appears in the paragraph, 0 otherwise
..........................
Problem seems to come from
MainMethodeSearchEngine in
org.eclipse.jdt.internal.ui.launcher
The Method
searchMainMethods
,there's
a to addSubTypes(List,
IProgressMonitor, IJavaSearchScope)
Method if includesSubtypes flag is
ON. This method add all types sub-
types as soon as the given scope
encloses them without testing if
sub-types have a main! After
IType[] before the
..........................
CLASS: MainMethodSearchEngine
METHOD: serachMainMethods
SCORE =
return
1+
(IProgressMonitor,
IJavaSearchScope, boolean)
excecution
call
0+ 1
% parameter = 100% -> s1= 1
= 2
Step 4: Heuristic based Filtering
We defined a set of heuristics to further filter the paragraphs associated with
methods that assign each paragraph a score:
We selected paragraphs that have:
1. s1 ≥ thP = 0.5
2. s2 + s3 + s4 ≥ thH = 1
SCORE = 1+ 0+ 1
% parameter = 100% -> s1= 1 ≥ 0.5
= 2 ≥ 1
a) Method parameters:% of parameters
s1= mentioned in the paragraphs. Value between 0 and 1
b) Syntactic descriptions (mentioning return values):check whether the paragraph contains the
s2= keyword “return”. If YES Value equal 1, 0 otherwise
1 if the methoddoes not
have parameters
Equal to one if the method is
void.
c) Overriding/Overloading:1 if any of the “overload” or
s3=“override” keywords appears in the paragraph, 0 otherwise
d) Method invocations:1 if any of the “call” or
s4=“excecute” keywords appears in the paragraph, 0 otherwise
Step 4: Heuristic based Filtering
We defined a set of heuristics to further filter the paragraphs associated with
methods that assign each paragraph a score:
OK
Step 5: Similarity based Filtering
We rank filtered paragraphs through their textual similarity with the method they are likely describing.
Removing:- English stop words;- Programming
language keywordsUsing:- Camel Case
splitting the on remaining words
- Vector Space Model
METHOD PARAGRAPH SCORE Similarity
Method_3 Paragraph_4 2.5 96.1%
Method_1 Paragraph_1 2.5 95.6%
Method_2 Paragraph_2 1.5 97.4%
Method_3 Paragraph_3 1.5 86.2%
Method_1 Paragraph_3 1.5 79.0%
Method_3 Paragraph_2 1.5 77.5%
Method_2 Paragraph_4 1.5 64.3%
Method_2 Paragraph_3 1.3 83.2%
Method_3 Paragraph_1 1.3 73.9%
Method_2 Paragraph_1 1.3 68.7%
Method_1 Paragraph_4 1.3 53.6%
th>=0.80
Empirical Study• Goal: analyze source code descriptions in developer
discussions
• Purpose: investigating how developer discussions describe methods of Java Source Code
• Quality focus: find good method description in messages exchanged among contributors/developers
• Context: Bug-report and mailing lists of two Java Project Apache Lucene and Eclipse
Context
Research Questions
RQ1 (method coverage): How many methods from
the analyzed software systems are described by the paragraphs identified by the proposed approach?
RQ2 (precision): How precise is the proposed approach
in identifying method descriptions?
RQ3 (missing descriptions): How many potentially
good method descriptions are missed by the approach?
RQ1: How many methods from the analyzed software systems are described by the paragraphs identified by the proposed approach?
RQ2: How precise is the proposed approach in identifying method descriptions?
We sampled 250 descriptions from each project
RQ3: How many potentially good method descriptions are missed by the approach?
TABLE IIIThe analysis of a sample of 100 paragraphs traced to methods,
but not satisfying the Step 4 heuristic
System True Negatives False Negatives
Eclipse 78 22
Apache Lucene 67 33
We sampled 100 descriptions from each project
Conclusion