Supervised Categorization of JavaScript^TM using Program Analysis Features

	Wei Lu            and      Min-Yen Kan
	luwei@nus.edu.sg                             kanmy@comp.nus.edu.sg   Computer Science Program                   School of Computing
	Singapore-MIT Alliance         National University of Singapore

Why to categorize JavaScript?

	Most crawlers and indexers ignore crucial information conveyed by external programs like JavaScript™
	Can we have them summarized automatically?

Categorization Scheme

Outline

	Introduction
	Categorization Scheme
	Feature Extraction Techniques
		Related Work
		Our Work
			Lexical Analysis
			Syntax Analysis
			Metrics Analysis
			Object Communication Analysis
			Contextual Analysis
	Conclusion

Related Work

	Source code categorization: Ugurel et.al.’s work published in 2002
		Seminal work on the subject
		Consists of two tasks
			Programming Language classification: find out the type of programming language – keywords, bi-grams… non-interesting!
			Topic classification: find out the topic related to the code - relies heavily on external resources (e.g. README file, code header), did not analyze source code itself much
	Can we extract features from source code itself?

Text Categorization Baseline

	Vector Space Model (VSM) is used in Text Categorization
	Bag of words approach
	Our baseline using Text Categorization Bag-of-Word approach: accuracy 87.47%
		A high baseline, but still has a gap of 12.5% to improve
		Also important to measure error reduction rate

Lexical Analysis

	How to tokenize JavaScript more reasonably?
	First attempt of improvement over baseline - we believe using a compiler-based approach is more reasonable
		Similar to POS tags used in Natural Language Processing , we introduce a tagset for JavaScript tokenization/tagging
		The tagging process comes with a token normalization
	Tags used:
	KEY:keyword; VAR:variable; SYM:symbol; NUM:number; STR:string; CMT:comment; REG:regexp

Evaluation on Lexical Analysis

Syntax Analysis

Evaluation on Syntax Analysis

Code Metrics Analysis

	Metrics: measurement of program source code
		Complexity: statistical measurements, e.g. CC, IFIN
			We proposed our own metrics in addition to published metrics

Evaluation on Metrics Analysis

Outline

	Introduction
	Categorization Scheme
	Feature Extraction Techniques
		Related Work
		Our Work
			Lexical Analysis
			Syntax Analysis
			Metrics Analysis
			Object Communication Analysis
			Contextual Analysis
	Conclusion

Object Communication Analysis

	var msg = "Welcome to this page";
	banner(0);
	function banner (index){
	var newWin = window.open();
	frm.txt.value="ok";
	window.status = msg.substring(0, index);
	index = index++;
	if (index >= msg.length) index = 0;
	window.setTimeout("banner("+index+" ) " , 100);
	}

Contextual Analysis

	Extracting information from the context of the enclosing web page

Evaluation on Object Communication and Contextual Analysis

Outline

	Introduction
	Categorization Scheme
	Feature Extraction Techniques
		Related Work
		Our Work
		Overall Evaluation
	Conclusion

Evaluation on All Components

Outline

	Categorization Scheme
	Feature Extraction Techniques
	Conclusion

Contributions

	Shown that program analysis can enhance source code categorization performance
		Both context-free and context-sensitive analysis
	Case study of JavaScript categorization
		New, functionality-based categorization
		Tool for feature extraction from JavaScript

Conclusions

	Limitations:
		Annotator Agreement
		Dynamic Analysis Incompleteness
		Choice of Classifier
	Future Work
		Source code classification of other languages
		Firefox extension / IE plug-in
	Dataset and system prototype available at:
	http://wing.comp.nus.edu.sg/~luwei/SMART

Question?

	Dataset and system prototype available at:
		http://wing.comp.nus.edu.sg/~luwei/SMART
	First author’s undergraduate honours year project thesis:
		http://wing.comp.nus.edu.sg/publications/theses/weiLuThesis.pdf
	Contacts:
		Wei LU: luwei@nus.edu.sg
		Min-Yen KAN: kanmy@comp.nus.edu.sg

Guidelines for talk

	20 minutes
	5 minutes for questions