1. Automatically detecting and describing high level actions within methods Presented by: Gayani Samaraweera

2. The problem  Given signature and the body of a method M, automatically discover each code fragment that implements a high level action comprising the overall algorithm of M, and accurately express each high level action as a succinct natural language description

4. Outline The problem Outline High level actions Method Evaluation Concerns Other uses Conclusion

5. High level actions Sequence fragment  A sequence of statements that when taken together represents a single high level action Conditional fragment  A conditional block that performs an action with subtle variations based on the condition Loop fragment  Code patterns that are commonly implemented using loop constructs that constitute a high level action

6. Method

7. Detecting high level actions Uses AST (Abstract Syntax Tree) CFG (Control Flow Gragh) Information from naming conventions and linguistic knowledge gained from observations of Java programs  Textual clues from SWUM (Software Word Usage Model)

8. cont.. Word usage information from identifiers Identifier splitting  Camel case splitting: on capital letters, underscores, numbers Eg: childXMLElement → child XML Element Expand identifier abbreviations  Eg: Button butSelectAll, MouseEvent evt SWUM → action, theme, optional secondary arguments of a statement grouping  Eg: list.add(Item i); → “add item to list” theme action secondary argument

9. Sequence as single action Sequence fragments Identifying sequences of statements with similar actions ◦ Indicated by similar method calls

10. cont.. Challenges Integrate to successor statement based on similarity  Different method names  Same method name different parameter types

11. cont.. Identifying fragments Statements with one or more method calls Add ended panel to content panel Add bid panel to content panel Verb → add → equals Head word of NP → panel → equals Preposition → to content panel → equals → integratable

12. cont.. Synthesizing descriptions If equal head word → plural Else Add okButton to content panel Head word of NP → different But if fields of same class → “all attributes” “different attributes”

13. Abstracting conditionals Challenges Integrating similar statements in different branches Integrating conditional statements guarding different branches Integrating return statements with literals or similar method calls

14. cont.. Identifying and describing conditionals Integrate statements of each block, compare each statement with statements of parent block  For method calls Singular

15. cont..  For return statements  For assignment statements Theme based on enclosing method Update, create or get

16. cont.. Describing conditional expressions Compare phrases as 'subject predicate object'  Subject and predicate are equal → based on what Based on what os name starts with If only head word of subject is equal Based on which Based on which radio button is selected

17. Finding traceable patterns in loops Challenges Common algorithms as finding, counting, copying Develop identification templates Develop heuristics to synthesize phrases for each template

18. cont.. Loop abstractions implemented Count Contains Find Copy Max-min

19. cont.. Identifying fragments and synthesize templates

20. cont.. Variations in synthesis templates 'find item (in collection) whose/which/such that ' in subject predicate object,  If item is subject → which  If an attribute of item is subject → whose  Default → such that

21. Evaluation Executed on 1.2 million methods across 1000 Java programs

22. cont.. How prevalent are the implemented high level methods? Sequence (methods with >= 10 statements 12.5%)  11% Conditional  40% of if-else  24% of switch Loop  51% of loops classified as iterating over all items in a collection  15% of iterator loops detected by implemented patterns

23. cont.. Potential reduction in reading detail Reduction in identified high level actions  Sequence → one phrase 22% of original size  Conditional → two phrases 29% of original size  Loop → varying # phrases 25% of original size

24. cont.. Precision of identification and description 15 human evaluators, each evaluating 15 code fragments (5-sequence, 5-conditional, 5-loop) 75 code fragments from 15 projects evaluated by 3 evaluators From methods with <= 20 statements 25-conditional, 25-sequence, 25-loop Loops: 5 fragments from each 5 patterns

25. cont.. Evaluators wrote an abstraction of the method Answered following based on 1 – strongly disagreeto5 – strongly agree identification description

26. cont.. Majority agreed or strongly agreed on both P1 and P2

27. Concerns.. May not generalize to other Java programs Results may vary on larger programs Results might not hold with novices Reduction in reading measurement may not hold with some developers

28. Improving client tools Extract method refactoring Create application based on what os starts with Set different attributes of SVGApplicationModel

29. cont.. Internal comment generation Instead of Extract Method refactoring, can add comments inline Add empty lines between related code fragments Suggesting more informative method names Improving automatically generated summary comments for a method

30. Conclusions First technique for identifying code fragments of statement sequences, conditionals and loops, that is abstracted to a high level action Automatically synthesizing natural language description

31. References 1. Giriprasad Sridhara, Lori Pollock, and K. Vijay-Shanker. Automatically detecting and describing high level actions within methods. In Proceeding of the 33rd international conference on Software engineering (ICSE '11). ACM, New York, NY, USA, 101- 110. 2. G. Sridhara, E. Hill, D. Muppaneni, L. Pollock, and K. Vijay-Shanker. Towards Automatically Generating Summary Comments for Java Methods. Intl. Conf on Automated Softw. Engg. (ASE’10), 2010. 3. Giriprasad Sridhara, Lori Pollock, K. Vijay-Shanker, "Generating Parameter Comments and Integrating with Method Summaries," International Conference on Program Comprehension, pp. 71-80, 2011 IEEE 19th International Conference on Program Comprehension, 2011 4. E. Hill. Integrating Natural Language and Program Structure Information to Improve Software Search and Exploration. PhD Dissertation, University of Delaware, 2010.

32. Thank you for listening!