Natural Language Analysis - Expanding Identifiers to Normalize Source Code Vocabulary

EXPANDING IDENTIFIERS TO NORMALIZING SOURCE

CODE VOCABULARYPRESENTED BY DAWN LAWRIE

LOYOLA UNIVERSITY MARYLAND

IN COLLABORATION WITH DAVE BINKLEY

Friday, October 7, 11

VOCABULARY MISMATCH

DIFFERENT VOCABULARY IN SOURCE CODE AND OTHER SOFTWARE ARTIFACTS

EXAMPLE

REQUIREMENT - “FEATURE LOCATION”

SOURCE CODE - “FEATURELOCATION”

OR WORSE “FLOC”


PURPOSE OF NORMALIZE

COPE WITH VOCABULARY MISMATCH

SOURCE CODE

OTHER SOFTWARE DOCUMENTS


EXAMPLE PROBLEMS

CONSIDER IDENTIFIERS

FEATURELOCATION

FLOC


EXAMPLE PROBLEMS


FEATURE LOCATION

FLOC

SPLITTING PROBLEM


EXAMPLE PROBLEMS


FEATURE LOCATION

F LOC

SPLITTING PROBLEM

SPLITTING PROBLEM


EXAMPLE PROBLEMS


FEATURE LOCATION

FEATURE LOCATION

SPLITTING PROBLEM

SPLITTING ANDEXPANSION PROBLEM


WHY NORMALIZE?

MANY SE PROBLEMS CAN BE ADDRESSED USING INFORMATION RETRIEVAL (IR) TECHNIQUES

UN-NORMALIZED CODE LEADS TO AN UNDER ESTIMATE OF THE IMPORTANCE OF CRUCIAL WORDS


NORMALIZE PROBLEM STATEMENT

FIND THE BEST EXPANSION OVERALL POSSIBLE SPLITS

FLOC FEATURE LOCATION


NORMALIZE ALGORITHM

TERMINOLOGY

HARD-WORD - WHITEHOUSE_LAWN

SOFT-WORD - WHITE-HOUSE_LAWN


NORMALIZE ALGORITHM

TERMINOLOGY



(2)


NORMALIZE ALGORITHM

TERMINOLOGY



(2)

(3)


NORMALIZE ALGORITHM


NORMALIZE ALGORITHM

STRLEN STRING LENGTH


MACHINE TRANSLATION APPROACH

EL PAPA VISITA LA IGLESIA




THEFATHERPOTATOPOPE

VISITSVISITORHIT

THE CHURCH




THEFATHERPOTATOPOPE

VISITSVISITORHIT

THE CHURCH




THEFATHERPOTATOPOPE

VISITSVISITORHIT

THE CHURCH

STRONG COHESION




THEFATHERPOTATOPOPE

VISITSVISITORHIT

THE CHURCH

STRONG COHESION


NORMALIZE ALGORITHM


NORMALIZE ALGORITHM

STRLEN


NORMALIZE ALGORITHM

STRLENS-TRLEN

ST-RLEN

STR-LENSTRL_ENSTRLE_NS_T_RLENS-TR-LENS_TRL_ENS_TRLE_NST_R_LENST_RL_ENST_RLE_NSTR_L_ENSTR_LE_NSTRL_E_NS_T_R_LENS_T_RL_ENS_T_RLE_NS_TR_L_ENS_TR_LE_NS_TRL_E_NST_R_L_ENST_R_LE_NST_RL_E_NSTR_L_E_NS_T_R_L_ENS_T_R_LE_NS_TR_L_E_NST_R_L_E_NS-T-R-L-E-N


NORMALIZE ALGORITHM

STRLENS-TRLEN

ST-RLEN


E(RLEN) = {RIFLEMEN}


NORMALIZE ALGORITHM

STRLENS-TRLEN

ST-RLEN


E(RLEN) = {RIFLEMEN}

WILDCARD EXPANSION

R*L*E*N*


NORMALIZE ALGORITHM

STRLENS-TRLEN

ST-RLEN


E(ST) = {SET, STOP, STRING}E(RLEN) = {RIFLEMEN}

E(STR) = {STEER, STRING}E(LEN) = {LENDER, LENGTH}


NORMALIZE ALGORITHM PART I

STRING STEER

VSSTR



STRING STEERLENDERLENGTH

LENDERLENGTH

VSSTR




LENDERLENGTH

VS

1. FIND COHESION BY SUMMING LOG OF PROBABILITIES OF WORD PAIRS

STR




LENDERLENGTH

VS

+ +


COHESIONBCOHESIONA

STR




LENDERLENGTH

VS

+ +


2. SELECT EXPANSION THAT MAXIMIZES COHESION

COHESIONBCOHESIONA

STR




LENDERLENGTH

VS

+ +



COHESIONBCOHESIONA

STR




LENDERLENGTH

VS

+ +



COHESIONBCOHESIONA

STRING

STR


NORMALIZE ALGORITHM PART II

VS

STR-LEN ST-RLEN



VS

STR-LEN ST-RLENSTRING LENGTH STOP RIFLEMEN



VS


1. FIND COHESION OVER EXPANSIONS



VS



2. SELECT EXPANSION OF THE SPLIT THAT MAXIMIZES COHESION



VS






VS




STRING LENGTH


ADDING CONTEXT


ADDING CONTEXT

DIR


ADDING CONTEXT

DIR E(DIR) = {DIRECTION, DIRECTORY}


ADDING CONTEXT


CONTEXT = {FORWARD, BACKWARD}


ADDING CONTEXT

FIND COHESION WITH CONTEXT WORDS IN ADDITION TO EXPANSIONS OF OTHER SOFT WORDS

USED IN BOTH PART 1 AND PART 2


CONTEXT = {FORWARD, BACKWARD}


NORMALIZE IMPLEMENTATION

USES GenTest TO SPLIT IDENTIFIERS

RETURNS MULTIPLE SPLITS

GOOGLE 5-GRAM DATASET


EVALUATION

Program Loc SLoc Unique Ids

which-2.20 3,670 2,293 487

a2ps-4.14 62,347 38,436 4,393

Program Selected Ids Hard Words Soft Words

which-2.20 487 903 1214

a2ps-4.14 211 459 618


EVALUATION

THREE GROUPS OF IDENTIFIERS

STANDARD LIBRARY CALLS

NAMES FROM STANDARD HEADER FILES / KEYWORDS

DOMAIN NAMES


EVALUATION




DOMAIN NAMES


DOMAIN NAMES


EVALUATION




DOMAIN NAMES


DOMAIN NAMES

Program Filtered Ids Reported Ids

which-2.20 152 335

a2ps-4.14 46 166


EXAMPLE EXPANSIONS

id Top 10 Expansion

Top Expansion

nextchar next_character next_character

indfound index_found_need index_found

optarg option_are_g optarg

itemno i_them_not itemno


RESEARCH QUESTIONS

WHAT IS THE OVERALL ACCURACY OF NORMALIZE?

DOES THE VOCABULARY USED HAVE A SIGNIFICANT IMPACT ON THE EXPANSION’S ACCURACY?

CAN THE EXPANDER INFORM THE SPLITTER?

CAN THE SPLITTER INFORM THE EXPANDER?


ACCURACY ON DOMAIN IDS


SOURCE OF EXPANSION WORDS

SOURCE CODE

INTERNAL DOCUMENTATION

MANUAL


BEST VOCABULARY SOURCE?


FUTURE WORK

EXPLORING DIFFERENT SOURCES OF CO-OCCURRENCE DATA

EXPLORING DIFFERENT WAYS OF CALCULATING PROBABILITIES

EXAMINING NORMALIZATION IN CONTEXT OF AN INFORMATION RETRIEVAL TASK


SUMMARY

IDENTIFIERS ARE WRITTEN DIFFERENTLY THAN OTHER SOFTWARE DOCUMENTS

DEGRADES PERFORMANCE OF IR TECHNIQUES

NORMALIZE CURRENTLY EXPANDS ABOUT HALF OF SOFT WORDS CORRECTLY


QUESTIONS?

Need an identifier split?GenTest Splitter available at

splitit.cs.loyola.edu


Technology

Natural Language Analysis - Expanding Identifiers to Normalize Source Code Vocabulary