Recovering Traceability Links between | a Simple Natural Language Sentence and Source Code | Using Domain Ontologies | Takashi Yoshikawa, Shinpei Hayashi, and Motoshi Saeki

Conclusion: Domain ontologies give us valuable guides for traceability-recovering and feature-location.

Department of Computer Science Tokyo Institute of Technology, Japan

1. Motivation

. .

3. Methodology

JDraw's website (http://jdraw.sf.net/?page=6)

• Need to recover doc-to-code traceability links

J-Domain [JDraw] - Opera http://jdraw.sourceforge.net/index.php?page=6

– for reducing maintenance costs – for easily reusing existing code

• Simple NL sentences as documents – some software products do not contain rich information – e.g., just feature names, agile software development – feature-location approach

The system overview Sentence

Extracting Words

splitting, stemming, extract identifiers removing stop words...

Source Code Domain Ontologies

method invocation analysis

Words in the Sentence Words in the Code (w/ scores)

Call-graph Functional Call-graphs

Traversing call-graph

Inputs

Prioritizing

Outputs Ordered Functional Call-graphs

Source Code

Documents: simple NL sentences

Call-graph traversing (Extracting func. call-graphs) 1. Root selection

2. Problem and Our Solution

2. Traversal

• Problems of two basic criteria for feature-location – Usage of word similarity: leads to false positives (A) / negatives (B) – Usage of method invocations: leads to false positives (C) – Combination of both approaches could not be a solution

Another criterion is needed – representing the characteristics of the product's problem domain NL sentence (a set of words)

Source code (a set of methods and their invocations)

User can draw a plain oval. Ontology canvas

C

draw Oval() The correct fragments

writeLog()

getCanvas()

getColor()

setPixel() draw

void setPixel(...) { ... draw ... }

DrawPanel() oval

color

A

False positives (method invocation)

B

False negatives (word similarity)

getColorPallete() OvalTool()

False positives (word similarity)

• Our solution: using domain ontology – Ontology = binary relationships between concepts – Representing the domain knowledge of the target product family – e.g., the draw function concerns the color concept

– Checking co-occurrence of relationships

Sa

{wa, wb } – Choose the methods having the words of the input sentence – The words become the methods' roles – Traverse method invocations in depthfirst order iff the invocation satisfies one of the traversal rules – Method roles are extracted by the rules

{wa, ...}

m1

role: {wa}

m2 Sb

3. Extracting results

m4

{wb, ...}

m5

role: {wb}

m6

– The traversed set of methods are the candidates of feature-related code

m3

m7

m8

Traversal rules Rule #1 (sentence) getCanvas()

Rule #2 (ontology) drawOval()

Rule #3 (inheritance) getColor()

role: {canvas}

role: {draw, oval}

role: {color}

draw

User can draw a plain oval.

canvas

drawPixel() role: {draw}

getCanvas()

getColorPallete()

role: {canvas}

role: {color}

Prioritizing • Using three criteria: 1. Emphasizing methods having important roles 2. Emphasizing method invocations matching to the relationships in the ontologies and/or the sentence 3. Normalizing the whole score of invocations by the inclusion rate • How many input words the candidate includes?

4. Case Study Experimental Results

• Evaluation target: JDraw 1.1.5 – Having 1,450 methods – Picked up 7 sentences from JDraw's manual on the Web – Prepared an ontology for painting tools (38 concepts, 45 relationships)

• Improved accuracy – (1st and 2nd cases) recalls increase (although precisions a little decrease) – (3rd case) newly detected

• Bad effects, no improvement: future work – (4th case) reason: usage of external modules – (7th case) reason: usage of compound words

with ontology Input sentences

Prec. Recall

1. "plain, filled and gradient filled rectangles" 2. "plain, filled and gradient filled ovals" 3. "image rotation" 4. "image scaling" 5. "save JPEGs of configurable quality" 6. "colour reduction" 7. "grayscaling"

0.83 0.82 1.00 0.22 0.40 0.74 -

0.94 0.98 0.35 0.68 1.00 0.95 -

without ontology F

0.88 0.89 0.52 0.33 0.57 0.83 -

Prec. Recall 1.00 1.00 0.00 1.00 0.67 0.74 -

0.19 0.21 0.00 0.58 1.00 0.95 -

F 0.32 0.35 0.73 0.80 0.83 -

Picked up the results having the highest F-measure from the results ranked in the top 5 F-measure (F) = 2 / (precision-1 + recall-1)