A Framework for Examning Topical Locality in Object-Oriented Software2012 IEEE International Conference on Computer Software and Applications

p76004546 江怡岑P76004685 王于庭

OUTLINE Introduction Background & Related work Framework Dataset and Experimental Procedure Static analysis results Conclusions

INTRODUCTION Program comprehension is a key

developer activity during software maintenance.

Topic models : rely on lexical information to identify topics that are semantically related to high-level domain concepts. LSI ( latent semantic indexing ) LDA ( latent Dirichlet allocation )

INTRODUCTION While topics reflect semantic

relatedness, it is believed that human evolves spatial cognition strategies to navigate the code base.

for object-oriented (OO) systems built on the principle of encapsulation, the entities should be spatially organized in a way that reflects the topics of software

INTRODUCTION the tenet of “topical locality”

spatial relatedness entails semantic relatedness So basic that in many cases it is not mentioned When the tenet is mentioned, its validity is not

measured explicitly.

our goal is to measure the extent to which this key tenet holds for OO systems. propose a framework to examine what extent three

relationships of topical locality hold in large-scale open-source projects.

BACKGROUND and Related WorkA. Way-finding in Code BaseB. Relating Spatial and Semantic

CuesC. Topical Locality Applied in

Software Engineering Tools

BACKGROUND and Related WorkA. Way-finding in Code Base

Developer comprehending a code base can therefore be thought of as continually trying to answer way-finding questions.

Moonen has examined way-finding in soft-ware and extended the concept of legibility to software.

BACKGROUND and Related Work B. Relating Spatial and Semantic Cues

We are interested in the interplay of different cues so that they can be effectively synthesized.

We focus on the relationship between two types of cues. Spatial. Semantic.

Spatial + Semantic = “topical locality” the software entities should be neither randomly

named nor randomly placed. Source code entities should be spatially organized to

reflect the semantics of software.

BACKGROUND and Related Work C. Topical Locality Applied in Software

Engineering Tools The idea of topical locality plays an important

role in building a number of software engineering tools.

Survey three tools Code Indexers Code Visualizers Code Summarizers

BACKGROUND and Related Work Code Indexers

An indexer takes source code and generates profiles of the code for later searching

Should index header comments ? we want to address how well name and

header comments represent the target code entity’s topic.

BACKGROUND and Related Work

BACKGROUND and Related Work Code Visualizers

Once a relevant code line is located , its surroundings provide valuable contextual information for the developer

examining topical locality of a contiguous fragment allows us to assess to what extent the code line indicates the topic of its surroundings.

BACKGROUND and Related Work

BACKGROUND and Related Work Code Summarizers

A summarizer generates a snapshot of the source code in order to reduce the cost for developers to read and understand the staggering amount of software repository information

Our contribution is to measure the degree of topical locality of the snapshot

BACKGROUND and Related Work

FRAMEWORKoverview Framework Overview

FRAMEWORKresearch questions Research questions

RQ1 : Which better conveys class body’s topic: class name, header comments, or a combination of both?

RQ2 : Can a code line indicate its surrounding’s topic?

RQ3 : Can a contiguous code fragment serve as a snapshot of the entire class?

FRAMEWORKmethod independent variables are concerned

with identifying spatial relationships dependent variable is about the

semantic relatedness Three measures:

TFIDF cosine similarity query term probability document overlap

We treat source code as document output score in the range [0, 1]

FRAMEWORKthree measures (1/3) TFIDF scheme – text mining model

𝑞𝑖 = ( )×𝑡𝑓𝑖 𝑄 𝑖𝑑𝑓𝑖 𝑤𝑖 = ( )×𝑡𝑓𝑖 𝑊 𝑖𝑑𝑓𝑖 𝑡𝑓𝑖 refers to the term frequency of 𝑡𝑒𝑟𝑚𝑖 𝑖𝑑𝑓𝑖 is the inverse document frequency, = 𝑖𝑑𝑓𝑖

2( +1/ ), where is the total number of 𝑙𝑜𝑔 𝑡 𝑑𝑓𝑖 𝑡documents in the corpus and is the number of 𝑑𝑓𝑖documents in which occurs.𝑡𝑒𝑟𝑚𝑖

FRAMEWORKthree measures (2/3) Query term probability

measures the likelihood of a term in the query/source being present in the target document.

FRAMEWORKthree measures (3/3) Document overlap

a set-based measure that quantifies the amount of overlap between two documents Q and W

Dataset and Experimental Procedure

LOC : the lines of code COM : the lines of comments CCs : the number of classes

Dataset and Experimental Procedure Use a source code indexer to process the code

base of the selected projects. The indexing process results in the profiles that

store partial and important information from the source code.

We calculate the three semantic relatedness measures (TFIDF-Cos, Prob and Overlap) based on the profiles.

RQ1 Can class name (N) and/or header

comment (H) convey the topic of class body(B) ?

Calculate the lexical similarity for (N,B), (H,B), (NH,B)

RQ2 Can a code line indicate the topic of its

surroundings? For randomly selected code line(L), we take a

contiguous code fragment of 30 lines as its surroundings (S) and select from the same file another 30-line contiguous code fragment(R)

Compare the lexical similarity of (L,S) with that of (L,R)

Those classes with at least 70 LOC are considered.

RQ3 Can a contiguous code fragment serve as a

snapshot of entire class? Form a code search perspective, the lexical

similarity of the snapshot should indicate the topical closeness of the classes

Randomly select a term w(‘data’ in Fig.4) to act as query keyword. The snapshot is extracted as 30-line contiguous code fragment.

Only consider classes with at least 60 LOC.

Static Analysis Results RQ1 : Name vs. Header RQ2:Code Line and Surroundings RQ3: Contiguous Fragment as a

Snapshot Threats to Validity

RQ1 : Name vs. Header NH is the closet to B in most cases, expect

MegaMek when measured by TFIDF, where NB is larger than HB and NHB.

=> MegaMek classes do not have useful header comments.

RQ1 : Name vs. Header Least Significant Distance(LSD) multiple comparison

test: a test places the combinations significantly different from others in separate groups, and allocates the best combination to ‘group A’.

The result classifies NH-B into ‘group A’, indicating that the similarity score of NH-B is significantly higher than N-B and H-B.

We conclude that if the class contains useful header comments, then it is important to combine the header comments with the class name in order to convey the topic of the class body.

RQ2:Code Line and Surroundings A code line indicates the topic of its

surroundings more than it indicates the topic of a random code fragment.

RQ3: Contiguous Fragment as a Snapshot We calculate the Pearson correlation coefficient,

which is a parametric statistic that shows the correlation between two variables.

From the viewpoint of distinguishing the topics of different classes, a contiguous code fragment can serve as a snapshot of the entire class.

Threats to Validity Construct Validity: the selection of 30-line

contiguous, non-empty, and comments-inclusive code fragment for addressing RQ2 and RQ3.

Empty lines contribute little to spatial and semantic information. All comments is a choice influenced by RQ1.

Internal validity : using three measures derived form different mathematical models diminished the measuring bias.

External validity : this analysis may not generalize to other software projects.

Conclusions In this paper, we contributed a novel

experimental framework for testing this tenet of “topical locality” and applied the framework to provide empirical evidence of topical locality in large-scale OO systems.

Our future work includes carrying out more empirical studies to examine other topical locality instances.

It is important to integrate the theoretical understandings and empirical findings to enhance the practical tool support for software developers.