Smart Knowledge Transfer using Google-like

Search
Srijoni Majumdar Partha Pratim Das
Advanced Department of
Technology Development Computer Science
Centre and Engineering
Indian Institute of Technology Indian Institute of Technology
Kharagpur-721302 Kharagpur-721302
[email protected] [email protected]
arXiv:2308.06653v1 [cs.SE] 12 Aug 2023

Abstract—To address the issue of rising software maintenance application specific entities, concepts have been located in
cost due to program comprehension challenges, we propose code comments based on enumerated domain concepts [12] or
SMARTKT (Smart Knowledge Transfer), a search framework, ontology [2], [34], [24], [20]. To extract project management
which extracts and integrates knowledge related to various
aspects of an application in form of a semantic graph. This graph details, comments are mined to track code changes in [1]
supports syntax and semantic queries and converts the process and [11] and software repositories are analysed to extract
of program comprehension into a google-like search problem. information related to bug history, version changes, developer
Index Terms—Program Comprehension, Knowledge Trans- and tester details and their interrelationships in [7], [29].
fer, Machine Learning, Natural Language Processing, Semantic Program comprehension can be aided by extracting relevant
Graph
knowledge from various sources (for a representative set, refer
In the last three decades, software maintenance cost has Table I) related to a working software. However, we observe
risen to 90% of the total Software Development Life Cycle that the available assistance tools consider only limited sources
(SDLC) cost [9], [10], [15]. Surveys conducted in [14], [30] and additionally there is an absence of an easy to use integrated
conclude that as 80% of the maintenance tasks are adaptive and framework based on these sources.
perfective [27], hence the absence of an integrated framework To analyse the comprehension challenges more specifically
or assistance for knowledge transfer (KT) to lessen program and understand the requirements for an effective design of
comprehension challenges contributes primarily to this rising an assistance framework, we conducted surveys and personal
cost. To execute a maintenance task, developers spend the interviews with a group of developers in a software company.
majority of their time to manually search and mine source files We present a representative scenario here (names have been
and other knowledge sources like design documents, defect changed for confidentiality): A developer Neha, working with
and version trackers, emails and the like, taking mental notes C++ and traditional Vi editor [16], is assigned to fix bug#67
or scribbling the mappings, in an attempt to infer an overall in ClearQuest [36], with error message “processing error :
knowledge about the design, behaviour and evolution of the unsigned 162 S1”. As she is new, she enquires from her
application [30], [25] so as to subsequently locate the relevant senior Sandra at every step. Sandra uses Cscope [35] tool
code sections and their dependencies. However, in most cases, in Vi to grep the code base with the error string and locates
documents are dated with missing information, tracker systems function VHDLPosedge#S2 in file VHDLPosedge.cc and
are not updated properly and help from earlier developers are provides to her. Neha asks Sandra for any similar defect.
scanty or not available. Due to these factors, coupled with Sandra searches ClearQuest, discovers bug#22 and searches
frequent interruptions [14] for attending calls or meetings, the the Microsoft Concurrent Versions System (CVS) [13] and
developers get involved in a tedious and inefficient process of emails to extract the bug related commit summary for code
building, revalidating and rebuilding their understanding of the level changes. The summary stated a change of data type from
application and resort to quick fixes which introduces hidden unsigned int to unsigned long int for variable
errors that cannot be removed by re-running the golden test var1, but in present code var1 has type – long int caus-
cases [30]. ing bug#67. Sandra then recalls this change as part of change
To address the program comprehension challenges, devel- request CR123 for optimisation of file VHDLPosedge.cc
opers extract software development related knowledge through a month ago. Sandra tells Neha to revert the datatype to
detection of low-level algorithm details using static instru- unsigned long int, as it would not affect the behaviour
mentation [3] or extraction of the control flow between run- of the code. As part of CR123, VHDLPosedge#S2 is called
time events using static analysis and dynamic profiling [33]. by a thread start function, so Sandra tells Neha to add mutex
Code search tools based on the abstract syntax tree have locks for read and writes in the function to prevent data-races.
been proposed in [31], [5], [23], [21], [22]. For extracting Sandra responds to queries of Neha based on multiple
relevant sources, and also provides additional important infor-
mation (Smart help). However due to evolving teams and frag-
mented task distribution, resources like Sandra who is aware
of various aspects of the application, are hardly available.

TABLE I
K NOWLEDGE T YPE – K NOWLEDGE S OURCES

Type, Description and Example Sources

Software Development: domain of program- Source Code, Runtime Trace,
ming like data structure, algorithm, memory, Code Comments, Version Trackers
concurrency. Example: variable, data race (SVN, github, CVS), Bug trackers
(JIRA, ClearQuest, Bugzilla),
Design Documents, KT sessions
Application Oriented: entities and actions Code Comments, Version Trackers,
of the specific application. Example: Convex Bug trackers, Design Documents,
Hull, bookmark KT sessions
Version Evolution: commits done for appli- Version Trackers, Code Comments,
cation along with summary. Example: Com- KT sessions
mit 71#: Code for new UI functions
Defect Evolution: bugs reported and fix Bug trackers, Code Comments, KT
summary and root cause analysis. Example: sessions
BugZilla#521: Data type mismatch
Project Management: mapping of developer Code Comments, Version Trackers
details to defects, version commits, source (SVN, github, CVS), Bug trackers,
code elements, most vulnerable module. Ex- Emails, KT sessions
ample: Developer A fixed bug BugZilla#521,
developed two modules
Business Specs.: client & company details. Induction manuals, emails, KT ses- Fig. 2. Architecture of S MART KT: Knowledge sources from Table I
Example: Business profile of clients sions

We propose a search framework named S MART KT for

single and multi-threaded C / C++ and python code-bases, S MART KT extracts knowledge from multiple sources (Ta-
designed to enact Sandra and respond to common queries ble I), associates them and represents them in form of a seman-
of maintenance engineers [32], [4] related to syntax and tic graph (Figure 1 corresponding to Sandra’s knowledge). The
semantics of a software. S MART KT supports four types of design (encompassing graph construction and query interface)
queries – a) Entity based [28]: Variable D?, Function A? b) of S MART KT (Figure 2) has been incorporated into five layers
List Search [28]: All static variables of file ftpety.c, All bugs which we discuss briefly below.
fixed on 12-03-2013 c) Template based [28]: Which is the
algorithm in function & what are the data-structures • Knowledge Primitive Extraction: In this layer, atomic
used? Function was effected by which bug numbers & units of knowledge (referred to as Primitives) like com-
how many were fixed by Developer ? d) Free-form english ment tokens, bug deployment date, global data write
queries: How many unsynchronised global variables are used event, program identifier are extracted from the relevant
to implement the UI Save button?. For each query, S MART KT sources using natural language processing [6] techniques
provides direct responses coupled with Smart information and instrumentation frameworks [19], [17].
like responding with additional data race alerts for queries • Feature Computation: In a software, application specific
on global variables. S MART KT additionally helps to bridge concepts are modeled in terms of software development
missing information across sources and validates application concepts (like data-structures, algorithm strategy, concur-
metadata (like comments). rency). In this layer, we infer each such concepts from
the primitives, using machine learning algorithms. For
example, using features extracted from runtime events,
structure of the application and an enumerated ontology
for algorithms, we can learn the classes (greedy, divide
and conquer) for the concept Algorithm Strategy.
• Inference and Knowledge Extraction: The inferred
concepts or knowledge primitives are mapped to
source code elements to construct knowledge triples
({primtive/inferred class, association, source code ele-
ment}) and form associations with each other. Example –
Bug#22, fixed on 12th July, 2015 (Project Management),
affects Foo1 (source code element) which is a virtual
member method (Software Development) of class F.
Fig. 1. Knowledge Graph (Note: VP stands for VHDLPosedge)
• Knowledge Graph Construction: A semantic graph is
constructed based on these triples and stored in RDF
 databases [26]. We use pagerank [38] and triangle count- [19] Chi-Keung Luk, Robert Cohn, Robert Muth, Harish Patil, Geoff Lowney,
ing [37] algorithms to learn from and extend the graph. Steven Wallace, Vijay Janapa Reddi, and Kim Hazelwood. Pin: building
customized program analysis tools with dynamic instrumentation. In
• Query Processing: We use SPARQL [8] to query graphs. Special Interest Group on programming languages notices (SIGPLAN),
For intelligent reponses, we extract related portion of the pages 190–200. ACM, 2005.
graphs based on word vector semantics [18]. For handling [20] Srijoni Majumdar, Ayush Bansal, Partha Pratim Das, Paul D Clough,
Kausik Datta, and Soumya Kanti Ghosh. Automated evaluation of
English language queries, we plan to employ syntax and comments to aid software maintenance. Journal of Software: Evolution
semantic matching to a corpus of queries. and Process, 34(7):e2463, 2022.
[21] Srijoni Majumdar, Nachiketa Chatterjee, Partha Pratim Das, and Amlan
A prototype of S MART KT to extract software development Chakrabarti. A mathematical framework for design discovery from
and application oriented knowledge with support for entity- multi-threaded applications using neural sequence solvers. Innovations
based, list and template-based queries has been developed and in Systems and Software Engineering, 17(3):289–307, 2021.
[22] Srijoni Majumdar, Nachiketa Chatterjee, Partha Pratim Das, and Amlan
is currently under trial by a group of professional developers. Chakrabarti. Dcube nn d cube nn: Tool for dynamic design discovery
from multi-threaded applications using neural sequence models. Ad-
R EFERENCES vanced Computing and Systems for Security: Volume 14, pages 75–92,
[1] Surafel Lemma Abebe, Sonia Haiduc, Andrian Marcus, Paolo Tonella, 2021.
and Giuliano Antoniol. Analyzing the evolution of the source code [23] Srijoni Majumdar, Nachiketa Chatterjee, Shila Rani Sahoo, and
vocabulary. In European Conference on Software Maintenance and Partha Pratim Das. D-cube: Tool for dynamic design discovery from
Reengineering (ESMR), pages 189–198. IEEE, 2009. multi-threaded applications using pin. In International Conference on
[2] Surafel Lemma Abebe and Paolo Tonella. Natural language parsing Software Quality, Reliability and Security (QRS), pages 25–32. IEEE,
of program element names for concept extraction. In International 2016.
Conference on Program Comprehension (ICPC), pages 156–159. IEEE, [24] Srijoni Majumdar, Shakti Papdeja, Partha Pratim Das, and Soumya Kanti
2010. Ghosh. Comment-mine—a semantic search approach to program com-
[3] Mustafa Basthikodi and Waseem Ahmed. Classifying a program code prehension from code comments. In Advanced Computing and Systems
for parallel computing against hpcc. In International Conference on for Security, pages 29–42. Springer, 2020.
Parallel, Distributed and Grid Computing (PDGC), pages 512–516. [25] Srijoni Majumdar, Papdeja Shakti, Partha Pratim Das, and Soumya
IEEE, 2016. Ghosh. Smartkt: A search framework to assist program comprehension
[4] Andrew Begel and Thomas Zimmermann. Analyze this! 145 questions using smart knowledge transfer. In International Conference on
for data scientists in software engineering. In International Conference Software Quality, Reliability and Security (QRS), pages 97–108. IEEE,
on Software Engineering (ICSE), pages 12–23. ACM, 2014. 2019.
[5] Nachiketa Chatterjee, Srijoni Majumdar, Shila Rani Sahoo, and [26] Eric Miller. An introduction to the resource description framework. Bul-
Partha Pratim Das. Debugging multi-threaded applications using pin- letin of the American Society for Information Science and Technology,
augmented gdb (pgdb). In International Conference on Software En- Wiley, 25(1):15–19, 1998.
gineering Research and Practice (SERP), pages 109–115. WorldComp- [27] Thomas M Pigoski. Practical software maintenance: best practices for
USA, 2015. managing your software investment. Wiley, 1996.
[6] Gobinda G Chowdhury. Natural language processing. Annual review of [28] Jeffrey Pound, Peter Mika, and Hugo Zaragoza. Ad-hoc object retrieval
information science and technology, Wiley, 37(1):51–89, 2003. in the web of data. In International conference on World Wide Web,
[7] Davor Čubranic and Gail C Murphy. Hipikat: Recommending pertinent pages 771–780. ACM, 2010.
software development artifacts. In International Conference on Software [29] Seher Razzaq, Yan-Fu Li, Chu-Ti Lin, and Min Xie. A study of the
Engineering (ICSE), pages 408–418. ACM, 2003. extraction of bug judgment and correction times from open source
[8] Richard Cyganiak. A relational algebra for sparql. Digital Media software bug logs. In International Conference on Software Quality,
Systems, HP Laboratories, 35:9, 2005. Reliability and Security Companion (QRS), pages 229–234. IEEE, 2018.
[9] Sayed Mehdi Hejazi Dehaghani and Nafiseh Hajrahimi. Which factors [30] Tobias Roehm, Rebecca Tiarks, Rainer Koschke, and Walid Maalej. How
affect software projects maintenance cost more? Acta Informatica do professional developers comprehend software? In International
Medica, The Academy of Medical Sciences of Bosnia and Herzegovina, Conference on Software Engineering (ICSE), pages 255–265. IEEE,
21(1):63–72, 2013. 2012.
[10] Len Erlikh. Leveraging legacy system dollars for e-business. IT [31] David Shepherd, Kostadin Damevski, Bartosz Ropski, and Thomas Fritz.
professional, IEEE, 2(3):17–23, 2000. Sando: an extensible local code search framework. In International
[11] Beat Fluri, Michael Wursch, and Harald C Gall. Do code and comments Symposium on the Foundations of Software Engineering (FSE), page 15.
co-evolve? on the relation between source code and comment changes. ACM, 2012.
In Working Conference on Reverse Engineering (WCRE), pages 70–79. [32] Jonathan Sillito, Gail C Murphy, and Kris De Volder. Asking and
IEEE, 2007. answering questions during a programming change task. IEEE Trans-
[12] Jose Luıs Freitas, Daniela da Cruz, and Pedro Rangel Henriques. actions on Software Engineering, IEEE, 34(4):434–451, 2008.
The role of concepts on program comprehension. In International [33] Jonas Trumper, Johannes Bohnet, and Jurgen Dollner. Understanding
Conference on Program Comprehension (ICPC). ACM, 2011. complex multithreaded software systems by using trace visualization.
[13] Dick Grune et al. Concurrent versions systems, a method for independent In International Symposium on Software Visualization (ISSV), pages
cooperation. VU Amsterdam. Subfaculteit Wiskunde en Informatica, 133–142. ACM, 2010.
1986. [34] Maria João Varanda Pereira, Mario Marcelo Beron, Daniela da Cruz,
[14] Andrew J Ko, Brad A Myers, Michael J Coblenz, and Htet Htet Aung. Nuno Oliveira, and Pedro Rangel Henriques. Problem domain oriented
An exploratory study of how developers seek, relate, and collect relevant approach for program comprehension. In OpenAccess Series in
information during software maintenance tasks. IEEE Transactions on Informatics (OASIcs), pages 21–37. Schloss Dagstuhl-Leibniz-Zentrum
software engineering, IEEE, 32(12):971–987, 2006. fuer Informatik, 2012.
[15] Jussi Koskinen. Software maintenance costs. Technical report, Informa- [35] Girish Venkatachalam. Vim for c programmers. Linux Journal, Slashdot
tion Technology Research Institute, University of Jyvaskyla, 2003. Group, 2005(140):9, 2005.
[16] Linda Lamb, Arnold Robbins, and Arthur Robbins. Learning the vi [36] Ueli Wahli et al. Software configuration management a clear case for
Editor. ” O’Reilly Media, Inc.”, 1998. ibm rational clearcase and clearquest ucm, dec. 2004. IBM,.
[17] Chris Lattner and Vikram Adve. The llvm compiler framework and [37] Emo Welzl. Partition trees for triangle counting and other range
infrastructure tutorial. In International Workshop on Languages and searching problems. In Annual symposium on Computational geometry,
Compilers for Parallel Computing (LCPC), pages 15–16. Springer, 2004. pages 23–33. ACM, 1988.
[18] Omer Levy and Yoav Goldberg. Dependency-based word embeddings. [38] Wenpu Xing and Ali Ghorbani. Weighted pagerank algorithm. In
In Annual Meeting of the Association for Computational Linguistics Communication Networks and Services Research, pages 305–314. IEEE,
(ACL), pages 302–308. ACM, 2014. 2004.