Application of back-translation: A transfer learning approach to identify ambiguous software requirements

Ishan Mani Subedi, Maninder Singh, Vijayalakshmi Ramasamy, Gursimran Singh Walia

Research output: Chapter in Book/Report/Conference proceedingConference contribution

5 Scopus citations

Abstract

Ambiguous requirements are problematic in requirement engineering as various stakeholders can debate on the interpretation of the requirements leading to a variety of issues in the development stages. Since requirement specifications are usually written in natural language, analyzing ambiguous requirements is currently a manual process as it has not been fully automated to meet the industry standards. In this paper, we used transfer learning by using ULMFiT where we pre-trained our model to a general-domain corpus and then fine-tuned it to classify ambiguous vs unambiguous requirements (target task). We then compared its accuracy with machine learning classifiers like SVM, Linear Regression, and Multinomial Naive Bayes. We also used back translation (BT) as a text augmentation technique to see if it improved the classification accuracy. Our results showed that ULMFiT achieved higher accuracy than SVM (Support Vector Machines), Logistic Regression and Multinomial Naive Bayes for our initial data set. Further by augmenting requirements using BT, ULMFiT got a higher accuracy than SVM, Logistic Regression, and Multinomial Naive Bayes classifier, improving the initial performance by 5.371%. Our proposed research provides some promising insights on how transfer learning and text augmentation can be applied to small data sets in requirements engineering.

Original languageEnglish (US)
Title of host publicationProceedings of the 2021 ACMSE Conference - ACMSE 2021
Subtitle of host publicationThe Annual ACM Southeast Conference
PublisherAssociation for Computing Machinery, Inc
Pages130-137
Number of pages8
ISBN (Electronic)9781450380683
DOIs
StatePublished - Apr 15 2021
Externally publishedYes
Event2021 ACM Southeast Conference, ACMSE 2021 - Virtual, Online, United States
Duration: Apr 15 2021Apr 17 2021

Publication series

NameProceedings of the 2021 ACMSE Conference - ACMSE 2021: The Annual ACM Southeast Conference

Conference

Conference2021 ACM Southeast Conference, ACMSE 2021
Country/TerritoryUnited States
CityVirtual, Online
Period4/15/214/17/21

Keywords

  • Machine learning
  • Neural networks
  • Requirement engineering and quality
  • Transfer learning

ASJC Scopus subject areas

  • Computational Theory and Mathematics
  • Computer Science Applications
  • Hardware and Architecture
  • Software

Fingerprint

Dive into the research topics of 'Application of back-translation: A transfer learning approach to identify ambiguous software requirements'. Together they form a unique fingerprint.

Cite this