Multilingual Intent Text Classification and Text Revision Model

This repository has been created in contribution to the final project titled Reproducibility Study of “Understanding Iterative Revision from Human-Written Text”, ACL Anthology 2022, for CS678 Advance Natural Language Processing under Professor Antonios Anastasopoulos and Professor Sina Ahmadi.

The paper aims to address the task of developing a comprehensive framework for modeling and understanding the iterative text revision process, to improve the quality and efficiency of the writing process through the creation of large-scale, multi-domain, edit-intention annotated corpora. The research question that the paper seeks to answer is how to develop a framework that can comprehensively model the iterative text revision process, including various domains of formal writing, edit intentions, revision depths, and granularities, and how to use this framework to create annotated corpora that can support computational modeling of iterative text revisions. Additionally, the paper investigates the relationship between edit intentions and writing quality and explores how incorporating annotated edit intentions can improve automatic evaluations of generative and edit-based text revision models. On these models, we perform Robustness and Multilinguality tests and document our results.

File Structure

• Report.pdf: Final documentation of the project
• code/
  • reproduction.ipynb: Notebook detailing the reproduction and experimentation of Understanding Iterative Revision from Human-Written Text
  • robustness-multilinguality.ipynb: Code to perform robustness and multilinguality tests on above
• cp1files: has all the files required for reproducing the paper “Understanding Iterative Revision from Human-Written Text” by Wanyu D et al. with additional models for comparision
• edit_generation: has shell scripts and python files for performing robustness and multilingual tests on the Pegasus and mT5 models for edit generation module
• intent_classification: has shell scripts and python files for performing robustness and multilingual tests on the RoBERTa and XLM-RoBERTa models for intent classification module
• multi_data: multilingual datasets with English, Hindi, French, Tamil and Chinese to perfrom zero shot and fully supervised analysis generated using Google API
• robust_data: noisy test data generated using "Beyond Accuracy: Behavioral Testing of NLP Models with CheckList" (Ribeiro et al., ACL 2020)

Team Members

• Pratish Mashankar G#: G01354094
• Sai Likhitha Allanki G#: G01336091
• William J David G#: G01129185

Declaration

Project submitted towards the partial procurement of credits for CS678 Advance Natural Language Processing at GMU under Professor Antonios Anastasopoulos and Professor Sina Ahmadi.