yash9439/code-mixed-machine-translation

The project focuses on addressing the challenges of code-mixed machine translation, specifically from Hinglish to English.

Code-Mixed Machine Translation Project

This repository contains the project outline for the Code-Mixed Machine Translation Project undertaken as part of the ANLP (Advanced Natural Language Processing) course at IIIT Hyderabad by Team The Triad: Yash Bhaskar, Sankalp Bahad, and Utsav Shekhar.

Table of Contents

• Introduction
• Scope of the Project
• Challenges
• Dataset Exploration
• Evaluation Metrics
• Literature Review
• Research Gap
• GitHub Repository Links for Reference Implementations
• Project Implementation Plan
• Conclusion

Introduction

Problem Statement

• Niche problem: Code-Mixed Machine Translation
• Code-mixed data is prevalent in the daily lives of multilingual individuals.
• Machine translation is crucial for bridging the linguistic gap in code-mixed data.
• Limited research exists in the field of code-mixed machine translation.

Scope of the Project

• Address the challenges of code-mixed machine translation, specifically from Hinglish to English.
• Improve upon baseline results using various techniques.
• Aim to achieve better scores on MT Evaluation Metric.

Challenges

• Informal nature of code-mixing.
• Lack of formal sources for code-mixed languages.
• Variability in transliteration.
• Code-switching patterns.

Dataset Exploration

Available Datasets

1. Hinglish-TOP dataset

   • Size: 180,000 generated utterances, 10,000 human-annotated utterances, 170,000 synthetically generated utterances
   • Related Publications: CST5: Code-Switched Semantic Parsing using T5
   • Description: Code-switched semantic parsing dataset with human-annotated and synthetically generated utterances.

2. Dakshina Dataset

   • Collection of text in 12 South Asian languages.
   • Includes native script Wikipedia text and parallel data in native script and Latin alphabet.

Feasibility Assessment

• Hinglish-TOP dataset is a valuable resource for code-switched semantic parsing tasks.
• Diverse data examples with human-annotated and synthetic utterances.
• Connection to CST5 technique suggests its relevance and potential impact.

Evaluation Metrics

• BLEU Score: Measures similarity between model output and reference translations.
• CoMeT Score: Specifically designed for code-mixed language translation.
• chrF++ (character F-score with better tokenization): Suitable for evaluating code-mixed translations at the character level.

Literature Review (Findings)

Paper 1: CoMeT: Towards Code-Mixed Translation Using Parallel Monolingual

• Challenges of code-mixing in informal conversations.
• Data preparation using the Hinglish-TOP dataset.
• Model fine-tuning and post-processing.
• Evaluation metrics: BLEU, CoMeT, chrF++.

Paper 2: CST5: Data Augmentation for Code-Switched Semantic Parsing

• Addressing the challenge of code-switched semantic parsing.
• Data preparation and preprocessing.
• Evaluation metrics: Naturalness, Semantic Equivalence.
• Impact of CST5 on semantic parsing performance.

Research Gap

• Identification of gaps in current research:
  1. Informal Nature of Code-Mixing.
  2. Lack of Formal Sources.
  3. Variability in Transliteration.
  4. Code-Switching Patterns.

GitHub Repository Links for Reference Implementations

Datasets

1. Code-Mixed Machine Translation Dataset: Contains raw data and preprocessed versions for training and evaluation.

2. Hinglish-TOP Dataset: A valuable resource for code-mixed language research.

Baseline Implementation

1. Indic-Trans: Transliteration library for Hindi to Roman conversion.

2. Fairseq: Toolkit for sequence-to-sequence tasks, including machine translation.

Project Implementation Plan

• Overview of the Proposed Approach
• Tentative Timeline
• Expected Challenges and Mitigation Strategies

Conclusion

• Summary of the Project Outline
• Importance of Addressing the Code-Mixed Machine Translation Problem
• Anticipated Project Outcomes

---

Note: The information presented here is part of the project outline and may be subject to updates and modifications during the project's progress.