aarondebattista09/ics5200

Dissertation repository.

ICS5200

This is the repository for the Msc dissertation, "Grammatical Error Correction for Maltese: A Deep Learning Solution for a Low Resource Scenario" by Aaron Debattista in partial fulfilment of the requirements for the degree of Master of Science in Artificial Intelligence (Universtity of Malta).

Contents

Common Voice

Folder containing PRE-TOKENISED and PRE-SYNTHESISED data files from Common Voice.

Evaluation Vocabs

Folder containing vocabularies for evaluating baseline models (baseline) and adapted models (22K).

Logs

Logs files from the training processes of the top performing projects during the study. These are mainly there to satisfy curiousity but are otherwise not important for assessment.

MLRS

Folder containing PRE-TOKENISED and PRE-SYNTHESISED data files from MLRS.

Overwrite

A folder containing files that need to overwrite existing files in third-party repositories for reasons of compatibility.

QariTalProvi_Batch_1

Contains the original files as well as the PRE-TOKENISED and CURATED datasets for Qari Tal-Provi (Batch 1).

QariTalProvi_Batch_2

Contains the original files as well as the PRE-TOKENISED and CURATED datasets for Qari Tal-Provi (Batch 2).

Template

Contains template files for plots/tables on Latex. It is not important for assessment.

Tools

Contains a number of tools used in support of the project.

• alignment.py - Generates alignment files.
• data_parser_QTP_Batch1.py - Was used to parse documents from QariTalProvi_Batch_1
• data_parser_QTP_Batch2.py - Was used to parse documents from QariTalProvi_Batch_2
• HF2M.py - Converts BERT model files into Marian-readable model files.
• M2_reference_generator.py - Used to generate reference M2 files for target files used in the final model (QariTalProvi_Batch_2)
• noiser.py - Calls methods from 'synthesis.py' in order to apply noise to the files from Common Voice and MLRS.
• synthesis.py - A library of synthesis methods.
• tokenisation.py - Contains Maltese tokeniser.