Code for my B.Sc. thesis on Natural Language Processing (NLP) for the Ancient Greek language. The paper can be found here.

With the rise of the transformer Neural Network, Language Models (LMs) have been created and trained for most known languages. The performance of those models when fine-tuned on downstream tasks, surpasses all previous baselines by a significant margin. Thus, it can be argued that pre-training LMs has now become of paramount importance when aiming to achieve good performance on NLP tasks.

Surprisingly though, not many LMs have been created for Ancient Greek. After a thorough research, only two models were found:

1. A character-level BERT by Brennan Nicholson, which can be found here.
2. A token-level BERT by Pranaydeep et al., which can be found here.

The former achieved decent results, but since it's character-level, it doesn't have many use cases apart from Masked Language Modelling. The latter also achieved good results when fine-tuned on the downstream task of Part-of-Speech (PoS) Tagging.

The goal of the thesis was to create an improved Ancient Greek LM, by leveraging new BERT-like architectures and pre-training methods. The results can be found in the paper.

It is strongly advised to use a Unix-like OS for this project, as most scripts have been adapted to them.

1. Install Anaconda or Miniconda.

2. Download the repo using the command

   git clone https://github.com/AndrewSpano/BSc-Thesis.git && cd BSc-Thesis

3. Create and activate a virtual environment using the command

   conda create --name ag-nlp-venv python=3.8 && conda activate ag-nlp-venv

4. Install the required packages using the command

   pip install -r requirements.txt

To download the data, clean it, preprocess it and train a BPE tokenizer, simply run the script

python download_and_process_data.py

This script will

1. Download, clean the text data and save it in data/plain-text/ (directory will be overwritten if needed).
2. Train a tokenizer and save it in objects/bpe_tokenizer/ (directory will be overwritten if needed).
3. Use the tokenizer to create input IDs from the plain-text data and save it in data/processed-data/ (directory will be overwritten if needed).
4. Train a sklearn LabelEncoder on the PoS tags and save it in objects/le.pkl (overwriting it if needed).
   
   

The RoBERTa family of models from the transformers open-source library has been used. For the implementation of the training process, two different frameworks were used. Specifically, PyTorch Lightning (PL) and the Trainer API from huggingface were used.

The RoBERTa model uses the auxiliary task of Masked Language Modelling (MLM) to pre-train a Language Model.

The PL model for MLM is implemented in the LitRoBERTaMLM class located in mlm_model.py. To pre-train the LM with PL, the script pl_train_mlm.py can be used. Its arguments can be found in the function parse_pl_mlm_input located in cmd_args.py. An example of running the script is:

python pl_train_mlm.py                                       \
    --logdir logs/pl-mlm/                                    \
    --config-path configurations/pl-mlm-example-config.ini   \
    --savedir objects/PL-AG-RoBERTa/                         \
    --plot-savepath plots/pl-mlm.png                         \
    --device cuda                                            \
    --distributed                                            \
    --seed random

To pre-train the LM with the Trainer API from huggingface, the script hf_train_mlm.py can be used. Its arguments can be found in the function parse_hf_mlm_input located in cmd_args.py. An example of running the script is:

python hf_train_mlm.py                                       \
    --logdir logs/hf-mlm/                                    \
    --config-path configurations/hf-mlm-example-config.ini   \
    --savedir objects/HF-AG-RoBERTa/                         \
    --plot-savepath plots/hf-mlm.png                         \
    --seed 3407

Once a LM has been pre-trained, its performance can be evaluated by fine-tuning it to a downstream task and assessing the results on it. In this repo, the downstream task that was chosen is Part-of-Speech (PoS) Tagging.

The PL model for PoS Tagging is implemented in the class PoSRoBERTa which is located in pos_model.py. To fine-tune the LM with PL, the script pl_train_pos.py can be used. Its arguments can be found in the function parse_pl_pos_input located in cmd_args.py. An example of running the script is:

python pl_train_pos.py                                       \
    --logdir logs/pl-pos/                                    \
    --config-path configurations/pl-pos-example-config.ini   \
    --pre-trained-model objects/PL-AG-RoBERTa/               \
    --savedir objects/PL-PoS-AG-RoBERTa                      \
    --plot-savepath plots/pl-pos.png                         \
    --confusion-matrix plots/pl-pos-cm.png                   \
    --device cuda                                            \
    --distributed

To fine-tune the LM on PoS Tagging with the Trainer API from huggingface, the script hf_train_pos.py can be used. Its arguments can be found in the function parse_hf_pos_input located in cmd_args.py. An example of running the script is:

python hf_train_pos.py                                       \
    --logdir logs/hf-pos/                                    \
    --config-path configurations/hf-pos-example-config.ini   \
    --pre-trained-model objects/HF-AG-RoBERTa/               \
    --savedir objects/HF-PoS-AG-RoBERTa                      \
    --plot-savepath plots/hf-pos.png                         \
    --confusion-matrix plots/hf-pos-cm.png                   \
    --seed 3

Hyperparameter tuning in order to minimize the MLM validation loss during pre-training has also been implemented for both frameworks. Specifically, the HyperOpt library was used. The search spaces can be found in the files pl_tune_mlm.py and hf_tune_mlm.py. The only command line argument that they accept is the number of maximum evaluations to perform, which by default is 100 if it's not provided. An example of running the scripts is:

1. PyTorch Lightning

   python pl_tune_mlm.py --max-evals 150

2. Huggingface

   python hf_tune_mlm.py