Assignment 3 for language analytics (S2023). This repository contains code for training a text generation model on comments from articles from The New York Times using recurrent neural networks. Furthermore, a script is provided for generating text from a user-suggested prompt.

The model is trained on comments from articles from The New York Times. The data is available here.

Note: The model in this repo was trained on 50,000 randomly chosen comments to avoid memory issues.

To reproduce the results follow the directions below. All terminal commands should be executed from the root of the repository.

1. Clone the repository
2. Download the data from Kaggle and place it in the data directory (see repository structure section below if in doubt)
3. Create a virtual environment and install the requirements

bash setup.sh

4. Train the model by running the following commands from the command-line

source ./env/bin/activate
python src/train_model.py --n_comments 50000

5. Generate text

python src/generate_text.py --prompt "Donald Trump wins"

The pipeline was developed and tested on uCloud(Ubuntu v22.10m, Coder python v1.77.3, python v3.10.7).

It is possible to specify the model you want to use for text generation, if you have several in the repository.

├── data                        <- data folder (not included in the github repo)
├── mdl
│   ├── model_seq_291.h5
│   └── tokenizer_seq_291.pickle 
├── src
│   ├── generate_text.py
│   └── train_model.py
├── assignment_description.md
├── README.md
├── requirements.txt
└── setup.sh

To evaluate the text-generating capabilities of the model trained on the comments, several prompts were provided to the model. The generated outputs from the model are presented in the table below:

Based on these test examples, it is evident that the model struggles to produce coherent sentences that are grammatically and semantically meaningful. Increasing the number of comments the model is trained on might increase its usability, or training for more epochs.