nlp-final-project

The repository contains the final project for the course "natural Language Processing" at RUG.
The project is base on esnli dataset and focuses on analysing the variants of explanation generation.

The dataset contains text pairs [premisis] and [hypothesis] and the [label] of the pair: entailment, contradiction or neutral. IN addition each pain acomplished with [explanation] on a natural language.
The task is to generate the explanation and label for the given pair.

The project is analysing three main approaches:

• Generating explanation using saparate model and then predict label using the explanation
• Jointly generate label and explanation using the same text generation model
• Predict label using saparate model and then generate explanation having label as input

Installation

Note: python 3.9 is required

Create a virtual environment and install the requirements

python -m venv .venv
python -m pip install -r requirements.txt

Cluster

The High Performance Computing cluster Hábrók of the RUG was used for models training and experiments.

The .sh scripts for running the experiments on the cluster are located in each scripts/ subfolder, the setup instruction is available as scripts/README.md.

Structure

• scripts/classification - contains training scripts for each any classification model used.
• scripts/generation - contains training scripts for each any explanation generation model used.
• scripts/pipeline - contains scripts to evaluate models than requires other models predictions as input.

Training scripts accept the following arguments:

• --base-model - model to finetune available at Hugginface Hub (e.g. roberta-base)
• --config-name - name of the config in params.json file (separate for each script)
• generation_type/classification_type - type of trainin process, whatever use or not explanation or lable in input data

See more python [script_name].py --help