Natural Language Generation (NLG) is the process of generating human-like language by machines. One of the key challenges in evaluating the quality of generated text is to compare it with 'gold standard' references.

However, obtaining human annotations for evaluation is an expensive and time-consuming process, making it impractical for large-scale experiments. As a result, researchers have explored alternative methods for evaluating the quality of generated text.

Two families of metrics have been proposed: trained metrics and untrained metrics. While trained metrics may not generalize well to new data, untrained metrics, such as word or character-based metrics and embedding-based metrics, offer a more flexible and cost-effective solution. To assess the performance of an evaluation metric, correlation measures such as Pearson, Spearman, or Kendall tests are used, either at the text-level or system-level.

This project aims to benchmark the correlation of existing metrics with human scores on generation task: translation or data2text generation or story generation.

1. First, you need to clone the repository and cd into it :

git clone https://github.com/lidamsoukaina/NLG_Evaluation_Metrics.git
cd NLG_EVALUATION_METRICS

2. Then, you need to create a virtual environment and activate it :

python3 -m venv venv
source venv/bin/activate

3. You need to install all the requirements using the following command :

pip install -r requirements.txt

4. [Optional] if you are using this repository in development mode, you can run the following command to set up the git hook scripts:

pre-commit install

4. You can now run the python files in the file name folder using the following commands :

cd cluster
python3 [file name]

To test the project, you can run the test.ipynb notebook.

TODO: List the tested metrics

TODO: Describe the criteria used to evaluate the results

TODO: Describe and analyse the results

TODO: List the next steps

TODO: List the references

• LETAIEF Maram
• LIDAM Soukaina