This repo contains the code, dataset and models for our ACL short paper (Stammbach et al., 2023)
Assuming Anaconda and linux, the environment can be installed with the following command:
conda create -n environmental_claims python=3.6
conda activate environmental_claims
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch
pip install -r requirements.txt
We also host the dataset and model on huggingface
In our paper and dataset, we discard sentences where 2 annotators say a sentence "is an environmental claim", but 2 annotators disagree and therefore we have a tie. We host the full dataset here, including all 3000 sentences, and agreement between annotators (either 0.5, 0.75 or 1.0). Labels are:
To predict environmental claims in custom data, we provide an inference script. For running the script with some data (either a "jsonl" file with a column "sentences" or "text", or a ".txt" file with one sentence by line"), simply run the following python command.
python src/inference_script.py --filename data/test.jsonl --model_name climatebert-environmental-claims --outfile_name environmental_claim_predictions.csv
To replicate the baseline experiments, run the following python script.
python src/baselines.py
(this prints the rows in our Table 2 for these experiments)
To fine-tune a climatebert model on our dataset, run the following python script.
python transformer_models.py --do_save --save_path climatebert-environmental-claims --model_name climatebert/distilroberta-base-climate-f
(this also saves the resulting fine-tuned model in directory --save_path)
If anything should not work or is unclear, please don't hesitate to contact the authors
This repo contains the code, dataset and models for our ACL short paper (Stammbach et al., 2023)
Assuming Anaconda and linux, the environment can be installed with the following command:
conda create -n environmental_claims python=3.6
conda activate environmental_claims
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch
pip install -r requirements.txt
We also host the dataset and model on huggingface
In our paper and dataset, we discard sentences where 2 annotators say a sentence "is an environmental claim", but 2 annotators disagree and therefore we have a tie. We host the full dataset here, including all 3000 sentences, and agreement between annotators (either 0.5, 0.75 or 1.0). Labels are:
To predict environmental claims in custom data, we provide an inference script. For running the script with some data (either a "jsonl" file with a column "sentences" or "text", or a ".txt" file with one sentence by line"), simply run the following python command.
python src/inference_script.py --filename data/test.jsonl --model_name climatebert-environmental-claims --outfile_name environmental_claim_predictions.csv
To replicate the baseline experiments, run the following python script.
python src/baselines.py
(this prints the rows in our Table 2 for these experiments)
To fine-tune a climatebert model on our dataset, run the following python script.
python transformer_models.py --do_save --save_path climatebert-environmental-claims --model_name climatebert/distilroberta-base-climate-f
(this also saves the resulting fine-tuned model in directory --save_path)
If anything should not work or is unclear, please don't hesitate to contact the authors