Unsupervised retrieval of events before and after an anchor event on news corpora using SBERT embeddings.
This project includes implmentations for a clustering approach to retrieve sets of events before and after an anchor on both the News Narratives and Gigaword corpora. They can be run after setting up a python virtualenv and installing the packages in requirements.txt.
python3.8 -m venv env
pip install -U pip
pip install -r requirements.txt
# To run the notebook
jupyter notebook --no-browser
Example outputs and further implementation details are available in the Project Overview linked above.
The allennlp package used for SRL parsing uses an older version of transformers that conflicts with the sentence-transformers package. For now I've commented out the allennlp package from the requirements file, however you can install allennlp and its required transformers version after performing all the necessary encoding over the corpus you are clustering.
The notebook for querying is implemented specifically for the News Narratives corpus, but should be able to be easily adapted for Gigaword or other corpora. I've included the conda environment specifications in the querying folder so a new conda environment can be constructed from that file including all required packages. The overall implementation is outlined in the Project Overview.