This repository contains the resources for a work in progress about extracting predicate-argument templates that correspond to DBpedia triples to be used as input for the rule-based generator FORGe. The project is currently on pause and may be continued in 2025.
For the moment, we only released the code for assessing the similarity between a triple and a candidate text. The Colab notebook allows for reproducing the experiments described in our Depling'23 paper. Unfortunately we lost the seeds of the original experiment, so the exact numbers reported in the paper may not be reproduced.
In the Notebooks folder, Sentence_similarity_WebNLG-fineTune-evaluate.ipynb contains:
- A part for creating data for fine-tuning a Transformers model on the triple/sentence task.
- A part for creating a fine-tuned model.
- A part for evaluating the fine-tuned model and the corresponding off-the-shelf model.