University of Pisa, UNIPI
Academic year 2021/22
Authors: Irene Pisani, Alice Bergonzini
September, 2022
- Dataset: Official KPA-2021 dataset - ArgKP-2021
- Key Point Matching:
KPA_KeyPointMatching.ipynb
- Key Point Generation:
KPA_KeyPointGeneration.ipynb
- Report: KPA_report.pdf
This work aims to describe simple approaches for solving Key Point Matching (KPM) and Key Point Generation (KPG) tracks proposed at Argument Mining 2021 in the context of the shared task on Quantitative Summarization and Key Point Analysis (KPA). The presented methods rely on the fine-tuning of some state-of-the-art pre-trained language models both for KPM and KPG subtasks. Regarding the KPM task all the models explored were validated using the Hold-Out validation technique and their results were compared to analyze their effectiveness within the task. Leveraging DeBERTa pre-trained transformer, our best model yields to competitive performance since it achieved on the test set a mAP Strict and mAP Relaxed score of, respectively, 0,7035 and 0,8857. For the KPG task, a simple baseline based on abstractive summarization approach was provided; our system takes advantage of the pre-trained Google mT5 transformer to generate several key points that are finally properly selected.