An effective disclosure of scientific knowledge and advancements to the general public is often hindered by the complexity of the technical language used in research which often results very difficult, if not impossible, for non-experts to understand. In this paper we present the approach developed by the SINAI team as the result of our participation in BioLaySumm shared task hosted by the BioNLP workshop at ACL 2024. Our approach stems from the experimentation we performed in order to test the ability of state-of-the-art pre-trained large language models, namely GPT 3.5, GPT 4 and Llama-3, to tackle this task in a few-shot manner. In order to improve this baseline, we opted for fine-tuning Llama-3 by applying parameter-efficient methodologies. The best performing system which resulted from applying self-play fine tuning method which allows the model to improve while learning to distinguish between its own generations from the previous step from the gold standard summaries. This approach achieved 0.4205 ROUGE-1 score and 0.8583 BERTScore.

This paper covers participation of the SINAI team in the shared task 1B: Radiology Report Summarization at the BioNLP workshop held on ACL 2023. Our proposal follows a sequence-to-sequence approach which leverages pre-trained multilingual general domain and monolingual biomedical domain pre-trained language models. The best performing system based on domain-specific model reached 33.96 F1RadGraph score which is the fourth best result among the challenge participants. This model was made publicly available on HuggingFace. We also describe an attempt of Proximal Policy Optimization Reinforcement Learning that was made in order to improve the factual correctness measured with F1RadGraph but did not lead to satisfactory results.

This paper covers participation of the SINAI team in Tasks 5 and 10 of the Social Media Mining for Health (#SSM4H) workshop at COLING-2022. These tasks focus on leveraging Twitter posts written in Spanish for healthcare research. The objective of Task 5 was to classify tweets reporting COVID-19 symptoms, while Task 10 required identifying disease mentions in Twitter posts. The presented systems explore large RoBERTa language models pre-trained on Twitter data in the case of tweet classification task and general-domain data for the disease recognition task. We also present a text pre-processing methodology implemented in both systems and describe an initial weakly-supervised fine-tuning phase alongside with a submission post-processing procedure designed for Task 10. The systems obtained 0.84 F1-score on the Task 5 and 0.77 F1-score on Task 10.