The aim of this study is to systematically examine the performance of transformer-based models for the detection of tumor morphology mentions in clinical documents in Spanish. For this purpose, we analyzed 3 transformer models supporting the Spanish language, namely multilingual BERT, BETO and XLM-RoBERTa. By means of a transfer- learning-based approach, the models were first pretrained on a collection of real-world oncology clinical cases with the goal of adapting trans- formers to the distinctive features of the Spanish oncology domain. The resulting models were further fine-tuned on the Cantemist-NER task, addressing the detection of tumor morphology mentions as a multi-class sequence-labeling problem. To evaluate the effectiveness of the proposed approach, we compared the obtained results by the domain-specific ver- sion of the examined transformers with the performance achieved by the general-domain version of the models. The results obtained in this pa- per empirically demonstrated that, for every analyzed transformer, the clinical version outperformed the corresponding general-domain model on the detection of tumor morphology mentions in clinical case reports in Spanish. Additionally, the combination of the transfer-learning-based approach with an ensemble strategy exploiting the predictive capabilities of the distinct transformer architectures yielded the best obtained results, achieving a precision value of 0.893, a recall of 0.887 and an F1-score of 0.89, which remarkably surpassed the prior state-of-the-art performance for the Cantemist-NER task.
Buscar en Dimension