Controlled LLM Correction and Multilingual Inference for Italian Sentiment Analysis in Narrative Medicine

Narrative Medicine complements structured clinical information by providing access to patients’ and healthcare professionals’ lived experiences. Sentiment analysis is a promising tool, yet its application is hindered by textual noise and by the limited availability of robust Italian-language resources.We study Large Language Models as controlled preprocessing modules for Italian narratives, using constrained correction and Italian-to-English translation designed to preserve meaning. Viewing narratives as an unstructured textual modality, we investigate how sentiment polarity and confidence estimates depend on preprocessing choices and generation-parameter calibration.We introduce a three-stage paired pipeline that computes sentiment on variants derived from the same input: preprocessed Italian text, LLM-corrected Italian text, and its English translation, enabling comparison between Italian-based and English-translation-based sentiment classification. The study is conducted on two labeled Italian benchmarks and an unlabeled narrative medicine corpus, contributing: (i) a controlled pipeline for LLM-assisted correction and translation in Italian sentiment analysis, (ii) a comparison between Italian-specific and translation-based multilingual inference, and (iii) an analysis of generation-temperature effects and prediction stability for decision-support use.