Call transcription classification using natural language processing and machine learning techniques

This project investigates the automatic detection of adverse events in pharmacy related telephone call transcripts using classical machine learning techniques. The problem is formulated as a supervised binary text classification task applied to long, unstructured, and noisy conversational data under realistic constraints in terms of dataset size and computational resources. The main objective is to identify the most appropriate classical machine learning model for this task in terms of performance, robustness, and reliability. To this end, a structured experimental pipeline is defined, including text preprocessing, feature representation, model training, and evaluation. Several classical text representations are examined, such as TF-IDF and Word2Vec based embeddings, together with multiple classifiers and ensemble methods. Model selection is guided by explicit decision criteria that prioritize not only overall effectiveness but also sensitivity to adverse events and stability across validation splits. The results show that classical machine learning models can achieve reliable performance for adverse event detection when supported by careful preprocessing and principled evaluation. Sparse lexical representations such as TF-IDF provide the most balanced behaviour, while dense Word2Vec based representations tend to favour higher sensitivity at the cost of reduced precision. Based on a multi criteria selection strategy, the best classical configuration is identified as bagging with k-nearest neighbours using TF-IDF features. A comparison with a complementary transformer based approach developed in parallel shows that transformer models outperform classical methods under the same evaluation protocol. Nevertheless, the selected classical model constitutes a strong and interpretable baseline under constrained conditions. Overall, the project demonstrates the feasibility of adverse event detection from conversational transcripts and provides a clear reference for future research and practical applications.