This study investigates deep-learning approaches to the problem of text classification, with a specific focus on an Italian-language corpus. Text classification is a central task in natural language processing in which predefined categories are assigned to units of text. Because of the discrete nature of language, effective pre-processing and the transformation of text into numerical representations are essential for enabling learning models to capture meaningful patterns. We examine multiple strategies for word representation and emphasise the role of word embeddings in model performance. In particular, two widely used pre-trained embedding methods—Word2Vec and ELMo—are compared, together with an alternative strategy in which embeddings are learned dynamically during training. For the classification task, both Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) recurrent networks are employed as feature extractors. Five models are trained and evaluated on a real-world multi-label dataset derived from Italian collective labour agreements. The results provide a comparative analysis of embedding techniques and neural architectures, highlighting their strengths and limitations and, in particular, the trade-off between predictive quality and computational cost.