EMOtivo: A Classifier for Emotion Detection of Italian Texts Trained on a Self-Labelled Corpus

Over the past decade, social media has led to the widespread use of natural language processing (NLP) techniques for studying digital communities. Initially, researchers focused on sentiment analysis, but later turned to more advanced approaches for detecting emotions. Unfortunately, most of these models are designed for English, neglecting languages such as Italian. Consequently, the scarcity of large annotated corpora for model training impairs the reliability of available models in real-world scenarios and hinders performance improvement. To address this issue, we propose a novel approach to develop an emotion detection classifier for Italian texts using a self-labelled corpus of downloaded tweets. Our proposed technique involves using emojis to assign emotional labels, as previous studies have shown a correlation between emojis and the overall emotional content of a message. We defined both a conceptual framework for assigning the labels and a model for classification. The proposed architecture employs a convolutional layer (CNN) and a long short-term memory (LSTM) layer. The classifier achieved an overall accuracy of 78.9% in accurately classifying 414 out of 525 tweets in the test set. Our approach shows promising results in capturing emotional meaning; however, we note that the quality of the classifier strongly depends on cleaning and pre-processing of input data.