THE TEXT CLASSIFICATION PIPELINE STARTING SHALLOW, GOING DEEPER

An increasingly relevant and crucial subfield of Natural Language Processing (NLP), tackled in this PhD thesis from a computer science and engineering perspective, is the Text Classification (TC). Also in this field, the exceptional success of deep learning has sparked a boom over the past ten years. Text retrieval and categorization, information extraction and summarization all rely heavily on TC. The literature has presented numerous datasets, models, and evaluation criteria. Even if languages as Arabic, Chinese, Hindi and others are employed in several works, from a computer science perspective the most used and referred language in the literature concerning TC is English. This is also the language mainly referenced in the rest of this PhD thesis. Even if numerous machine learning techniques have shown outstanding results, the classifier effectiveness depends on the capability to comprehend intricate relations and non-linear correlations in texts. In order to achieve this level of understanding, it is necessary to pay attention not only to the architecture of a model but also to other stages of the TC pipeline. In an NLP framework, a range of text representation techniques and model designs have emerged, including the large language models. These models are capable of turning massive amounts of text into useful vector representations that effectively capture semantically significant information. The fact that this field has been investigated by numerous communities, including data mining, linguistics, and information retrieval, is an aspect of crucial interest. These communities frequently have some overlap, but are mostly separate and do their research on their own. Bringing researchers from other groups together to improve the multidisciplinary comprehension of this field is one of the objectives of this dissertation. Additionally, this dissertation makes an effort to examine text mining from both a traditional and modern perspective. This thesis covers the whole TC pipeline in detail. However, the main contribution is to investigate the impact of every element in the TC pipeline to evaluate the impact on the final performance of a TC model. It is discussed the TC pipeline, including the traditional and the most recent deep learning-based models. This pipeline consists of State-Of-The-Art (SOTA) datasets used in the literature as benchmark, text preprocessing, text representation, machine learning models for TC, evaluation metrics and current SOTA results. In each chapter of this dissertation, I go over each of these steps, covering both the technical advancements and my most significant and recent findings while performing experiments and introducing novel models. The advantages and disadvantages of various options are also listed, along with a thorough comparison of the various approaches. At the end of each chapter, there are my contributions with experimental evaluations and discussions on the results that I have obtained during my three years PhD course. The experiments and the analysis related to each chapter (i.e., each element of the TC pipeline) are the main contributions that I provide, extending the basic knowledge of a regular survey on the matter of TC.