Abstract
dc:descriptionThis doctoral thesis focuses on leveraging information derived from unstructured and semi-structured text, with the aim of extracting knowledge from natural language. The primary objective is the development of methodologies that either transform unstructured information into a format suitable for applying traditional analysis techniques or enable direct processing without requiring a predetermined structure. The proposed methodologies support categorization of information, detection of topics and trends, real-time text streaming monitoring, prediction of sentiment changes, and retrieval of critical information from free text, contributing to the substantial understanding and utilization of complex information flows. These elements constitute important tools in complex environments where business information analysis, technical report analysis, project descriptions, or field recording analysis are required. In the First Chapter, a concise overview of written discourse forms is presented, examining their evolution from ancient times to the present and their significance for knowledge preservation and advancement across various fields. This is followed by an analysis of technological developments that have created new possibilities for leveraging textual data. The different textual forms identified are examined alongside the fundamental preprocessing stages that prepare text for processing using machine learning methods. Furthermore, emphasis is placed on the subset of machine learning known as deep learning, which is based on neural networks, with a detailed presentation of the relevant algorithms. The chapter concludes with a description of the evaluation metrics used to assess the effectiveness of the proposed techniques in achieving quality results. In the Second Chapter, a literature review is conducted, beginning with a concise historical overview of Natural Language Processing, the first methods developed in the field, and their gradual evolution over time. Subsequently, research works that utilize different preprocessing stages are presented, aimed at optimizing the accuracy and efficiency of NLP systems. Text classification applications for real-world problems are examined, along with studies based on topic modeling to reveal hidden conceptual structures. Furthermore, text visualization and analytical exploration techniques that aim to improve understanding and interpretation of data and results are described. Finally, approaches for information extraction from unstructured text are included, with the goal of transforming it into structured form that can be used further for analysis. In the Third Chapter, a two-step classification algorithm is designed, accompanied by a two-cycle annotation process, with the aim of accelerating and improving the accuracy of the classification procedure. In the first step, useful information is filtered from the total volume of data, while in the second step, this information is classified into thematic categories according to its content. The annotation process begins with a small dataset labeled by experts, which is used for the initial training of the algorithm. Subsequently, the experts review and correct the model's predictions on new, unlabeled data, thereby enhancing the speed of system development. The results on real data from the maritime transport sector show that the two-step classification algorithm outperforms the simple one-step classification. In the Fourth Chapter, an algorithm is developed that is based on the automatic identification of topics or clusters within an extensive set of documents from different time periods. This algorithm combines topic modeling techniques, geometric features, and Large Language Models, with the aim of supporting systematic exploration and recognizing trends over time. Simultaneously, it functions as a valuable tool for assisting experts, enabling them to focus on specific, homogeneous areas within scientific fields or thematic clusters. The method has been successfully applied in the literature through various variations in the fields of time series, civil engineering, and medicine, demonstrating its adaptability and effectiveness across different scientific domains. In the Fifth Chapter, monitoring data streams in real-time is critical, particularly when dealing with unstructured texts, which have now become an integral part of our daily activities. In this context, two approaches are proposed that include a method for dynamic monitoring of evolving text streams and a method for predicting the sentiment polarity of texts. The first method combines natural language processing and text visualization techniques with traditional process monitoring algorithms, aiming to improve the analysis and understanding of evolving information flows. The second method incorporates sentiment analysis techniques with time series methodologies, seeking to detect long-term trends and develop predictive models that capture the dynamic evolution of sentiment content. Overall, the proposed methods constitute a comprehensive and flexible methodological framework for the dynamic analysis and monitoring of temporally varying unstructured data, enhancing informed decision-making and promoting deeper understanding of evolutionary trends in business and social environments. In the Sixth Chapter, an innovative two-stage approach is proposed, which incorporates machine learning techniques with the aim of extracting and transforming information from unstructured text into structured form. In the first stage, a sentence filtering process is applied for the purpose of selecting sentences that contain useful information for analysis. This is followed by the second stage, in which word classification is implemented through feature extraction techniques into predetermined categories of interest that have been defined by experts. The proposed method aims to develop a flexible and high-accuracy framework for processing and analyzing unstructured text, with the ultimate goal of transforming it into usable, structured information. Its effectiveness is demonstrated through its application to real datasets in the fields of medicine and shipping. In the Seventh Chapter, the conclusions arising from the different methods that were developed are summarized. The study findings demonstrate that leveraging unstructured information is not only feasible but also particularly accurate, offering significant advantages. The ability to extract substantial information from unstructured data constitutes a strategic tool for businesses, organizations, and researchers, as it accelerates the recognition of topics of interest and enhances their understanding. This leads to faster and more effective organization of information, improving competitiveness and decision-making. In a world where the volume of unstructured information is increasing exponentially, the need for such methods is more urgent than ever, making their utilization a critical success factor in the modern data era.
Degree
thesis:*- Grantor dc:publisher
- University of Patras
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Papageorgiou, Grigorios
- Παπαγεωργίου, Γρηγόριος
Subjects
dc:subject × 16- Επεξεργασία φυσικής γλώσσας
- Μεγάλα γλωσσικά μοντέλα
- Εξόρυξη πληροφορίας κειμένου
- Μηχανική μάθηση
- Βαθιά μάθηση
- Natural language processing
- Large language models
- Text mining
- Machine learning
- Deep learning
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60386