National Technical University of Athens (NTUA)
Αυτόματη ταξινόμηση κειμένων με χρήση αυτο-οργανούμενων χαρτών και μεθόδων μηχανικής μάθησης
Abstract
dc:descriptionThe growth of the World Wide Web both in terms of the number of connected nodes and the volume of information it contains has led to difficulties in effectively searching and retrieving information from end users. Similarly, on a smaller scale, in the context of a Digital Library or an Institutional Repository, increasing the volume of information tends to reduce search efficiency. Thus, the need arose for the development of new ways of representing the available information, accessing it and finally turning it into knowledge. Machine Learning techniques and more specifically techniques based on Unsupervised and Deep Neural Networks are considered as the most appropriate technology for the effective search and retrieval of information from texts. These techniques have the ability to discover related texts automatically using vector similarity measures. In particular, Unsupervised Learning techniques are, in this case, superior to the corresponding Supervised Learning techniques, as the latter require an extensive, meticulously labeled data set, which is usually difficult to develop in real applications. This doctoral dissertation is part of the broader research field of “Automatic Information Extraction from texts” using Machine Learning Techniques and deals with open topics in this area. Specifically, this dissertation approaches the popular problem of automatically extracting information from texts, with methods / approaches that are roughly divided into four main categories: a) extraction, b) assignment, c) hybrid and d) prediction methods. Methods of extracting information from texts are very diverse and are applied in a number of fields with a variety of applications. Having first introduced a number of different applications, where information extraction methods have been successfully adopted, the advantages of using such methods, especially in Digital Libraries, are examined. Then the problem of automatically classifying a news data set is approached, which is modeled as a multi-label classification problem. In this case, a Self-Organized Maps (SOM) Neural Network is used, while a simple but effective procedure is proposed that transforms the multi-label problem into a multi-class classification problem. In addition, an intelligent algorithm for selecting labels is proposed, in order to show that the neighboring nodes in the Map affect the selection of tags for a specific node. Finally, a heuristic method is used to select the size of the SOM. The extensive experimental analysis carried out showed that the proposed solution improves the efficiency of the classification, not only in terms of accuracy, but also in terms of computational resources and time required for the training of the Network. In the framework of this dissertation, an overview of multi-class classification methods is also carried out, while a process for the automatic classification of e-books by extracting information from the table of contents of the books is proposed. In this case an unsupervised machine learning neural network (SOM) and two Deep Neural Network architectures were used under different configuration scenarios. The aim of this process was to study the feasibility of developing a recommendations system to support students and teachers in identifying relevant sources based on a detailed thematic description (e.g., the summary or table of contents of a book) instead of some keywords based on the experimental analysis performed. Finally, in the context of this dissertation, the creation of a Linked Data portal using Semantic Web technologies is proposed, with the aim of integrating information extraction mechanisms and their results, with the ultimate goal of enriching metadata , in order to assist efficient search and retrieval of information from end users in the collections of a Digital Library.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2021
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Giannopoulou, Eleni
- Γιαννοπούλου, Ελένη
Subjects
dc:subject × 30- Εξαγωγή πληροφορίας
- Τεχνητή νοημοσύνη
- Στατιστική Ανάλυση Φυσικής Γλώσσας
- Αλγόριθμοι ταξινόμησης
- Ταξινόμηση Πολλαπλής Ετικέτας
- Αυτό-οργανούμενοι χάρτες
- Μη εποπτευόμενη Μηχανική Μάθηση
- Νευρωνικά Δίκτυα Βαθιάς Μάθησης
- Εμπλουτισμός Μεταδεδομένων
- Ψηφιακές βιβλιοθήκες
- Ταξινόμηση Βιβλίων
- Διανυσματικοποίηση Πίνακα Περιεχομένων
- Information extraction
- Artificial intelligence
- Statistical Natural Language Processing
- Classification algorithms
- Multi-label classification
- Self-organizing maps
- Unsupervised Machine Learning
- Deep neural networks
- Metadata enrichment
- Digital libraries
- Book Classification
- ToC Vectorization
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Μηχανική και συστήματα επικοινωνιών, Τηλεπικοινωνίες
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Communication engineering and systems, Telecommunications
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/49986