University of Thessaly (UTH)
A hybrid and scalable framework for sentiment and opinion mining in greek short texts: from domain-specific lexicons and hybrid machine learning to active learning and large language models
Abstract
dc:descriptionSentiment analysis has become a fundamental task in Natural Language Processing (NLP), supporting the automatic interpretation of opinions and emotions expressed in textual data. Although recent advances in deep learning have significantly improved classification performance, low-resource languages continue to face substantial challenges arising from limited annotated corpora, scarce semantic resources, and the inability of generic lexical knowledge to capture domain-specific sentiment variations. Existing approaches typically address these challenges in isolation, lacking a unified methodology that systematically integrates semantic knowledge throughout the sentiment analysis process. This thesis proposes a unified knowledge-driven lifecycle for sentiment analysis in lowresource language environments. The proposed lifecycle systematically integrates the construction, acquisition, representation, and utilization of domain-specific semantic knowledge within a coherent methodological framework. Domain-specific semantic knowledge is first constructed through automatically generated sentiment lexicons, efficiently expanded using Active Learning to reduce manual annotation effort, transformed into structured and interpretable semantic representations, and finally exploited by deep learning models for sentiment classification. The proposed methodology is experimentally validated using multiple Greek social media datasets spanning diverse thematic domains, providing a realistic evaluation under low-resource conditions. The research introduces a reproducible methodology for constructing domain-specific semantic lexicons through the proposed Scorew and Weightw metrics, enabling the systematic quantification of both sentiment polarity and domain-specific lexical importance. It further demonstrates that Active Learning substantially reduces annotation effort while maintaining annotation quality, allowing semantic knowledge to evolve throughout the learning process. Building upon these resources, the proposed topic-aware structured semantic representation enables explicit semantic knowledge to be integrated with neural learning while preserving interpretability. The experimental results demonstrate that systematically constructed, efficiently acquired, and explicitly represented semantic knowledge can effectively complement data-driven learning in low-resource language environments. Beyond the performance of an individual sentiment classifier, the principal contribution of this thesis is the introduction and validation of a unified knowledge-driven lifecycle that establishes a coherent methodological foundation for integrating evolving semantic knowledge into sentiment analysis and provides a basis for future knowledge-driven research in low-resource Natural Language Processing
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Skoularikis, Kyriakos
- Σκουλαρίκης, Κυριάκος
Subjects
dc:subject × 24- Ανάλυση συναισθήματος
- Γλώσσες Χαμηλών Πόρων
- Εξειδικευμένη Σημασιολογική Γνώση Ανά Θεματικό Πεδίο
- Ενεργητική μάθηση
- Σημασιολογική Αναπαράσταση
- Βαθιά μάθηση
- Επεξεργασία φυσικής γλώσσας
- Μηχανική μάθηση
- Πλαίσιο Καθοδηγούμενο από τη Γνώση
- Text sentiment analysis
- Low-Resource Languages
- Domain Specific Semantic Knowledge
- Active learning
- Topic-Aware Semantic Representation
- Deep learning neural networks
- Natural Language Processing (NLP)
- Machine learning
- Knowledge-Driven Framework,
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/62690