Sentiment analysis for financial news
| dc.contributor.advisor | Filippakis, Michael | |
| dc.contributor.advisor | Φιλιππάκης, Μιχαήλ | |
| dc.contributor.author | Savvas, Spyros | |
| dc.contributor.author | Σάββας, Σπύρος | |
| dc.date.accessioned | 2025-10-15T07:39:18Z | |
| dc.date.available | 2025-10-15T07:39:18Z | |
| dc.date.issued | 2025 | |
| dc.identifier.uri | https://dione.lib.unipi.gr/xmlui/handle/unipi/18232 | |
| dc.format.extent | 108 | el |
| dc.language.iso | en | el |
| dc.publisher | Πανεπιστήμιο Πειραιώς | el |
| dc.rights | Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα | * |
| dc.rights.uri | http://creativecommons.org/licenses/by-nc-nd/3.0/gr/ | * |
| dc.title | Sentiment analysis for financial news | el |
| dc.type | Master Thesis | el |
| dc.contributor.department | Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών Συστημάτων | el |
| dc.description.abstractEN | This thesis explores the development and application of sentiment analysis techniques specifically tailored for financial news headlines, utilizing a dataset of 4,838 unique instances. The primary objective is to systematically evaluate and compare the performance of diverse methodologies, ranging from traditional lexicon-based approaches to state-of-the-art transformer architectures, for accurate sentiment classification (positive, negative, neutral) within the financial domain. Methodologically, the study implements and contrasts three distinct categories of models: (1) Lexicon-based classifiers utilizing the general-purpose VADER and the domain-specific Loughran-McDonald financial sentiment dictionary; (2) Fine-tuned Bidirectional Encoder Representations from Transformers (BERT) models, exploring variations in sentence representation through CLS, Mean, and Max pooling strategies, evaluated on a common imbalanced test set; and (3) The Gemma 7B-IT large language model, fine-tuned as a sequence classifier using parameter-efficient techniques (4-bit quantization and Low-Rank Adaptation - LoRA), also evaluated on the same common imbalanced test set. The models' performance is rigorously assessed using standard classification metrics, with a particular focus on the Macro F1-score to address observed class imbalance in the dataset. Experimental results unequivocally demonstrate the significant performance advantage of transformer-based models (BERT and Gemma) over lexicon-based approaches, which struggle with nuance and domain specificity. Fine-tuned BERT models achieve strong results (e.g., the CLS pooling strategy with non-stratified training data reached up to 88.2\% accuracy and a 0.87 Macro F1-score). The Gemma 7B-IT model, fine-tuned as a sequence classifier with stratified training data, also demonstrated comparable top-tier performance, achieving 87.2\% accuracy and a 0.87 Macro F1-score. This research contributes a valuable comparative benchmark for financial sentiment analysis, underscoring the effectiveness of modern, efficiently fine-tuned large language models and established transformer architectures for tackling domain-specific natural language processing tasks. | el |
| dc.contributor.master | Πληροφοριακά Συστήματα και Υπηρεσίες | el |
| dc.subject.keyword | Sentiment analysis | el |
| dc.subject.keyword | Transformer models | el |
| dc.subject.keyword | Natural Language Processing (NLP) | el |
| dc.subject.keyword | Large Language Models (LLMs) | el |
| dc.subject.keyword | BERT | el |
| dc.subject.keyword | Parameter-Efficient Fine-Tuning (PEFT) | el |
| dc.date.defense | 2025-07-08 |
Αρχεία σε αυτό το τεκμήριο
Αυτό το τεκμήριο εμφανίζεται στις ακόλουθες συλλογές
-
Τμήμα Ψηφιακών Συστημάτων
Department of Digital Systems
Εκτός από όπου διευκρινίζεται διαφορετικά, το τεκμήριο διανέμεται με την ακόλουθη άδεια:
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα