Σχολή Επιστημών και Τεχνολογίας της Πληροφορίας

Μόνιμο URI για αυτήν την κοινότηταhttps://pyxida.aueb.gr/handle/123456789/2

Η Σχολή Επιστημών και Τεχνολογίας της Πληροφορίας περιλαμβάνει τα Τμήματα: - Τμήμα Πληροφορικής - Τμήμα Στατιστικής

Περιήγηση

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω

Τώρα δείχνει 1 - 1 από 1

Brand-based sentiment analysis
(2021) Ormyliotou, Georgia; Ορμυλιώτου, Γεωργία; Pavlopoulos, Ioannis; Athnes University of Economics and Business, Department of Informatics
In this thesis, machine and deep learning models were applied in order to improve sentiment analysis on brand-specific texts. As a side task, Named-Entity Recognition experiments took place in order to analyze the brand name detection and possibly use the model as a Brand-Entity Recognition system, a brand-focused version of NER. For the purposes of NER, the pre-trained algorithms of SpaCy and Greek BERT-NER were used and, then, the Greek BERTbased model was fine-tuned using our dataset. With respect to sentiment analysis, the following machine learning algorithms were initially trained, Random Forests, Multinomial Naive Bayes, Logistic Regression and Linear SVC. Then, deep learning architectures were implemented such as BiGRU and CNN and in the end, we experimented with transformers such as BERT and XLM-RoBERTa (XLM-R). Considering this as the text-level sentiment estimation, a brand-level sentiment estimation was developed making use of the brand information in two ways. The first one is by feeding the brand to the model (input) and the second one is estimating the brand (output) alongside sentiment, using multi-task learning. The results have shown that there is a correlation between the brand and the sentiment, which as a feature should not be defied.