Deep sequence model for genre classification

Rapesis, Roberto; Ράπεσης, Ρομπέρτο

dc.contributor.advisor	Sotiropoulos, Dionysios
dc.contributor.advisor	Σωτηρόπουλος, Διονύσιος
dc.contributor.author	Rapesis, Roberto
dc.contributor.author	Ράπεσης, Ρομπέρτο
dc.date.accessioned	2025-07-11T09:29:33Z
dc.date.available	2025-07-11T09:29:33Z
dc.date.issued	2025-06
dc.identifier.uri	https://dione.lib.unipi.gr/xmlui/handle/unipi/17937
dc.format.extent	74	el
dc.language.iso	en	el
dc.publisher	Πανεπιστήμιο Πειραιώς	el
dc.title	Deep sequence model for genre classification	el
dc.title.alternative	Βαθιά ακολουθιακά μοντέλα για την ταξινόμηση στίχων	el
dc.type	Bachelor Dissertation	el
dc.contributor.department	Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Πληροφορικής	el
dc.description.abstractEN	This thesis advances the field of Music Information Retrieval (MIR) through a systematic investigation of deep learning approaches for lyrics-based music genre classification. Our principal contributions include a comprehensive lyrical corpus with genre annotations, deliberately stripped of non-lyrical metadata to isolate and evaluate the predictive power of lyrical content alone. We present a novel hybrid architecture integrating BERT's semantic extraction capabilities with a Multi-Layer BiLSTM with an attention mechanism, rigorously evaluated through both frozen and fine-tuned feature approaches. Additionally, we provide an open-source modular framework enabling reproducible experimentation via deterministic and configurable training pipelines, along with problem identification that explicitly discriminates between inherent task ambiguities and model-specific limitations. Through two rigorously documented experiments on lyrical genre classification, we comparatively evaluate standalone and hybrid modeling approaches. Our results not only validate previous claims in MIR literature through controlled experimentation but also empirically demonstrate our architectural innovations. This work delivers both practical tools for music classification and fundamental insights into the ontological nature and computational modeling of musical genres.	el
dc.subject.keyword	Genre	el
dc.subject.keyword	Genre classification	el
dc.subject.keyword	Classification	el
dc.subject.keyword	Deep neural network	el
dc.subject.keyword	Deep sequence model	el
dc.subject.keyword	Hybrid model	el
dc.subject.keyword	MIR	el
dc.subject.keyword	Lyrics	el
dc.subject.keyword	Song	el
dc.subject.keyword	Songs	el
dc.subject.keyword	LSTM	el
dc.subject.keyword	BiLSTM	el
dc.subject.keyword	Genres	el
dc.subject.keyword	Music	el
dc.subject.keyword	BERT	el
dc.subject.keyword	Song classification	el
dc.subject.keyword	Lyrics classification	el
dc.subject.keyword	Identification	el
dc.subject.keyword	Attention	el
dc.subject.keyword	Experiment	el
dc.subject.keyword	Experiments	el
dc.subject.keyword	Music information retrieval	el
dc.subject.keyword	Text classification	el
dc.subject.keyword	Natural language processing	el
dc.subject.keyword	Text embedding	el
dc.subject.keyword	Sentiment analysis	el
dc.subject.keyword	Semantic analysis	el
dc.subject.keyword	Transfer learning	el
dc.subject.keyword	NLP	el
dc.subject.keyword	Transformer models	el
dc.subject.keyword	Transformers	el
dc.subject.keyword	Framework	el
dc.date.defense	2025-06-16

Files in this item

Files	Size	Format	View
There are no files associated with this item.

This item appears in the following Collection(s)

Τμήμα Πληροφορικής

Show simple item record