Leveraging dynamic masked softmax and shared hidden layers for hierarchical text-based product classification with bert

Gross, Lotte

http://hdl.handle.net/10362/173787

Utilize este identificador para referenciar este registo.

Nome:	Descrição:	Tamanho:	Formato:
51403_Master_Thesis.pdf		1.87 MB	Adobe PDF	Ver/Abrir

Contacte-nos

Autores

Gross, Lotte

Orientador(es)

Han, Qiwei

Resumo(s)

This study explores the transformative impact of BERT and its variants, particularly RoBERTa, on hierarchical multi-class product classification. Leveraging the bidirectional nature of BERT, the research evaluates flat and hierarchical model architectures, revealing RoBERTa's superiority due to its nuanced understanding of diverse language styles in product titles. The hierarchical model, incorporating dynamic masked softmax, achieves a remarkable 96% accuracy in layer 2, showcasing efficient category handling. Despite longer training times, the innovative approach mitigates error propagation. The study emphasizes the trade-off between computational cost and interpretability, providing insights for future NLP research.