Deep Learning for Natural Language Processing: Novel State-of-the-art Solutions in Summarisation of Legal Case Reports

Piscaglia, Nicola (2020) Deep Learning for Natural Language Processing: Novel State-of-the-art Solutions in Summarisation of Legal Case Reports. [Laurea magistrale], Università di Bologna, Corso di Studio in Ingegneria e scienze informatiche [LM-DM270] - Cesena
Documenti full-text disponibili:
[img] Documento PDF (Thesis)
Disponibile con Licenza: Creative Commons: Attribuzione - Non commerciale - Non opere derivate 3.0 (CC BY-NC-ND 3.0)

Download (11MB)


Deep neural networks are one of the major classification machines in machine learning. Several Deep Neural Networks (DNNs) have been developed and evaluated in recent years in recognition tasks, such as estimating a user base and estimating their interactivity. We present the best algorithms for extracting or summarising text using a deep neural network while allowing the workers to interpret the texts from the output speech. In this work both extractive and abstractive summarisation approaches have been applied. In particular, BERT (Base, Multilingual Cased) and a deep neural network composed by CNN and GRU layers have been used in the extraction-based summarisation while the abstraction-based one has been performed by applying the GPT-2 Transformer model. We show our models achieve high scores in syntactical terms while a human evaluation is still needed to judge the coherence, consistency and unreferenced harmonicity of speech. Our proposed work outperform the state of the art results for extractive summarisation on the Australian Legal Case Report Dataset. Our paper can be viewed as further demonstrating that our model can outperform the state of the art on a variety of extractive and abstractive summarisation tasks. Note: The abstract above was not written by the author, it was generated by providing a part of thesis introduction as input text to the pre-trained GPT-2 (Small) Transformer model used in this work which has been previously fine-tuned for 4 epochs with the”NIPS 2015 Papers” dataset.

Tipologia del documento
Tesi di laurea (Laurea magistrale)
Autore della tesi
Piscaglia, Nicola
Relatore della tesi
Corso di studio
Ordinamento Cds
Parole chiave
legal case reports,text mining,legal documents,text summarisation,legal case reports summarisation,deep learning,Natural Language Processing,NLP,Machine Learning,automatic summarisation,Riassunto automatico,Riassunto automatico di sentenze,Riassunto automatico di rapporti di casi giudiziari
Data di discussione della Tesi
19 Marzo 2020

Altri metadati

Statistica sui download

Gestione del documento: Visualizza il documento