Title Transformer neuronske mreže za obradu prirodnog jezika
Title (english) Transformer neural networks for natural language processing
Author Darko Petrović
Mentor Emil Dumić (mentor)
Committee member Mario Periša (predsjednik povjerenstva)
Committee member Robert Geček (član povjerenstva)
Committee member Emil Dumić (član povjerenstva)
Granter University North (University centre Varaždin) (Department of Multimedia, Design and Application) Koprivnica
Defense date and country 2022-09-29, Croatia
Scientific / art field, discipline and subdiscipline TECHNICAL SCIENCES Graphic Technology Processes of Graphic Reproduction
Abstract U ovom radu bit će ispitano nekoliko tipova transformer neuronskih mreža (Neural Networks, NN) za obradu prirodnog jezika. U radu će biti opisano područje obrade prirodnog jezika (Natural Language Processing, NLP) koja predstavlja vezu između računala i ljudskih jezika. Bit će objašnjeni i povezani pojmovi, shvaćanje prirodnog jezika (Natural Language Understanding) i stvaranje prirodnog jezika (Natural Language Generation, NLG). Najčešće se za obradu prirodnog jezika koriste različite metode strojnog učenja. Posebno će se opisati različiti tipovi dubokih neuronskih mreža koji mogu biti iskorišteni i za obradu prirodnog jezika: RNN (Recurrent Neural Network), ograđena RNN (gated RNN), LSTM (Long Short Term Memory) te transformeri. Zatim će se opisati modeli neuronskog strojnog prevođenja bazirani na pozornosti (attention-based neural machine translation). Potom će se detaljnije opisati transformer tip neuronske mreže, također baziran na pozornosti, te nekoliko tipova takvih istreniranih neuronskih mreža: BERT (Bidirectional Encoder Representations from Transformers), RoBERTa (Robustly optimized BERT pretraining approach), ALBERT (A Lite BERT) i XLNet. Usporedbu različitih metoda za strojni prijevod jednog jezika u drugi možemo napraviti koristeći BLEU ocjenu (Bilingual evaluation understudy). U praktičnom dijelu rada bit će prikazano nekoliko tipova transformera i njihovi rezultati za područje obrade prirodnog jezika. Za usporedbu će biti korištena neka od postojećih baza poput skup podataka knjiga Opus (eng. Opus books dataset) za prijevod, koristeći engleski i francuski jezik. Usporedba različitih tipova neuronskih mreža će biti dana koristeći BLEU ocjenu.
Abstract (english) In this paper, several types of transformer neural networks (NN) for natural language processing will be examined. The paper will describe the field of natural language processing (NLP), which represents the connection between computers and human languages. Related concepts, Natural Language Understanding and Natural Language Generation (NLG) will be explained. Different machine learning methods are most often used for natural language processing. Different types of deep neural networks that can be used for natural language processing will be described in particular: RNN (Recurrent Neural Network), fenced RNN (gated RNN), LSTM (Long Short Term Memory) and transformers. Then attention-based neural machine translation models will be described. Then the transformer type of neural network, also based on attention, and several types of such trained neural networks will be described in more detail: BERT (Bidirectional Encoder Representations from Transformers), RoBERTa (Robustly optimized BERT pretraining approach), ALBERT (A Lite BERT) and XLNet. A comparison of different methods for machine translation from one language to another can be made using the BLEU evaluation (Bilingual evaluation understudy). In the practical part of the work, several types of transformers and their results for the field of natural language processing will be presented. For comparison, some of the existing databases, such as the Opus books dataset, will be used for translation, using English and French. A comparison of different types of neural networks will be given using the BLEU score.
Keywords
Neuronska Mreža
Duboko Učenje
Transformer Neuronska Mreža
BERT
BLEU
Keywords (english)
Neural Network
Deep Learning
Transformer Neural Network
BERT
BLEU
Language croatian
URN:NBN urn:nbn:hr:122:471182
Study programme Title: Multimedia Science Study programme type: university Study level: graduate Academic / professional title: magistar/magistra inženjer/inženjerka grafičke tehnologije (magistar/magistra inženjer/inženjerka grafičke tehnologije)
Type of resource Text
File origin Born digital
Access conditions Open access
Terms of use
Created on 2023-01-27 12:55:51