Teksto suvokimas pasitelkiant neuroninius tinklus /

This work analyzes most popular and effective neural network architecture efficiency solving sentiment analysis and punctuation error detection tasks. The biggest problems preventing the investigation of such tasks were the lack of structurized datasets which were created as a part of this paper. Criteria were first established and the natural language processing tasks were identified before collecting the datasets. Two datasets were collected as a part of this research paper. The first dataset was online product review dataset for sentiment analysis task. Second dataset was collected from wikipedia lithuanian article corpus. It was later used to create a dataset for punctuation error detection. A brief overview was given on the latest achievements in the natural language space in transformer based neural networks like BERT. Comparison between the older Recurrent artificial neural network architecture and the newer transformer based neural networks is given. In the second part of the paper the analysis if sentiment analysis task is given, best neural network configuration and training methods are provided as well as comparison between recurrent neural network and transformer based neural network performance. The second analysis is given of punctuation error detection in lithuanian. The most accurate automatic punctual error correction tool in lithuanian is created based on a mixed explicit programmed rule and transformer based neural network approach according to authors knowledge. The results of the work help advance natural language processing sphere in lithuanian language and commercial tools could be created by fine tuning the approaches presented in this research paper. The work could also help to solve natural language processing problems in other less popular languages like latvian, estonian, etc.

Grantor dc:publisher

Institutional Repository of Vilnius University

Year dc:date

2020

Author dc:creator

• Barčauskas, Mindaugas,

Contributors dc:contributor

• Mirzianov, Oleg

Statement dc:rights

• info:eu-repo/semantics/openAccess

Language dc:language

lit

Repository record dc:identifier

https://repository.vu.lt/VU:ELABAETD81705781&prefLang=en_US

OAI identifier oai:identifier

oai:vu.lt:elaba:81705781