magistrsko delo
Martin Božič (Author), Marko Robnik Šikonja (Mentor)

Abstract

Strojno popravljanje slovničnih napak v slovenskem jeziku je še ne rešen problem. Rešitev bi olajšala pisno komunikacijo. Problem v okviru magistrske naloge razdelimo na podprobleme: popravljanje zapisa besed, zaznavanje napačno zapisanih besed, popravljanje pregibanja besed in popravljanje vrstnega reda besed. Najboljše rezultate dosežemo z izpopolnjevanjem slovenskega SloT5 modela. Najboljše modele uporabimo pri izdelavi spletne aplikacije. Ugotovimo, da je pri reševanju problema popravljanja slovničnih napak najbolj pomembna izbira ustreznega osnovnega jezikovnega modela in izgradnja kvalitetne učne množice. Pri izgradnji učne množice skušamo zajeti čim več kvalitetnih in realnih slovničnih napak, ne da bi pri tem spremenili ali pokvarili izvorni pomen besedila.

Keywords

veliki jezikovni modeli;slovnične napake;popravljanje slovničnih napak;slovnični popravki;nevronske mreže;model T5;transformerji;model SloBERTa;računalništvo in informatika;magisteriji;

Data

Language: Slovenian
Year of publishing:
Typology: 2.09 - Master's Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [M. Božič]
UDC: 004.8:81'322(043.2)
COBISS: 168228099 Link will open in a new window
Views: 55
Downloads: 14
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Adaptation of large language models for grammar correction in Slovene
Secondary abstract: Machine correction of grammatical errors in the Slovenian language is still an unsolved problem. Its solution would improve written communication. We divide the problem into subproblems: correcting word spelling, detecting misspelled words, correcting word inflection and correcting word order. The best results are achieved by finetuning the Slovenian SloT5 model. We use the best models in a web application. We conclude that in correcting grammatical errors, the most important consideration is the choice of a large language model and construction of a learning set. When building the learning set, we try to capture as many realistic grammatical errors as possible, without changing the meaning of the text.
Secondary keywords: large language models;grammar correction;grammatical corrections;neural networks;machine learning;Slovene;model T5;transformers;model BERT;model SloBERTa;computer science;computer and information science;master's degree;Računalniško jezikoslovje;Strojno učenje;Slovenščina;Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS): Master's thesis/paper
Study programme: 1000471
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 95 str.
ID: 19933610