Pretvaranje opisa zadanog prirodnim jezikom u naredbe jezika SQL

master's thesis

• Title (english): Translation Of Natural Language Description into SQL instructions
• Author: Sara Borzić
• Mentor: Goran Delač (mentor)
• Committee member: Goran Delač (predsjednik povjerenstva)
• Committee member: Klemo Vladimir (član povjerenstva)
• Committee member: Marin Šilić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-04, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Prevođenje pitanja prirodnog jezika u SQL upit je zadatak semantičkog parsiranja koji privlači mnogo pozornosti. Od pojave SPIDER skupa podataka se zadatak još više popularizirao te unaprijedio zbog zahtjevnosti SPIDER skupa podataka. Dobre performanse na mjeri točnog podudaranja u SPIDER skupu podataka pokazuju vještinu dobre generalizacije modela. RAT-SQL+GAP model je baziran na arhitekturi Transformera s dodanim okvirom za generaciju dodatnih podataka za unaprijedno treniranje. Model predstavljen u radu je model koji prevodi pitanje hrvatskog jezika u SQL upit. Testiran je s različitim modelima za unaprijedno treniranje: BART, mBART50MtoM i BERTić na prilagođenom Spider skupu podataka. Rezultati su pokazali kako se najbolji rezultati mogu dobiti s dotreniravanjem BART modela za hrvatski jezik. Sljedeći najbolji rezultati testiranja na hrvatskim podatcima su prisutni za korištenje MBART50MtoM modela, treniranog na spojenim skupovima podataka hrvatskih i engleskih pitanja. Rezultat ovog rada je prvi model koji koristi pitanja prirodnog jezika na hrvatskom jeziku za zadatak pretvorbe pitanja prirodnog jezika u SQL upite.

Abstract (english)

Translating natural language question in SQL, task known as text-to-SQL, is semantic parsing task which is attracting a lot of attention lately. Since the appearance of the SPIDER data set, the task has become even more popular and improved due to the complexity of the SPIDER data set. Good performance on the exact set match without values in the SPIDER dataset, demonstrates the skill of good model generalization. The RAT-SQL + GAP model is based on the Transformer architecture with an added framework for generating additional data for advanced training. The model presented in this thesis is a model that translates the Croatian language question into an SQL query. It has been tested with different advanced training models: BART, mBART50MtoM and BERTić on a custom Spider dataset. The experiments showed that the best results on Croatian dataset can be obtained by training the BART model for the Croatian language. The following best test results on Croatian data are present for the use of the MBART50MtoM model, trained on merged data sets of Croatian and English questions. The result of this paper is the first model that uses natural language questions in Croatian for the task of converting natural language questions into SQL queries.

• Keywords: SQL; obrada prirodnog jezika; hrvatski jezik; transformeri; semantičko parsiranje; samopažnja; BART; mBART50MtoM; BERTić; RAT-SQL; Spider
• Keywords (english): SQL; natural language processing; Croatian language; English language; transformers; semantic parsing; self-attention; BART; mBART50MtoM; BERTić; RAT-SQL; Spider
• Language: croatian
• URN:NBN: urn:nbn:hr:168:837199
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-14 08:33:09