Avtomatska gradnja korpusa in ekstrakcija relacij v slovenščini : magistrsko delo

Štravs, Miha (author), Žitnik, Slavko (mentor)

Iskanje relacij med entitetami v besedilu je področje obdelave naravnega jezika. Pri iskanju relacij želimo v stavku: "Ljubljana je glavno mesto Slovenije" odkriti, da med entitetama Ljubljana in Slovenija nastopa relacija glavno mesto. V zaključnem delu smo najprej naredili pregled metod za učenje modelov za napovedovanje relacij. Nato smo si izbrali tri metode z različnimi pristopi za učenje modelov, ki napovedujejo relacije. Metodo s povratno nevronsko mrežo z dolgim kratkoročnim spominom, metodo z vložitvami BERT in metodo RECON, ki uporabi grafovsko nevronsko mrežo s pozornostjo. Za učenje modelov smo uporabili slovenski korpus, ki smo ga polavtomatsko generirali iz besedil slovenske Wikipedije. Naučene modele smo nato testirali na testnem korpusu besedil slovenske Wikipedije in testnem korpusu člankov strani 24ur.com. Na testnem korpusu slovenske Wikipedije so vse tri metode dosegle visoke priklice in točnosti, najbolje se je odrezala metoda RECON. Veliko slabše rezultate so dosegle na testni množici člankov 24ur.com, kjer se je še najbolje izkazala metoda z vložitvami BERT, ko je uporabila vložitve CroSloEngual.

• Language: Slovenian
• Keywords: ekstrakcija relacij, ekstrakcija informacij, globoko učenje, grafovske mreže pozornosti, BERT, LSTM
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FMF - Faculty of Mathematics and Physics; FRI - Faculty of Computer and Information Science
• Year: 2022
• PID: 20.500.12556/RUL-138295
• UDC: 004
• COBISS.SI-ID: 115087619
• Publication date in RUL: 14.07.2022

Secondary language

• Language: English
• Title: Automatic corpus construction and relation extraction for Slovene

Abstract

Finding relations between entities in a text is an area of natural language processing. In the sentence: "Ljubljana is the capital of Slovenia" we want to find the relation capital between entities Ljubljana and Slovenia. We first start with a review of the methods used for training models to predict relations. We then chose three methods with different approaches. The method with long short-term memory neural network, method which uses BERT encoder representations and method RECON which uses graph attention networks. To train the models, we used the Slovenian corpus which was generated semi-automatically from the text of the Slovenian Wikipedia. We test the models on a test corpus of Slovenian Wikipedia and the test corpus of articles on 24ur.com. All three methods achieved high recall and precision for the test corpus of the Slovenian Wikipedia and the RECON method performed best. Results were worse on the test set of 24ur.com articles, where the method which used BERT encoder representations CroSloEngual achieved the best results.

• Keywords: relation extraction, information extraction, deep learning, graph attention networks, BERT, LSTM