Detekcija antonimov z vektorskimi vložitvami besed

Pegan, Jasmina (author), Robnik Šikonja, Marko (mentor), Gantar, Apolonija (co-mentor)

Cilj diplomske naloge je razvoj klasifikatorja za prepoznavo protipomenk. Za izdelavo rešitve je bila uporabljena baza vnaprej pripravljenih vektorskih vložitev besed za slovenščino. Najprej smo sestavili učno množico protipomenk in sopomenk. Sledilo je iskanje čimbolj ustreznega klasifikacijskega modela. Ogledali smo si nekaj modelov metode podpornih vektorjev in nekaj globokih nevronskih mrež. Izbranim besedam smo poiskali pomensko sorodne besede in na njih uporabili naučeni model. Tako smo pridobili kandidate za pare protipomenk in sopomenk. Točnost rezultatov smo ocenili na testni množici. Najbolje ocenjeni model dosega klasifikacijsko točnost 0.70.

• Language: Slovenian
• Keywords: protipomenke, sopomenke, vektorske vložitve besed, strojno učenje, klasifikacija
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2019
• PID: 20.500.12556/RUL-110533
• COBISS.SI-ID: 1538361795
• Publication date in RUL: 16.09.2019

Secondary language

• Language: English
• Title: Antonym detection with word embeddings

Abstract

This thesis aims to develop a classifier for antonym detection. A database of pre-made word embeddings for Slovene was used to create the solution. First we collected a learning set consisting of synonyms and antonyms. Then we searched for the most appropriate classification model. We observed some support vector machine models and some deep neural networks. We applied the learned model to groups of words closest to the selected words. Thus, we obtained candidates for pairs of synonyms and antonyms. The accuracy of the results set was evaluated on the test set. The top rated model reaches classification accuracy of 0.70.

• Keywords: antonyms, synonyms, word embeddings, machine learning, classification