Uporaba globokih konvolucijskih nevronskih mrež na jezikovnih problemih : diplomsko delo

Pušnik, Žiga (author), Robnik Šikonja, Marko (mentor)

Cilj diplomske naloge je preizkusiti učenje jezikovnih problemov s pomočjo globokih konvolucijskih nevronskih mrež. Konvolucijske nevronske mreže so bile razvite predvsem za področje umetnega zaznavanja in delujejo na podlagi konvolucije. Naučili smo jih, da so na podlagi kratkega povzetka besedila napovedale razred, h kateremu spada. Drugi problem, ki smo ga reševali je postavljanje vejic v slovenskem jeziku. Konvolucijsko nevronsko mrežo smo sprogramirali s programskim jezikom python. Uporabili smo knjižnjico Theano. Izhajali smo iz že obstoječih raziskav. Opišemo način, kako smo obdelali podatkovne množice, da so primerne za naš model. Opravili smo več poskusov. Primerjali smo lematizacijo in krnjenje ter predstavitev besedila z vektorizacijo in predstavitev z bitnim poljem. Zadovoljive rezultate smo dobili, če smo besedilo kvantizirali, kjer smo črke vektorizirali z 1 do m kodiranjem. Naši rezultati pri postavljanju vejic so primerljivi z rezultati drugih raziskav.

• Language: Slovenian
• Keywords: strojno učenje, obdelava naravnega jezika, nevronska mreža, nevron, konvolucija, konvolucijska nevronska mreža, klasifikacija, klasifikacijski model, klasifikator, klasifikacijska točnost, jezik, besedilo, vejica, lema, krn, moment, gradientni spust, vzratno širjenje napake, jezikovni korpus, atribut
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Publisher: [Ž. Pušnik]
• Year: 2015
• Number of pages: 47 str.
• PID: 20.500.12556/RUL-72286
• COBISS.SI-ID: 1536476611
• Publication date in RUL: 10.09.2015

Licences

• License: CC BY-SA 2.5 SI, Creative Commons Attribution-ShareAlike 2.5 Slovenia
• Link: https://creativecommons.org/licenses/by-sa/2.5/si/deed.en
• Description: You are free to reproduce and redistribute the material in any medium or format. You are free to remix, transform, and build upon the material for any purpose, even commercially. You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use. If you remix, transform, or build upon the material, you must distribute your contributions under the same license as the original. You may not apply legal terms or technological measures that legally restrict others from doing anything the license permits.

Secondary language

• Language: English
• Title: Using deep convolutional neural networks on natural language problems

Abstract

The thesis examines the learning of language problems with convolutional neural networks. Convolutional neural networks were developed for machine vision. We used them to classify short abstracts and to learn a comma placement in Slovenian language. We programmed our convolutional neural network in programming language python with Theano library. Our work is based on existing research. We describe adaptation of datasets to our model. Several experiments were conducted and we compared lemmatization versus stemming and vector representation of text versus byte array representation. The best results were obtained with text quantized with 1 to m encoding. Comma placing results are comparable with other machine learning approaches.

• Keywords: machine learning, natural language processing, neural network, neuron, convolution, convolutional neural netvork, clasification, clasification model, clasificator, clasification accuracy, language, text, comma, lemma, stemm, momentum, gradient descent, backpropagation, text corpus, attribute