Bayesovsko duboko učenje za klasifikaciju teksta

master's thesis

• Title (english): Bayesian Deep Learning for Text Classification
• Author: Antonio Šajatović
• Mentor: Jan Šnajder (mentor)
• Committee member: Jan Šnajder (predsjednik povjerenstva)
• Committee member: Bojana Dalbelo-Bašić (član povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2019-07-11, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Modeli dubokog učenja postižu vrhunske performanse u klasifikaciji teksta, jednom od zadataka u obradi prirodnog jezika. Međutim, kako ovi modeli koriste točkovne procjene svojih parametara, oni nisu u stanju modelirati prediktivnu nesigurnost. Na duboko učenje je primijenjena Bayesova teorija vjerojatnosti, što je rezultiralo područjem koje se zove Bayesovsko duboko učenje, omogućujući dubokim modelima da modeliraju nesigurnost na principijelan način. Cilj ovog rada bio je ispitati performanse... More Bayesovih dubokih modula za klasifikaciju teksta i nove primjene nesigurnosti modela. Bayes by Backprop metoda korištena je za implementaciju nekoliko modela klasifikacije teksta, koji su postigli performanse usporedive s modelima s točkovnim procjenama parametara. Modeli su evaluirani na tri različita skupa podataka za klasifikaciju na engleskom jeziku. Naposljetku, prediktivna nesigurnost koja se upotrebljava za suzdržavanje od klasifikacije uspoređena je s jednostavnim bazičnim pristupom u eksperimentima selektivne klasifikacije. Less

Abstract (english)

Deep learning models achieve state of the art performance in text classification, a task in natural language processing. However, as these models use point estimates of their parameters, they are not able to model the predictive uncertainty. Bayesian probability theory was applied to deep learning, resulting in a field called Bayesian deep learning, enabling deep learning models to incorporate uncertainty in a principled way. To goal of this thesis was to examine the performance of Bayesian... More deep models for text classification and novel applications of model uncertainty. The Bayes by Backprop method was used to implement several text classification models that were shown to achieve performance comparable with point estimate models. The models were evaluated on three different English text classification datasets. Finally, predictive uncertainty used to abstain from classification was compared with a simple baseline approach in the selective classification experiments. Less

• Keywords: duboko učenje; Bayesovsko duboko učenje; obrada prirodnog jezika; klasifikacija teksta; selektivna klasifikacija
• Keywords (english): deep learning; Bayesian deep learning; natural language processing; text classification; selective classification
• Language: english
• URN:NBN: urn:nbn:hr:168:759583
• 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: 2020-01-07 22:14:10