Analýza textu pomocí neuronových sítí

Abstract

This diploma thesis is concerning with text classification and text topic analysis. The aim was to first write and experimentally compare classic and neural models for mentioned disciplines. Total of six datasets were acquired for the purpose of experimentation. Five classical models and seven neural network architectures were used for the experiments. Model LDA and experimental autoencoder, created as part of this thesis which was using weights of learned autoencoder as topic model, were compared in case of topic analysis. The results show that in case of classification classical methods can achieve very good accuracy that can be compared to neural networks and with much less required time however neural networks are more accurate than classical models. In case of topic analysis, the classical model LDA was creating better topics however neural model was in some cases capable of better clustering of analyzed documents and its topic quality was not very distant from LDA method. This thesis introduces an interesting alternative to classical model LDA.