Using Deep Learning for Sentiment Analysis of Expressions in Croatian Language

Uobičajeni postupci analize sentimenta temelje se na rječniku apriornog sentimenta. Problem predstavlja modeliranje sentimenta višerječnih izraza poput ``poprilično dobar'' ili ``nimalo loš'', ali i većih jezičnih jedinica. Opisan je i implementiran postupak učenja reprezentacija riječi prema metodi Colloberta i dr.~(2011) te postupak analize sentimenta višerječnih izraza hrvatskoga jezika modelom zasnovanom na rekurzivnoj neuronskoj mreži prema radu Sochera i dr.~(2012). Navedeni su rezultati učenja reprezentacija riječi na temelju dva različita korpusa i rezultati evaluacije modela za analizu sentimenta nad tri različita skupa za učenje.Methods of sentiment analysis are usually based upon a sentiment-labeled lexicon. Problems arise when trying to model the sentiment of multi-word phrases such as ``pretty good'' or ``not bad'', but also of entire sentences. A word-representation training method based on Collobert et al.~(2011) is defined and implemented, as is a method for sentiment analysis of multi-word phrases written in Croatian using a model based on recursive neural networks according to Socher et al.~(2012). We state the results of word-representation training on two different corpora and the results of evaluating the sentiment analysis model on three different training sets