Analiza sentimenta zvočnih posnetkov in njihovih transkriptov

Analiziranje sentimenta s pomočjo metod strojnega učenja je ena bolj raziskanih tem na področju obdelave naravnega jezika. Večina raziskav se osredotoča na analiziranje pisanega besedila kot so članki ali knjige. V primeru govorjenega besedila pa se poleg transkriptov posnetkov lahko analizira tudi sama zvočna datoteka posnetka. V diplomski nalogi smo raziskali in naučili različne modele strojnega učenja za analizo sentimenta na transkriptih posnetkov, nato pa poskusili izboljšati rezultate tekstovnih modelov z modeli, zgrajenimi na podatkih pridobljenih iz zvočnih datotek posnetkov. Za združevanje ter izboljšanje napovedi besedilnih in zvočnih modelov smo uporabili metodo zlaganja modelov. V delu smo raziskali in implementirali celoten cevovod za predprocesiranje podatkov, generiranje značilk ter učenje in testiranje besedilnih in zvočnih modelov ter meta modela z metodo zlaganja. Analyzing sentiment using machine learning methods is one of the most researched topics in the field of natural language processing. Most research focuses on analyzing written text such as articles or books. In the case of spoken text, in addition to the transcripts of the recordings, the audio file of the recording itself can also be analyzed. In this thesis, we researched and trained different machine learning models for sentiment analysis on recording transcripts, and then tried to improve the results of text-based models with models built on data obtained from audio files of recordings. We use stacking to combine and improve the predictions of text and audio models. In this work we explored and implemented a complete pipeline for data preprocessing, feature generation and learning and testing of text and audio models and a meta model using stacking.