YouTube Movie Reviews: In, Cross, and Open-domain Sentiment Analysis in an Audiovisual Context

In this contribution we focus on the task of automatically analyzing a speaker’s sentiment in on-line videos containing movie reviews. In addition to textual information, we consider adding audio features as typically used in speech-based emotion recognition as well as video features encoding valuable valence information conveyed by the speaker. We combine this multi-modal experimental setup with a detailed analysis of different methods for linguistic sentiment analysis by gradually increasing the level of domain-independence: First, we consider in-domain analysis by examining a cross-validation setup applied on a novel database named Multi-Modal Movie Opinion (ICT-MMMO) corpus. Next, we concentrate on cross-domain analysis by using a large corpus of written movie reviews for training. Finally, we explore the application of on-line knowledge sources for inferring the speaker’s sentiment. Our experimental results indicate that training on written movie reviews is a promising alternative to exclusively using (spoken) in-domain data for building a system that analyses spoken movie review videos and that language-independent audiovisual analysis can compete with linguistic