Emotions in film reviews
Lars Buitinck (UvA Informatics Institute and Netherlands eScience Centre) developed a method for automatically identifying emotions in amateur film reviews, together with communication scientists at the University of Amsterdam. Buitinck presented his method during the 37th edition of the international European Conference on Information Retrieval in Vienna.
Social scientists research the emotions which films arouse in viewers in order to better understand the role of media in society and culture. Mapping the emotions which are invoked by a film could also prove useful for recommending films which a user might find interesting (like Netflix does, for example). One challenge when mapping the emotions of viewers is the fact that it is difficult and expensive to measure these emotions. The method developed by Buitinck and colleagues measures the emotions which are evoked by films by analysing amateur film reviews.
Multiple emotions within a single sentence
Previous research into emotion recognition mainly focused on recognising emotions at sentence level and was limited to a binary distinction: a sentence will either express a positive or a negative emotion. Buitinck's new method can identify multiple emotions within a single sentence. The method uses statistical patterns in word usage in order to identify expressed emotions and also takes correlations between human emotions into account (which emotions occur together relatively frequently?).
Text analysis of film reviews on the internet
In order to develop this method, the researchers picked film reviews from the internet in which they manually identified expressed emotions (which words express an emotion?) and then classified them (which emotion is being expressed?). Using this collection of reviews and emotions, the next task was to develop an algorithm which was capable of detecting the emotions in reviews as effectively as the communication scientists themselves. Buitinck's algorithm was able to detect the same emotions as the scientists in 85% of cases.
As well as presenting two new algorithms for detecting multiple expressed emotions within a sentence, Buitinck has also published the data set of reviews and identified emotions so that other scientists can use this information. Access the data set here