Main Article Content
Abstract
The classification of emotions is a hot topic of research in the industrial and academic field. The
main methods of classifying emotions are based on machine learning and treat the classification of
emotions as a text classification problem. However, the classification of emotions is widely
recognized as a highly domain-dependent task. The mood classifier trained in one domain may not
work properly in another domain. A simple solution to this problem is to train a domain-specific
emotion classifier for each domain. However, it is difficult to identify enough data for each domain
because they are in a large amount. In addition, this method omits mood information in other
domains. In this document, we propose to collaboratively train multi-domain sentiment classifiers
based on learning multiple tasks. In particular, we divide the mood classifier into two domains in
each domain, one generic and one domain specific. The general opinion classifier can capture the
global opinion information and is trained in various areas to obtain a better generalizability. The
domain-specific opinion classifier is trained using data labeled in a domain to capture the domainspecific
opinion information. In addition, we examine two types of relationships between domains,
one based on the text content and the other based on the distribution of mood words. We create a
domain similarity diagram using domain relationships and encode it in our approach as
regularization for domain- specific opinion classifiers.