A Bilingual Study for Personalized Sentiment Model PERSEUS

This paper investigates the significance of analyzing language preferences in personalized sentiment analysis. Motivated by the considerable amount of text generated by multilingual speakers on social platforms, we focus on constructing a single model that is able to analyze sentiments in a multilingual environment. In particular, Twitter texts are used in this research where the choice of language can be switched at a message-, sentence-, word- or topic-level. To represent and analyze the text, we extract concepts and main topics from the text and apply a recurrent neural network with attention mechanism in order to learn the relation between the lexical choices and the opinions of each sentiment holder. The personalized sentiment model PERSEUS is applied as the central structure of this research. Moreover, a language index is added to each concept to enable multilingual analysis, which provides a solution for analyzing code-switching in the text as well. In this work, English and German are chosen for a pilot study, and an artificial corpus is created to evaluate the situation with multilingual speakers.