Abstract:
Subjectivity and sentiment analysis research has gained increasing attention in the recent years like many language technologies. Its aim is to investigate and to develop techniques to recognize subjectivity or sentiment in human-generated content such as text, speech or image. While subjectivity and sentiment detection tasks are necessarily related to each other, subjectivity detection is relatively understudied and needs more attention, being a challenging problem even for humans. For capturing subjectivity clues in the text, various linguistic properties are made use of and for predicting the subjectivity of an unknown piece of text, machine learning methods are applied. In this respect, the subjectivity detection problem can be reduced to a text classi cation problem. A set of texts evaluated for some prede ned clues of subjectivity, are input to a learning module, which will predict if a given unknown piece of text is subjective or objective. In this work, we study subjectivity detection in news items using machine learning methods and develop a framework that runs at the document-level. We assume that the descriptive features of expressions is a good candidate to capture the subjective tone in texts and based on this premise, propose a novel feature set for subjectivity classi cation. We implement a supervised scheme and extensively evaluate it on a dataset which we have collected and annotated. Our ndings present new directions and useful contributions to the subjectivity detection literature. We introduce the rst subjectivity detection system in Turkish language, present our new database with annotations and report high accuracy in subjectivity detection.
