Abstract

Expressing the polarity of sentiment as 'positive' and 'negative' usuallyhave limited scope compared with the intensity/degree of polarity. These twotasks (i.e. sentiment classification and sentiment intensity prediction) areclosely related and may offer assistance to each other during the learningprocess. In this paper, we propose to leverage the relatedness of multipletasks in a multi-task learning framework. Our multi-task model is based onconvolutional-Gated Recurrent Unit (GRU) framework, which is further assistedby a diverse hand-crafted feature set. Evaluation and analysis suggest thatjoint-learning of the related tasks in a multi-task framework can outperformeach of the individual tasks in the single-task frameworks.

 

Quick Read (beta)