High Interpretable Transfer Network for Aspect Level Sentiment Classification

Anonymous

Abstract: Aspect-level affective classification (ASC)aims to detect the affective polarity of agiven viewpoint target in a sentence. In theASC method based on neural network,most of the work uses the attentionmechanism to capture the sentiment wordscorresponding to the opinion target, andthen gather them as evidence to infer theemotion of the target. However, due to thecomplexity of annotation, the scale ofaspect level data sets is relatively small.Data scarcity leads to the attentionmechanism sometimes unable to payattention to the sentiment wordscorresponding to the target, which finallyweakens the performance of the neuralmodel. In order to solve this problem, thispaper proposes a complete HighInterpretable Transfer Network transferlearning framework (HITN), which adoptsmethods such as data enhancement,attention adjustment and transfer toeffectively improve the performance ofASC model. A large number ofexperimental results show that our methodhas always been all the previous migrationmethods in this field, even compared withsome complex models.