Domain Specific Learning for Sentiment Classification and Activity Recognition

Abstract

A deep neural network, while avoiding its complex process of feature selection, requires sufficient training samples to learn those connection weights of adjacent layers. However, in many real applications, not enough training samples are available in all cases. This paper suggests a universal-todomain-specific learning method based on recurrent neural network for cross-domain sentiment classification and activity recognition. In the situation of having only a small amount of training samples that is in available, the structure of its network model can be adjusted flexibly according to the needs of a target domain classification or recognition. Where there are two points of our concern as follows: 1) the finetune and regular constraints can increase its training efficiency by updating in a small local area, namely, sharing these parameters between input and hidden layers with a target domain and 2) then, a linear output network moves on its implementing amelioration from subtlety as an exploration or exploitation in order to mitigate the phenomenon of over-fitting. Aiming at an actual situation, this domain-specific learning model with a slide window of instances and features is designed and implemented for a good long-short term memory. Finally, the two strategies are applied into IMDB reviews, Amazon product reviews, and human activities recognition collected by the built-in gyroscope sensors data, and the experimental results verify their validity.

Keywords

• Personalized activity recognition
• recurrent neural networks
• sentiment classification