Public sentiment reflects the people’s attitude to society and politics. Through a large amount of observation, we found that the trends of the development of many typical web public sentiments can be divided into three phases: relatively stable phase, rapidly increased phase and rapidly declined phase. The correct phase detection and prediction win precious time for relevant department to formulate corresponding policies to deal with the situation, and for mainstream media to steer the public sentiment. In this paper we present a complete approach for automated phase detection and prediction of web public sentiment. The main idea of this paper is first to apply a dynamic social network to model Web public sentiment, and then use some combined approach to predict several important parameters of the built social network, and finally detect the phase of web public sentiment through smooth the predicted curve and the voting method. The experimental result shows that the approach is qualitatively quite useful when used to analyze, monitor and even steer the information on the internet.