The contribution of mutual information in the intonational phrase prediction in Chinese text

The contribution of mutual information (MI) in the intonational phrase (IP) prediction is analyzed and verified. The basic idea of employing MI in IP prediction is that people are likely to pause between the less correlated words, where the MI value is low. We present a decision tree based predictor which adopts POS as the main feature firstly as the baseline, and then we analyze the correlation between MI and IP. The approach which only bases on the MI to predict the IP boundary is demonstrated, and three methods combining the MI and POS in the predictor is presented too. In the MI based approach, a considerable performance (F-Score: 64.2%) is achieved, and 3.4% promotion from the baseline is achieved after combining MI and POS in our experiment. All our work indicates that MI is an effective feature in the prosodic phrase boundaries prediction in Chinese text, and combining the MI and POS in the predictor is valuable.