Expand CRF to Model Long Distance Dependencies in Prosodic Break Prediction

Intonation phrase length distribution is important information for prosodic break prediction. However, existing CRF frameworks cannot make full use of it. An expanded CRF is proposed in this paper to tackle this problem. Its lattice carries the location of previous intonation phrase (L3) break, and consequently makes it possible to support various dynamic features, such as the number of syllables from the previous L3 break and the POS of word after the previous L3 break. Remarkable improvements are obtained with the expanded CRF for L3 break prediction task. It is also promising to benefit other tasks containing long distance dependencies.