Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

Prediction of Prosodic Phrase Boundaries Using Stochastic Context-Free Grammar

Shigeru Fujio, Yoshinori Sagisaka, Norio Higuchi

ATR Interpreting Telecommunications Research Labs., Kyoto, Japan

We propose a method of predicting prosodic phrase boundaries using a SCFG (stochastic context-free grammar) for non-bracketed input word attributes. In this paper, prosodic phrase boundaries are predicted using information generated by a SCFG trained using phrase dependency bracketings and prosody bracketings. Using the Inside-Outside algorithm for training, corpora with phrase dependency brackets are first used to train the SCFG from scratch. Next, this SCFG is re-trained using the same corpora with prosody brackets. Then, the probability of each bracketing structure is computed using the SCFG, and is used as a parameter in the prediction of the prosodic phrase boundaries. To examine the effect of these parameters, prosodic phrase boundaries were predicted. Held-out data: 92.9% of the prosodic phrase boundaries, and 83.1% of the non-prosodic phrase boundaries, were correctly predicted.

Full Paper

Bibliographic reference.  Fujio, Shigeru / Sagisaka, Yoshinori / Higuchi, Norio (1994): "Prediction of prosodic phrase boundaries using stochastic context-free grammar", In ICSLP-1994, 839-842.