ishine/crf_prosody_prediction

demo code of prosodic events prediction applying CRFs

crf_prosody_prediction

demo code of prosodic events prediction applying CRFs

./scripts

Tools here can help with pre-processing the original data in diretory "./data/txt_with_stress". One example of the original data looks like:

https://github.com/kokoinyannn/crf_prosody_prediction/blob/master/data/txt_with_stress/S_PH_B_1_006.txt

in which the numbers stand for the stress level of the corresponding syllable, and the "|" (vertical line) stands for phrase boundary.

By applying label_gen and sent_gen, the expected output looks like

[['Bright', 'NNP', 'N', 'NPS', 'NSS'], ['sunshine', 'NN', 'S', 'PS', 'NSS'], ['shimmers', 'NNS', 'N', 'NPS', 'NSS'], ['on', 'IN', 'N', 'NPS', 'NSS'], ['the', 'DT', 'N', 'NPS', 'NSS'], ['ocean', 'NN', 'N', 'PS', 'SS'], ['.', '.', '', 'NPS', 'NSS']]

in which the first label in each element is token itself, following POS tagging, slash information (S for having slash and N for no slash), phrase stress information and sentence stress information.

Data like this are ready for the crf training.

./crf

The core programs of predicting prosodic events using crf and detailed description can be found here.

publication

This research was poster presented at 2018 ASJ autumn meeting. The paper can be found here: http://gavo.t.u-tokyo.ac.jp/~xinyi/ASJ_paper.pdf