An Optimal Prediction Method Using Hierarchical N-gram Based on Bayesian Decision Theory | Toshiyasu Matsushima Laboratory

Title	An Optimal Prediction Method Using Hierarchical N-gram Based on Bayesian Decision Theory (in Japanese)
Authors	末永高志、松嶋敏泰
Released Year	2012
Format	Journal
Category	Knowledge information processing
Jounal Name
Jounal Page	vol.6, no.1, pp.102-110
Published Year	2013
Published Month	3
Abstract （English）	Predictive word is an input technology showing candidate words which a system predict by user partial input. We treat predictive methods using an N-gram model. The model is generally produced by analyzing train data. The data is more sparse in proportion to an N-gram order, because of enormous combinations of words in the sequences. An issue of producing the model is how to combine a lower order model into a higher order one. Many researchers proposed models composed of weighed each-order one, such as a mixture distribution or an interpolation created by discount parameters considering about extremely lower frequent sequence. But these methods have no theoretical guarantee about prediction errors. In this paper, we treat the issue as a statistical problem that the model order is unknown, and discuss prediction errors from a point of view about Bayesian decision theory. We present that an optimal prediction method with reference to the Bayes criterion for minimizing the errors. Experimental results using Japanese documents show that our method performs good predictive words.
Note （English）	3
Manuscript
Presentation