| 摘要 |
PROBLEM TO BE SOLVED: To estimate the structure of a finite state transducer (FST) by machine learning using given input series and output series.SOLUTION: Prior probability (foresight knowledge) based on a Dirichlet process is introduced, and the structure of an FST is estimated as sampling from a probability process. With an arc of the FST assumed to be (p,i,o,q) (p: transition source state, i: input symbol, o: output symbol, q: transition destination state, (s, k): arc variable, s: transition source state identifier, k: serial number of transition source state arc of identifier s), the prior probability of the arc is expressed as P(p,i,o,q|α,G)=1(p=s)×F(i,o,q;α,G) by using a base measurement G, a discrete probability distribution F(i,o,q) sampled from Dirichlet process DP (α,G) determined by a concentration degree parameter α, and an indicator function 1(C) of a condition C. |
