A new method of error correction using 2nd-order Markov models of forward, backward and middle type in Japanese syllables

2nd-order Markov model in Japanese is useful to correct errors in "kanji-kana" "bunsetsu" and reducing the ambiguity of speech recognition candidates (syllable) in "bunsetsu", but not to detect and correct errors located at boundaries of "bunsetsu" in a sentence. This paper proposes a new method to correct such erroneous syllables using three types of 2nd-order Markov model, forward, backward and middle types. In an experiment, it is found that H_2f=H_2b≥H_2m, where H_2f, H_2b and H_2m denotes the entropy of forward, backward and middle type respectively. Corresponding to the location of an erroneous syllable in bunsetsu, the accumulative accuracy rate, which denotes the rate of correct syllables included in 1st candidate or up to 10th candidates is presented as follows: 1. in the case of the top location in bunsetsu (type S₁): C_2b≥C_2m≥C_2f; 2. in the case of the second location in bunsetsu (type S₂): C_2m≥C_2f≥C_2b; 3. in other cases (type S₃): C_2f≥C_2m≥C_2b where C_2f, C_2b or C_2m is accuracy rate by using forward, backward or middle Markov models respectively. From this result, it is found that the error correction capability is improved by applying backward type to S₁, middle type to S₂, and forward type to S₃.