On-line postprocessing differs from off-line postprocessing in that only one character is processed at a time, and that the available contextinformation is restricted to the result that is already confirmed. In this thesis, efficient postprocessing methods are proposed for on-line Hangul recognition using domain specific knowledge useful for the postprocessing in data entry applications such as form fillings. Here we define two types of domain specific knowledge : structural knowledge and statistical knowledge.
Structural knowledge is defined as structured strings of limited character set, size or pattern of occurrences, as in address. This type of knowledge is represented as a hierarchical tree and used in limiting perplexity by determining a set of possible characters to appear in the next position. Statistical knowledge is information of relative frequency of appearance of a character in a position. The knowledge is utilized to determine likelihood of a character class from many candidates. Structural knowledge is eligible when contextual information is tightly constrained, while statistical knowledge is when loosely constrained.
To evaluate the proposed methods, two typical examples of knowledge, address and name, were chose. The recognition experiment were performed with and without postprocessing. The results show that the proposed method improves substantially over those without postprocessing.
