A method for forming a maximal tree by searching statistical data of symbols from training data. A table is formed including a string having first and second ends, and a symbol of the training data in the first end and a minimal number of preceding symbols in the training data to make a context unique for the symbol. The contexts are sorted in lexical order with priority on the first end. A device for forming a maximal tree by searching statistical data of symbols from training data. The device forms a table including a string having first and second ends, and a symbol of the training data in the first end and a minimal number of preceding symbols in the training data to make the context unique for the symbol. The device sorts the contexts in lexical order with priority on the first end.