Recognition character candidates and their similarities for each character obtained by a character recognition section from an input character string are stored in a first recognition result memory, and recognition character candidates obtained by rotating the corresponding characters through 180 degrees and their similarities are stored in a second recognition result memory. Address pointers for accessing the first and second recognition result memories are stored in an address pointer memory. The first recognition result memory is accessed in accordance with the address pointers read out from the address pointer memory in an ascending order, and the second recognition result memory is accessed in accordance with the address pointers read out from the address pointer memory in a descending order. Coincidences between recognition candidates read out from the first and second recognition result memories and character strings of dictionary words read out from a dictionary memory are computed by a coincidence computing section. A recognition result of the input character string is obtained based on the coincidences.