In a Thai text, a delimiter for indicating the word boundary is not explicitly used. Many tasks of Natural Language Processing (NLP) such as Thai-English machine translation, Thai speech synthesis and spelling correction require boundaries of words. Therefore, word segmentation is one of the main problems in NLP. There are two main problems in word segmentation. The first is the ambiguity problem and the second is the unknown word boundary problem. Many approaches such as longest matching, maximal matching and trigram model have been proposed. However, these approaches can not give high accuracy because longest matching and maximal matching use only heuristics and trigram model consider only two previous context words for solving the problems. The accuracy in solving ambiguity problem is about 53% and 73% for maximal matching and trigram model respectively. This thesis proposes to use a feature-based approach with two learning algorithms namely RIPPER and Winnow in solving the problems in Thai word segmentation. A feature can be anything that tests for specific information in the context around the word in question, such as context words and collocations. In the experiment we train the system by using RIPER and Winnow algorithm separately, on an 80% of part-of-speech tagged corpus and the rest is used for testing. We divided the evaluation into two parts. One is the accuracy in solving the ambiguity problem and the other is the accuracy in solving the unknown word boundary problem. The accuracy using RIPPER and Winnow in solving the ambiguity problem is more than 85% and 90% respectively. On the other hand, the accuracy in solving the unknown word boundary problem is more than 70% and 80% for RIPPER and Winnow respectively. The experiment results show the feature-based approach outperform trigram model and maximal matching, and Winnow is superior to RIPPER for extracting the features from the corpus.
