A Comparison of Distributed, PAM, and Trie Data Structure Dictionaries in Automatic Spelling Correction for Indonesian Formal Text

Spelling errors can be divided into two groups, non-word errors and word errors. A non-word errors produce words that do not exist in dictionary, while word errors is a real word but not the right word. In this work, we address the non-word errors spelling correction for Indonesian formal text. The objective of our work is to compare the effectiveness of three kinds of dictionary structure for spelling correction, distributed dictionary, PAM (partition around medoids) dictionary, and dictionary using trie data structure, with the baseline of a simple flat dictionary. We conduct experiments with two variations of edit distances, i.e. Levenshtein and Damerau-Levenshtein, and utilized n-grams for ranking suggestion. We also build a gold standard of 200 sentences that consists of 4,323 tokens with 288 of them are non-word errors. Among the various combinations of dictionary type and edit distance, the trie data structure with Damerau-Levenshtein distance gets the best accuracy to produce candidate correction, i.e. 95.89% in 45.31 seconds. Furthermore, the combination of trie data structure with Damerau-Levenshtein distance also gets the best accuracy in choosing the best candidate, i.e. 73.15%.