Modification of Chu-Liu/Edmonds algorithm and MIRA learning algorithm for dependency parser on Indonesian language

Among dependency parsing algorithms available in MALTParser and MSTParser, the best accuracy for parsing Indonesian language is achieved by Chu-Liu-Edmonds algorithm. This is due to the long distance relation between head and dependent in Indonesian sentences. Most of inaccuracy parsing results is caused by the non-verb sentence root score where there are many cases in Indonesian sentence having a non-verb as the sentence root. We proposed several modifications on the Chu-Liu-Edmonds parsing algorithm and MIRA learning algorithm and conducted the experiments for Indonesian language using Universal Dependencies Corpus. The modifications are the rule on candidate tree selection, loss function, constraint inequation and root score calculation. We conducted experiments on single modification and combined one. We found that there are result differences between sentences with conjunction and without it. Using 4,477 sentences as the training data and 557 sentences as the testing data, taken from Universal Dependencies corpus, the combined technique enhance the baseline accuracy by 0.5–0.6%.