Influence Analysis of Smoothing Algorithms in Language Modelling for Indonesian Statistical Machine Translation

Abstract

The diversity of languages makes the need for translation so that communication between individuals of different languages can be appropriately established. The statistical translator engine (SMT) was a translator engine based on a statistical approach to parallel corpus analysis. One crucial part of SMT was language modeling (LM). LM was the calculation of word probability from a corpus-based on n-grams. There was a smoothing algorithm in LM where this algorithm will bring up the probability of a word whose value was zero. This study compares the use of the best smoothing algorithm from each of the three LM according to the standard Moses, namely KenLM, SRILM, and IRSTLM. For SRILM using smoothing algorithm interpolation with Witten-bell and interpolation with Ristads natural discounting, for KenLM using interpolation with modified Kneser-ney smoothing algorithm, and for IRSTLM using modified Kneser-ney and Witten-bell algorithm which was referenced based on previously researched. This study uses a corpus of 10,000 sentences. Tests carried out by BLEU and testing by Melayu Sambas linguists. Based on the results of BLEU testing and linguist testing, the best smoothing algorithm was chosen, namely modified Kneser-ney in KenLM LM, where the average results of automated testing, for Indonesian-Melayu Sambas and vice versa were 41. 6925% and 46. 66%. Moreover, for testing linguists, the accuracy of the Indonesian-Melayu Sambas language and vice versa was 77. 3165% and 77. 9095%