COMPARISON OF PERFORMANCE DIRECT AND DATA DRIVEN MACHINE TRANSLATION FOR LAMPUNGINDONESIAN LANGUAGE

• Main Author: ABIDIN (NIM: 23515015), ZAENAL
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/31866
• Institution: Institut Teknologi Bandung

In this research, automatically Lampung language translation into Indonesian <br /> <br /> language was using three approach methods, namely DMT, direct machine <br /> <br /> translation approach, SMT, statistical machine translation approach, and NMT <br /> <br /> attention, neural machine translation attention approach. DMT approach method <br /> <br /> has been applied by breaking the Lampung language sentence into several parts <br /> <br /> according to its space. The result is several words that will be matched into the <br /> <br /> dictionary database key. The result of that matching process is a value, a word in <br /> <br /> Indonesian language. From that result, the dictionary database key will be <br /> <br /> rearranged into a sentence in Indonesian language. SMT approach method will do <br /> <br /> in several phases. Starting from pre-proceed phases that is the beginning phase <br /> <br /> preparing the parallel corpus. Then the phase continuing into training phase, a <br /> <br /> phase that managing the parallel corpus to obtain the language model and the <br /> <br /> translation model. After that it must be run a testing phase and ended with an <br /> <br /> evaluation phase. NMT is such a new approach method in machine translation <br /> <br /> technology, that has worked by combining the encoder. Encoder is a recurrent <br /> <br /> neural network component that encrypt the source language to several lengthstable <br /> <br /> vectors. And decoder is a recurrent neural network component that generate <br /> <br /> translation result integratedly. NMT research has begun with creating a pair of <br /> <br /> 3000 parallel sentences of lampung language (api dialect) and Indonesian <br /> <br /> language. Then it is continue to decide the NMT parameter model for the data <br /> <br /> training process. The next step is building NMT model and evaluate it. <br /> <br /> The testing of those three methods has used 25 single sentences without out-ofvocabulary <br /> <br /> (OOV), 25 single sentences with OOV, 25 plural sentences without <br /> <br /> OOV, and 25 plural sentences with OOV. The testing translation result using DMT <br /> <br /> method shows the bilingual evalution understudy (BLEU) average value is 13,72 <br /> <br /> %, the BLEU average value using SMT method is 77,51 % and the last, the BLEU <br /> <br /> average value using NMT method is 51, 96 %.