Word Sense Disambiguation and Location Named Entity Recognition for Madurese-Indonesian Rule-based Machine Translation

This paper addresses the polysemy and location entity recognition problems on Madurese-Indonesian Rulebased Machine Translation (RBMT) by implementing Word Sense Disambiguation (WSD) and Location Named Entity Recognition (NER).For WSD, we develop a method that implements a weighted Simplified Lesk Algorithm (SLA) using the Collection Frequency Factor (CFF).We also develop a rule-based location recognition method for Location NER using a combination of NER and N-gram.We use two lexical resources, i.e., the Madurese-Indonesian database and the Madurese corpus of sense-annotated disambiguation examples.Experiment results show that combining WSD and NER can enhance the quality of Madurese-Indonesian RBMT by solving the polysemy and location entity recognition problems.Combining lexical resources is recommended since machine translation is developed primarily for human use.At the same time, implementing a weighted SLA outperforms the original by handling the common word problem that causes noise in the context-sense comparison.For the low-resource Madurese-Indonesian language pairs, RBMT is more effective than Neural Machine Translation (NMT) because RBMT's expert-defined rules yield higher translation accuracy, unlike NMT's unpredictable results.When evaluated on four test data, our proposed methods' maximum BLEU are respectively 0.8123, 0.6035, 0.7651, and 0.7375.