Low Complexity Named-Entity Recognition for Indonesian Language using BiLSTM-CNNs

Abstract

Named-Entity Recognition (NER) is a fundamental task in Natural Language Processing (NLP) and information extraction. NER is used to extract information such as names of the people, organizations, and places. NER has been used in many fields of work, one of which is in chatbot development. NLP and machine learning approaches enable a smarter chatbot with better personal analysis to users. This research builds a NER model in Indonesian Language using Bidirectional Long Short-Term Memory (BiLSTM) and Convolutional Neural Networks (CNNs) model architecture. Unlike the former research, this model only uses word-level embedding in the CNNs layer to keep the model simple. The Named-Entities (NEs) used in this study are limited to the name of the person, organization, location, quantity, and time using the BILOU labeling format. The performance of the model built is measured using micro-averaged f1 score evaluation metric. The BiLSTM-CNNs + pretrained word2vec embedding model provides good performance compared to other models with an f1 score of 71.37%.