BiLSTM with CRF Part-of-Speech Tagging for Khasi language

Abstract

Part-of-speech (POS) tagging is one of the tasks involved in sequence labeling which recognizes different word classes or tags in a sentence. In this paper, a deep learning approach is used, and to perform a comparative performance analysis, each model is first executed in its setting. These settings include Long-Short-Term Memory (LSTM), its variant, Bidirectional Long-Short-Term Memory (BiLSTM), and a Conditional Random Field (CRF) network. Additionally, these models are executed by combining LSTM and CRF to create an LSTM-CRF network and BiLSTM and CRF (BiLSTM-CRF) network. And although some word classes in the Khasi language are context-sensitive, the adopted models, supported by a pretrained word embedding model, can contextually categorize words into their appropriate categories. Word embedding is also used to aid in the part-of-speech (PoS) tagging task, where words are encoded with real-valued vectors, with similar vector representations of words having similar meanings; this minimizes the likelihood that the words in a sentence will be incorrectly tagged. In addition, the capability of the BiLSTM layer to effectively exploit both past and future input features, in conjunction with the CRF layer’s ability to apply sentencelevel tags, contributes to the model’s achievement with an encouraging degree of accuracy.