Part-of-Speech Tagging of Odia Language Using Statistical and Deep Learning Based Approaches

Automatic part-of-speech (POS) tagging is a preprocessing step of many natural language processing tasks, such as named entity recognition, speech processing, information extraction, word sense disambiguation, and machine translation. It has already gained promising results in English and European languages. However, in Indian languages, particularly in the Odia language, it is not yet well explored because of the lack of supporting tools, resources, and morphological richness of the language. Unfortunately, we were unable to locate an open source POS tagger for the Odia language, and only a handful of attempts have been made to develop POS taggers for the Odia language. The main contribution of this research work is to present statistical approaches such as the maximum entropy Markov model and conditional random field (CRF), as well as deep learning based approaches, including the convolutional neural network (CNN) and bidirectional long short-term memory (Bi-LSTM) to develop the Odia POS tagger. A publicly accessible corpus annotated with the Bureau of Indian Standards (BIS) tagset is used in our work. However, most of the languages around the globe have used the dataset annotated with the Universal Dependencies (UD) tagset. Hence, to maintain uniformity, the Odia dataset should use the same tagset. Thus, following the BIS and UD guidelines, we constructed a mapping from the BIS tagset to the UD tagset. The maximum entropy Markov model, CRF, Bi-LSTM, and CNN models are trained using the Indian Languages Corpora Initiative corpus with the BIS and UD tagsets. We have experimented with various feature sets as input to the statistical models to prepare a baseline system and observed the impact of constructed feature sets. The deep learning based model includes the Bi-LSTM network, the CNN network, the CRF layer, character sequence information, and a pre-trained word vector. Seven different combinations of neural sequence labeling models are implemented, and their performance measures are investigated. It has been observed that the Bi-LSTM model with the character sequence feature and pre-trained word vector achieved a result with 94.58% accuracy.