Deep Learning Based Grammar Checker for Kannada

Abstract

Language is the most basic and traditionally natural means of communication in the present day to day conduct. And grammar places a vital role in the success of a language. As Humans have been trained throughout out life with a abundance data that is accumulated, refined over course of time with rules and compression of relevance to enable us to understand and converse between one another. But to incorporate such under-standing to a machine , to be able to evaluate and differentiate contextual information into proper grammatical form hence to validate that the information is in the right form is also equally important in the present day as it well as it a complex chore . The paper addresses this problem and proposes the development of such grammar checking tool for the Dravidian language Kannada. One of the first consideration is that the complexity of the language poses a challenge and opting to use a rule based approach is a easier solution and allows to identify flagged errors efficiently. It requires a linguistic expert to draw out hundreds of sequential rules that is complex to maintain. Here, a model is proposed that uses a deep learning approach to train a LSTM (Long Short Term Memory) neural model trained over a large data set to achieve the required classification, using a context retaining representation of the data achieved through Word2Vec along TensorFlow and Keras packages. The proposed model is capable of efficiently performing Grammatical error detection (GED)

  Index Terms: Natural language processing, LSTM, GED, Word2Vec, Deep learning, Neural Network, Word Embedding