On-Device Major Indian Language Identification Classifiers to Run on Low Resource Devices

Abstract

Language Identification acts a first and necessary step in building intelligent Natural Language Processing (NLP) systems that handle code mixed data. There is a lot of work around this problem, but there is still scope for improvement, especially for local Indian languages. Also, earlier works mostly concentrates on just accuracy of the model and neglects the information like, whether they can be used on low resource devices like mobiles and wearable devices like smart watches with considerable latency. Here, this paper discusses about both binary classification and multiclass classification using character grams as the features. Considering total nine languages in this classification which includes, eight code mixed Indian languages with English (Hindi, Bengali, Kannada, Tamil, Telugu, Gujarati, Marathi, Malayalam) and standard English. Binary classifier discussed in this paper will classify Hinglish (Hindi when written using English script is commonly known as Hinglish) from seven other code-mixed Indian Languages with English and standard English. Multiclass classifier will classify the previously mentioned languages. Binary classifier gave an accuracy of 96% on the test data and the size of the model was 1.4 MB and achieved an accuracy of 87% with multiclass classifier on same test set with model size of 3.6 MB.