An efficient system for grammatical error correction on mobile devices

As per the analysis by lemongrad, among 1,121 million English speakers globally, there are 743 million nonnative and 378 million native English speakers (7:3 Ratio). With the increasing number of non-native English speakers, there has been a lot of ongoing research on automated Grammatical Error Correction (GEC). Despite the advances in this field, a GEC system for mobile devices with a low model size, quick inference time and high accuracy is a pressing need. Existing solutions for mobile devices are server-based, which poses a potential risk of data privacy to the user. To this end, we propose an on-device hybrid system which consists of various components, such as preprocessor with spell-checker, Deep Neural Network (DNN) model, N-gram language model, rule-corrector for GEC. User's input text is passed through the proposed system sequentially to get the grammatically correct, contextually enhanced and profanity checked output sentence. Our novel system is at par with other available models for mobile devices with inference time of 95ms, module size of 11.6 MB and F
_0.5
score of 52.3 on CoNLL-2014 [1] (test).