Grammatical correctness of word predictors

Abstract

Word predictors are frequently used software helping millions of people on a daily with typing on their mobile devices. They do however have their flaws, as they are not always grammatically correct in their predictions. The aim of this research is to create a way to increase the grammatical correctness of other models and see how that affects the performance and whether they can be implemented on a mobile device. Two models have been created, an n-gram model and an RNN model. They are compared to one another and then a POS tagger has been added to both models and compared to each other and those results with the that of the initial models. A dataset of sentences extracted from resources on social media services has been used to train the models on. The base models’ performances on predictions are unfortunately quite disappointing, mainly due to a too small of a dataset. Additionally, it is found that adding the POS tagger does not improve the main performance issue. Furthermore, the created models are not optimized and thus not suited for mobile devices due to the large size and high response times. The grammatical correctness of the predictions, however, did increase with the usage of the POS taggers, albeit to a much lesser degree than was initially hoped for.