Automatic Sentence Grammaticality Judgment

Nowadays, many natural language processing tasks are well studied, such as text-classification tasks, summarization tasks, machine translation tasks, etc. There are also a lot of fine question-answering models. However, new tasks appear all the time. The task I faced with is automatic grammaticality judgement of sentences, written in English. For lots of machine learning tasks, there is a lot of data available in the public domain that can be used to conduct the experiments. But the task of automatic grammaticality judgement is much more complicated. It is difficult to find sentences, which are supposed to be grammatically incorrect. In my work, I consider ways of generating such sentences, and based on them I prepare a dataset. These data are used to experiment with different classifier models, such as recurrent neural networks, convolutional neural networks and fully-connected classifiers on top of the BERT and RoBERTa models.