Abstract
Article and determiner errors are common in the writing of English language learners,\r but automated systems of detecting and correcting them can be challenging to build as the\r context in which an article is found can be ambiguous. Here, I investigate the use of recurrent\r neural networks to detect and correct such errors separately, treating it as a sequence labeling\r task. The NUS Corpus of Learner English is used for training and evaluation. A maximum\r precision of 76%, recall of 41%, and F0.5 score of 46.5% were achieved in the error detection\r task using a bidirectional LSTM. A maximum precision of 51%, recall of 34%, and F0.5\r score of 47% were achieved in the error correction task using a bidirectional LSTM, which is\r competitive with previous results on this corpus and recent results using other neural network\r models. Furthermore, the technique used relies only on lexical items, with no additional\r features necessary. These results, in conjunction with clear venues for improvement, show\r that the method is a promising one for the task.