An analysis of the utility of negative examples to improve the syntactic ability of neural language models