Results on the RULEC-GEC test sample. The best values for the precision, recall, and F0.5 are in bold and the second-best values are in italics.

Contexts in source publication

Context 1

... experimental results are presented in Table 2 On its own, our approach performs worse than Yandex.Speller, especially in terms of recall. However, when used as a postprocessing step, it increases the recall, albeit at the cost of some decrease in precision, showing a higher value of the F 0.5 measure. ...

Context 2

... these rules admit exceptions, their application results in improved metrics on the learner corpus, see Table 2, lines 2-3, 9-10, and 16-17. This suggests that contexts presenting counterexamples to these rules rarely occur in L2 writing. ...

Context 3

... different prepositions can fit in place of the mask, it is important to set a high threshold, so as to avoid unnecessary changes in the sentence. The results obtained after this operation are presented in Table 2, lines 4, 11, and 18. ...

Context 4

... then choose the that maximizes the increase in probability. The error correction results for each of the agreement and control error types discussed above are presented in Table 2, lines 5, 12, and 19. Here again, we observe an increase in F 0.5 , as well in both precision and recall. ...

Context 5

... the results demonstrated by the other approaches as shown in lines 22-27 of Table 2, we see that, in terms of F 0.5 , our results are second to the machine-translation model from [5] with fine-tuning, which, unlike our model, needs a large amount of labeled data for training. Our model demonstrates a significantly lower recall, but its precision is slightly better than that of a finetuned model and much better than that of the model without fine-tuning. ...