誤り理由を考慮したニューラル文法誤り訂正

抄録

Neural machine translation (NMT) methods with an attention mechanism are promising for automated grammatical error correction compared to other statistical machine translation methods. However, current NMT-based grammatical error correction models have at least two issues: (i) it is difficult to identify why error corrections are made, i.e., correction models are black boxes and (ii) the attention of each correction does not depend on error types. To resolve these difficulties, we propose a multi-attention based neural grammatical error correction model, which utilizes an appropriate attention for error correction. We evaluated our proposed model and the baseline single-attention model with the CoNLL-2014 shared task dataset, and found that F0.5 scores are comparable.