Tendulkar’s Cat and Schrodinger’s Bat—Knowledge-Enhanced Real-Word Error Correction

Real-word errors are those errors in which a word is lexically correct or syntactically correct in the sentence it belongs to, but is contextually incorrect. As a result, they are difficult to identify and arguably even more difficult to correct. In this work, we attempt to create a model based on total word similarity which is able to correct real-word errors in sentences based on the simple assumption that a real-word error will not be semantically cohesive with respect to its context, and that the actual word which was intended, will be. The main aspects of our approach are the usage of semantic context, knowledge context, and syntactic context, all together at once. To this approach, we also retain the concept of edit penalty, which is an integral component of most spelling checkers. We modified the Microsoft Research Sentence Completion task of 2011 for measuring the performance of real-word error correction models and achieve 80% accuracy.