Discriminative reranking for context-sensitive spell-checker

Nowadays, a large amount of documents is generated daily. These documents may contain some spelling errors which should be detected and corrected by using a proofreading tool. Therefore, the existence of automatic writing assistance tools such as spell-checkers/correctors could help to improve their quality. Spelling errors could be categorized into five categories. One of them is real-word errors, which are misspelled words that have been wrongly converted into another word in the language. Detection of such errors requires discourse analysis rather than just checking the word in a dictionary. We propose a discourse-aware discriminative model to improve the results of context-sensitive spell-checkers by reranking their resulted n-best list. We augment the proposed reranker into two existing context-sensitive spell-checker systems; one of them is based on statistical machine translation and the other one is based on language model. We choose the keywords of the whole document as contextual features of the model and improve the results of both systems by employing the features in a log-linear reranker system. We evaluated the system on two different languages: English and Persian. The results of the experiments in English language on the Wall street journal test set show improvements of 4.5% and 5.2% in detection and correction recall, respectively, in comparison to the baseline method. The mentioned improvement on recall metric was achieved with comparable precision. We also achieve state-of-the-art performance on the Persian language.