Automatic Extraction of Correction Patterns from Expert-Revised Corpora

In this paper, we first present the task of automatically extracting correction patterns from texts which have been manually revised by domain experts. In real industrial scenarios, raw texts obtained via surveys or web crawling often require manual intervention to flatten word capitalization, punctuation, linguistic variability and entity naming. In this context, we propose a distributional and language-independent approach that learns revision rules that also manages errors introduced by the experts themselves. We extensively evaluated our approach on more than 300,000 expert-revised sentences.