... keystrokethe correct candidate words from appearing in thelist of candidates in their native scripts, thereby pre-venting them from being entered altogether. Finally, we can collect a large ... common): follow-ing previous work (Zheng et al., 2011), we ob-tained word pairs from Wikipedia3and SpellGood4. We lower-cased the entries from these sources, re-moved the duplicates and the ... word. In contrast, inen common, all error types are more prone to oc-cur word- medially. This means that errors at word edges are corrected more often than the word- internal errors, which can...