Error Recovery in Natural Language Parsing with a Level-Synchronous Approach

A level-synchronous probabilistic scoring function, which takes advantage of the wide-scope contextual information in the same phrase-level, is proposed in this paper to detect and recover the errors in ill-formed inputs. Traditionally, partial parsing is used in dealing with ill-formed inputs without fixing the errors. However, such an approach only provides coarse and limited information, which prevents the sentence from being processed further (such as translation). To fix the errors in ill-formed inputs, a recovery mechanism using the probabilistic scoring function is proposed in this paper. Experimental results show that 35% of the ill-formed inputs can be recovered to well-formed parses. The recall of constituent brackets is also significantly improved from 68.49% to 76.60%, while the precision of brackets is slightly improved from 79.49% to 80.69%.