ReliAble dependency arc recognition

We propose a novel natural language processing task, ReliAble dependency arc recognition (RADAR), which helps high-level applications better utilize the dependency parse trees. We model RADAR as a binary classification problem with imbalanced data, which classifies each dependency parsing arc as correct or incorrect. A logistic regression classifier with appropriate features is trained to recognize reliable dependency arcs (correct with high precision). Experimental results show that the classification method can outperform a probabilistic baseline method, which is calculated by the original graph-based dependency parser. As a fundamental task of natural language processing, dependency parsing has become increasingly popular in recent years. It aims to find a dependency parse tree among words for a sentence. Fig. 1 shows an example of dependency parse tree for a sentence, where sbj is a subject, obj is an object, etc. (Johansson & Nugues, 2007). Dependency parsing are widely used: in biomedical text However, when we migrate dependency parsing systems from laboratory demonstrations to high-level applications, even the best parser available today still encounter some serious difficulties. First of all, parsing performance usually dramatically degrades in real fields because of domain migration. Secondly, since every parser inevitably will make some mistakes during decoding, outputs from any dependency parser are always fraught with a variety of errors. Thus, in some high-level applications which expect to use correct parsing results, it is extremely important to be able to predict the reliability of the auto-parsed results. If these applications just use correct parsing results and ignore incorrect results, their performances may be improved further. For instance, if an entity relation extraction (a kind of information extraction) system, which depends on parsing results heavily (Zhang, Zhang, Su, & Zhou, 2006), only extracts relations from correct parsing sentences, then the system can extract more accurate relations and import less wrong relations through incorrect parsing results. Although some implied relations in those incorrect parsing sentences are missed, these missing relations may be extracted from other sentences that can be parsed correctly while zooming in the data to the whole Web. Most large-margin based training algorithm for dependency parsing output models that predict a single parse tree of the input sentence, with no additional confidence information about the correctness of it. Therefore, an interesting problem is how to judge a parsing result as correct or not. However, it is difficult to obtain a parse tree in which all sub-structures are parsed correctly. …