Multiple Feature-sets Algorithm Of Dependency Parsing

Dependency parsing is a competitive parsing technique compared with traditional parsers like constituent grammar parsers. Its advantages are two-fold, namely, simplicity and generality. Dependency parsing makes it possible for people without much linguistics knowledge to understand the structure of a sentence easily. Besides, dependency parsing doesn’t add additional information for a specific language. It focus on the modification relationship in the sentence (for example, adjective words usually modify the noun words right behind them).Traditional dependency parsing algorithms can be broadly divided into three categories: graph-based methods, transition-based methods and semi-supervised methods. Each method has its strengths and weaknesses. However, long-distance modification is a common challenge for all of them. The main thrust of graph-based methods is to introduce high-order information and the hot point of transition-based methods lies in expanding the search space. Besides, for semi-supervised methods, researchers have been trying to introduce external information into the system.We proposed a simple and effective method for combating long-distance modification problem based on chunking information. Firstly, we trained a chunker using a open-source implementation of a first-order Conditional Random Fields; Second, we built a clause-chunk tree for a given sentence according to the punctuation marks and chunking information; Finally, we integrated features extracted from clause-chunk trees into the baseline parsers. Experiments shows that our algorithm significantly outperform the baseline parsers (the MSTParser algorithm and the Carreras algorithm) without increasing the time complexity. Given the chunking information, our parser increase the parsing accuracy from91.36%and92.20to93.19%and93.89%, respectively.