| Link-parser is a syntactic parser based on link grammar, a formal representation of syntactical structures of natural language consisting of two parts:the dictionary and the algorithm. The dictionary records all the possible grammatical relations each word can occur in and the algorithm parses a sentence according to the rules described in the dictionary to find out whether it includes a complete linkage. An ungrammatical sentence will always contain some word(s) outside the linkage, thus failing the parser. However, because the dictionary of link-parser is compiled manually, it has the following shortcomings:(1) the dictornary is based on the compiler’s intuition rather than large-scale corpora, thus not systematic and consistent;(2) the rules in some entries of the dictonary are not complete so the parser cannot detect some grammatical errors such as subject-verb disagreement、 verb-preposition and verb-complement misuse. To overcome the above-mentioned shortcomings, we decide to reconstruct the link-parser dictionary through inducing lexicon-based grammars from typed dependency treebanks.We reconstruct the dictionary of link-parser using the100million words Penn Treebank as training set in the following way:(1) Convert phrase-structure format of Penn Treebank into dependency format;(2) Improve the current annotation scheme of dependency treebank and refine the dependency relations in the dependency treebank;(3) Extract the link-grammar dictionary from the treebank and evaluate the performance of the newly-constructed dictionary with504sentences sampled from Chinese Learner English Corpus as test set.The results show that the newly-constructed link-parser dictionary is more systemic and consistent and has broader coverage of the lexical-grammar rules than the original one. When applied to detect the grammatical errors in the test set, the accuracy, recall and F-score of the new dictionary increases by5.9%,19.9%and13.4%respectively compared with the original dictionary. |
PDF Full Text Request