GazaParser: a high-performance natural language parser.The best empirical performance (evalb) on English and Chinese treebanks: English WSJ Section 23: F1=91.86% Chinese CTB5.1 Article 271-300: F1=85.58% The dependency accuracy (converted with Penn2Malt and evaluated with MaltEval): English WSJ Section 23: UAS = 93.2% Chinese CTB5.1 Article 271-300: UAS = 87.1% Note that this parser does not use any parser combination technique or any addtional training data. Source codes are ready for download. Many thanks to the great Berkeley parser, for providing me a lot of useful source codes. I hope this parser can also be useful to someone else. Currently, the parsing speed is not very fast. I wish I can fix it in the future works. Users may try to set a higher pruning threshold "-dpt -14.0". This parser can run much faster with a minor loss of performance (F1=91.79% for English). We appreciate anyone who uses this parser to acknowledge our work: Xiao Chen and Chunyu Kit, 2011, Improving POS tagging for context-free parsing, In Proceedings of 5th International Joint Conference on Natural Language Processing (IJCNLP 2011). Chiangmai, Thailand. Xiao Chen, 2012, Discriminative Constituent Parsing with Localized Features. PhD Thesis, City University of Hong Kong. Xiao Chen and Chunyu Kit, 2012, Higher-Order Constituent Parsing and Parser Combination. In Proceedings of ACL 2012。
|