Discriminative Learning of Syntactic and Semantic Dependencies

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 218–222 Manchester, August 2008 Discriminative Learning of Syntactic and Semantic Dependencies Lu Li1, Shixi Fan2, Xuan Wang1, Xiaolong Wang1 Shenzhen Graduate School, Harbin Institute of Technology, Xili, Shenzhen 518055, China 1{lli,wangxuan,wangxl}@insun.hit.edu.cn 2fanshixi@hit.edu.cn Abstract A Maximum Entropy Model based system for discriminative learning of syntactic and semantic dependencies submitted to the CoNLL-2008 shared task (Surdeanu, et al., 2008) is presented in this paper. The sys- tem converts the dependency learning task to classification issues and reconstructs the dependent relations based on classification results. Finally F1 scores of 86.69, 69.95 and 78.35 are obtained for syntactic depen- dencies, semantic dependencies and the whole system respectively in closed chal- lenge. For open challenge the correspond- ing F1 scores are 86.69, 68.99 and 77.84. 1 Introduction Given sentences and corresponding part-of-speech of each word, the learning of syntactic and seman- tic dependency contains two separable goals: (1) building a dependency tree that defines the syn- tactic dependency relationships between separated words; (2) specifying predicates (no matter verbs or nouns) of the sentences and labeling the seman- tic dependents for each predicate. In this paper a discriminative parser is pro- posed to implement maximum entropy (ME) mod- els (Berger, et al., 1996) to address the learning task. The system is divided into two main subsys- tems: syntactic dependency parsing and semantic dependency labeling. The former is used to find a well-formed syntactic dependency tree that occu- pies all the words in the sentence. If edges are added between any two words, a full-connected c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. graph is constructed and the dependency tree could be found using a maximum spanning tree (MST) algorithm (McDonald, et al., 2005). The latter fo- cuses on separable predicates whose semantic de- pendents could be determined using classification tools, such as ME models 1 etc.. We participated in both closed and open chal- lenge of the CoNLL-2008 shared task (Surdeanu, et al., 2008). Results are reported on both develop- ment and test sets in this paper. 2 System Description 2.1 Syntactic Parsing The goal of syntactic parsing is to create a la- beled syntactic dependency parse y for input sen- tence x including words and their parts of speech (POS). Inspired by the parsing model that imple- ments maximum spanning tree (MST) algorithm to induce the dependency parsing tree (McDonald, et al., 2005), the system employs the same frame- work. The incorporated features are defined over parts of speech of words occurring between and around a possible head-dependent relation. Suppose G = (V, E) is a directed graph, where V is the set of vertices denoting the words in sen- tence x and E is the set of directed edges between any two vertices with some scores. The MST al- gorithm is to find the most probable subgraph of G that satisfies tree constraints over all vertices. The score function of the parsing …

Content truncated to 500 words