The Integration of Dependency Relation Classification and Semantic Role Labeling Using Bilayer Maximum Entropy Markov Models

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 243–247 Manchester, August 2008 The Integration of Dependency Relation Classification and Semantic Role Labeling Using Bilayer Maximum Entropy Markov Models Weiwei Sun and Hongzhan Li and Zhifang Sui Institute of Computational Linguistics Peking University {weiwsun, lihongzhan.pku}@gmail.com, szf@pku.edu.cn Abstract This paper describes a system to solve the joint learning of syntactic and seman- tic dependencies. An directed graphical model is put forward to integrate depen- dency relation classification and semantic role labeling. We present a bilayer di- rected graph to express probabilistic re- lationships between syntactic and seman- tic relations. Maximum Entropy Markov Models are implemented to estimate con- ditional probability distribution and to do inference. The submitted model yields 76.28% macro-average F1 performance, for the joint task, 85.75% syntactic depen- dencies LAS and 66.61% semantic depen- dencies F1. 1 Introduction Dependency parsing and semantic role labeling are becoming important components in many kinds of NLP applications. Given a sentence, the task of de- pendency parsing is to identify the syntactic head of each word in the sentence and classify the rela- tion between the dependent and its head; the task of semantic role labeling consists of analyzing the propositions expressed by some target predicates. The integration of syntactic and semantic parsing interests many researchers and some approaches has been proposed (Yi and Palmer, 2005; Ge and Mooney, 2005). CoNLL 2008 shared task pro- poses the merging of both syntactic dependencies and semantic dependencies under a unique unified representation (Surdeanu et al., 2008). We explore 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. the integration problem and evaluate our approach using data provided on CoNLL 2008. This paper explores the integration of depen- dency relation classification and semantic role la- beling, using a directed graphical model that is also known as Bayesian Networks. The directed graph of our system can be seen as one chain of obser- vations with two label layers: the observations are argument candidates; one layer’s label set is syn- tactic dependency relations; the other’s is semantic dependency relations. To estimate the probability distribution of each arc and do inference, we im- plement a Maximum Entropy Markov Model (Mc- Callum et al., 2000). Specially, a logistic regres- sion model is used to get the conditional probabil- ity of each arc; dynamic programming algorithm is applied to solve the ”argmax” problem. 2 System Description Our DP-SRL system consists of 5 stages: 1. dependency parsing; 2. predicate prediction; 3. syntactic dependency relation classification and semantic dependency relation identifica- tion; 4. semantic dependency relation classification; 5. semantic dependency relation inference. 2.1 Dependency Parsing In dependency parsing stage, MSTParser1 (Mc- Donald et al., 2005), a dependency parser that searches for maximum spanning trees over di- rected graphs, is used. we use MSTParser’s default 1http://www.seas.upenn.edu/ strctlrn/MSTParser/MSTParser.html 243 Lemma and its POS tag Number of children Sequential POS tags of children Lemma and POS of Neighboring words Lemma and POS of parent …

Content truncated to 500 words