Discriminative vs. Generative Approaches in Semantic Role Labeling

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 223–227 Manchester, August 2008 Discriminative vs. Generative Approaches in Semantic Role Labeling Deniz Yuret Koc¸ University dyuret@ku.edu.tr Mehmet Ali Yatbaz Koc¸ University myatbaz@ku.edu.tr Ahmet Engin Ural Koc¸ University aural@ku.edu.tr Abstract This paper describes the two algorithms we developed for the CoNLL 2008 Shared Task “Joint learning of syntactic and se- mantic dependencies”. Both algorithms start parsing the sentence using the same syntactic parser. The first algorithm uses machine learning methods to identify the semantic dependencies in four stages: identification and labeling of predicates, identification and labeling of arguments. The second algorithm uses a generative probabilistic model, choosing the seman- tic dependencies that maximize the proba- bility with respect to the model. A hybrid algorithm combining the best stages of the two algorithms attains 86.62% labeled syntactic attachment accuracy, 73.24% la- beled semantic dependency F1 and 79.93% labeled macro F1 score for the combined WSJ and Brown test sets1. 1 Introduction In this paper we describe the system we developed for the CoNLL 2008 Shared Task (Surdeanu et al., 2008). Section 2 describes our approach for iden- tifying syntactic dependencies. For semantic role labeling (SRL), we pursued two independent ap- proaches. Section 3 describes our first approach, where we treated predicate identification and la- beling, and argument identification and labeling as 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. 1These numbers are slightly higher than the official results due to a small bug in our submission. four separate machine learning problems. The fi- nal program consists of four stages, each stage tak- ing the answers from the previous stage as given and performing its own identification or labeling task based on a model generated from the train- ing set. Section 4 describes our second approach where we used a generative model based on the joint distribution of the predicate, the arguments, their labels and the syntactic dependencies con- necting them. Section 5 summarizes our results and suggests possible improvements. 2 Syntactic dependencies We used a non-projective dependency parser based on spanning tree algorithms. The parameters were determined based on the experimental results of the English task in (McDonald et al., 2005), i.e. we used projective parsing and a first order feature set during training. Due to the new representation of hyphenated words in both training and testing data of our shared task and the absence of the gold part of speech (GPOS) column in the test data, the for- mat of the CoNLL08 shared task is slightly differ- ent from the format of the CoNLL05 shared task, which is supported by the McDonald’s parser. We reformatted the data accordingly. The resulting la- beled attachment score on the test set is 87.39% for WSJ and 80.46% for Brown. 3 The 4-stage discriminative approach Our first approach to SRL consists of four distinct stages: (1) predicate identification, (2) predicate labeling, (3) argument identification, and (4) argu- ment labeling. A discriminative machine …

Content truncated to 500 words