Dependency-Based Semantic Role Labeling using Convolutional Neural Networks

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 279–288, Denver, Colorado, June 4–5, 2015. Dependency-Based Semantic Role Labeling using Convolutional Neural Networks William R. Foland Jr. OKRobotGo, Ltd. 5345 Dunraven Circle Golden, Co, 80403, USA bill.foland@okrobotgo.com James H. Martin Department of Computer Science and Institute of Cognitive Science University of Colorado Boulder, CO 80309 James.Martin@colorado.edu Abstract We describe a semantic role labeler with state- of-the-art performance and low computational requirements, which uses convolutional and time-domain neural networks. The system is designed to work with features derived from a dependency parser output. Various system options and architectural details are discussed. Incremental experiments were run to explore the benefits of adding increasingly more com- plex dependency-based features to the system; results are presented for both in-domain and out-of-domain datasets. 1 Introduction Semantic role labeling (Gildea and Jurafsky [2002]), the task of identifying and classifying the semantic arguments of verbal and nominal predicates in text, represents one of the most complex NLP tasks to be addressed by supervised machine learning tech- niques. In the standard supervised approach to building SRL systems, collections of multiway clas- sifiers are trained using annotated corpora such as PropBank (Palmer et al. [2005]). In this approach, classifiers are trained using features derived directly from the original source text, as well as from subse- quent syntactic and semantic processing. As reported in several shared tasks (Carreras and M`arquez [2004],Carreras and M`arquez [2005],Hajiˇc et al. [2009]), SRL systems trained in this manner can achieve high performance. State-of-the-art sys- tems employ classifiers such as support vector ma- chines trained with large numbers of relatively com- plex combinations of features, often combined with re-ranking based on multiple syntactic analyses. Un- fortunately, these approaches have a number of non- trivial limitations including the computational cost of the syntactic parsing and the sparse nature of the complex features on which they rely. This latter lim- itation is particularly critical since it leads to signif- icant degradation in performance when the trained system is applied to texts from new domains. However, recent results using multilayer neu- ral networks and pre-trained word embeddings have demonstrated high performance using a much smaller number of minimalist features. The archi- tecture described by Collobert et al. [2011] com- bines time delay convolutional neural networks (Waibel et al. [1989]) and pre-trained word embed- dings for a number of NLP tasks. They develop four components and compare their performance to pre- vious benchmarks, one of which is an SRL system which uses features derived from a phrase-structure parse as input, based on the CoNLL 2005 shared task (Carreras and M`arquez [2005]). The work described here adopts the basic archi- tecture from Collobert et al. [2011] and explores issues related to the use of this architecture in the context of the CoNLL 2009 shared task. In partic- ular, we present Daisy, a system that (1) employs features derived from dependency parse as input, (2) assigns semantic roles to both verbal and nom- inal predicates, and (3) …

Content truncated to 500 words