Enhancing Active Learning for Semantic Role Labeling via Compressed Dependency Trees

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 183–191, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Enhancing Active Learning for Semantic Role Labeling via Compressed Dependency Trees Chenhua Chen, Alexis Palmer, and Caroline Sporleder Computational Linguistics and Phonetics Saarland University, Saarbr¨ucken, Germany ch.chenua@googlemail.com, {apalmer, csporled}@coli.uni-sb.de Abstract This paper explores new approaches to ac- tive learning (AL) for semantic role label- ing (SRL), focusing in particular on com- bining typical informativity-based sam- pling strategies with a novel measure of representativeness based on compressed dependency trees (CDTs). In essence, the compressed representation encodes the target predicate and the key dependents of the verb complex in the sentence. We first present our method for producing CDTs from the output of an existing dependency parser. The compressed trees are used as features for training a supervised SRL sys- tem. Second, we present a study of AL for SRL. We investigate a number of dif- ferent sample selection strategies, and the best results are achieved by incorporating CDTs for example selection based on both informativity and representativeness. We show that our approach can reduce by up to 50% the amount of training data needed to attain a given level of performance. 1 Introduction The focus of this paper is active learning for se- mantic role labeling, a little-studied intersection of two rather substantial bodies of work. One aim of active learning (AL) is to reduce the number of labeled training instances required to reach a given performance level using super- vised machine learning techniques. This is accom- plished by allowing the learner to guide the selec- tion of examples to be annotated and added to the training set; at each iteration the learner queries for the example (or set of examples) that will be most informative to its present state. AL is an attrac- tive idea for natural language processing (NLP) because of its potential to dramatically reduce the need for expensive expert annotation, and it has been successfully applied in various areas of natu- ral language processing (Tang et al., 2002; Settles and Craven, 2008), including named entity recog- nition (Shen et al., 2004),text classification (Yang et al., 2009), image retrieval (Zhou, 2006), part- of-speech tagging (Ringger et al., 2007), mor- pheme glossing (Baldridge and Palmer, 2009), and syntactic parsing (Hwa, 2004; Osborne and Baldridge, 2004). The problems of scarce annotated data and the expense of annotating new data are at least as rel- evant for semantic role labeling (SRL) as for the above-mentioned areas of NLP. Existing work on automatic SRL usually explores supervised ma- chine learning approaches to mark the semantic roles of predicates automatically by training clas- sifiers using large annotated corpora.1 Although such approaches can achieve reasonably good per- formance, annotating a large corpus is still ex- pensive and time consuming. Moreover, the per- formance of trained classifiers may degrade re- markably when they are applied to out-of-domain data (Johansson and Nugues, 2008a). There is very little work on AL for SRL (e.g. …

Content truncated to 500 words