Adapting Text instead of the Model: An Open Domain Approach

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 229–237, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics Adapting Text instead of the Model: An Open Domain Approach Gourab Kundu, Dan Roth University of Illinois at Urbana Champaign Urbana, IL 61801 {kundu2,danr}@illinois.edu Abstract Natural language systems trained on labeled data from one domain do not perform well on other domains. Most adaptation algorithms proposed in the literature train a new model for the new domain using unlabeled data. How- ever, it is time consuming to retrain big mod- els or pipeline systems. Moreover, the domain of a new target sentence may not be known, and one may not have significant amount of unlabeled data for every new domain. To pursue the goal of an Open Domain NLP (train once, test anywhere), we propose ADUT (ADaptation Using label-preserving Transfor- mation), an approach that avoids the need for retraining and does not require knowledge of the new domain, or any data from it. Our ap- proach applies simple label-preserving trans- formations to the target text so that the trans- formed text is more similar to the training do- main; it then applies the existing model on the transformed sentences and combines the predictions to produce the desired prediction on the target text. We instantiate ADUT for the case of Semantic Role Labeling (SRL) and show that it compares favorably with ap- proaches that retrain their model on the target domain. Specifically, this “on the fly” adapta- tion approach yields 13% error reduction for a single parse system when adapting from the news wire text to fiction. 1 Introduction In several NLP tasks, systems trained on annotated data from one domain perform well when tested on the same domain but adapt poorly to other do- mains. For example, all systems of CoNLL 2005 shared task (Carreras and Màrquez, 2005) on Se- mantic Role Labeling showed a performance degra- dation of almost 10% or more when tested on a dif- ferent domain. Most works in domain adaptation have focused on learning a common representation across train- ing and test domains (Blitzer et al., 2006; DauméIII, 2007; Huang and Yates, 2009). Using this represen- tation, they retrain the model for every new domain. But these are not Open Domain Systems since the model needs to be retrained for every new domain. This is very difficult for pipeline systems like SRL where syntactic parser, shallow parser, POS tagger and then SRL need to be retrained. Moreover, these methods need to have a lot of unlabeled data that is taken from the same domain, in order to learn mean- ingful feature correspondences across training and test domain. These approaches cannot work when they do not have a lot of unlabeled data from the test domain or when the test domain in itself is very di- verse, e.g., the web. The contribution of this paper is a new frame- work for adaptation. We propose ADUT (ADap- tation Using label-preserving Transformation) as a framework …

Content truncated to 500 words