This paper introduces structural correspondence learning (SCL) as a method for domain adaptation in natural language processing (NLP). SCL aims to automatically induce correspondences among features from different domains by modeling their correlations with pivot features—features that behave similarly in both domains. The method is tested on part-of-speech (PoS) tagging, where it shows performance improvements for varying amounts of source and target training data, as well as improvements in target domain parsing accuracy using the improved tagger. Discriminative learning methods are widely used in NLP, but they perform best when training and test data are from the same distribution. In many NLP tasks, however, we face new domains with scarce or no labeled data. SCL addresses this by using unlabeled data from both domains to learn a common feature representation that is meaningful across both. This representation is learned using a method called structural correspondence learning (SCL), which identifies correspondences among features by modeling their correlations with pivot features. SCL is a general technique applicable to any feature-based discriminative learner. It involves defining pivot features that occur frequently in both domains and using them to learn a mapping from the original feature spaces to a shared, low-dimensional real-valued feature space. The algorithm then uses this mapping to train a classifier on the source domain data, which is expected to generalize well to the target domain. The paper demonstrates SCL's effectiveness in transferring a PoS tagger from the Wall Street Journal (WSJ) to MEDLINE (biomedical abstracts), which use very different vocabularies. The results show improved PoS accuracy and end-to-end parsing accuracy when using the improved tagger. SCL also performs well in scenarios with no labeled target domain data, outperforming both supervised and semi-supervised baselines. The paper also discusses related work in domain adaptation, highlighting the importance of using unlabeled data and the differences between various approaches. It concludes that SCL is a promising method for domain adaptation, offering a general technique that can be applied to various NLP tasks. The results show that SCL consistently outperforms other methods in terms of performance and generalization across different domains.This paper introduces structural correspondence learning (SCL) as a method for domain adaptation in natural language processing (NLP). SCL aims to automatically induce correspondences among features from different domains by modeling their correlations with pivot features—features that behave similarly in both domains. The method is tested on part-of-speech (PoS) tagging, where it shows performance improvements for varying amounts of source and target training data, as well as improvements in target domain parsing accuracy using the improved tagger. Discriminative learning methods are widely used in NLP, but they perform best when training and test data are from the same distribution. In many NLP tasks, however, we face new domains with scarce or no labeled data. SCL addresses this by using unlabeled data from both domains to learn a common feature representation that is meaningful across both. This representation is learned using a method called structural correspondence learning (SCL), which identifies correspondences among features by modeling their correlations with pivot features. SCL is a general technique applicable to any feature-based discriminative learner. It involves defining pivot features that occur frequently in both domains and using them to learn a mapping from the original feature spaces to a shared, low-dimensional real-valued feature space. The algorithm then uses this mapping to train a classifier on the source domain data, which is expected to generalize well to the target domain. The paper demonstrates SCL's effectiveness in transferring a PoS tagger from the Wall Street Journal (WSJ) to MEDLINE (biomedical abstracts), which use very different vocabularies. The results show improved PoS accuracy and end-to-end parsing accuracy when using the improved tagger. SCL also performs well in scenarios with no labeled target domain data, outperforming both supervised and semi-supervised baselines. The paper also discusses related work in domain adaptation, highlighting the importance of using unlabeled data and the differences between various approaches. It concludes that SCL is a promising method for domain adaptation, offering a general technique that can be applied to various NLP tasks. The results show that SCL consistently outperforms other methods in terms of performance and generalization across different domains.