Domain Adaptation with Latent Semantic Association for Named Entity Recognition

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 281–289, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Domain Adaptation with Latent Semantic Association for Named Entity Recognition Honglei Guo Huijia Zhu Zhili Guo Xiaoxun Zhang Xian Wu and Zhong Su IBM China Research Laboratory Beijing, P. R. China {guohl, zhuhuiji, guozhili, zhangxx, wuxian, suzhong}@cn.ibm.com Abstract Domain adaptation is an important problem in named entity recognition (NER). NER classi- fiers usually lose accuracy in the domain trans- fer due to the different data distribution be- tween the source and the target domains. The major reason for performance degrading is that each entity type often has lots of domain- specific term representations in the different domains. The existing approaches usually need an amount of labeled target domain data for tuning the original model. However, it is a labor-intensive and time-consuming task to build annotated training data set for every target domain. We present a domain adapta- tion method with latent semantic association (LaSA). This method effectively overcomes the data distribution difference without lever- aging any labeled target domain data. LaSA model is constructed to capture latent seman- tic association among words from the unla- beled corpus. It groups words into a set of concepts according to the related context snip- pets. In the domain transfer, the original term spaces of both domains are projected to a con- cept space using LaSA model at first, then the original NER model is tuned based on the se- mantic association features. Experimental re- sults on English and Chinese corpus show that LaSA-based domain adaptation significantly enhances the performance of NER. 1 Introduction Named entities (NE) are phrases that contain names of persons, organizations, locations, etc. NER is an important task in information extraction and natu- ral language processing (NLP) applications. Super- vised learning methods can effectively solve NER problem by learning a model from manually labeled data (Borthwick, 1999; Sang and Meulder, 2003; Gao et al., 2005; Florian et al., 2003). However, em- pirical study shows that NE types have different dis- tribution across domains (Guo et al., 2006). Trained NER classifiers in the source domain usually lose accuracy in a new target domain when the data dis- tribution is different between both domains. Domain adaptation is a challenge for NER and other NLP applications. In the domain transfer, the reason for accuracy loss is that each NE type often has various specific term representations and context clues in the different domains. For ex- ample, {“economist”, “singer”, “dancer”, “athlete”, “player”, “philosopher”, ...} are used as context clues for NER. However, the distribution of these representations are varied with domains. We expect to do better domain adaptation for NER by exploit- ing latent semantic association among words from different domains. Some approaches have been pro- posed to group words into “topics” to capture im- portant relationships between words, such as Latent Semantic Indexing (LSI) (Deerwester et al., 1990), probabilistic Latent Semantic Indexing (pLSI) (Hof- mann, 1999), Latent …

Content truncated to 500 words