Inducing Domain-Specific Semantic Class Taggers from (Almost) Nothing

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 275–285, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Inducing Domain-specific Semantic Class Taggers from (Almost) Nothing Ruihong Huang and Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 {huangrh,riloff}@cs.utah.edu Abstract This research explores the idea of inducing domain-specific semantic class taggers us- ing only a domain-specific text collection and seed words. The learning process be- gins by inducing a classifier that only has access to contextual features, forcing it to generalize beyond the seeds. The contex- tual classifier then labels new instances, to expand and diversify the training set. Next, a cross-category bootstrapping pro- cess simultaneously trains a suite of clas- sifiers for multiple semantic classes. The positive instances for one class are used as negative instances for the others in an it- erative bootstrapping cycle. We also ex- plore a one-semantic-class-per-discourse heuristic, and use the classifiers to dynam- ically create semantic features. We eval- uate our approach by inducing six seman- tic taggers from a collection of veterinary medicine message board posts. 1 Introduction The goal of our research is to create semantic class taggers that can assign a semantic class label to ev- ery noun phrase in a sentence. For example, con- sider the sentence: “The lab mix was diagnosed with parvo and given abx”. A semantic tagger should identify the “the lab mix” as an ANIMAL, “parvo” as a DISEASE, and “abx” (antibiotics) as a DRUG. Accurate semantic tagging could be beneficial for many NLP tasks, including coref- erence resolution and word sense disambiguation, and many NLP applications, such as event extrac- tion systems and question answering technology. Semantic class tagging has been the subject of previous research, primarily under the guises of named entity recognition (NER) and mention de- tection. Named entity recognizers perform se- mantic tagging on proper name noun phrases, and sometimes temporal and numeric expressions as well. The mention detection task was introduced in recent ACE evaluations (e.g., (ACE, 2007; ACE, 2008)) and requires systems to identify all noun phrases (proper names, nominals, and pro- nouns) that correspond to 5-7 semantic classes. Despite widespread interest in semantic tag- ging, nearly all semantic taggers for comprehen- sive NP tagging still rely on supervised learn- ing, which requires annotated data for training. A few annotated corpora exist, but they are rela- tively small and most were developed for broad- coverage NLP. Many domains, however, are re- plete with specialized terminology and jargon that cannot be adequately handled by general-purpose systems. Domains such as biology, medicine, and law are teeming with specialized vocabulary that necessitates training on domain-specific corpora. Our research explores the idea of inducing domain-specific semantic taggers using a small set of seed words as the only form of human su- pervision. Given an (unannotated) collection of domain-specific text, we automatically generate training instances by labelling every instance of a seed word with its designated semantic class. We then train a classifier to do …

Content truncated to 500 words