Classifying Message Board Posts with an Extracted Lexicon of Patient Attributes

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1557–1562, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Classifying Message Board Posts with an Extracted Lexicon of Patient Attributes Ruihong Huang and Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 {huangrh, riloff}@cs.utah.edu Abstract The goal of our research is to distinguish vet- erinary message board posts that describe a case involving a specific patient from posts that ask a general question. We create a text classifier that incorporates automatically gen- erated attribute lists for veterinary patients to tackle this problem. Using a small amount of annotated data, we train an information extrac- tion (IE) system to identify veterinary patient attributes. We then apply the IE system to a large collection of unannotated texts to pro- duce a lexicon of veterinary patient attribute terms. Our experimental results show that us- ing the learned attribute lists to encode pa- tient information in the text classifier yields improved performance on this task. 1 Introduction Our research focuses on the problem of classify- ing message board posts in the domain of veterinary medicine. Most of the posts in our corpus discuss a case involving a specific patient, which we will call patient-specific posts. But there are also posts that ask a general question, for example to seek advice about different medications, information about new procedures, or how to perform a test. Our goal is to distinguish the patient-specific posts from general posts so that they can be automatically routed to dif- ferent message board folders. Distinguishing patient-specific posts from general posts is a challenging problem for two reasons. First, virtually any medical topic can appear in either type of post, so the vocabulary is very similar. Second, a highly skewed distribution exists between patient- specific posts and general posts. Almost 90% of the posts in our data are about specific patients. With such a highly skewed distribution, it would seem logical to focus on recognizing instances of the minority class. But the distinguishing characteristic of a general post is the absence of a patient. Two nearly identical posts belong in different categories if one mentions a patient and the other does not. Consequently, our aim is to create features that iden- tify references to a specific patient and use these to more accurately distinguish the two types of posts. Our research explores the use of information ex- traction (IE) techniques to automatically identify common attributes of veterinary patients, which we use to encode patient information in a text classifier. Our approach involves three phases. First, we train a conditional random fields (CRF) tagger to iden- tify seven common types of attributes that are of- ten ascribed to veterinary patients: SPECIES/BREED, NAME, AGE, GENDER, WEIGHT, POSSESSOR, and DISEASE/SYMPTOM. Second, we apply the CRF tagger to a large set of unannotated message board posts, collect its extractions, and harvest the most frequently extracted terms to create a Veterinary Pa- tient Attribute (VPA) Lexicon. Finally, …

Content truncated to 500 words