Learning Local Content Shift Detectors from Document-level Information

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 759–770, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Learning Local Content Shift Detectors from Document-level Information Rich´ard Farkas Institute for Natural Language Processing University of Stuttgart farkas@ims.uni-stuttgart.de Abstract Information-oriented document labeling is a special document multi-labeling task where the target labels refer to a specific information instead of the topic of the whole document. These kind of tasks are usually solved by look- ing up indicator phrases and analyzing their local context to filter false positive matches. Here, we introduce an approach for machine learning local content shifters which detects irrelevant local contexts using just the origi- nal document-level training labels. We handle content shifters in general, instead of learn- ing a particular language phenomenon detec- tor (e.g. negation or hedging) and form a sin- gle system for document labeling and content shift detection. Our empirical results achieved 24% error reduction – compared to supervised baseline methods – on three document label- ing tasks. 1 Introduction There are special document multi-labeling tasks where the target labels refer to a specific piece of information extractable from the document instead of the overall topic of the document. In these kinds of tasks the target information is usually an attribute or relation related to the target entity (usually a per- son or an organisation) of the document in question, but the task is to assign class labels at the document (entity) level. For example, the smoking habits of the patients are frequently discussed in the textual parts of clinical notes (Uzuner et al., 2008). In this case the task is to find specific information in the text – i.e. the patient in question is a smoker, past smoker, non-smoker – but at the end an applica- tion has to assign labels to the documents(patients). Similarly, the soccer club names where a sportsman played for are document(sportman)-level labels in Wikipedia articles expressed by the Wikipedia cat- egories. The target information in these tasks is usually just mentioned in the document and much of the document is irrelevant for this information request in contrast to standard document classifi- cation tasks where the goal is to identify the top- ics of the whole document. On the other hand, they are not a standard information extraction task as the task is to assign class labels to documents, and the training dataset contains labels just at this level. These special tasks lie somewhere between information extraction and document classification and require special approaches to solve them. We will call them Information-oriented document label- ing throughout this paper. There are several appli- cation areas where information-oriented document labels are naturally present in an enormous amount like clinical records, Wikipedia categories and user- generated tags of news. Previous evaluation campaigns (Uzuner et al., 2008; Pestian et al., 2007; Uzuner, 2009) demon- strated that information-oriented document labeling can be effectively performed by looking up indicator phrases which can be gathered by hand, by corpus …

Content truncated to 500 words