Improving Recall of Regular Expressions for Information Extraction

Abstract

Learning or writing regular expressions to identify instances of a specific concept within text documents with a high precision and recall is challenging. It is relatively easy to improve the precision of an initial regular expression by identifying false positives covered and tweaking the expression to avoid the false positives. However, modifying the expression to improve recall is difficult since false negatives can only be identified by manually analyzing all documents, in the absence of any tools to identify the missing instances. We focus on partially automating the discovery of missing instances by soliciting minimal user feedback. We present a technique to identify good generalizations of a regular expression that have improved recall while retaining high precision. We empirically demonstrate the effectiveness of the proposed technique as compared to existing methods and show results for a variety of tasks such as identification of dates, phone numbers, product names, and course numbers on real world datasets.