In this paper, we describe the lessons we learned in developing AgentBuilder, a commercial system for rapidly creating agents that extract information from web sites. AgentBuilder employs a Programming-by-Example (PBE) paradigm, where users train the system by showing the system examples of web pages and the information to be extracted from these pages. The system uses a sophisticated machine learning algorithm for inducing extraction rules from examples and eventually creates web agents that navigate through a site and extract information. Previous work on Programming-by-Example has discussed the importance of felicity conditions, which simplify training so that the learner can more readily understand what is being taught. In this paper, we show that, in addition, developers must design the learning system to insure that the teacher (i.e. the user) can train the system without becoming frustrated. We discuss these characteristics, which we refer to as “trainability”, and show how...
Steven Minton, Sorinel I. Ticrea, Jennifer Beach