User Driven IE is a concept only at the moment: its aim is to address several areas of research such as how to use machine learning techniques to allow a system to be adapted to a new domain without expert intervention, and how the user will interact with the system. Below we discuss the structures that must be learned and proposed strategies for learning them, and the proposed interaction with users that we envision will be necessary to customize a system to a particular application.
A number of issues arise in connection with designing user-driven IE. First, the quality of the system depends partly on the quality of the training data it is provided with (cf. the above figure on the low-quality of much of the human MUC data, compared with the best human data). This makes the provision of tools to involve users in this process as part of their normal work-flow important see e.g. [24]. Secondly, the type of the learned data structures impact the maintainability of the system. Stochastic models, for example, perform well in certain cases, but cannot be hand-tailored to squeeze out a little extra performance, or to eliminate an obvious error. This is an advantage of error-driven transformation-based learning of patterns for IE with a deterministic automaton-based recognition engine, such as the work of the team at MITRE [1], following the work of Brill [9], as well as for all work done in the ILP paradigm.