An overall aim of UDIE would be to find the right compromise for a user of IE between automatic and user-driven methods. An important aspect of UDIE that supplements the use of learning methods is a user interface quite different from developer-orientated interfaces such as GATE . There will be a range of ways in which a user can indicate to the system their interests, in advance of any automated learning or user feedback, since it would be foolish to ignore the extent to which a user may have some clear notions of what is wanted from a corpus. However, and this is an important difference from classic IE descending from MUC, we will not always assume in what follows that the user does have ``templates in mind", but only that there are facts of great interest to the user in a given corpus and it can be the job of this system interface to help elicit them in a formal representation.
It is crucial to recall here that one of the few productive methods for optimising traditional IR in the last decade has been the use of user- feedback methods, typically ones where a user can indicate from a retrieved document set that, say, this ten are good and this ten bad. These results are then fed back to optimise the retrieval iteratively by modifying the request. It is not easy to adapt this methodology directly to IE, even though now, with full text IR available for large indexed corpora, one can refer to sentence documents being retrieved by IR, documents of precisely the span of a classic IE template, so that one might hope for some transfer of IR optimisation methods.
However, although the user can mark sentences so retrieved as good or bad, the ``filled template" part of the document, filled template pairings cannot be so marked by a user who, by definition, is not assumed to be familiar with template formalisms. In this section of the work we shall mention ways in which a user can indicate preferences, needs and choices at the interface that contribute to template construction whose application he can assess, though not their technical structure.
Doing this will require the possibility of a user marking, on the screen, key portions of text, ones that contain the desired facts; as well as the ability to input, in some form of an interface language (English, Italian etc.), concepts in key facts or template content including predicates and ranges of fillers). This aspect of the paper is complementary to supervised learning methods for templates, lexicons and KR structures, none of which need assume that the user does have a full and explicit concept of what is wanted from a corpus.