Generally Applicable Legal Expert Systems (GALES)

Starting from certain properties of a previously developed system, the viability of rule-based legal expert systems shall be inquired. For the envisaged system the following properties should hold: 1.The system is restricted in its base structure (rules, inference engine) to classical deductive logics and a semantics for constructed languages (Carnap). At the same time, it ought to simulate at the user interface level the typical natural language case solving behaviour of lawyers. For this purpose, an ontology-neutral formal structure for all types of legal argumentation will be implemented. 2.Knowledge acquisition ought to be possible without involving a knowledge engineer, i.e., in direct communication between natural and artificial expert. As a basic method for this purpose deductive data-bases are envisaged, which map the formal structures as mentioned above (1) and which achieve the relations to natural language texts. 3.Expert systems of this kind are to be enabled to communicate, via according interfaces and scripts, to compare diverging opinions among them and eventually continue work on solving cases which have been started by other expert systems of the same kind. 4.It is obvious that the employment of such communicating rule-based expert systems for co-operative work in the Internet is highly promising. 5.In order to test such expert systems for their fitness to be employed in practical work, prototypes in the field of Austrian administrative law will be implemented.

Can artificially intelligent systems be made to evaluate given information and revise such evalua-tions on the basis of new information "reasonably" beyond mere logical validity? This question could be answered affirmatively for the case of syntactically elementary propositions. The case of complex propositions, however, leads into theoretical questions of a new kind. 1 Objective The project`s objective was primarily to explorate the foundtions for a new kind of expert systems (bridging the gap between artificial intelligence and arificial rationality) and eventually test accord-ing prototypes against selected examples drawn from administrative law. 2 Innovative aspects The innovative kernel of the investigations consists in a strict distiction of logical and epistemic operations on shared propositions: while the logical operations serve for deriving and consistency checking, degrees of trust or "affirmation" are calculated with the help of an affirmation function on the involved propositions. The system has learning abilities via appropriate interfaces. Trainers may be either human (prefer-entally experts in a given field of knowledge) or analogous, already trained, systems. 3 Main results Based on a non-probabilistic affirmation measure, the chosen approach has delivered good results for logically elementary propositions, where it adopts and revises degrees of affirmation as ex-pected from a rational reasoner. The case of complex propositions uncovered the following problem: without the introduction of further restricting conditions, attributions of degrees of affirmation are unequivocally defined but not intuitively reasonable in all cases, while the introduction of such restricting conditions, though assuring intuitive reasonableness, were not unequivocally defined. Further investigations into this problem showed that in addition to a memory of logical and concep-tual rules (as well as of the given propositions), a system of the intended kind will also need a memory of its own previous learning processes.