Machine Intelligibility and the Duality Principle

The scale and diversity of networked sources of data and computer programs is rapidly swamping human abilities to digest and even locate relevant information. The high speed of computing has compounded this problem by the generation of even larger amounts of data, derived in ways that are generally opaque to human users. The result is an increasing gulf between human and computer abilities. Society's ever wider-scale dependence on rapidly growing networked sources of software threatens severe breakdowns if machine intelligibility issues are not given high priority. In this paper we argue that lack of machine intelligibility in human-computer interactions can be traced directly to present approaches to software design. According to the duality principle in this paper, software involved in human-computer interaction should contain two distinct layers: a declarative knowledge-level layer and a lower-level functional or procedural-knowledge layer. This extends the formal methods separation of speciication and implementation by requiring that the declarative layer be capable of extensive human interrogation at runtime. The declarative layer should support simple deductive and inductive inference. The ease with which declarative knowledge can be translated to natural language could be used to provide a human-comprehensible \window" into the properties of the underlying functional layer. Adaptation of the declarative knowledge in response to human interaction could be supported by modern machine learning mechanisms. In addition, declarative knowledge could be used to facilitate human-comprehensible communication between programs. Existing well-developed technologies can be used to implement the declarative layer. The obvious language of choice is pure 1 Prolog, augmented with machine learning mechanisms based on Inductive Logic Programming. The underlying functional layer would be composed of normal procedurally encoded computer programs. It is argued that the duality principle in software design is a necessity for dealing with the demands of wide-scale computer usage in the information age and should be an urgent goal for computer science research at the start of the 21st century.