We sketch the framework for a theoretical and applied system that we are developing which uses argumentation schemes and dialogue games to support dialectical interaction between people and machine learning systems. The goal is to support the automated justi cation and explanation of decisions made by AI systems, through a natural, human-oriented interface, in response to contemporary societal concerns about the impact of AI decisions upon individuals.
Decisions are increasingly made by intelligent machines. Whilst many of these
decisions are operational, for example trading decisions made by nancial
technology (FinTech) software, the A.I. wavefront, constituting the range of problems
that can be decided by machines, is accelerating. It is likely that coming years will
bring new decision making machines acting within at least the Banking, Medical,
Insurance, Legal, and Transport sectors. Sectors that are heavily regulated and
in which there is frequently a legal mandate that the output of predictive models,
and the decisions that follow from them, can be scrutinised and interpreted.
Traditionally the trade-o has been towards simpler linear predictive models,
sacri cing some accuracy for simplicity and hence understandability of the model.
However this has changed in recent years as ndings in Machine Learning (ML)
[
When interpretation of results is di cult and understandability of the model is poor there is a consequent barrier to both responsible and legal use of those systems as well as wider, more general acceptance of the systems. There is also the additional risk that a form of \moral outsourcing" occurs in which operators of intelligent systems derogate from their responsibilities as a consequence of delegating to the machine, despite people legally remaining the \problem holders". Additionally, the vocal positions of a number of public commentators has fed both a popular and a legislative wariness of A.I. This has lead to, perhaps premature, moves towards legal regulation of intelligent machines, autonomous robots, and ML algorithms. Many of those calling for legislation, and many of the drafted regulations, propose explainable A.I.; systems that can generate explanations for their decisions in concert with the search for solutions. However this is not a capability of the intelligent systems that are at stake.
Conversely, techniques such as argumentative dialogue, are human friendly in a way that ML and DL based techniques are not. People communicate with each other to share knowledge using language, they argue in order to persuade others, they present arguments to justify their positions, they engage in dialogues to explore alternative positions, and they do these things frequently throughout their lives. Argumentative dialogue is thus a primary interface between people, and possesses attributes that are required for the interface between people and intelligent machines. Approaches to computational argumentation includes techniques for reasoning defeasibly about knowledge, and for managing various types of dialogue between arguers.
We propose that the explanatory role within A.I. should be taken up by the application of computational argument. Speci cally that argumentative dialogue has the potential not merely to be a human computer interface but to be a humane interface between people and intelligent machines. Argumentative dialogue is a unifying technology that can enable intelligent machines not only to explain but also to justify their decisions and, when necessary, to select those arguments necessary to persuade others that the decision made was right under the circumstances. In this paper we demonstrate how the construction of such a humane interface draws together computational argument research at the logical, dialectical, procedural, and heuristic layers and applies it to the problem of understanding, explaining, and justifying decision making within intelligent machines based upon di ering formalisms.
In this paper we present an overview of research to bring together increasingly capable ML techniques with increasingly sophisticated human-oriented, dialogue-based interfaces. 2
Dual process models have been used in psychology for many years to account for the multiple ways in which thinking happens. Whilst various authors have delineated the boundary between types of reasoning in a variety of ways, the basic split is between unconscious, automatic reasoning processes as exempli ed by the majority of problems that are solved by the current generation of ML systems and conscious, controlled reasoning processes, the type of slow, rule-based, explicit, language linked logical reasoning that is exempli ed by traditional symbolic AI, targeting domains such as natural language, or strategic planning. This approach accounts partly for the problem of explainable AI. Great advances have been made in solving problems associated with unconscious reasoning processes, but we conjecture that human acceptance of the results is often dependent upon conscious reasoning, an explanation of the reasoning in terms that people would usually use to explain their own behaviour to others. Furthermore, people often don't accept the base explanations for another's behaviour and require a defense of the behaviour at issue. Such verbal interactions, where the interlocutors may choose the right thing to say in order to persuade their opponent, are inherently strategic, and thus explainable AI systems, when acting in human society, will often need to go beyond mere explanation, to justify their behaviour given the context in which it occurs. To di erentiate this speci c subset of dialogical interactions, we refer to such Justi catory and Explanatory Techniques for Dialogues as \JET Dialogues".
Clearly a link between unconscious, ML-based, AI systems, and conscious
reasoning processes to support JET dialogues is required. We propose that
argumentation schemes [
General Premise: B is generally indicated as true when its sign, A, is true. Conclusion: B is true in this situation.
Schemes capture the idea that whilst the expression of a given instance of
reasoning is often unique to the situation in which it occurs, the framework
of expression, how the data (premises) are marshalled and linked to the result
(conclusion), frequently ts into a nite range of possibilities. Furthermore, some
instances of fallacious reasoning can often be equated with poor application
of a scheme. Schemes have been catalogued yielding collections of hundreds of
individual schemes[
Input
ML Process Premises
Reasoning Pattern
ML System
Output
Conclusion Argumentation Scheme
Key Data Flow Argumentative Relationship Correspondence Boundary
To this end we are aligning existing, identi ed argumentation schemes with
published ML research, using the OpenML repository [
There are a number of threads of further work which roughly correspond to various sub-systems illustrated in Fig. 2. Firstly, mapping from ML representations to knowledge structures. Secondly, natural language generation. Thirdly, strategic and contextual personalisation of dialogue game interactions. This captures the notion that whilst an explanation will be fairly static, a factual statement of the reasoning process, a justi cation, which ideally persuades the other party, but at least is acceptable to them, may di er depending upon the circumstances
ML Process Reasoning Pattern
Dialogue Transcript
Input updates
ML Agent Dialogue
Participant(s) in which the dialogue takes place, and the nature of the other party. Persuasive arguments are very susceptible to the disposition, knowledge, and circumstances of the person to whom they are directed, choosing the right argument for the right person, strategy, is, as a result, extremely important in successful dialogical interaction. As these research threads are drawn together, so a humane interface to the underlying ML system is built enabling more rich, human-oriented interactions to occur.