EMIL: Extracting Meaning from Inconsistent Language

There are well-developed formal and computational theories of argumentation to reason in the face of inconsistency, some with implementations; there are recent efforts to extract arguments from large textual corpora. Both developments are leading towards automated processing and reasoning with inconsistent, linguistically expressed knowledge in order to provide explanations and justifications in a form accessible to humans. However, there remains a gap between the knowledge-bases of computational theories of argumentation, which are generally coarse-grained and semi-structured (e.g. propositional logic), and inferences from knowledge-bases derived from natural language, which are fine-grained and highly structured (e.g. predicate logic). Arguments that occur in textual corpora are very rich, highly various, and incompletely understood. We identify several subproblems which must be addressed in order to bridge the gap, requiring the development of a computational foundation for argumentation coupled with natural language processing. For the computational foundation, we provide a direct semantics, a formal approach for argumentation, which is implemented and suitable to represent and reason with an associated natural language expression for defeasibility. It has attractive properties with respect to expressivity and complexity; we can reason by cases; we can structure higher level argumentation components such as cases and debates. With the implementation, we output experimental results which emphasise the importance of our efficient approach. To motivate our formal approach, we identify a range of issues found in other approaches. For the natural language processing, we adopt and adapt an existing controlled natural language (CNL) to interface with our computational theory of argumentation; the tool takes natural language input and automatically outputs expressions suitable for automated inference engines. A CNL, as a constrained fragment of natural language, helps to control variables, highlights key problems, and provides a framework to engineer solutions. The key adaptation incorporates the expression ‘it is usual that’, which is a plausibly ‘natural’ linguistic expression of defeasibility. This is an important, albeit incremental, step towards the incorporation of linguistic expressions of defeasibility; yet, by engineering such specific solutions, a range of other, relevant issues arise to be addressed. Overall, we can input arguments expressed in a controlled natural language, translate them to a formal knowledge base, represent the knowledge in a rule language, reason with the rules, generate argument extensions, and finally convert the arguments in the extensions into natural language. Our approach makes for fine-grained, highly structure, accessible, and linguistically represented argumentation evaluation. The overall novel contribution of the paper is an integrated, end-to-end argumentation system which bridges a gap between automated defeasible reasoning and a natural language interface. The component novel contributions are the computational theory of ‘direct semantics’, the motivation for our theory, the results with respect to the direct semantics, the implementation, the experimental results, the tie between the formalisation and the CNL, the adaptation of a CNL defeasibility, and an ‘engineering’ approach to fine-grained argument analysis.