From the textual description of an accident to its causes

摘要

Every human being, reading a short report concerning a road accident, gets an idea of its causes. The work reported here attempts to enable a computer to do the same, i.e. to determine the causes of an event from a textual description of it. It relies heavily on the notion of norm for two reasons:•The notion of cause has often been debated but remains poorly understood: we postulate that what people tend to take as the cause of an abnormal event, like an accident, is the fact that a specific norm has been violated.•Natural Language Processing has given a prominent place to deduction, and for what concerns Semantics, to truth-based inference. However, norm-based inference is a much more powerful technique to get the conclusions that human readers derive from a text. The paper describes a complete chain of treatments, from the text to the determination of the cause. The focus is set on what is called “linguistic” and “semantico-pragmatic” reasoning. The former extracts so-called “semantic literals” from the result of the parse, and the latter reduces the description of the accident to a small number of “kernel literals” which are sufficient to determine its cause. Both of them use a non-monotonic reasoning system, viz. LPARSE and SMODELS.Several issues concerning the representation of modalities and time are discussed and illustrated by examples taken from a corpus of reports obtained from an insurance company.