Why did the accident happen? A norm-based reasoning approach

To understand well a text written in natural language (NL), we need our knowledge about the norms of its domain. By the word “norm”, we mean here the normal and expected course of events in the absence of exceptions [6]. This type of knowledge enables us to infer richer conclusions than those given by means of truthpreserving entailments, for example, from the text: “Mon véhicule se trouvait arrêté à un stop, quand un véhicule m'a heurté à l'arrière”, (My vehicle was stopped at a stop sign, when a vehicle struck me at the back). Norms provide conclusions like: vehicle A and me were in the same file and direction, vehicle A had to stop to avoid the shock... None of these conclusions is explicit. However, any reader infers them immediately. Conclusions obtained by using norms can in general be defeasible, but they are accepted as long as the text does not contradict them. Often, narrative texts do not describe norms explicitly. They focus rather on their violations, by describing generally abnormal situations. In the light of this main remark, our goal consists in looking for the cause of an accident from its textual description by hypothesizing that the searched cause (called the primary anomaly) is the violation of the most specific norm in the text [3]. The other violations of norms result from the first one and are called derived anomalies. We are working on a corpus of 60 car crash reports written in French. Each report is a small text describing briefly the circumstances of an accident. To validate our approach, the reasoning system must find for each text the same answer given by an ordinary human reader to the question: “what is the most specific violated norm which can considered as the plausible cause of the accident ?”. These answers are determined manually for each text at the beginning of the process.