Norm Based Causal Reasoning in Textual Corpus

Truth based entailments are not sufficient for a good comprehension of NL. In fact, it can not deduce implicit information necessary to understand a text. On the other hand, norm based entailments are able to reach this goal. Let us consider this text [1][2]: ”the vehicle in front of me braked”. Using a truth based approach; we can obtain all the logical consequences of a formula such as: (∃v, t) V ehicle(v)∧Instant(t)∧In−Front−Of(v, me, t)∧break(v, t). While norms provide further conclusions like: v and me were in the same direction, no vehicle was between v and me, I had to brake when v braked . . . This idea was behind the development of Frames [3] and Scripts [6][5] in the 70’s. But these theories are not formalized enough and their adaptation to new situations is far from being obvious. Actually, no repository of norms is available for a given domain. Moreover, norms are seldom made explicit in texts, because as Schank noticed, texts do not describe the normal course of events but focus rather on the description of abnormal situations. The motivation of the present work is to extract norms by detecting their violations in the texts. We are working on a corpus of 60 texts describing car crashes. For each text, we are searching the cause of the accident as perceived by a standard reader. We hypothesize that the perceived cause of an abnormal event is the violation of a norm (anomaly). Among all the anomalies evoked by a text, one of them is considered as ’primary’. It represents the most plausible cause of the accident. The other anomalies result from the primary one and are called derived anomalies.