Automatic extraction of causal knowledge from natural language texts

This thesis studies the automatic recognition of implicit causal relations between clauses. Previous work, although more accurate than a random baseline, does not achieve sufficient accuracy for practical use. First we show, using annotation experiments, that recognising implicit causation is a subjective task, in spite of its association with several observable features. We then propose an evaluation protocol that takes the subjectivity of the task into account. Previous linguistics work uses the notion of world knowledge. We show that the most likely feature for representing this knowledge -verb pairs- is not predictive of causation in practice. We then show that the current state of the art is not sufficient to allow us to represent nor to acquire the world knowledge that is necessary for this task. We conclude that the field cannot make important progress without first solving the problem of abstract eventuality representation and clustering.