Towards a Probabilistic Model for Lexical Entailment

While modeling entailment at the lexical-level is a prominent task, addressed by most textual entailment systems, it has been approached mostly by heuristic methods, neglecting some of its important aspects. We present a probabilistic approach for this task which covers aspects such as differentiating various resources by their reliability levels, considering the length of the entailed sentence, the number of its covered terms and the existence of multiple evidence for the entailment of a term. The impact of our model components is validated by evaluations, which also show that its performance is in line with the best published entailment systems.