Automatic Answer Validation using COGEX

This paper reports the performance of Language Computer Corporation's natural language logic prover for the English and Spanish subtasks of the Answer Validation Exercise. Cogex takes as input a pair of plain English text snippets, transforms them into semantically rich logic forms, automatically generates natural language axioms which will be used during the search for a proof, and determines the degree of entailment between the entailing text snippet and a possible relaxed version of the entailed text. The system labels an input pair as positive entailment if its proof score is above the threshold learned during training. Our semantic logic-based approach achieves 43.93% F-measure for the English data and 60.63% on Spanish.