Anaphora Resolution for Off-line Answer Extraction using Instances

We investigate to what extent it helps to use an automatically generated instance list for resolving anaphoric definite NPs within the framework of an off-line answer extraction system. We compare three extraction methods. First, we looked at the extraction method that uses no anaphora resolution. Secondly, we determined the precision of the facts acquired by using instances for anaphora resolution. And thirdly, as a baseline we extracted facts by selecting the most recent named entity as the antecedent of the definite NP. From our experiments we can conclude that anaphora resolution using instances yields a high precision. Overall this precision is only 1% lower than the precision of the original answer extraction method without applying anaphora resolution. We outperformed the baseline of which the precision was not even half as high as the other two extraction methods.