WordNet for “ Easy ” Textual Inferences

This paper presents a WordNet-based automatic approach for calculating “easy” inferences. We build a rule-based system which extracts the pairs of the SICK corpus whose sentences only differ by zero or one word and then identifies which inference relation (i.e. entailment, contradiction, neutrality) exists between these words, based on WordNet relations. Since the sentences of those pairs only differ by the words of the comparison, the inference relation found between the words is taken to apply to the whole sentences of the pair. For some cases not dealt by WordNet we use our own heuristics to label the inference type. With this approach we accomplish three goals: a) we manage to correct the annotations of a part of the SICK corpus and provide the corrected corpus, b) we evaluate the coverage and relation-completeness of WordNet and provide taxonomies of its strengths and weaknesses and c) we observe that “easy” inferences are a suitable evaluation technique for lexical resources and suggest that more such methods are used in the task. The outcome of our work can help improve the SICK corpus and the WordNet resource and it also introduces a new way of dealing with lexical resources evaluation tasks.