Learning to mine all minimal evidences for unverified claims

Mining evidences is crucial in checking unverified claims. Most existing methods usually find a single evidence expressed by a set of sentences to verify a given claim. However, treating a set of sentences as unique evidence is insufficient or even misleading, e.g., when it involves both supporting and refuting information. Besides, gathering evidence from different perspectives helps us better analyze and understand the claim. In this article, we suggest mining all irreducible evidences for supporting or refuting a claim, where we treat a minimal set of sentences in the given text corpus for either the support or the refutation as an irreducible point of view and call it a minimal evidence. We develop a neural-symbolic framework to mine all minimal evidences. It exploits a logical algorithm to compute all minimal evidences one by one, where every minimal evidence is computed through two neural models scorer and reasoner learnt from the annotated minimal evidences. Experimental results demonstrate that our framework is effective in finding multiple minimal evidences for both textual and structural claims. Furthermore, we investigate several implementation combinations for scorers and reasoners so as to seek the best practice on our framework.