Abstract: A computer-implemented method for learning unknown concepts during natural language processing is disclosed, including identifying a sentence associated with an unknown concept, selecting a first sequential set of sentences from a first document, including the sentence associated with the unknown concept, one sentence prior, and subsequent to the sentence associated with the unknown concept, selecting a second sequential set of sentences from a second document, including a sentence associated with a known concept, and one sentence prior and subsequent to the sentence associated with the known concept, comparing concepts associated with the first sequential set of sentences and second sequential set of sentences, determining whether an inference can be made between the unknown concept associated with the sentence from the first document and the sentence associated with the known concept associated with the sentence from the second document, and tagging the unknown concept associated with the known concept.