Abstract:
While information extraction and textual entailment systems have greatly enhanced the amount of knowledge available from a text corpus, the next generation of natural language understanding systems -- such as Machine Reading systems -- will need to employ dedicated mechanisms to ensure that acquired knowledge is consistent with the previous commitments encoded in the system's knowledge base. This paper introduces a new technique for knowledge acquisition for Machine Reading systems, known as Reading by Entailment (RbE), which uses two complementary systems for recognizing entailment relationships in order to identify the set of hypotheses that can be inferred from a text corpus. Entailment hypotheses are generated by a random walk model that operates on texts and informs a plausible model of knowledge entailment. Knowledge entailment is cast as a maximal knowledge relevance problem which operates on knowledge mappings.