Robust textual inference via learning and abductive reasoning

摘要: We present a system for textual inference (the task of inferring whether sentence follows from another text) that uses learning and logical-formula semantic representation the text. More precisely, our begins by parsing then transforming sentences into logical formula-like similar to one used (Harabagiu et al., 2000). An abductive theorem prover tries find minimum "cost" set assumptions necessary show statement other. These costs reflect how likely different are, are learned automatically using information syntactic/semantic features linguistic resources such as WordNet. If other given only highly plausible, low cost assumptions, we conclude it can be inferred. Our approach viewed combining statistical machine classical reasoning, in hope marrying robustness scalability with preciseness elegance proving. give experimental results recent PASCAL RTE 2005 challenge competition on recognizing inferences, where this algorithm achieved highest confidence weighted score.