A Hybrid System to apply Natural Language Inference over Dependency Trees

In the last decade, there has been a surge of interest in the problem of textual inference, which systems can use to automatically determine whether a hypothesis, H , can be inferred from a given text, T . A variety of approaches have been explored ranging from shallow-but-robust to deep-but-brittle. Systems that have tried to avoid semantic representations have applied shallow techniques on natural language snippets, such as measuring lexical overlaps (Jijkoun and Rijke 2005), extracting pattern-based relations (Romano et al. 2006), or applying approximate matching to predicateargument structure (Hickl et al. 2006). Although these methods are robust and effective, they are not suitable for problems that require multi-steps of inference, such as in the FraCaS multi-premise problems that follow, because the task that they are designed to address simply does not involve applying sequences of rules: