|
Abstract : |
r.gaizauskas,k.humphreys We describe an approach to finding literal answer strings to natural language questions in large text collections. The approach involves linking an IR system with an NLP system that performs reasonably thorough linguistic analysis. The IR system treats the question as a query and returns a set of top ranked documents or passages. The NLP system parses the question and analyses the top ranked documents or passages returned by the IR system, yielding a ?meaning representation ? of each. It then instantiates a privileged query variable in the semantic representation of the question against the semantic representation of the analysed documents or passages to discover the answer, using a general purpose coreference mechanism. The approach has been evaluated in the TREC-8 question and answer (Q & A) track evaluation. While initial overall success is limited, it is sufficient to warrant further investigation of the approach. In particular this work will shed light on the interesting question that the TREC Q & A task poses: to what extent are ?deeper ? models of language processing necessary to perform question answering against large text collections?, |
