Model-based Answer Selection

Obtaining informative answer passages and ranking them appropriately has previously been error prone for complex, non-factoid questions related to action and event occurrences, causes, and spatiotemporal attributes. A fundamental problem that has hampered the efforts to date has been the inability to extract relations of interest that determine the search for relevant answer passages. We report on a model-based approach to answer selection based on the idea that relations relevant to a question are best captured by an expressive model of events. We outline the essential attributes of the model and report on its application to the AQUAINT focused data corpus for Question Answering.