Answer Retrieval From Extracted Tables

Question answering (QA) on table data, which contains densely packed information in two-dimensional form, is a challenging information retrieval task. Data can be placed at a distance from the metadata describing it. The metadata itself can be difficult to identify given the layout of a particular table. This paper describes a QA system for tables created with both machine learning and heuristic table extraction methods. Our approach creates a cell document for each table cell. A probabilistic language model selects the most likely cell documents for the information need. The performance of the system is tested with government statistical data, and errors are analyzed in order to improve the system. We also apply these improvements on another type of table data set and show the experimental results.