Graph-Based Answer Passage Ranking for Question Answering

Passage retrieval of Question Answering (QA) systems aims to find the text segments or sentences that may contain the exact answers for the given question. Previous studies on passage retrieval are mostly utilized a single function to calculate the relevance scores of passages. However, some research has proved that the relations between passages can be utilized to improve the accuracy of relevance evaluation. Hence, a passage retrieval method based on passage-passage graph model is proposed. A KNN-based question expansion method is proposed and then the candidate answer passages are retrieved based on the expanded question model. The passage graph is constructed based on the similarities between the candidate answer passages. Finally, a graph-based ranking model is utilized to re-calculate the relevance scores of the answer passages and the ranking parameter is trained using the learning method. Experiment results show that our method can significantly increase the MRR and TRDR performances compared to the baseline methods.