UTD HLTRI at TREC 2017: Complex Answer Retrieval Track

This paper presents our Complex Answer PAragraph Retrieval (CAPAR) system designed for our participation in the TREC Complex Answer Retrieval (CAR) track. Because we were provided with a massive training set consisting of complex questions as well as the paragraphs that answered each aspect of the complex question, we cast the paragraph ranking as a learning to rank (L2R) problem, such that we can produce optimal results at testing time. We considered two alternative Learning to Rank (L2R) approaches for obtaining the relevance scores of each paragraph: (1) the Siamese Attention Network (SANet) for Pairwise Ranking and (2) AdaRank. The evaluation results obtained for CAPAR revealed that the Siamese Attention Network (SANet) for Pairwise Ranking outperformed AdaRank as the L2R approach for CAPAR.