Exploring Shallow Answer Ranking Features in Cross-Lingual and Monolingual Factoid Question Answering

Answer ranking is critical to a QA (Question Answering) system because it determines the final system performance. In this paper, we explore the behavior of shallow ranking features under different conditions. The features are easy to implement and are also suitable when complex NLP techniques or resources are not available for monolingual or cross-lingual tasks. We analyze six shallow ranking features, namely, SCO-QAT, keyword overlap, density, IR score, mutual information score, and answer frequency. SCO-QAT (Sum of Co-occurrence of Question and Answer Terms) is a new feature proposed by us that performed well in NTCIR CLQA. It is a co-occurrence based feature that does not need extra knowledge, word-ignoring heuristic rules, or special tools. Instead, for the whole corpus, SCO-QAT calculates co-occurrence scores based solely on the passage retrieval results. Our experiments show that there is no perfect shallow ranking feature for every condition. SCO-QAT performs the best in C-C (Chinese-Chinese) QA, but it is not a good choice in E-C (English-Chinese) QA. Overall, Frequency is the best choice for E-C QA, but its performance is impaired when translation noise is present. We also found that passage depth has little impact on shallow ranking features, and that a proper answer filter with fined-grained answer types is important for E-C QA. We measured the performance of answer ranking in terms of a newly proposed metric EAA (Expected Answer Accuracy) to cope with cases of answers that have the same score after ranking.