AMIF: A Hybrid Model for Improving Fact Checking in Product Question Answering

Fact checking in product-related community question answering is the task of verifying the truthfulness of an answer towards a given question, where the study has just begun. Most existing related work has focused on tailoring solutions to shallow feature fusion for the single-text claim involved with fact-checked evidence, limiting their success and generality in such answer truthfulness prediction task on E-commerce platforms. In this study, we propose an attention-based hybrid framework for multi-feature interaction fusion to determine the truthfulness of the answer towards a product-related question in E-commerce, which could not only support fine-grained semantic calibration between question-answer pairs for better understanding of the target answers, but also substantially cross-check all retrieved evidence to mine coherent opinions towards the pair. In addition, our framework further integrates non-textual features from metadata for improving performance. Extensive experiments conducted on real-world representative benchmark data show that our proposed model achieves superior performance on the task of answer veracity prediction.