A Logic Aware Neural Generation Method for Explainable Data-to-text

The most notable neural data-to-text approaches generate natural language from structural data relying on the surface form of the structural content, which ignores the underlying logical correlation between the input data and the target text. Moreover, identifying such logical associations and explaining them in natural language is desirable but not yet studied. In this paper, we introduce a practical data-to-text method for the logic-critical scenario, specifically for anti-money laundering applications. It involves detecting risks from input data and explaining any abnormal behaviors in natural language. The proposed method is a Logic Aware Neural Generation framework (LANG), which is a preliminary attempt to explore the integration of logic modeling and text generation. Concretely, we first convert expert rules to a logic graph. Then, the model utilizes meta path based encoder to exploit the expert knowledge. Besides, a retriever module with the encoded logic knowledge is used to bridge the gap between numeric input and target text. Finally, a rule-constrained loss is leveraged to improve the generation probability of tokens in rule recalled statements to ensure accuracy. We conduct extensive experiments on anti-money laundering data. Results show that the proposed method significantly outperforms baselines in both objective measures with relative 35% improvements in F1 score and subjective measures with 30% improvement in human preference.