Mutual Information Guided Financial Report Generation with Domain Adaption

The conditional long text generation is an important yet challenging task especially for some domain-specific applications such as financial report generation. Generally, this emerging financial report generation is to generate long and informative text provided by a piece of short news, where the domain knowledge obviously plays an important role. Apparently, such domain knowledge dynamically varies but could be learned from reports written by human specialists to address multiple viewpoints contained in the original news. To address this issue, this article proposes the mutual information guided financial report generation approach with domain adaption. Particularly, we first extract the underlying domain spaces from the input news as well as their corresponding reports, respectively. Then, these domains are aligned with each other in a mutual manner via the proposed domain adaption component. To further enhance the model performance, the mutual information loss is designed to guarantee the semantic meaning of the generated report is close to that of the input news. Extensive experiments are conducted on three public datasets and the proposed approach achieves the state-of-the-art model performance.