Generating long macro reports from a piece of breaking news is quite a challenging task. Essentially, this task is a long text generation problem from short text. Apparently, the difficulty of this task lies in the logic inference of human beings. To address this issue, this paper proposes a novel hybrid deep generative neural model which first learns the outline of the input news and then generates macro financial reports from the learnt outline. In the outline generation component, we generate the outline text using the framework of Pointer-Generator network with attention mechanism. In the target report generation component, we generate the macro financial reports by the revised VAE model. To train our end-to-end model, we have collected the experimental dataset containing over one hundred thousand pairs of news-report data. Extensive experiments are then evaluated on this dataset. The proposed model achieves the SOTA performance against both the baseline models and the state-of-the-art models with respect to evaluation criteria BLEU, ROUGE and human scores. Although the readability of the generated reports by our approach is better than that of the rest models, it remains an open problem which needs further efforts in the future.

从一条突发新闻生成长篇宏报告是一项非常具有挑战性的任务。本质上，这个任务是一个从短文本生成长文本的问题。显然，这项任务的难点在于人类的逻辑推理。为了解决这个问题，本文提出了一种新的混合深度生成神经模型，该模型首先学习输入新闻的轮廓，然后从学习到的轮廓生成宏观财务报告。在轮廓生成组件中，我们使用具有注意力机制的Pointer-Generator网络框架生成轮廓文本。在目标报告生成组件中，我们通过修改后的 VAE 模型生成宏观财务报告。为了训练我们的端到端模型，我们收集了包含超过十万对新闻报道数据的实验数据集。然后在这个数据集上评估了广泛的实验。所提出的模型在评估标准 BLEU、ROUGE 和人类分数方面，相对于基线模型和最先进的模型都实现了 SOTA 性能。尽管我们的方法生成的报告的可读性优于其他模型，但它仍然是一个悬而未决的问题，需要在未来进一步努力。