基于细粒度主题建模的数据到文本生成模型

DATA-TO-TEXT GENERATION MODEL BASED ON FINE-GRAINED TOPIC MODELING

数据到文本生成是自然语言生成领域中的一个重要任务。端到端的神经网络结构的性能仅依赖于模型自身的表示学习能力,而忽略了文本和数据之间的主题一致性,进而导致生成文本的质量不高。为建模文本和数据间的主题关联,提出一种基于细粒度主题建模的数据到文本生成模型。基于两个真实数据集上的实验结果表明,该模型取得了最优的性能效果。

Data-to-text generation is a critical task in natural language generation field.The end-to-end neural network architectures generally rely on their representation learning ability to improve the performance of generation,and ignore the topic consistency between the generated text and the input structure data,which leads the poor text quality.Therefore,this paper proposes a fine-grained topic modeling based data-to-text generation model to learn the topic relevance between text and structured data.Experimental results on two real datasets show that the proposed model achieves the best performance.