Make Templates Smarter: A Template Based Data2Text System Powered by Text Stitch Model

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1057–1062 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1057 Make Templates Smarter: A Template Based Data2Text System Powered by Text Stitch Model Bingfeng Luo∗, Zuo Bai, Kunfeng Lai∗, Jianping Shen PingAn Life Insurance, Shenzhen, China {luobingfeng981,baizuo822}@pingan.com.cn {laikunfeng597,shenjianping324}@pingan.com.cn Abstract Neural network (NN) based data2text models achieve state-of-the-art (SOTA) performance in most metrics, but they sometimes drop or modify the information in the input, and it is hard to control the generation contents. More- over, it requires paired training data that are usually expensive to collect. Template-based methods have good fidelity and controllability but require heavy human involvement. We pro- pose a novel template-based data2text system powered by a text stitch model. It ensures fi- delity and controllability by using templates to produce the main contents. In addition, it re- duces human involvement in template design by using a text stitch model to automatically stitch adjacent template units, which is a step that usually requires careful template design and limits template reusability. The text stitch model can be trained in self-supervised fash- ion, which only requires free texts. The ex- periments on a benchmark dataset show that our system outperforms SOTA NN-based sys- tems in fidelity and surpasses template-based systems in diversity and human involvement. 1 Introduction Data2Text takes structured data like key-value pairs as inputs and generates corresponding texts. It has been used in various applications like automatically generating weather reports (Angeli et al., 2010), restaurant descriptions (Duˇsek et al., 2019), etc. Recent works on Data2Text mainly focus on neu- ral network (NN) based models (Liu et al., 2018; Puduppully et al., 2019). Despite their great suc- cess,their fidelity (express all data correctly) and controllability (control the generated contents) are always their main issues. For example, for most ∗The work was done when Bingfeng Luo and Kunfeng Lai were with PingAn Life. Bingfeng Luo is now at ByteDance (Email: luobingfeng@bytedance.com), Kunfeng Lai is now at Huawei (Email: laikunfeng@huawei.com) NN-based generation models, we cannot guarantee that the generated texts do not drop or modify the information in the input. Besides, it is hard to fix errors made by NN-based generation models. Considering these issues, template-based data2text models (Reiter and Dale, 1997; Bouayad- Agha et al., 2011; Dou et al., 2018) are still widely used in real-world applications. As shown in Fig. 1, one of the most commonly used templates is the slot-filling style template (Dou et al., 2018). It does not require any linguistic knowledge. Users just write a normal sentence but leave the changeable parts as slots. While this kind of system can produce faithful and controllable texts in specific domains, writing these templates typically requires a lot of human efforts, especially when we want to have both variety and fluency in the output texts. A main reason for intensive human involvement is the template reusability. It is difficult to reuse sentence-level templates in new tasks that usually have different requirements. Therefore, people usu- ally break long templates …

Content truncated to 500 words