Enhancing Neural Data-To-Text Generation Models with External Background Knowledge

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3022–3032, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 3022 Enhancing Neural Data-To-Text Generation Models with External Background Knowledge Shuang Chen1∗, Jinpeng Wang2, Xiaocheng Feng1, Feng Jiang1,3, Bing Qin1, Chin-Yew Lin2 1 Harbin Institute of Technology, Harbin, China 2 Microsoft Research Asia 3 Peng Cheng Laboratory hitercs@gmail.com, {jinpwa, cyl}@microsoft.com, {xcfeng, qinb}@ir.hit.edu.cn, fjiang@hit.edu.cn Abstract Recent neural models for data-to-text genera- tion rely on massive parallel pairs of data and text to learn the writing knowledge. They of- ten assume that writing knowledge can be ac- quired from the training data alone. However, when people are writing, they not only rely on the data but also consider related knowl- edge. In this paper, we enhance neural data-to- text models with external knowledge in a sim- ple but effective way to improve the fidelity of generated text. Besides relying on parallel data and text as in previous work, our model attends to relevant external knowledge, en- coded as a temporary memory, and combines this knowledge with the context representa- tion of data before generating words. This al- lows the model to infer relevant facts which are not explicitly stated in the data table from an external knowledge source. Experimen- tal results on twenty-one Wikipedia infobox- to-text datasets show our model, KBAtt, con- sistently improves a state-of-the-art model on most of the datasets. In addition, to quantify when and why external knowledge is effective, we design a metric, KBGain, which shows a strong correlation with the observed perfor- mance boost. This result demonstrates the rel- evance of external knowledge and sparseness of original data are the main factors affecting system performance. 1 Introduction Automatic text generation from structured data (data-to-text) is a classic task in natural language generation which aims to automatically gener- ate fluent, truthful and informative texts based on structured data (Kukich, 1983; Holmes-Higgin, 1994; Reiter and Dale, 1997). Data-to-text is of- ten formulated into two subproblems: content se- lection which decides what contents should be in- cluded in the text and surface realization which ∗Contribution during internship at Microsoft Research. Subject Relation Object Guelma (Q609871) country (P17) Algeria (Q262) defender (Q336286) instance of (P31) association football position (Q4611891) … … … MC El Eulma (Q2742749) league (P118) Algerian Ligue Professionnelle 1 (Q647746) External Background Knowledge from Wikidata Nacer Hammami (born December 28, 1980) is an Algerian football player who is currently playing for MC El Eulma in the Algerian Ligue Professionnelle 1. Description Infobox Entity Linking Figure 1: An example of generating description from a Wikipedia infobox. External background knowledge expanded from the infobox is helpful for generation. determines how to generate the text based on se- lected contents. Traditionally, these two sub- problems have been tackled separately. In recent years, neural generation models, especially the encoder-decoder model, solve these two subprob- lems jointly and have achieved remarkable suc- cesses in several benchmarks (Mei et …

Content truncated to 500 words