Step-by-Step: Separating Planning from Realization in Neural Data-to-Text Generation

Proceedings of NAACL-HLT 2019, pages 2267–2277 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2267 Step-by-Step: Separating Planning from Realization in Neural Data-to-Text Generation∗ Amit Moryossef† Yoav Goldberg†‡ Ido Dagan† amitmoryossef@gmail.com, {yogo,dagan}@cs.biu.ac.il †Bar Ilan University, Ramat Gan, Israel ‡Allen Institute for Artificial Intelligence Abstract Data-to-text generation can be conceptually divided into two parts: ordering and struc- turing the information (planning), and gener- ating fluent language describing the informa- tion (realization). Modern neural generation systems conflate these two steps into a sin- gle end-to-end differentiable system. We pro- pose to split the generation process into a sym- bolic text-planning stage that is faithful to the input, followed by a neural generation stage that focuses only on realization. For training a plan-to-text generator, we present a method for matching reference texts to their correspond- ing text plans. For inference time, we de- scribe a method for selecting high-quality text plans for new inputs. We implement and evalu- ate our approach on the WebNLG benchmark. Our results demonstrate that decoupling text planning from neural realization indeed im- proves the system’s reliability and adequacy while maintaining fluent output. We observe improvements both in BLEU scores and in manual evaluations. Another benefit of our ap- proach is the ability to output diverse realiza- tions of the same input, paving the way to ex- plicit control over the generated text structure. 1 Introduction Consider the task of data-to-text generation, as ex- emplified in the WebNLG corpus (Colin et al., 2016). The system is given a set of RDF triplets describing facts (entities and relations between them) and has to produce a fluent text that is faith- ful to the facts. An example of such triplets is: John, birthPlace, London John, employer, IBM With a possible output: ∗This research was supported in part by the German Re- search Foundation through the German-Israeli Project Coop- eration (DIP, grant DA 1600/1-1) and by a grant from Theo Hoffenberg and Reverso. 1. John, who was born in London, works for IBM. Other outputs are also possible: 2. John, who works for IBM, was born in London. 3. London is the birthplace of John, who works for IBM. 4. IBM employs John, who was born in London. These variations result from different ways of structuring the information: choosing which fact to mention first, and in which direction to express each fact. Another choice is to split the text into two different sentences, e.g., 5. John works for IBM. John was born in London. Overall, the choice of fact ordering, entity order- ing, and sentence splits for these facts give rise to 12 different structures, each of them putting the fo- cus on somewhat different aspect of the informa- tion. Realistic inputs include more than two facts, greatly increasing the number of possibilities. Another axis of variation is in how to verbalize the information for a given structure. For example, (2) can also be verbalized as 2a. John works for IBM and was born in London. and …

Content truncated to 500 words