Table-to-Text Generation with Accurate Content Copying
Abstract Table-to-text generation is an important task in natural language generation that aims to generate smooth, informative text based on structured data. In this paper, we propose a novel transformer-based autoregressive model that incorporates table content copying and language model based generation. At first, we propose a word transformation method to process a target text. By using target text containing fields and position information, we can help the model learn the relationship between target text and table and gain the position of where to copy. We then propose two auxiliary learning goals: table-text constraint loss and copy loss. Table-text constraint loss is introduced to effectively model table inputs, whereas copy loss is exploited to precisely copy word fragments from a table. In addition, we change the maximization-based text search strategy to reduce the probability of problems such as sentence repetition and inconsistency. On the WIKIBIO dataset, our model improves its BLUE scores from 45.47 to 46.87 and ROUGE scores from 41.54 to 42.28, outperforming state-of-the-art baseline models on automatic evaluation metrics. On the ROTOWIRE test set, compared with the best baseline model, our model gets 4.29% higher on CO metric, and 1.93 points higher on BLEU.