An Improved Coarse-to-Fine Method for Solving Generation Tasks

An Improved Coarse-to-Fine Method for Solving Generation Tasks Wenyu Guan1,2, Qianying Liu3, Guangzhi Han2, Bin Wang4 and Sujian Li1 1 Key Laboratory of Computational Linguistics, MOE, Peking University 2 School of Software and Microelectronics, Peking University 3 Graduate School of Informatics, Kyoto University 4 Xiaomi AI Lab, Xiaomi Inc. {guanwy, hanguangzhi10, lisujian}@pku.edu.cn ying@nlp.ist.i.kyoto-u.ac.jp, wangbin11@xiaomi.com Abstract The coarse-to-fine (coarse2fine) methods have recently been widely used in the generation tasks. The methods first generate a rough sketch in the coarse stage and then use the sketch to get the final result in the fine stage. However, they usually lack the correction abil- ity when getting a wrong sketch. To solve this problem, in this paper, we propose an improved coarse2fine model with a control mechanism, with which our method can con- trol the influence of the sketch on the final re- sults in the fine stage. Even if the sketch is wrong, our model still has the opportunity to get a correct result. We have experimented our model on the tasks of semantic parsing and math word problem solving. The results have shown the effectiveness of our proposed model. 1 Introduction The coarse-to-fine (coarse2fine) methods have been applied in many generation tasks such as machine translation (Xia et al., 2017) , abstract writing (Wang et al., 2018b) and semantic pars- ing (Dong and Lapata, 2018). They have shown excellent performances but still have many disad- vantages. Traditional coarse2fine models usually tackle one task in two stages. In the first stage (coarse stage), a low-level seq2seq model is used to generate a rough sketch, which makes the data more compact and alleviates the problem of data sparsity. Some examples of sketches are shown in Table 1. Besides, Sketches in this stage are also easier to generate. Then, in the fine stage, both text and previous sketches will be input to another high-level seq2seq model to predict the final result so that the high-level model can produce a precise output. In the coarse2fine models, the concept of tem- plate sketch provides a new view of compiling a rough template, but how to guarantee its quality is still a problem. Meanwhile, details from the fine stage will be filled into sketches to produce the final result, so the quality of sketches serves an essential influence on the result. If the generated sketches are in high quality, the coarse2fine model performs well. Otherwise, we fail to get an excel- lent output. The main reason is that the sketch is misleading and has no possibility to be corrected once it is wrong. In this paper, we propose an improved coarse2fine model to solve this problem. It has a similar framework which consists of two levels of seq2seq models. First, the model predicts a rough sketch in the coarse stage. In the fine stage, com- pared with traditional coarse2fine model, we use the generated sketches in the coarse stage as as- sistant information to help the decoder. Besides, We set a weight to control the degree of how …

Content truncated to 500 words