Stick to the Facts: Learning towards a Fidelity-oriented E-Commerce Product Description Generation

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4959–4968, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4959 Stick to Facts: Towards Fidelity-oriented Product Description Generation Zhangming Chan1,2,∗,†, Xiuying Chen1,2,∗, Yongliang Wang3, Juntao Li1,2, Zhiqiang Zhang3, Kun Gai3, Dongyan Zhao1,2 and Rui Yan1,2,‡ 1 Center for Data Science, AAIS, Peking University 2 Wangxuan Institute of Computer Technology, Peking University 3 Alibaba Group, Beijing {zhangming.chan,xy-chen,ljt,zhaody,ruiyan}@pku.edu.cn {yongliang.wyl,zhang.zhiqiang,jingshi.gk}@alibaba-inc.com Abstract Different from other text generation tasks, in product description generation, it is of vital importance to generate faithful descriptions that stick to the product attribute informa- tion. However, little attention has been paid to this problem. To bridge this gap we pro- pose a model named Fidelity-oriented Product Description Generator (FPDG). FPDG takes the entity label of each word into account, since the product attribute information is al- ways conveyed by entity words. Specifically, we first propose a Recurrent Neural Network (RNN) decoder based on the Entity-label- guided Long Short-Term Memory (ELSTM) cell, taking both the embedding and the en- tity label of each word as input. Second, we establish a keyword memory that stores the entity labels as keys and keywords as values, and FPDG will attend to keywords through attending to their entity labels. Experiments conducted a large-scale real-world product de- scription dataset show that our model achieves the state-of-the-art performance in terms of both traditional generation metrics as well as human evaluations. Specifically, FPDG in- creases the fidelity of the generated descrip- tions by 25%. 1 Introduction The effectiveness of automatic text generation has been proved in various natural language process- ing applications, such as neural machine transla- tion (Luong et al., 2015; Zhou et al., 2017), di- alogue generation (Tao et al., 2018a; Hu et al., 2019), and abstractive text summarization (Chen et al.; Gao et al., 2019). One such task, product description generation has also attracted consid- erable attention (Lipton et al., 2015; Wang et al., ∗Equal contribution. Ordering is decided by a coin flip. † Contribution during internship at Alibaba Group. ‡ Corresponding author. Input high-waist; straight; jeans; blue; ZARA Bad output This UNIQLO low-waist jeans are suitable for the curvature of the waistline and will not slip when worn for a long time. The straight version and black color together makes you even slimmer. Good output This blue jeans look personalized and chic, making you more distinctive. With a high-waist design, the legs look even more slender, and the whole person looks taller. The straight design hides the fat in the legs and is visually slimmer. Table 1: Example of product description generation. The text with underline demonstrates faithful descrip- tion, and text with deleteline demonstrates wrong de- scription. 2017; Zhang et al., 2019a). An accurate and attrac- tive description not only helps customers make an informed decision but also improves the likelihood of purchase. Concretely, the product description generation task takes several keywords describing the attributes of a …

Content truncated to 500 words