Coherent Comments Generation for Chinese Articles with a Graph-to-Sequence Model

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4843–4852 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 4843 Coherent Comment Generation for Chinese Articles with a Graph-to-Sequence Model Wei Li1, Jingjing Xu1, Yancheng He2, Shengli Yan2, Yunfang Wu1, Xu Sun1,3 1MOE Key Lab of Computational Linguistics, School of EECS, Peking University 2Platform & Content Group, Tencent 3Deep Learning Lab, Beijing Institute of Big Data Research, Peking University {liweitj47, jingjingxu}@pku.edu.cn {collinhe, victoryyan}@tencent.com {wuyf, xusun}@pku.edu.cn Abstract Automatic article commenting is helpful in en- couraging user engagement and interaction on online news platforms. However, the news documents are usually too long for traditional encoder-decoder based models, which often results in general and irrelevant comments. In this paper, we propose to generate comments with a graph-to-sequence model that models the input news as a topic interaction graph. By organizing the article into graph structure, our model can better understand the internal structure of the article and the connection be- tween topics, which makes it better able to understand the story. We collect and release a large scale news-comment corpus from a popular Chinese online news platform Tencent Kuaibao.1 Extensive experiment results show that our model can generate much more coher- ent and informative comments compared with several strong baseline models.2 1 Introduction Online news platform is now a popular way for people to get information, where users also make comments or read comments made by others, making the comments very valuable resource to attract user attention and encourage interactions among users (Park et al., 2016). The ability to automatically generate comments is desirable for online news platforms, especially comments that can encourage user engagement and interactions, serving as one form of intelligent chatbot (Shum et al., 2018). Important as the comment generation task is, it is still relatively new. Qin et al. (2018) proposed the problem of automatic article com- ment generation, which is to generate comments 1https://kuaibao.qq.com/ 2Code for the paper is available at https://github.com/lancopku/ Graph-to-seq-comment-generation given the title and content of the article (An ex- ample is shown in Table 1). They only proposed the task, but did not propose a specially designed solution to the problem other than sequence-to- sequence paradigm (Sutskever et al., 2014). Ma et al. (2018) proposed a retrieval based model that uses variational topic model to find comments that are related to the news in an unsupervised fashion. Lin et al. (2018) proposed to refer to the retrieved comments during generation, which is a combina- tion of retrieval and generation based model. Pure generation based model remains challenging, yet is a more direct way to solve the problem. Addi- tionally, when the article is very different from the historical ones, there may not be appropriate com- ments to refer to. In this work, we would like to explore a generation model that better exploits the news content to solve the problem. Different from the scenarios where sequence- to-sequence models achieve great success like ma- chine translation (Bahdanau …

Content truncated to 500 words