Answer-guided and Semantic Coherent Question Generation in Open-domain Conversation

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5066–5076, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5066 Answer-guided and Semantic Coherent Question Generation in Open-domain Conversation Weichao Wang , Shi Feng , Daling Wang , Yifei Zhang School of Computer Science and Engineering, Northeastern University, Shenyang, China wangwecha@gmail.com {fengshi,wangdaling,zhangyifei}@cse.neu.edu.cn Abstract Generating intriguing question is a key step to- wards building human-like open-domain chat- bots. Although some recent works have fo- cused on this task, compared with questions raised by humans, significant gaps remain in maintaining semantic coherence with the post- s, which may result in generating dull or de- viated questions. We observe that the answer has strong semantic coherence to its question and post, which can be used to guide ques- tion generation. Thus, we devise two meth- ods to further enhance semantic coherence be- tween post and question under the guidance of answer. First, the coherence score between generated question and answer is used as the reward function in a reinforcement learning framework, to encourage the cases that are consistent with the answer in semantic. Sec- ond, we incorporate adversarial training to ex- plicitly control question generation in the di- rection of question-answer coherence. The ex- tensive experiments show that our two meth- ods outperform state-of-the-art baseline algo- rithms with large margins in raising semantic coherent questions. 1 Introduction Neural question generation (NQG) has been ex- tensively studied because of its broad application in question answering (QA) systems (Tang et al., 2018; Duan et al., 2017), reading comprehension (Kim et al., 2019; Sun et al., 2018), and visu- al question answering (Fan et al., 2018). How- ever, the research of open-domain conversation- al question generation (CQG) (Wang et al., 2018; Hu et al., 2018) is still in its infancy. Differen- t from traditional NQG task, the goal of CQG is to enhance the interactiveness and persistence of chit-chatting. Raising appropriate questions in the conversational system is essential because a ma- ture system should well interact with users by not only responding posts but also asking questions (Li et al., 2016b). Furthermore, asking intrigu- ing questions can proactively probe the users to go deeper and further into the conversational top- ic (Yu et al., 2016). The existing CQG methods usually suffer from dullness and deviation problem, because the se- mantic coherence is easily disrupted. Wang et al. (2018) first studied the CQG task with PMI model to predict topic words in the inference, which may incorporate noise and generate deviat- ed questions. Hu et al. (2018) proposed an aspect- based question generation model with an encoder- decoder method that did not consider the diversity and might raise dull questions. Taking Table 1 as a motivating example, the top- ic of this conversation is about cooking and dishes. For the initial post about enjoying cooking, the ap- propriate response is a question asking about the special dishes or making appetizers, and …

Content truncated to 500 words