There Is No Standard Answer: Knowledge-Grounded Dialogue Generation with Adversarial Activated Multi-Reference Learning

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1878 - 1891 December 7-11, 2022 ©2022 Association for Computational Linguistics There Is No Standard Answer: Knowledge-Grounded Dialogue Generation with Adversarial Activated Multi-Reference Learning Xueliang Zhao1†, Tingchen Fu2†, Chongyang Tao3, Rui Yan2∗ 1Wangxuan Institute of Computer Technology, Peking University 2Gaoling School of Artificial Intelligence, Renmin University of China 3Microsoft Corporation {zhaoxlpku,lucas.futingchen,chongyangtao}@gmail.com ruiyan@ruc.edu.cn Abstract Knowledge-grounded conversation (KGC) shows excellent potential to deliver an engag- ing and informative response. However, exist- ing approaches emphasize selecting one golden knowledge given a particular dialogue context, overlooking the one-to-many phenomenon in dialogue. As a result, the existing paradigm limits the diversity of knowledge selection and generation. To this end, we establish a multi- reference KGC dataset and propose a series of metrics to systematically assess the one-to- many efficacy of existing KGC models. Fur- thermore, to extend the hypothesis space of knowledge selection to enhance the mapping relationship between multiple knowledge and multiple responses, we devise a span-based variational model and optimize the model in a wake-sleep style with an ameliorated evi- dence lower bound objective to learn the one- to-many generalization. Both automatic and human evaluations demonstrate the efficacy of our approach. 1 Introduction Maintaining appropriate human-computer dialogue is an important task leaping toward advanced artifi- cial intelligence and external knowledge is a key in- gredient to engaging and meaningful responses (Di- nan et al., 2019). To this end, the research area of knowledge-grounded conversation (KGC) has been explored with great interest. In recent years, a number of methods (Lian et al., 2019; Kim et al., 2020; Zhao et al., 2020a,b) and benchmarks (Di- nan et al., 2019; Zhou et al., 2018) have been pro- posed. These methods mainly follow the two-step paradigm proposed by Dinan et al. (2019): Given a dialogue context and a candidate knowledge pool, †The first two authors contribute equally. Xueliang Zhao is responsible for the design of the methodology and algo- rithm. Tingchen Fu is responsible for the implementation and experiment. The order is decided by a coin flip. *Corresponding author: Rui Yan (ruiyan@ruc.edu.cn). There is no blue pigmentation in blue eyes. The eyes appear blue as a result of rayleigh scattering, the same process that makes the sky appear blue. Response R1: Daddy why is the sky blue? Same reason your eyes are, son . R2: It sounds really stupid to me to say ' the sky isn't actually lllllblue, but dyed by rayleigh scattering of sunlinght or something ||||like that. I mean, the sky is blue. It's blue, guys. R3: I even know someone whose two eyes in different colors, they lllllcall it heterochromia. Do they have no pigmentation in eyes as lllllwell? R4: Is it affected by generic and inheritable from parents? K1: This would be analogous to the change in the color of the sky , llllllfrom the blue given by the rayleigh scattering of sunlight by llllllsmall gas molecules when the sky is clear, to the gray caused by llllllmie scattering …

Content truncated to 500 words