Sampling Matters! An Empirical Study of Negative Sampling Strategies for Learning of Matching Models in Retrieval-based Dialogue Systems

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1291–1296, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1291 Sampling Matters! An Empirical Study of Negative Sampling Strategies for Learning of Matching Models in Retrieval-based Dialogue Systems Jia Li1, Chongyang Tao1, Wei Wu2, Yansong Feng1, Dongyan Zhao1,3 and Rui Yan1,3∗ 1Wangxuan Institute of Computer Technology, Peking University, Beijing, China 2Microsoft Corporation, Beijing, China 3Center for Data Science, Peking University, Beijing, China 1,3{lijiaa,chongyangtao,fengyansong,zhaody,ruiyan}@pku.edu.cn 2{wuwei}@microsoft.com Abstract We study how to sample negative examples to automatically construct a training set for ef- fective model learning in retrieval-based dia- logue systems. Following an idea of dynami- cally adapting negative examples to matching models in learning, we consider four strategies including minimum sampling, maximum sam- pling, semi-hard sampling, and decay-hard sampling. Empirical studies on two bench- marks with three matching models indicate that compared with the widely used random sampling strategy, although the first two strate- gies lead to performance drop, the latter two ones can bring consistent improvement to the performance of all the models on both bench- marks. 1 Introduction In this work, we study the problem of response se- lection as an approach to implementing a retrieval- based dialogue system (Ji et al., 2014; Wang et al., 2013). A key step in response selection is measur- ing the matching degree between a conversation context and a response candidate. Existing studies focus on constructing a matching model with so- phisticated neural architectures (Lowe et al., 2015; Zhou et al., 2016; Yan et al., 2016; Wu et al., 2017; Zhang et al., 2018; Zhou et al., 2018; Tao et al., 2019), but pay little attention to how to effectively learn such architectures from data. On the one hand, it is well known that learning of complicated neural architectures requires large-scale high qual- ity training data; on the other hand, since human labeling is expensive and exhausting, most of the existing work just adopts a simple heuristic to au- tomatically build a training set where human re- sponses are treated as positive examples and neg- ative response candidates are randomly sampled. ∗Corresponding author: Rui Yan (ruiyan@pku.edu.cn). Such a training set might contain many false neg- atives and trivial true negatives that are very easy to distinguish from those true positives. As a re- sult, models with advanced architectures can only reach sub-optimal performance after learning (Wu et al., 2018). In this paper, instead of configuring new archi- tectures, we investigate how to improve the per- formance of existing matching models with a bet- ter learning method. A learning method usually involves choice of loss functions and construction of training data, and we are particularly interested in automatic training data construction, as data are often more crucial to the performance of mod- els. The key problem in training data construc- tion lies in how to properly choose negative exam- ples, and our idea is that negative examples should adapt …

Content truncated to 500 words