A Prospective-Performance Network to Alleviate Myopia in Beam Search for Response Generation

Proceedings of the 27th International Conference on Computational Linguistics, pages 3608–3618 Santa Fe, New Mexico, USA, August 20-26, 2018. 3608 A Prospective-Performance Network to Alleviate Myopia in Beam Search for Response Generation Zongsheng Wang1, Yunzhi Bai2∗, Bowen Wu1, Zhen Xu3∗, Zhuoran Wang1,4, Baoxun Wang1 1Tricorn (Beijing) Technology Co., Ltd, Beijing, China 2Telecom ParisTech, Paris, France 3Harbin Institute of Technology, Harbin, China 4Institute of Internet Industry, Tsinghua University, Beijing, China 1 {wangzongsheng, wubowen, wangzhuoran, wangbaoxun}@trio.ai 2 yunzhi bai@outlook.com 3 zxu@insun.hit.edu.cn Abstract Generative dialog models usually adopt beam search as the inference method to generate re- sponses. However, small-width beam search only focuses on the limited current optima. This deficiency called myopic bias ultimately suppresses the diversity and probability of generated responses. Although increasing the beam width mitigates the myopic bias, it also proportionally slows down the inference . To alleviate the myopic bias in small-width beam search, this paper proposes a Prospective-Performance Network (PPN) to predict the future reward of the given partially-generated response, and the future reward is defined by the expectation of the partial response appearing in the top-ranked responses given by a larger-width beam search. Enhanced by PPN, the decoder can promote the results with great potential during the beam search phase. The experimental results on both Chinese and English corpora show that our method is capa- ble of increasing the quality and diversity of generated responses, with inference efficiency well maintained. 1 Introduction In recent years, Neural Response Generation (NRG) (Vinyals and Le, 2015; Shang et al., 2015) with sequence-to-sequence (Seq2Seq) structures (Sutskever et al., 2014a; Bahdanau et al., 2015) has been widely studied and adopted in open-domain dialog systems such as XiaoIce (Shum et al., 2018). Many studies have been done to generate target responses meeting desirable proprieties including emo- tion (Zhou et al., 2017), diversity (Li et al., 2016), etc., or to explore issues in decoding step such as exposure bias, loss-evaluation mismatch (Ranzato et al., 2016) and label bias (Wiseman and Rush, 2016). Most NRG systems adopt the beam search algorithm to generate responses given queries. In brief, beam search explores the possible responses by storing only top-ranked ones as candidates at each time step. Though it is a useful prediction strategy, depending on beam width, beam search more or less suffers from its nature of focusing solely on current optimal results. Consequently, beam search tends to ignore some partial sequences which might lead to better future outcomes, especially if its beam width is too small. This deficiency is called the myopic bias (He et al., 2017). Recently, He et al. (2017) and Li et al. (2017) proposed methods that take account of future BLEU of partial sequences as the future reward during beam search, to alleviate the myopic bias in Neural Machine Translation (NMT). Their experiments show that such methods are capable of improving the BLEU scores of generated translations. However, several studies point out that BLEU is weakly correlated with human judgments in response generation tasks (Liu et al., 2016; Mou et …

Content truncated to 500 words