Morphological Inflection: A Reality Check

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 562–574 July 9-14, 2023 ©2023 Association for Computational Linguistics Evaluating Open-Domain Dialogues in Latent Space with Next Sentence Prediction and Mutual Information Kun Zhao1∗, Bohao Yang2∗, Chenghua Lin2†, Wenge Rong3,4, Aline Villavicencio2, Xiaohui Cui1 † 1 Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, School of Cyber Science and Engineering, Wuhan University, China 2 Department of Computer Science, The University of Sheffield, United Kingdom 3 State Key Laboratory of Software Development Environment, Beihang University, China 4 School of Computer Science and Engineering, Beihang University, China {zhaokun, xcui}@whu.edu.cn, w.rong@buaa.edu.cn {byang27, c.lin, a.villavicencio}@sheffield.ac.uk Abstract The long-standing one-to-many issue of the open-domain dialogues poses significant chal- lenges for automatic evaluation methods, i.e., there may be multiple suitable responses which differ in semantics for a given conversational context. To tackle this challenge, we propose a novel learning-based automatic evaluation metric (CMN), which can robustly evaluate open-domain dialogues by augmenting Condi- tional Variational Autoencoders (CVAEs) with a Next Sentence Prediction (NSP) objective and employing Mutual Information (MI) to model the semantic similarity of text in the latent space. Experimental results on two open- domain dialogue datasets demonstrate the su- periority of our method compared with a wide range of baselines, especially in handling re- sponses which are distant to the golden refer- ence responses in semantics. 1 Introduction Open-domain dialogue generation is a prominent research direction in conversational AI due to a wide range of useful applications that it can facili- tate, such as for personal digital assistants and cus- tomer service (Sai et al., 2020; Huang et al., 2020; Wang et al., 2021; Tang et al., 2023). While evalu- ating Natural Language Generation (NLG) systems is notoriously difficult, evaluation of open-domain dialogue generation introduces an extra layer of complexity, as a variety of responses can be gener- ated, each semantically different and yet valid in the given context (Li et al., 2016; Gu et al., 2019; Qiu et al., 2019). For example, given the conver- sational context “Iverson is my all-time favourite player.”, responses such as “He is my favourite player too.” or “Yes, his quickness is amazing!” are both contextually relevant, yet semantically differ- ent. ∗ Equal contribution. † Corresponding authors. Existing approaches for evaluating open-domain dialogue systems can be broadly divided into two different categories: reference-based and reference- free approaches. The reference-based metrics typ- ically score a system by computing how simi- lar an output response is compared to the gold- standard reference. Popular metrics under this category may rely on surface-form similarity by counting the n-gram overlap between the response candidate and the reference (e.g., BLEU (Pap- ineni et al., 2002), ROUGE (Lin, 2004), and ME- TEOR (Banerjee and Lavie, 2005)), or by calculat- ing the similarity based on embedding representa- tions such as Embedding-Average (Wieting et al., 2016), or even via high-dimensional representa- tions learned for the response and the reference such as BERTScore (Zhang et al., 2020). One …

Content truncated to 500 words