Learning Neural Templates for Recommender Dialogue System

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7821–7833 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 7821 Learning Neural Templates for Recommender Dialogue System Zujie Liang1∗, Huang Hu2∗, Can Xu2, Jian Miao2, Yingying He2, Yining Chen2, Xiubo Geng2, Fan Liang1, Daxin Jiang2† 1Sun Yat-sen University, Guangzhou, China 2Microsoft Corporation, Beijing, China 1{liangzj9@mail2.sysu.edu.cn, isslf@mail.sysu.edu.cn} 2{huahu,caxu,jianm,yingyhe,yinichen,xigeng,djiang}@microsoft.com Abstract Though recent end-to-end neural models have shown the promising progress on Conversa- tional Recommender System (CRS), two key challenges still remain. First, the recom- mended items cannot be always incorporated into the generated replies precisely and appro- priately. Second, only the items mentioned in the training corpus have a chance to be recommended in the conversation. To tackle these challenges, we introduce a novel frame- work called NTRD for recommender dialogue system that decouples the dialogue genera- tion from the item recommendation. NTRD has two key components, i.e., response tem- plate generator and item selector. The for- mer adopts an encoder-decoder model to gen- erate a response template with slot locations tied to target items, while the latter fills in slot locations with the proper items using a sufficient attention mechanism. Our approach combines the strengths of both classical slot filling approaches (that are generally control- lable) and modern neural NLG approaches (that are generally more natural and accurate). Extensive experiments on the benchmark RE- DIAL show our NTRD significantly outper- forms the previous state-of-the-art methods. Besides, our approach has the unique advan- tage to produce novel items that do not ap- pear in the training set of dialogue corpus. The code is available at https://github. com/jokieleung/NTRD. 1 Introduction Building an intelligent dialogue system that can freely converse with human, and fulfill complex tasks like movie recommendation, travel planning and etc, has been one of longest standing goals of natural language processing (NLP) and artifi- cial intelligence (AI). Thanks to the breakthrough ∗Equal contribution. Work performed when Zujie Liang was an intern at Microsoft STCA. †Corresponding author: djiang@microsoft.com. in deep learning, the progress on dialogue system has been greatly advanced and brought into a new frontier over the past few years. Nowadays, we are witnessing the booming of virtual assistants with conversational user interface like Microsoft Cortana, Apple Siri, Amazon Alexa and Google Assistant. The recent large-scale dialogue mod- els such as DialoGPT (Zhang et al., 2020), Meena (Adiwardana et al., 2020) and Blender (Roller et al., 2021), demonstrate the impressive performance in practice. Besides, the social bots such as Xi- aoIce (Shum et al., 2018) and PersonaChat (Zhang et al., 2018a) also exhibit the great potential on the emotional companion to humans. The conversational techniques shed a new light on the search and recommender system, as the users can seek information through interactive dia- logues with the system. Traditional recommender systems often rely on matrix factorization meth- ods (Koren et al., 2009; Rendle, 2010; Wang et al., 2015; He et al., 2017), and suffer from the cold- start problem (Schein et al., 2002; Lika et al., 2014) when …

Content truncated to 500 words