[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-seo-307669-105":3,"doc-detail-307669-en":80,"detail-sidebar-cat-1-en-105":97},{"code":4,"msg":5,"data":6},0,"ok",{"site_id":7,"language":8,"slug":9,"title":10,"keywords":11,"description":12,"schema_data":13,"social_meta":73,"head_meta":75,"extra_data":77,"updated_unix":79},105,"en","dual-feedback-knowledge-retrieval-for-task-oriented-dialogue-systems","Dual-Feedback Knowledge Retrieval for Task-Oriented Dialogue Systems","","Efficient knowledge retrieval is essential for end-to-end task-oriented dialogue systems, enabling selection of relevant information to satisfy user requests. Existing methods often combine retrieval and response generation, creating scalability issues on large knowledge bases and offering weak supervision for retriever learning. This work proposes a retriever-generator architecture and trains the retriever using generator feedback as pseudo-labels. A dual-feedback mechanism produces both positive and negative signals from generator outputs, improving performance on three benchmark datasets.",{"@graph":14,"@context":72},[15,34,55],{"@type":16,"itemListElement":17},"BreadcrumbList",[18,23,27,31],{"item":19,"name":20,"@type":21,"position":22},"https://docshare.wps.com","Home","ListItem",1,{"item":24,"name":25,"@type":21,"position":26},"https://docshare.wps.com/template/","Template",2,{"item":28,"name":29,"@type":21,"position":30},"https://docshare.wps.com/template/general/","General",3,{"item":32,"name":10,"@type":21,"position":33},"https://docshare.wps.com/template/dual-feedback-knowledge-retrieval-for-task-oriented-dialogue-systems/307669/",4,{"url":32,"name":10,"@type":35,"image":36,"author":41,"headline":10,"publisher":44,"fileFormat":47,"inLanguage":8,"description":12,"dateModified":48,"datePublished":49,"encodingFormat":47,"isAccessibleForFree":50,"interactionStatistic":51},"DigitalDocument",{"url":37,"@type":38,"width":39,"height":40},"https://docshare.wps.com/thumbnails/dual-feedback-knowledge-retrieval-for-task-oriented-dialogue-systems/307669.png","ImageObject",442,249,{"name":42,"@type":43},"Asher","Person",{"url":19,"name":45,"@type":46},"DocShare","Organization","application/pdf","2026-09-20","2026-09-19",true,{"@type":52,"interactionType":53,"userInteractionCount":26},"InteractionCounter",{"@type":54},"ViewAction",{"@type":56,"mainEntity":57},"FAQPage",[58,64,68],{"name":59,"@type":60,"acceptedAnswer":61},"What problem does the paper address for task-oriented dialogue systems?","Question",{"text":62,"@type":63},"The paper addresses scalability and supervision challenges in knowledge retrieval for task-oriented dialogue, especially when knowledge bases are large. It focuses on training an effective retriever without retriever-specific labels.","Answer",{"name":65,"@type":60,"acceptedAnswer":66},"How does the proposed retriever-generator architecture work?",{"text":67,"@type":63},"The retriever explicitly selects relevant information from the knowledge base, while the generator uses dialogue context plus retrieved information to produce the response. This separates retrieval from generation for better control.",{"name":69,"@type":60,"acceptedAnswer":70},"What is the dual-feedback mechanism?",{"text":71,"@type":63},"The dual-feedback mechanism generates both positive and negative feedback based on generator outputs. This helps the retriever learn which retrieved entities are truly relevant.","https://schema.org",{"og:url":32,"og:type":74,"og:title":10,"og:site_name":45,"og:description":12},"article",{"robots":76,"canonical":32},"index,follow",{"doc_id":78,"site_id":7},307669,1789852898,{"code":4,"msg":81,"data":82},"success",{"doc_id":78,"user_id":83,"nickname":42,"user_avatar":84,"doc_module":22,"category_id":85,"category_name":29,"doc_title":10,"doc_description":12,"doc_content":86,"file_id":87,"file_url":88,"file_type":89,"file_size":90,"view_count":26,"is_deleted":4,"is_public":22,"is_downloadable":22,"audit_status":22,"page_count":91,"language":92,"language_code":8,"site_id":7,"html_lang":8,"table_of_contents":93,"faqs":94,"seo_title":95,"seo_description":12,"update_tm":79,"read_time":96},687197207639,"https://ap-avatar.wpscdn.com/davatar_a8503ba1806abce46bf441b54a3ca4cd",158,"Dual-Feedback Knowledge Retrieval for Task-Oriented Dialogue Systems  \nTianyuan Shi 1 , Liangzhi Li2∗, Zijian Lin 1 , Tao Yang 1 , Xiaojun Quan 1 Qifan Wang3  \n1 School of Computer Science and Engineering, Sun Yat-sen University, China  \n2Meetyou AI Lab, 3Meta AI  \n{shity6, linzj27, [yangt225}@mail2.sysu.edu.cn](yangt225}@mail2.sysu.edu.cn), [quanxj3@mail.sysu.edu.cn](quanxj3@mail.sysu.edu.cn), [liliangzhi@xiaoyouzi.com](liliangzhi@xiaoyouzi.com), [wqfcr@fb.com](wqfcr@fb.com)  \nAbstract  \nEfficient knowledge retrieval plays a pivotal role in ensuring the success of end-to-end taskoriented dialogue systems by facilitating the selection of relevant information necessary to fulfill user requests. However, current approaches generally integrate knowledge retrieval and response generation, which poses scalability challenges when dealing with extensive knowledge bases. Taking inspiration from open-domain question answering, we propose a retrievergenerator architecture that harnesses a retriever to retrieve pertinent knowledge and a generator to generate system responses. Due to the lack of retriever training labels, we propose relying on feedback from the generator as pseudo-labels to train the retriever. To achieve this, we introduce a dual-feedback mechanism that generates both positive and negative feedback based on the output of the generator. Our method demonstrates superior performance in task-oriented dialogue tasks, as evidenced by experimental results on three benchmark datasets. Our code is available at [https://github.com/Stycoo/](https://github.com/Stycoo/)[ ](https://github.com/Stycoo/)Dual-Feedback-TOD.  \n1 Introduction  \nTask-oriented dialogue (TOD) systems (Eric et al., 2020) are designed to fulfill specific tasks, such as hotel bookings, through natural language conversations with users. These systems can be integrated into applications such as chatbots and voice assistants, serving various industries like hospitality, e-commerce, and customer service. To generate informative system responses, TOD systems typically rely on an external knowledge base (KB) to retrieve relevant entity information. While large language models (LLMs) like ChatGPT have demonstrated impressive capabilities in understanding multi-turn dialogues and generating fluent responses, there are still cases where they require access to localized KBs to handle specific tasks. Therefore, knowledge  \n∗ Corresponding authors.  \nRetriever  \nGenerator  \n\n| \u003Cbr>USER | I need a moderately priced Chinese restaurant located in the east of the city. |\n| --- | --- |\n\n\n| Chesteron may be what you're looking for. Would you like to know more about it? | \u003Cbr>SYS |\n| --- | --- |\n\nRetrieved Entities  \nFigure 1: Visualization of erroneous feedback from the generator for retriever training. Entities with darker colors indicate higher relevance scores. The generator mistakenly identifies “Milton” as the most relevant entity, whereas the correct entity should be “Chesterton”.  \nretrieval remains a critical component that necessitates long-term research in dialogue systems.  \nTraditional pipeline approaches in TOD systems involve multiple modules such as dialogue state tracking and dialogue policy learning, which heavily rely on annotated belief states for knowledge retrieval (Lei et al., 2018 ; Yang et al., 2021) . In contrast, end-to-end task-oriented dialogue (E2ETOD) systems aim to generate responses in a single step without the need for intermediate retrieval annotations, thereby highlighting the importance of external knowledge retrieval. Existing E2ETOD systems can be classified into three categories based on their utilization of external knowledge. Firstly, memory networks are employed to store the knowledge, and multi-hop interactions are designed to aggregate relevant information (Madotto et al., 2018 ; Qin et al., 2020 ; Raghu et al., 2021) . Secondly, pre-trained language models encode linearized KB records, which are then used as input for the response genera","cbCaibvohzDnQZ2X","https://ap.wps.com/l/cbCaibvohzDnQZ2X","pdf",487696,15,"English","# Abstract\n# 1 Introduction\n# Task-oriented dialogue systems and knowledge retrieval\n# Related work on knowledge usage in E2ETOD","[{\"question\":\"What problem does the paper address for task-oriented dialogue systems?\",\"answer\":\"The paper addresses scalability and supervision challenges in knowledge retrieval for task-oriented dialogue, especially when knowledge bases are large. It focuses on training an effective retriever without retriever-specific labels.\"},{\"question\":\"How does the proposed retriever-generator architecture work?\",\"answer\":\"The retriever explicitly selects relevant information from the knowledge base, while the generator uses dialogue context plus retrieved information to produce the response. This separates retrieval from generation for better control.\"},{\"question\":\"What is the dual-feedback mechanism?\",\"answer\":\"The dual-feedback mechanism generates both positive and negative feedback based on generator outputs. This helps the retriever learn which retrieved entities are truly relevant.\"}]","Dual-Feedback Knowledge Retrieval for Task-Oriented Dialogue Systems | PDF",5,{"code":4,"msg":81,"data":98},[99,104,109,114,118,123,128,133,137],{"id":100,"doc_module":22,"doc_module_name":25,"category_name":101,"show_sort_weight":102,"slug":103},11,"Presentations",90,"presentations",{"id":105,"doc_module":22,"doc_module_name":25,"category_name":106,"show_sort_weight":107,"slug":108},12,"Resumes",80,"resumes",{"id":110,"doc_module":22,"doc_module_name":25,"category_name":111,"show_sort_weight":112,"slug":113},14,"Invoices",70,"invoices",{"id":91,"doc_module":22,"doc_module_name":25,"category_name":115,"show_sort_weight":116,"slug":117},"Posters",60,"posters",{"id":119,"doc_module":22,"doc_module_name":25,"category_name":120,"show_sort_weight":121,"slug":122},16,"Social Media",50,"social-media",{"id":124,"doc_module":22,"doc_module_name":25,"category_name":125,"show_sort_weight":126,"slug":127},17,"Forms",40,"forms",{"id":129,"doc_module":22,"doc_module_name":25,"category_name":130,"show_sort_weight":131,"slug":132},18,"Letters",30,"letters",{"id":134,"doc_module":22,"doc_module_name":25,"category_name":135,"show_sort_weight":96,"slug":136},21,"Paper Templates","papers-templates",{"id":85,"doc_module":22,"doc_module_name":25,"category_name":29,"show_sort_weight":4,"slug":138},"general-158"]