Query Enhanced Knowledge-Intensive Conversation via Unsupervised Joint Modeling

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1730–1745 July 9-14, 2023 ©2023 Association for Computational Linguistics Query Enhanced Knowledge-Intensive Conversation via Unsupervised Joint Modeling Mingzhu Cai Siqi Bao Xin Tian Huang He Fan Wang Hua Wu Baidu Inc., China {caimingzhu, baosiqi, tianxin06, hehuang, wang.fan, wu_hua}@baidu.com Abstract In this paper, we propose an unsupervised query enhanced approach for knowledge- intensive conversations, namely QKConv. There are three modules in QKConv: a query generator, an off-the-shelf knowledge selector, and a response generator. QKConv is opti- mized through joint training, which produces the response by exploring multiple candidate queries and leveraging corresponding selected knowledge. The joint training solely relies on the dialogue context and target response, get- ting exempt from extra query annotations or knowledge provenances. To evaluate the effec- tiveness of the proposed QKConv, we conduct experiments on three representative knowledge- intensive conversation datasets: conversational question-answering, task-oriented dialogue, and knowledge-grounded conversation. Experi- mental results reveal that QKConv performs better than all unsupervised methods across three datasets and achieves competitive perfor- mance compared to supervised methods. 1 Introduction In addition to open-domain chitchat, there exist various knowledge-intensive conversations, such as conversational question-answering, task-oriented dialogue, and knowledge-grounded conversation. Although large-scale language models can implic- itly store common knowledge within parameters (Petroni et al., 2019; Zhao et al., 2020b), they are known to suffer from producing plausible state- ments with factual errors (a.k.a. knowledge halluci- nation) (Roller et al., 2021; Marcus, 2020). There- fore, there is a trend to rely on external resources, such as Wikipedia databases or search engine re- sults, to facilitate knowledge-intensive conversa- tions (Dinan et al., 2019; Komeili et al., 2022). In knowledge-intensive conversations, the most straightforward way to retrieve external knowledge is to take the dialogue context as the query and use an off-the-shelf retriever to return the knowledge entry. However, it encounters some difficulties in retrieving appropriate knowledge (Shuster et al., 2021). As the focus or topic changes along with the conversation flow, the outdated information in the dialogue context brings extra noise to the re- triever, resulting in obsolete or irrelevant knowl- edge retrieved. Moreover, the dialogue context has a native misalignment with the short and interroga- tive query preferred in existing retrievers. Some methods choose to finetune a task-specific retriever to enhance the performance of knowledge selection (Guu et al., 2020; Shuster et al., 2021; Glass et al., 2022). However, this strategy is usu- ally computationally expensive (e.g., finetuning a dense retriever requires constant recomputation for massive knowledge entries) or even infeasible for complex retrieval systems (e.g., retraining a search engine is impractical). Some other methods choose to generate a self-contained query based on the dia- logue context (Yu et al., 2020; Anantha et al., 2021; Chen et al., 2022). This strategy relies on careful query annotations to guarantee the completeness of essential information extraction and the adaptation to the knowledge selector. In this paper, we introduce a novel unsupervised query enhanced approach for knowledge-intensive conversations, namely QKConv. …

Content truncated to 500 words