Initiative-Aware Self-Supervised Learning for Knowledge-Grounded Conversations

Abstract

In the knowledge-grounded conversation (KGC) task systems aim to produce more informative responses by leveraging external knowledge. KGC includes a vital part, knowledge selection, where conversational agents select the appropriate knowledge to be incorporated in the next response. Mixed initiative is an intrinsic feature of conversations where the user and the system can both take the initiative in suggesting new conversational directions. Knowledge selection can be driven by the user's initiative or by the system's initiative. For the former, the system usually selects knowledge according to the current user utterance that contains new topics or questions posed by the user; for the latter, the system usually selects knowledge according to the previously selected knowledge. No previous study has considered the mixed-initiative characteristics of knowledge selection to improve its performance.In this paper, we propose a mixed-initiative knowledge selection method (MIKe) for KGC, which explicitly distinguishes between user-initiative and system-initiative knowledge selection. Specifically, we introduce two knowledge selectors to model both of them separately, and design a novel initiative discriminator to discriminate the initiative type of knowledge selection at each conversational turn. A challenge for training MIKe is that we usually have no labels for indicating initiative. To tackle this challenge, we devise an initiative-aware self-supervised learning scheme that helps MIKe to learn to discriminate the initiative type via a self-supervised task. Experimental results on two datasets show that MIKe significantly outperforms state-of-the-art methods in terms of both automatic and human evaluations, indicating that it can select more appropriate knowledge and generate more informative and engaging responses.

Enhanced Knowledge Selection for Grounded Dialogues via Document Semantic Graphs

“…For our human evaluation, we randomly sample 200 turns from the output of MIKe (Meng et al, 2021), our ranking model and our graphbased model. Annotators are asked to select which system's response is the best among the three (allowing for ties), and which system's knowledge is the most relevant.…”

Construction of Hierarchical Structured Knowledge-based Recommendation Dialogue Dataset and Dialogue System

“…We encode the knowledge type history via the transformer encoder (Vaswani et al, 2017). This transformer encoder (we call this "knowledge type encoder") adds a positional embedding for each turn (= turn embedding) to the input so that the model reflects in which turn each knowledge type was used (Meng et al, 2021). We concatenate the last output of this encoder h k l−1 t trans with the hidden state of the dialogue context h x l y l−1 as the query, and regard {h kt,m } M m=1 as the key.…”

“…We use copy mechanism (Gu et al, 2016;See et al, 2017) to make it easier to generate knowledge words and follow the method used in Meng et al (2021).…”

Teach Me What to Say and I Will Learn What to Pick: Unsupervised Knowledge Selection Through Response Generation with Pretrained Generative Models

“…It uses a BERT encoder. MIKe (Meng et al, 2021): Further improves KS by explicitly distinguishing between user-initiative and system-initiative knowledge selection. It uses a BERT encoder.…”

“…Attempts to improve on these benchmarks can mostly be divided into two categories, based on their point of focus. Selection oriented methods focus on enhancing the KS task, usually by introducing additional learning signals like the priorposterior discrepancy (Lian et al, 2019; or long-term structural traits of conversations like flow and initiative changes (Kim et al, 2020;Zhan et al, 2021;Meng et al, 2021;Zheng et al, 2020). Generation oriented methods on the other hand, try to mitigate the selection bottleneck by employing more powerful methods to incorporate knowledge in the generation process, thus reformulating the KS problem as an adaptive fine-grained selection to be dealt with in decoding (Zheng and Zhou, 2019).…”