CET2: Modelling Topic Transitions for Coherent and Engaging Knowledge-Grounded Conversations

TL;DR

A Coherent and Engaging Topic Transition (CET2) framework to model topic transitions for selecting knowledge that is coherent to the context of the conversations while providing adequate knowledge diversity for topic development.

Abstract

Knowledge-grounded dialogue systems aim to generate coherent and engaging responses based on the dialogue contexts and selected external knowledge. Previous knowledge selection methods tend to rely too heavily on the dialogue contexts or over-emphasize the new information in the selected knowledge, resulting in the selection of repetitious or incongruous knowledge and further generating repetitive or incoherent responses, as the generation of the response depends on the chosen knowledge. To address these shortcomings, we introduce a Coherent and Engaging Topic Transition (CET2) framework to model topic transitions for selecting knowledge that is coherent to the context of the conversations while providing adequate knowledge diversity for topic development. Our CET2 framework considers multiple factors for knowledge selection, including valid transition logic from dialogue contexts to the following topics and systematic comparisons between available knowledge candidates. Extensive experiments on two public benchmarks demonstrate the superiority and the better generalization ability of CET2 on knowledge selection. This is due to our well-designed transition features and comparative knowledge selection strategy, which are more transferable to conversations about unseen topics. Analysis of fine-grained knowledge selection accuracy also shows that CET2 can better balance topic entailment (contextual coherence) and development (knowledge diversity) in dialogue than existing approaches.

Figures (5)

• Figure 1: Two conversation examples of knowledge selections lacking in diversity (a) or coherence (b) and leading to non-engaging or incoherent conversations. $r_{t-1}$ and $u_t$ are utterances from dialogue history. Knowledge bases are the knowledge candidates to be chose from. $k_j$ was chosen for $r_{t-1}$ and $r_t$ in (a), leading to repetitive responses. $k_m$ and $k_n$ are chosen for $r_{t-1}$ and $r_t$ in (b),
• Figure 2: Mechanisms for selecting knowledge in response considering both dialogue history and knowledge candidates.
• Figure 3: The architecture of CET2. The Sentence Encoder outputs representations for the dialogue context and knowledge candidates. Two special candidates are shown: the $i$-th knowledge $d_i$ (also denoted $\Tilde{d}$), is the gold knowledge for the current turn and the $j$-th knowledge $d_j$ (denoted $\Tilde{d}^{last}$) is the gold knowledge of last turn. The Topic Transition Modeling acquires transition features. The Comparative Knowledge Selection (in blue) outputs the selected knowledge. The Variance Aware Training depicts our training strategy of controlling the knowledge variance between turns with Topic Shift Constraints.
• Figure 4: Knowledge Selection accuracy overturns. The horizontal axis is the turn index, and the vertical axis is knowledge selection accuracy. The knowledge selection accuracy drops as conversations progress longer.
• Figure 5: Generations of different methods, "[$k_i$]" indicates the chosen knowledge. Blue words indicate repetition and red words are related to incoherence. The numbers following blue words point to their repeated parts with the same numbers in a dialogue history. Post i is the dialogue history at turn i. At each turn i, models try to predict the response of GOLD.