DialoGraph: Incorporating Interpretable Strategy-Graph Networks into Negotiation Dialogues

Abstract

To successfully negotiate a deal, it is not enough to communicate fluently: pragmatic planning of persuasive negotiation strategies is essential. While modern dialogue agents excel at generating fluent sentences, they still lack pragmatic grounding and cannot reason strategically. We present DIALOGRAPH, a negotiation system that incorporates pragmatic strategies in a negotiation dialogue using graph neural networks. DIALOGRAPH explicitly incorporates dependencies between sequences of strategies to enable improved and interpretable prediction of next optimal strategies, given the dialogue context. Our graph-based method outperforms prior state-of-the-art negotiation models both in the accuracy of strategy/dialogue act prediction and in the quality of downstream dialogue response generation. We qualitatively show further benefits of learned strategy-graphs in providing explicit associations between effective negotiation strategies over the course of the dialogue, leading to interpretable and strategic dialogues. 1 1 Code, data and a demo system is released at https://github.com/rishabhjoshi/ DialoGraph_ICLR21

Sentiment Aided Graph Attentive Contextualization for Task Oriented Negotiation Dialogue Generation

“…Although this method clearly incorporates pragmatic tactics, it does not take advantage of the expressive power of neural networks. In contrast, Joshi et al (2021) tried to capture the interplay between the negotiation strategies in successive turns using Graph Attention Networks(GAT), and finally used a different decoder module to generate responses. In addition to the possibility of error propagation, they haven't taken any step to make sure the agent would provide a perfect counteroffer during negotiation.…”

“…However, rather than composing dialogue strategies on-the-fly, they have been used to encode external knowledge (Tuan et al, 2019;Zhou et al, 2018) or speaker information (Ghosal et al, 2019). The first person to combine GATs with hierarchical pooling was Joshi et al (2021), who also used a conversation system to acquire practical dialogue techniques. Their model can be plugged into other models as an explicit sequence encoder, unlike earlier works.…”

“…Nego-GAT is our proposed work. We compare Nego-GAT's performance to that of Dialo-Graph (Joshi et al, 2021), the most comprehensive model for the negotiating task now available, which also employs GAT to store dialogue actions and strategy sequences. The FST-enhanced hierarchical encoder-decoder model (FeHED) (Zhou et al, 2019b), which was the previous state-ofthe-art model, was also used to compare our agent's performance.…”

“…The majority of earlier research on negotiating dialogues really concentrated on improving conversation techniques, ranging from high-level task-specific strategies (Lewis et al, 2017) to more precise task execution planning (He et al, 2018), to fine-grained planning of language outputs given strategic options (Zhou et al, 2019a). Joshi et al (2021) proposed a framework for modeling intricate negotiation strategies that makes use of intermediate structures and Graph Attention Networks (GAT) to make the model understandable. Their model makes use of the recently proposed hierarchical graph pooling-based methodology to comprehend the relationships between negotiating techniques, such as conceptual and verbal strategies and conversation activities, and their relative value in selecting the best sequence.…”

“…In keeping with the concept put forward by Joshi et al (2021) and Yang et al (2021), we present our model Nego-GAT, which integrates GAT (Graph attention) embeddings into the large pre-trained model GPT-2 (Radford et al, 2019) to enhance contextualization while producing negotiation answers. On the other hand, in the case of negotiation dialogues, sentiment plays an essential role while generating responses.…”