DiffuCOMET: Contextual Commonsense Knowledge Diffusion

TL;DR

This work develops a series of knowledge models that leverage diffusion to learn to reconstruct the implicit semantic connections between narrative contexts and relevant commonsense knowledge, and introduces new metrics for commonsense inference that more closely measure knowledge diversity and contextual relevance.

Abstract

Inferring contextually-relevant and diverse commonsense to understand narratives remains challenging for knowledge models. In this work, we develop a series of knowledge models, DiffuCOMET, that leverage diffusion to learn to reconstruct the implicit semantic connections between narrative contexts and relevant commonsense knowledge. Across multiple diffusion steps, our method progressively refines a representation of commonsense facts that is anchored to a narrative, producing contextually-relevant and diverse commonsense inferences for an input context. To evaluate DiffuCOMET, we introduce new metrics for commonsense inference that more closely measure knowledge diversity and contextual relevance. Our results on two different benchmarks, ComFact and WebNLG+, show that knowledge generated by DiffuCOMET achieves a better trade-off between commonsense diversity, contextual relevance and alignment to known gold references, compared to baseline knowledge models.

Figures (8)

• Figure 1: Overview of our diffusion-based contextual commonsense knowledge generation.
• Figure 2: Knowledge diffusion based on facts or entities. Dashed arrows denote the forward process used for constructing gold references at the training phase. Solid arrows denote the reverse process used for generating knowledge with attention to the narrative context.
• Figure 3: Illustration of clustering-based evaluation metrics for contextual commonsense knowledge generation.
• Figure 4: DiffuCOMET performance at different diffusion steps during inference. Both DiffuCOMET-Fact and DiffuCOMET-Entity are developed based on BART-large and tested on the ROCStories portion of $\mathcal{C}{om}\mathcal{F}{act}$. Beam-Comet performance is shown as a baseline, with the number of decoded facts set to match DiffuCOMET-Entity at each diffusion step.
• Figure 5: Range selection (red square) of DBSCN clustering thresholds for our proposed metrics.