EventGround: Narrative Reasoning by Grounding to Eventuality-centric Knowledge Graphs

TL;DR

EventGround tackles the challenge of narrative reasoning by grounding free-text narratives to eventuality-centric knowledge graphs. It introduces a three-stage pipeline: (1) extract, normalize, and abstract events via semantic role labeling, pronoun replacement, and partial information extraction; (2) align partial events to an eventuality-centric KG (ASER-core-100) to form anchor-directed subgraphs; (3) perform reasoning over the grounded subgraphs using either graph neural networks or large language models, with results showing state-of-the-art performance on SCT-v1.0, SCT-v1.5, and MCNC. Key contributions include a practical grounding framework that addresses event representation and sparsity, an effective grounding strategy with semantic matching and constrained subgraph retrieval, and demonstrable interpretability through the grounded evidence. The approach offers a scalable, interpretable path toward knowledge-enhanced narrative understanding with potential for broader applications in story analysis and commonsense reasoning.

Abstract

Narrative reasoning relies on the understanding of eventualities in story contexts, which requires a wealth of background world knowledge. To help machines leverage such knowledge, existing solutions can be categorized into two groups. Some focus on implicitly modeling eventuality knowledge by pretraining language models (LMs) with eventuality-aware objectives. However, this approach breaks down knowledge structures and lacks interpretability. Others explicitly collect world knowledge of eventualities into structured eventuality-centric knowledge graphs (KGs). However, existing research on leveraging these knowledge sources for free-texts is limited. In this work, we propose an initial comprehensive framework called EventGround, which aims to tackle the problem of grounding free-texts to eventuality-centric KGs for contextualized narrative reasoning. We identify two critical problems in this direction: the event representation and sparsity problems. We provide simple yet effective parsing and partial information extraction methods to tackle these problems. Experimental results demonstrate that our approach consistently outperforms baseline models when combined with graph neural network (GNN) or large language model (LLM) based graph reasoning models. Our framework, incorporating grounded knowledge, achieves state-of-the-art performance while providing interpretable evidence.

Figures (13)

• Figure 1: Given a piece of story, our goal is to ground it to eventuality-centric KGs to retrieve contextualized background world knowledge for better narrative understanding.
• Figure 2: An overview of EventGround.
• Figure 3: A comparison on the event grounding performance under different settings. The bar plot (with $y$-axis on the left) shows the percentage hit rate of event matching. The lines show the average L2 distance $\bar{d}$. We do not conduct normalization for "w/o extract.".
• Figure 4: The F1-score to threshold curves. They reflect the event matching performance under different threshold $l$.
• Figure 5: An example from SCT-v1.0. The top-10 node attention weights are shown in the barplot. The top-3 nodes are bolded and underlined .