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.
Introduction
Narrative reasoning—the ability to predict, explain, or infer the sequence of events in a story—is a fundamental component of human cognition. For a machine to understand a simple narrative like "Tom was tired and wanted to have fun. He bought a movie ticket for Harry Potter," it must possess "world knowledge" that connects these disparate facts. It needs to know that being tired might lead to seeking entertainment, and that buying a ticket is a prerequisite for watching a movie.
Historically, researchers have approached this problem in two ways. Implicit modeling involves training large language models (LLMs) on vast corpora, hoping they pick up these patterns as "parametric memory." While effective, these models are often "black boxes" that lack explicit logical structures. Explicit modeling involves building structured Knowledge Graphs (KGs) like ASER or ATOMIC, which map out millions of relations between eventualities. However, "grounding" (linking) messy, free-form natural language to these rigid structures is notoriously difficult. This paper introduces EventGround, a framework designed to bridge this gap by grounding narratives into eventuality-centric KGs to provide machines with interpretable background knowledge.
The EventGround Framework
The authors identify two primary hurdles in the grounding process: event representation (events in text are often entangled with specific context and pronouns) and sparsity (a specific sentence like "Tom bought a ticket for Harry Potter" is unlikely to exist verbatim in a KG). EventGround addresses these through a multi-stage pipeline.
Extraction and Normalization
The first step involves breaking down complex sentences into standardized event units. The system uses Semantic Role Labeling (SRL) to extract (verb, argument) structures. To maintain narrative consistency across sentences, it employs coreference resolution, replacing specific names and pronouns with standardized placeholders (e.g., "Tom" and "he" both become [P0]). This process, known as semantic normalization, ensures that the model recognizes different mentions of the same actor as the same entity.
Partial Information Extraction (PIE)
To solve the sparsity problem, the authors propose Partial Information Extraction (PIE), a "controllable abstraction" method. Humans naturally abstract details to find commonalities; we know that "evacuating to a house" is a specific instance of "evacuating." If a specific event isn't found in the KG, PIE iteratively drops less critical arguments—such as modifiers, locations, or secondary objects—to find a more general "anchor event" that exists in the graph.
Subgraph Retrieval and Reasoning
Once events are matched to KG nodes using semantic similarity (SBERT), the system retrieves the shortest paths between these nodes in the KG. This creates a "reasoning subgraph" containing relevant background knowledge. This subgraph is then fed into a reasoning module. The authors evaluated two versions: a GNN-based model using Relational Graph Convolutional Networks and an LLM-based model that prompts ChatGPT with a serialized version of the graph.
Experimental Evaluation
The authors evaluated EventGround on three major narrative reasoning benchmarks: the Story Cloze Test (SCT) v1.0 & v1.5, which involves predicting the correct ending of a story, and the Multiple Choice Narrative Chain (MCNC), a task to predict the next event in a sequence. They used ASER, a massive KG containing millions of events and discourse relations (e.g., Precedence, Reason, Condition), as their knowledge source.
As shown in the results, EventGround consistently outperformed strong baselines. When combined with a DeBERTa-large backbone, EventGround achieved 98.29% accuracy on SCT-v1.0 and 66.05% on MCNC.
Key Insights
- Abstraction is Critical: Ablation studies showed that removing the PIE module caused a significant drop in performance. This confirms that handling event sparsity through abstraction is the "make or break" factor for effective KG grounding.
- Semantic Matching vs. Token Matching: Traditional methods often fail by relying on exact word matches. By using SBERT, EventGround can link "purchased a ticket" to "bought a ticket," significantly increasing the grounding success rate.
- Interpretability: Unlike vanilla LMs, EventGround provides a "paper trail." The retrieved subgraphs show exactly which pieces of world knowledge (e.g.,
[P0] study-> Reason ->[P0] pass the test) led to the model's prediction, offering a level of transparency missing from black-box models.
Conclusion
EventGround represents a significant step forward in narrative reasoning. By grounding unstructured text into structured eventuality KGs, it provides a framework that is both high-performing and interpretable. While currently focused on personal entities, future work could extend normalization to locations and objects. Additionally, exploring more complex retrieval methods, such as Reinforcement Learning, may further enhance the relevance of the retrieved background knowledge.
Summary updated