Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fine-Grained Modeling of Narrative Context: A Coherence Perspective via Retrospective Questions

Liyan Xu, Jiangnan Li, Mo Yu, Jie Zhou

TL;DR

NarCo introduces a fine-grained narrative-context graph where passages are nodes and backward-facing retrospective questions form edges, capturing explicit coherence signals beyond end-to-end reasoning. It realizes this graph with a two-stage LLM prompting pipeline (question generation followed by self-verification) without human annotations, and validates its utility through three studies on edge efficacy, node augmentation, and broad QA tasks. Across recap, plot retrieval, and long-document QA, NarCo-enhanced representations and edges improve performance over baselines, demonstrating practical benefits for retrieval-augmented narrative understanding. The work presents a scalable, task-agnostic paradigm that complements traditional end-to-end models and opens avenues for integrating fine-grained narrative context into pretraining and downstream QA systems.

Abstract

This work introduces an original and practical paradigm for narrative comprehension, stemming from the characteristics that individual passages within narratives tend to be more cohesively related than isolated. Complementary to the common end-to-end paradigm, we propose a fine-grained modeling of narrative context, by formulating a graph dubbed NarCo, which explicitly depicts task-agnostic coherence dependencies that are ready to be consumed by various downstream tasks. In particular, edges in NarCo encompass free-form retrospective questions between context snippets, inspired by human cognitive perception that constantly reinstates relevant events from prior context. Importantly, our graph formalism is practically instantiated by LLMs without human annotations, through our designed two-stage prompting scheme. To examine the graph properties and its utility, we conduct three studies in narratives, each from a unique angle: edge relation efficacy, local context enrichment, and broader application in QA. All tasks could benefit from the explicit coherence captured by NarCo.

Fine-Grained Modeling of Narrative Context: A Coherence Perspective via Retrospective Questions

TL;DR

NarCo introduces a fine-grained narrative-context graph where passages are nodes and backward-facing retrospective questions form edges, capturing explicit coherence signals beyond end-to-end reasoning. It realizes this graph with a two-stage LLM prompting pipeline (question generation followed by self-verification) without human annotations, and validates its utility through three studies on edge efficacy, node augmentation, and broad QA tasks. Across recap, plot retrieval, and long-document QA, NarCo-enhanced representations and edges improve performance over baselines, demonstrating practical benefits for retrieval-augmented narrative understanding. The work presents a scalable, task-agnostic paradigm that complements traditional end-to-end models and opens avenues for integrating fine-grained narrative context into pretraining and downstream QA systems.

Abstract

This work introduces an original and practical paradigm for narrative comprehension, stemming from the characteristics that individual passages within narratives tend to be more cohesively related than isolated. Complementary to the common end-to-end paradigm, we propose a fine-grained modeling of narrative context, by formulating a graph dubbed NarCo, which explicitly depicts task-agnostic coherence dependencies that are ready to be consumed by various downstream tasks. In particular, edges in NarCo encompass free-form retrospective questions between context snippets, inspired by human cognitive perception that constantly reinstates relevant events from prior context. Importantly, our graph formalism is practically instantiated by LLMs without human annotations, through our designed two-stage prompting scheme. To examine the graph properties and its utility, we conduct three studies in narratives, each from a unique angle: edge relation efficacy, local context enrichment, and broader application in QA. All tasks could benefit from the explicit coherence captured by NarCo.
Paper Structure (47 sections, 4 equations, 5 figures, 6 tables)

This paper contains 47 sections, 4 equations, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Questions Under Discussion
  4. Narrative Comprehension Assessments
  5. LLM Understanding and Reasoning
  6. Structured Representation
  7. NarCo: Narrative Cognition Graph
  8. Graph Formulation
  9. Nodes
  10. Edges
  11. Graph Realization
  12. Question Generation
  13. Self Verification
  14. Study I: Edge Efficacy
  15. Approach
  16. ...and 32 more sections

Figures (5)

  • Figure 1: Our proposed NarCo graph described in Section \ref{['sec:graph']}, with retrospective questions connecting two nodes.
  • Figure 2: Three presented studies leveraging NarCo.
  • Figure 3: Prompt for Question Generation (turn 1). Slots in blue refer to the input texts.
  • Figure 4: Prompt for Question Generation (turn 2). Slots in blue refer to the input texts.
  • Figure 5: Prompt for Question Filtering via back verification. Slots in blue refer to the input texts.