Enhancing Narrative Commonsense Reasoning With Multilevel Causal Knowledge.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Narrative understanding relies on causality, which exists at both event and sentence levels.
• Prior research has focused on either sentence-level or event-level causality, limiting comprehensive narrative comprehension.
• Effective narrative commonsense reasoning requires integrating diverse causal information.

Purpose of the Study:

• To develop a two-stage framework that leverages both sentence-level and event-level causalities for enhanced narrative understanding.
• To address the challenge of event sparsity in capturing causal semantics within narratives.
• To improve narrative commonsense reasoning capabilities through a hierarchical knowledge graph.

Main Methods:

• A two-stage approach was devised: first, injecting sentence-level causalities into pretrained language models (PLMs) via posttraining tasks.
• Second, refining narrative commonsense reasoning using event causalities, addressing event sparsity by breaking down events into word components to capture word-word relations.
• Constructing a hierarchical knowledge graph (KG) using event-level causalities and word-level relations, followed by a KG-based reasoning process.

Main Results:

• The proposed framework effectively integrates sentence-level and event-level causal relationships.
• The method successfully alleviates the event sparsity problem by utilizing word-level relations.
• Experimental results demonstrate significant improvements in narrative commonsense reasoning.

Conclusions:

• The developed two-stage approach offers a comprehensive method for exploiting multi-level causalities in narratives.
• Integrating sentence and event causalities, along with word-level relations, enhances commonsense reasoning.
• The hierarchical knowledge graph provides a robust knowledge ground for advanced narrative understanding.