[Paper Review] ELG: An Event Logic Graph

Xiao Ding, Zhongyang Li+2 more|arXiv (Cornell University)|Jul 18, 2019

Topic Modeling20 references4 citations

TL;DR

This paper introduces Event Logic Graph (ELG), a directed cyclic knowledge graph that models sequential, causal, conditional, and hierarchical (is-a) relations between events to capture real-world event evolution patterns. Constructed from large-scale text using event extraction and relation recognition, ELG significantly improves script event prediction, achieving 54.93% accuracy with a multi-model ensemble, outperforming baselines by up to 3.2%.

ABSTRACT

The evolution and development of events have their own basic principles, which make events happen sequentially. Therefore, the discovery of such evolutionary patterns among events are of great value for event prediction, decision-making and scenario design of dialog systems. However, conventional knowledge graph mainly focuses on the entities and their relations, which neglects the real world events. In this paper, we present a novel type of knowledge base - Event Logic Graph (ELG), which can reveal evolutionary patterns and development logics of real world events. Specifically, ELG is a directed cyclic graph, whose nodes are events, and edges stand for the sequential, causal, conditional or hypernym-hyponym (is-a) relations between events. We constructed two domain ELG: financial domain ELG, which consists of more than 1.5 million of event nodes and more than 1.8 million of directed edges, and travel domain ELG, which consists of about 30 thousand of event nodes and more than 234 thousand of directed edges. Experimental results show that ELG is effective for the task of script event prediction.

Motivation & Objective

• To address the lack of event-centric knowledge in traditional knowledge graphs, which focus on entities and their relations.
• To model real-world event evolution patterns—such as temporal sequences and causal dependencies—using a structured knowledge base.
• To develop a scalable framework for constructing ELG from unstructured text, enabling commonsense reasoning about event logic.
• To evaluate ELG's effectiveness in downstream tasks, particularly script event prediction, using neural graph-based models.

Proposed method

• Define ELG as a directed cyclic graph where nodes represent events and edges represent sequential, causal, conditional, or hypernym-hyponym relations.
• Extract events from raw text using trigger words (e.g., verbs) and represent them as semantic tuples (S, P, O), where P is the action.
• Recognize and label relations between event pairs using NLP techniques, distinguishing directionality in sequential and causal relations.
• Merge semantically similar events and generalize specific events to form a unified, abstracted event graph.
• Construct domain-specific ELGs: a financial domain ELG with 1.5M events and 1.8M edges, and a travel domain ELG with 30K events and 234K edges.
• Train and evaluate graph-based models (e.g., SGNN, EventComp, PairLSTM) on script event prediction using ELG as input.

Experimental results

Research questions

• RQ1 Can a knowledge base that models event evolution patterns improve script event prediction performance?
• RQ2 How do different structural representations—pairwise, chain-based, or graph-based—impact event prediction accuracy?
• RQ3 To what extent does incorporating hierarchical (is-a) and causal relations enhance event reasoning in ELG?
• RQ4 Can graph neural networks effectively leverage the relational structure of ELG for improved event prediction?

Key findings

• The SGNN model achieved the highest script event prediction accuracy of 52.45%, a 3.2% improvement over the best baseline (PairLSTM).
• Combining SGNN with EventComp and PairLSTM boosted performance to 54.93%, demonstrating complementary benefits across model architectures.
• Graph-based models (SGNN) outperformed pair-based (PairLSTM) and chain-based (EventComp) models, confirming the value of structural context in event reasoning.
• The attention mechanism in SGNN-attention improved performance over standard SGNN, indicating that context events have varying relevance to prediction.
• SGNN reached stable high accuracy faster than PairLSTM, as shown in learning curves, indicating superior training efficiency and convergence.
• ELG construction achieved high scalability, with a financial domain ELG containing over 1.5 million event nodes and 1.8 million edges.