Learning to Update Knowledge Graphs by Reading News

Jizhi Tang, Yansong Feng, Dongyan ZhaoEMNLP 2019 · OpenReview

Abstract

Knowledge Graphs (KGs) are essential for structured information but struggle with real-time updates. This paper introduces GUpdater, a framework that leverages news snippets to update KGs. By using a text-guided graph attention mechanism and temporary "shortcut" edges, GUpdater captures both explicit and implicit changes, significantly outperforming traditional rule-based and information extraction methods.

1. Introduction

Knowledge Graphs—structured databases of real-world facts like "LeBron James plays for the Lakers"—are incredibly useful but notoriously difficult to keep current. When a news event occurs, such as a player trade, the KG must be updated immediately to remain accurate. Current approaches face two major hurdles:

  1. Rule Rigidity: Most systems use Information Extraction (IE) to pull facts from text and then apply pre-defined rules to update the KG. These rules are hard to maintain and fail to generalize across different domains.
  2. Implicit Information: News articles rarely list every consequence of an event. If a player moves from Team A to Team B, the text might not explicitly state they are no longer teammates with their former roster. Capturing these "hidden" updates is a significant challenge for standard IE tools.
The transition of subgraphs during a knowledge graph update, illustrating how news text influences the structural state of the graph.

2. Methodology: The GUpdater Framework

The authors propose GUpdater, an encoder-decoder model built on Graph Neural Networks (GNNs). Its architecture features three key innovations designed to bridge the gap between unstructured text and structured graph data.

2.1 Relational Graph Attention Layer (R-GAT)

This layer combines the strengths of Relational Graph Convolutional Networks (R-GCNs), which handle diverse relationship types, with Graph Attention Networks (GATs), allowing the model to focus on the most relevant neighboring nodes during the update process.

2.2 Text-based Attention

This serves as the "guidance" system. The model processes news text using a bi-directional Gated Recurrent Unit (bi-GRU). The resulting context is used to calculate attention weights for the graph, effectively telling the model which "messages" are important to pass between nodes based on the news content.

2.3 Shortcuts and Decoding

In large KGs, entities mentioned in the same news snippet might be many "hops" away. GUpdater creates temporary "shortcut" edges between all entities mentioned in the text, allowing information to flow directly between them. The decoder then uses DistMult, a standard triple-scoring function, to determine which links should exist in the final, updated version of the graph.

The GUpdater pipeline: from the initial graph state with shortcuts to the updated state via R-GAT encoding and DistMult decoding.

3. The NBAtransactions Dataset

To evaluate GUpdater, the researchers introduced NBAtransactions, a benchmark dataset consisting of 4,100 news-transaction pairs from 2010 to 2019. The dataset covers eight types of transactions, from simple free agency signings to complex multi-player trades. Notably, only about 5.6% of the necessary changes are explicitly mentioned in the text, making it a rigorous test for capturing implicit information.

4. Key Results and Insights

GUpdater significantly outperformed several strong baselines, including state-of-the-art relation extraction models (PCNN) and standard R-GCNs.

  • Overall Accuracy: On the 1-hop subgraph setting, GUpdater achieved an F1 score of 0.9866, compared to 0.9429 for the best baseline (IE-gold + R-GCN).
  • Implicit Changes: In the "Added Accuracy" metric (measuring the ability to find new links not explicitly stated), GUpdater scored 0.8926, nearly doubling the performance of the IE-gold + R-GCN baseline (0.4681).
  • Ablation Success: Removing text-based attention or shortcuts led to a massive drop in performance, confirming that the integration of text and graph structure is vital.

5. Conclusion

GUpdater demonstrates that the graph structure itself holds the key to filling in the blanks left by news text. By using text as a "gatekeeper" for graph attention, the model avoids meaningless computation and reduces incorrect updates. Future work will focus on handling successive events in sequences and managing the introduction of entirely new entities into the graph set.

Summary updated

BibTeX
@inproceedings{tang2019learning,
  author = {Jizhi Tang and Yansong Feng and Dongyan Zhao},
  title = {{Learning to Update Knowledge Graphs by Reading News}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2019},
  url = {https://openreview.net/forum?id=z-kjKHzlnqW},
}
Figure