Integrating Relational Knowledge With Text Sequences for Script Event Prediction.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Script event prediction involves inferring future events from incomplete event sequences.
• Current models often treat scripts as simple sequences or graphs, failing to capture complex relational and semantic event information.
• There is a need for models that can jointly understand event semantics and their inter-relationships.

Purpose of the Study:

• To propose a novel script representation called relational event chain.
• To develop a new model, relational-transformer, for learning event embeddings that incorporate both semantic and relational knowledge.
• To improve the accuracy of script event prediction by leveraging relational information.

Main Methods:

• Formalizing scripts as relational event chains by extracting event relationships from an event knowledge graph.
• Utilizing a relational-transformer model that combines transformers and graph neural networks (GNNs).
• Learning event embeddings that encode both semantic and relational properties.

Main Results:

• The proposed relational-transformer model significantly outperforms existing baselines on both one-step and multi-step script event prediction tasks.
• Demonstrated the effectiveness of encoding relational knowledge into event embeddings for improved prediction accuracy.
• Analysis of different model structures and relational knowledge types provided insights into model performance.

Conclusions:

• The relational event chain representation and the relational-transformer model offer a more comprehensive approach to script event prediction.
• Incorporating relational knowledge is crucial for enhancing the understanding and prediction of event sequences.
• The findings validate the approach and suggest potential for further advancements in event understanding.