Article: Integrating rich event-level and schema-level information for script event prediction Journal: International Journal of Computational Science and Engineering (IJCSE) 2024 Vol.27 No.6 pp.691 - 702 Abstract: A script consists of a series of structured event sequences extracted from the texts. Given historical scripts, script event prediction aims to predict the subsequent event. The critical aspect in script event prediction is how to effectively represent events, which plays an important role in making accurate predictions. Most existing methods describe events through verbs and a few core independent arguments (i.e., subject, object, and indirect object), which lack the capability to deal with complex and sparse event data. In this paper, we propose a hierarchical event prediction (HEP) model, which integrates information from both event-level and schema-level. At the event level, HEP enriches the existing event representation with extra arguments (i.e., time and place) and modifiers, which provides in-depth event information. At the schema level, it induces the sparse events into conceptual schema, which improves the model's generalisation ability to make more reasonable predictions. To more effectively integrate these two layers of information, we propose an event multi-layer encoder, which can effectively integrate information between different layers. Experimental results on New York Times and CNN corpora demonstrate the effectiveness and superiority of HEP. Inderscience Publishers - linking academia, business and industry through research

Title: Integrating rich event-level and schema-level information for script event prediction

Authors: Wei Qin; Xiangfeng Luo; Hao Wang

Addresses: School of Computer Engineering and Science, Shanghai University, Shanghai, China ' School of Computer Engineering and Science, Shanghai University, Shanghai, China ' School of Computer Engineering and Science, Shanghai University, Shanghai, China

Abstract: A script consists of a series of structured event sequences extracted from the texts. Given historical scripts, script event prediction aims to predict the subsequent event. The critical aspect in script event prediction is how to effectively represent events, which plays an important role in making accurate predictions. Most existing methods describe events through verbs and a few core independent arguments (i.e., subject, object, and indirect object), which lack the capability to deal with complex and sparse event data. In this paper, we propose a hierarchical event prediction (HEP) model, which integrates information from both event-level and schema-level. At the event level, HEP enriches the existing event representation with extra arguments (i.e., time and place) and modifiers, which provides in-depth event information. At the schema level, it induces the sparse events into conceptual schema, which improves the model's generalisation ability to make more reasonable predictions. To more effectively integrate these two layers of information, we propose an event multi-layer encoder, which can effectively integrate information between different layers. Experimental results on New York Times and CNN corpora demonstrate the effectiveness and superiority of HEP.

Keywords: script event prediction; SEP; event-level; schema-level; contrast learning; bimodal cross attention; BCA.

DOI: 10.1504/IJCSE.2024.142834

International Journal of Computational Science and Engineering, 2024 Vol.27 No.6, pp.691 - 702

Received: 06 Jun 2023
Accepted: 16 Nov 2023
Published online: 28 Nov 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: Integrating rich event-level and schema-level information for script event prediction

Keep up-to-date