NarGNN: Narrative Graph Neural Networks for New Script Event Prediction Problem
| Autor: | Shuang Yang, Cong Xue, Zhihao Tang, Fali Wang, Daren Zha |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Artificial neural network
Computer science Event (computing) business.industry Graph theory 02 engineering and technology Semantics computer.software_genre Scripting language 020204 information systems 0202 electrical engineering electronic engineering information engineering Task analysis Graph (abstract data type) Experiential knowledge 020201 artificial intelligence & image processing Artificial intelligence business computer Natural language processing |
| Zdroj: | ISPA/BDCloud/SocialCom/SustainCom |
| DOI: | 10.1109/ispa-bdcloud-socialcom-sustaincom51426.2020.00086 |
| Popis: | Scripts encode world knowledge that can help text understanding, and script learning is to automatically learn this knowledge from unstructured text. Previous script event prediction only considers predicting subsequent events given an existing event sequence. However, in real life, we want to predict events at any location rather than just the last event, because we can control subsequent events by preventing or promoting intermediate events. To this end, we introduce a novel approach, named NarGNN, which can integrate the factual knowledge and experiential knowledge to predict the intermediate event. Specifically, we first construct an event evolutionary graph from the newswire corpus. Then we use Fact Encoder Layer to encode existing event facts, including the semantic information of the event itself and the sequence information among events. Third, we use Fusion Layer to fuse the graph information containing experiential knowledge and the embeddings of existing facts obtained from the previous layer. Fourth, Attention Layer is used to choose the most reasonable result. Finally, our proposed model is evaluated on widely used New York Times corpus and the results demonstrate significant improvements compared with state-of-the-art methods. Also, it is worth noting that NarGNN can be naturally extended to address the previous task with better performances than other methods. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|