Multilingual Event Causality Identification via Meta-learning with Knowledge
Shen Wang, Fei Cai, Mengxi Zhang · 2024
Multilingual Event Causality Identification (Multilingual ECI) is the task of detecting causal relations between events mentioned in the multilingual text. The biggest challenge of this task is the small size of the minor language corpus, making it very difficult to train a model with good performance. To address this issue, we propose a meta-learning method that integrates multilingual knowledge, which mainly includes two modules: the multilingual knowledge obtain module, which alleviates data scarcity by introducing multilingual knowledge from outside; The meta-learning module utilizes ProtoMAML combined with l₂ standardization to extract language independent causal features from text and learn better event representations in different languages. The experimental results on the latest publicly available MECI dataset show that our model outperforms the current best baseline, with an average F1 improvement of 3.9% in different languages.