The paper proposes an end-to-end Transformer-based network, ECEformer, to learn the evolutionary chain of events (ECE) for temporal knowledge graph reasoning (TKGR). The key contributions are:
ECEformer consists of two novel modules: ECE Representation Learning (ECER) and Mixed-Context Knowledge Reasoning (MCKR). ECER employs a Transformer encoder to embed each event in the ECE, exploring the internal structure and semantic relationships within individual quadruples. MCKR induces the embeddings of each event and enhances interaction within and between quadruples via an MLP-based information mixing layer.
To enhance the timeliness of the events, ECEformer devises an additional time prediction task to imbue effective temporal information within the learned unified representation.
Extensive experiments on six benchmark datasets demonstrate the state-of-the-art performance and the effectiveness of the proposed ECEformer.
To Another Language
from source content
arxiv.org
Głębsze pytania