.png){kind=icon}
返回
Inductive reasoning with type-constrained encoding for emerging entities
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neunet.2024.106468.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Knowledge graph reasoning, vital for addressing incompleteness and supporting applications, faces challenges with the continuous growth of graphs. To address this challenge, several inductive reasoning models for encoding emerging entities have been proposed. However, they do not consider the multi-batch emergence scenario, where new entities and new facts are usually added to knowledge graphs (KGs) in multiple batches in the order of their emergence. To simulate the continuous growth of knowledge graphs, a novel multi- batch emergence (MBE) scenario has recently been proposed. We propose a path-based inductive model to handle multi-batch entity growth, enhancing entity encoding with type information. Specifically, we observe a noteworthy pattern in which entity types at the head and tail of the same relation exhibit relative regularity. To utilize this regularity, we introduce a pair of learnable parameters for each relation, representing entity type features linked to the relation. The type features are dedicated to encoding and updating the features of entities. Meanwhile, our model incorporates a novel attention mechanism, combining statistical co-occurrence and semantic similarity of relations effectively for contextual information capture. After generating embeddings, we employ reinforcement learning for path reasoning. To reduce sparsity and expand the action space, our model generates soft candidate facts by grounding a set of soft path rules. Meanwhile, we incorporate the confidence scores of these facts in the action space to facilitate the agent to better distinguish between original facts and rule-generated soft facts. Performances on three multi-batch entity growth datasets demonstrate robust performance, consistently outperforming state-of-the-art models.
Keyword:
Knowledge graphs
Inductive reasoning
Emerging entities encoding
期刊
IF:
6.3
论文数:
8.2K
被引数:
3.0W
机构
引用论文
“Pomacytosis”—Semi-extracellular phagocytosis of cyanobacteria by the smallest marine algae.png)
PLOS Biology
OAAI
PLOS BiologyIF0
OAAI分享收藏分享收藏Bayesian hypernetwork collaborates with time-difference evolutional network for temporal knowledge predictionNEURAL NETWORKS
NEURAL NETWORKSIF6.3
PREAI
分享收藏A new dipnoan genus,Ichnomylax, from the Lower Devonian of Victoria, Australia来自澳大利亚维多利亚州下泥盆省的一个新的dipnoan属，Ichnomylax
PREAI
分享收藏
查看更多内容