Efficient Entity Resolution via Hierarchical Graph Attention and Semantic Blocking

In entity resolution, co-occurring entities are often semantically related. However, existing methods typically assume that each entity is unique within the data source, overlooking the dependencies between attributes. This leads to limitations in handling complex relationships. This paper employs LangChain and FAISS as blockers to quickly filter out obviously mismatched pairs, followed by a hierarchical graph attention-based Transformer network to generate high-dimensional embeddings of entities, modeling the complex dependencies between entities to determine whether the remaining pairs refer to the same realworld entity. Experiments demonstrate that this method exhibits significant advantages in handling complex entity relationships and large-scale data.