Name
Abstract | ||
|---|---|---|
Knowledge graph embedding aims to learn representations of entities and relations in a knowledge graph. Recently, QuatE has introduced the graph embeddings into the quaternion space. However, there are still challenges in dealing with complex patterns, including 1-N, N1, and multiple-relations between two entities. Since the learned entity embeddings tend to overlap with each other in the first two cases, and the learned relation embeddings tend to overlap with each other in the last case. To deal with these issues, we propose QuatSE, a novel knowledge embedding model that adjusts graph embeddings via spherical linear interpolation (Slerp) of entities and relations. For a triple (head entity, relation, tail entity), QuatSE calculates Slerp between each entity and its relation, and adds the normalized interpolation to the corresponding entity. The operation avoids the problem of embedding overlap and ensures the information of original entity is not missed. We further compare the effect of interpolation using different normalization strategies (L1 or L2) for Slerp. Several experiments suggest that QuatSE works well in 1-N, N-1 and multiple-relations pattern. QuatSE outperforms the existing quaternion-based models. |
Year | DOI | Venue |
|---|---|---|
2022 | 10.1007/978-3-031-17120-8_17 | NATURAL LANGUAGE PROCESSING AND CHINESE COMPUTING, NLPCC 2022, PT I |
Keywords | DocType | Volume |
Knowledge graph embedding, Spherical linear interpolation, Quaternion | Conference | 13551 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 4 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Jiang Li | 1 | 0 | 0.34 |
| Xiangdong Su | 2 | 0 | 1.01 |
| Xinlan Ma | 3 | 0 | 0.34 |
| Guanglai Gao | 4 | 78 | 24.57 |