Knowledge Graph Embedding With Interactive Guidance From Entity Descriptions

Abstract

Knowledge Graph (KG) embedding aims to represent both entities and relations into a continuous low-dimensional vector space. Most previous attempts perform the embedding task using only knowledge triples to indicate relations between entities. Entity descriptions, although containing rich background information, have not been well utilized in these methods. In this paper, we propose Entity Descriptions-Guided Embedding (EDGE), a novel method for learning the knowledge graph representations with semantic guidance from entity descriptions. EDGE enables an embedding model to learn simultaneously from 1) knowledge triples that have been directly observed in a given KG, and 2) entity descriptions which have rich semantic information about these entities. In the learning process, EDGE encodes the semantics of entity descriptions to enhance the learning of knowledge graph embedding, and integrates such learned KG embedding to constraint their corresponding word embeddings in entity descriptions. Through this interactive procedure, semantics of entity descriptions may be better transferred into the learned KG embedding. We evaluate EDGE in link prediction and entity classification on Freebase and WordNet. Experimental results show that: 1) with entity descriptions injected, EDGE achieves significant and consistent improvements over state-of-the-art baselines; and 2) compared to those one-time injection schemes studied before, the interactive guidance strategy maximizes the utility of entity descriptions for KG embedding, and indeed achieves substantially better performance.INDEX TERMS Knowledge graph embedding, entity descriptions, interactive guidance.

A knowledge graph completion model integrating entity description and network structure

“…Entity description information is the most used external information in the knowledge graph completion model. To effectively use entity description information, researchers have Knowledge graph completion model designed different fusion methods, which include DKRL (Xie et al, 2016b), EDGE (Zhou et al, 2019) and KG-BERT (Yao et al, 2020). DKRL uses continuous-bag-of-words (CBOW) and CNN to represent entity descriptions, which is then applied to TransE's scoring function to train and optimize the representations of entities and relationships.…”

“…First, compared with methods based on network structure such as PairRE (Chao et al, 2021), DMACM (Huang et al, 2021), RNNLogic (Qu et al, 2021) and HyperGEL (Zeb et al, 2021), the EDA-KGC model has significantly better performance due to the external information. Second, compared with methods that leverage entity description, such as DKRL (Xie et al, 2016b) and EDGE (Zhou et al, 2019), the proposed model uses the comprehensive integration of entity description and network structure to capture the potential semantic information. Third, compared with the models that are jointly pre-trained on the knowledge graph and the text corpus, such as KG-BERT (Yao et al, 2020), BLP (Daza et al, 2021) and GenKGC (Xie et al, 2022), EDA-KGC saves the expensive model overhead as it requires less training data and training time.…”

“…In contrast, they ignore much information from other sources, such as entity descriptions and attributes. To address the above issue, several models such as DKRL (Xie et al, 2016b), EDGE (Zhou et al, 2019), KG-BERT (Yao et al, 2020) and SDT (Chen et al, 2020) have attempted to introduce the entity description, attributes and other information into knowledge graph representation learning.…”

Utilizing Textual Information in Knowledge Graph Embedding: A Survey of Methods and Applications

“…Wang et al [29] propose Entity Descriptions-Guided Embedding (EDGE) to encode the entity description. BiLSTM is introduced in EDGE to well handle the context and word sequence in the entity description.…”

“…More importantly, it is slightly affected by the spare connectivity in KG. As a result, methods that utilize the textual information in KG embedding start to get attention [25]- [29]. We category these works according to if the representation is built from the textual informaiton: (i) Text-based KG embedding: The textual information are encoded to represent the entities and relations.…”