Cross-lingual entity alignment via joint attribute-preserving embedding

Proposed a joint attribute-preserving embedding model for cross-lingual entity alignment, where it has two main modules / embeddings to learn the embeddings of relationship and attribute triples in two KBs:

  1. Structure embedding. This focuses on modelling relationship structures of two KBs and uses existing alignment to overlap common structures

  2. Attribute embedding. This captures the correlations of attributes between entities and clusters entities based on this

Once we have both the structure embedding and attribute embedding, we would combine them to jointly embed all entities in two KBs into the same vector space.

Future Work

  1. Extend embedding with cross-lingual hyperplane projection to alleviate the problem of multi-mapping relations

  2. Incorporate attribute values into our attribute embeddings. Currently we discard them due to sparsity and cross-linguality

  3. Assess the robustness of our model and evaluate our approach on more heterogeneous KBs

Co-training embeddings of knowledge graphs and entity descriptions for cross-lingual entity alignment

Proposed KDCoE to enhance the semi-supervised learning of multilingual KG embeddings. KDCoE iteratively trains two component embeddings models:

  1. KG embedding model jointly trains a translational knowledge model with a linear-transformation alignment model to capture KG structure

  2. Description embedding model uses attentive GRU and multi-lingual word embeddings to encode multi-lingual entity descriptions

The co-training is processed on a large Wikipedia-based trilingual KG. This method alleviate the challenge of incomplete inter-lingual links (ILLs) that match cross-lingual counterparts of entities.

Future Work

  1. Explore the boosting approach for cross-lingual KG completion

  2. Explore the effect of other forms of knowledge models in KGEM for encoding each language-specific KG structure

Ryan

Ryan

Data Scientist

Leave a Reply