Leveraging deep neural networks and knowledge graphs for entity disambiguation

Presents a novel deep semantic relatedness model (DSRM) based on deep neural networks and semantic knowledge graphs to measure entity semantic relatedness for topical coherence as topical coherence is important for entity disambiguation! The DSRM model is trained on large knowledge graphs where it maps different types of knowledge of entities from different knowledge graphs such that the distance between two semantically similar entities is minimised!

Future Work

  1. Direct encoding of semantic paths from KGs to neural networks

  2. Design a joint model for entity disambiguation and other entity-related measurement so that the model can benefit from different tasks and generates dynamic context-aware entity relatedness scores

Entity disambiguation by knowledge and text jointly embedding

Proposed EDKate, which represent entities and mentions into low-dimensional space using both the knowledge base and text and use these embeddings to design effective features and build a two-layer disambiguation model. Entity disambiguation method typically consists of three stages:

  1. Constructing representations for mentions / entities from raw data

  2. Extract features from disambiguation models using mentions and entities representations

  3. Optimising the disambiguation model by empirically setting / learning weights on extracted features

Most existing methods focus on stages 2 and 3 with stage 1 mostly using BoW representations. This paper provides a solution for stage 1.

Iterative entity alignment via joint knowledge embeddings

Proposed a joint knowledge embeddings for entity alignment that jointly encodes both entities and relations of different KGs into a unified low-dimensional semantic space using a small seed set of aligned entities. WIth this, we could align entities according to their semantic distance. We also present an iterative and parameter sharing method to improve alignment performance. Overall, the proposed method consists of three components:

  1. Knowledge embeddings using translation-based KRL methods

  2. Joint embeddings where we map knowledge embeddings of different KGs into a joint semantic space

  3. Iterative alignment where we iteratively align entities and their counterparts and update the joint knowledge embeddings

Future Work

  1. Incorporate richer external information of KGs for entity alignment (rather than just internal structural information)

  2. Experiment with different KRL methods and see which one is more effective for entity alignment

Ryan

Ryan

Data Scientist

Leave a Reply