Knowledge acquisition tasks are divided into three categories:

  1. Knowledge Graph Completion

  2. Entity Discovery

  3. Relation Extraction

The first is used to expand existing knowledge graph whereas the last two is used to discover new knowledge from unstructured text.

Knowledge Graph Completion

This is the area where you aim to add new triples to a knowledge graph. There are many subtasks such as:

  1. Link prediction

  2. Entity prediction

  3. Relation prediction

Early research focuses on learning representation for triple predictions but most of the work failed to capture multi-step relationships. Triplet classification is a related KGC task whereby we evaluate the correctness of a factual triplets. The two methods below are:

  1. Embedding-based Ranking

  2. Relation Paths – Reasoning, Reinforcement Learning (RL), Rule-based Reasoning

  3. Meta Relational Learning

  4. Triple Classification

Embedding-based Models

As you can see, embedding-based ranking methods require you to learn embedding vectors based on existing triples. KRL models and joint learning models can be used for KGC. Many embeddings methods do not differentiate entities and relation prediction and so SENN aims to tackle this. SENN distinguishes three KGC subtasks explicitly using a unified shared embedding with adaptive loss function to learn the different features for each subtask.

In addition, existing methods rely heavily on existing connections and fail to capture the evolution of factual knowledge or entities with little connections. You could also explore discriminative and generative methods.

Relation Path Reasoning

Embedding methods fail to model complex relation paths. This is where relation path reasoning comes in, taking advantage of the path information within a knowledge graph. Random walk prediction has been widely investigated. Path-Ranking Algorithm (PRA) is one such example where you choose a relational path under a combination of path constraints and perform maximum-likelihood classification. There are many ways to improve path reasoning, for example, incorporating auxiliary information such as text descriptions or model multi-hop relational path (chain-of-reasoning).

RL-based Path Finding

Reinforcement learning can be used for multi-hop reasoning by formulating the path finding problem as a Markov Decision Process (MDP). Below are examples of different RL methods!

Rule-based Reasoning

Here, we can use logical rule learning on the symbolic nature of knowledge. Rule is defined by the head and body, whereby head <- body. Head is an atom (a fact with variable subjects and objects) and the body can be a set of atoms. Most research focuses on adding logical rule to embedding methods to improve reasoning. KALE is an example of this. This combination of neural and symbolic methods have led to implementing rule-based reasoning. pLogicNet is an example of this.

Meta Relational Learning

This requires models to predict new relational facts with only few samples (few-shot learning). This is to mimic real-world example where knowledge is dynamic and usually requires you to predict unseen triples (expansion of knowledge).

Triple Classification

As mentioned, this is the task of determining whether a fact is correct or not (binary classification). This is usually determined by a scoring function with a pre-specified threshold.

Ryan

Ryan

Data Scientist

Leave a Reply