There are many subtasks within entity discovery:
This is also known as named entity recognition (NER). Early work focuses on hand-crafted text features whereas recent works have been focusing more on applying seq2seq models to learn character and word-level features.
This task involves linking entity mentions with the corresponding entities in the knowledge graph.
This involves coarse and fine-grained types. Fine-grained types has a hierarchical structure to type categorisation and it’s usually treated as multi-class / multi-label classification.
This task involves fusing knowledge from different knowledge graphs. Given two entity sets, we want to find entities whereby e1 and e2 are equivalent! In practice, we use seed words (synonyms for example) to kickstart the alignment process. Embedding-based approach involves calculating the similarity between different entity embeddings.
This is where we want to extract the relationships between different entities. This task often face the problem of lack of labelled relational data, therefore, leading to mostly distant supervision. Distant supervision is where you use rule-based algorithms to create training data by making some hypotheses about your data. Traditional method involves feature engineering but latest approach focuses on the interconnection between features. Below is an overview of the neural relation extraction space.
Neural Relation Extraction
Both CNNs and RNNs are used for relation extraction. Distance features between entities was first explored for relation classification.
Both CNNs and RNNs were enhanced using attention mechanism. For example, word-level attention mechanism was used to capture semantic information of words.
Graph Convolutional Networks
GCNs is used to encode the dependency tree of sentences or to learn KG embeddings to leverage relational knowledge for building sentence representation.
This is where we have a generator and a discriminator, where the generator is responsible for generating fake noisy data, in our case, adding noisy data to our word embeddings, and the discriminator is responsible for distinguish between real positive and fake positive!
There are other advances such as using reinforcement learning, deep neural network, transfer learning, and few shot relation classification. (Need to explore more – look at those advances below!)