Relation extraction with explanation
Annotated the test set from FB-NYT dataset with sentence-level explanations to evaluate the quality of explanations derived by relation extraction models. The paper examines two strong baseline relation extraction models to evaluate the quality of their explanation. We found two methods to improve relation extraction:
Replacing entity mentions with their fine-grained entity types for sentence representation leads to improvement in extract accuracy and model explainability
Augmenting the sentence dataset with automatically generated “distractor” sentences (contain no supporting information for relation) to train the model to ignore irrelevant information
Few-shot relation extraction via bayesian meta-learning on relation graphs
Proposed a novel Bayesian meta-learning approach to learn the posterior distribution of the prototype vectors of relations for few-shot relation extraction. The initial prior of prototype vectors are created using a graph neural network on the global relation graph. The paper uses stochastic gradient Langevin dynamics to optimise the posterior distribution of the prototype vectors. The idea behind meta-learning is to train models with many different tasks, each of which only has few examples so that the trained model can quickly generalise to new tasks with limited data. To better generalise to new relations, we decided to use a global graph between different relations.
Automate the process of learning the structure of relation graph by using existing studies
Apply the proposed approach to other applications such as few-shot image classification