Curriculum-based adaptation of language model for knowledge graph: Handling low and noisy data

This paper presents a novel algorithm to adapt language models to the task of knowledge graphs, formulated as a staged RL problem. The methodology incorporates curriculum learning while asserting control over the variance of the model. Additionally, it leverages adaptive sampling and normalized loss to reduce the level of human intervention. The algorithm has proven robust and exhibited insensitivity to hyperparameters. Experiments on graph question answering datasets have demonstrated that the approach improves performance on benchmarks with 40% less data, highlighting the effectiveness of selective training approaches.