Evaluating Prompt-Based Question Answering for Object Prediction in the Open Research Knowledge Graph

Abstract

Recent investigations have explored prompt-based training of transformer language models for new text genres in low-resource settings. This approach has proven effective in transferring pre-trained or fine-tuned models to resource-scarce environments. This work presents the first results on applying prompt-based training to transformers for scholarly knowledge graph object prediction. Methodologically, it stands out in two main ways: 1) it deviates from previous studies that propose entity and relation extraction pipelines, and 2) it tests the method in a significantly different domain, scholarly knowledge, evaluating linguistic, probabilistic, and factual generalizability of large-scale transformer models. Our findings demonstrate that: i) out-of-the-box transformer models underperform on the new scholarly domain, ii) prompt-based training improves performance by up to 40% in relaxed evaluation, and iii) tests of the models in a distinct domain reveals a gap in capturing domain knowledge, highlighting the need for increased attention and resources in the scholarly domain for transformer models.