Relation Extraction using Few-Shot Entailment on Conversational Data

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
Relation Extraction is an important task for personal Natural Language Understanding, and especially so in conversational applications where Knowledge Graphs, built from such relations, are essential for knowledge storage. Recent advances in Natural Language Understanding have shown that pre-trained Language Models tend to be the best at solving various Language Understanding tasks. Most of the time, they are fine-tuned on the specific downstream task, which relies on a large enough, high-quality dataset for the task. To overcome this issue and improve model efficiency, the novel approach of Few-Shot Learning is explored. Combined with the task of Natural Language Entailment, recent research has shown that models using Few-Shot Learning and Inference can outperform fully fine-tuned State-of-the-Art models on Relation Extraction tasks (and others). Since relations are so important for conversational applications, the question is in how far the approach of performing Relation Extraction with Entailment-based Few-Shot Learning can be applied to conversational domains. Therefore, this thesis investigates in how far this is applicable. The obtained results of the Few-Shot Entailment models tested do not reach state-of-the-art approaches on relatively comparable tasks. Still, one main conclusion is that entity type information is potentially an important factor for accurate relation extraction, which is recommended for further research.
Faculteit der Sociale Wetenschappen