Unsupervised Open Information Extraction with Large Language Models
This master’s thesis explores the intersection of Open Information Extraction (OpenIE) and linguistic acceptability. The research investigates whether linguistic acceptability can be used to rank extracted triples in OpenIE tasks. The proposed method uses a lightweight linguistic acceptability model to rank triples generated by DeepEx, achieving competitive results with significantly less training data. Uses only 8,500 sentences compared to DeepEx’s 6.2M sentences for ranking, The thesis demonstrates that augmenting probability scores with acceptability measures produces strong results across diverse evaluation benchmarks including NYT, PENN, WEB, and CaRB, highlighting the potential of linguistic acceptability as a ranking mechanism for information extraction tasks.