Plausibility assessment of triples with distant supervision

Abstract

Building knowledge bases by extracting triples from text has attracted significant attention but introduced another fundamental problem: rampant triples with erroneous expressions which are rarely found in human expressions. Recent validation research has not paid much attention to the generators of expressions when validating triples. Machine-extracted expressions should be validated in a different way unlike human-generated expressions because they are extracted without intention or consciousness. Focusing on the plausibility assessment of triples, this research proposes a new plausible/nonsensical framework overlaid with a true/false framework. Then it conceptualizes the validation of machine-extracted triples as a two-step procedure: a domain-independent plausibility assessment and a domain-dependent truth validation only for plausible triples. Furthermore, this research introduces two learning methods. A distant supervision method consistently builds both positive and negative training data, eliminating the need for indefinable but obligatory negative training data. A lazy learning algorithm skips the generation of pre-defined models that have difficulty in dealing with triples various expressions. These algorithms also learn some form of semantic relationships that improve the performance of plausibility assessment. The experimental results support the proposed approach, which outperformed several unsupervised baselines. The proposed approach can be used to filter out newly extracted nonsensical triples and existing nonsensical triples in knowledge bases. It can be used on its own, or it can complement existing truth validation process. Extending background knowledge for better coverage and implementing converging algorithms remain for future investigation.