KBPT: knowledge-based prompt tuning for zero-shot relation triplet extraction.

Area of Science:

• Natural Language Processing
• Information Extraction
• Knowledge Representation

Background:

• Knowledge representation is crucial for information extraction but often overlooked in zero-shot settings.
• Zero-shot relation triplet extraction (RTE) aims to identify relationships not present in training data, facing challenges with data scarcity.

Purpose of the Study:

• To develop a novel framework, knowledge-based prompt tuning (KBPT), for zero-shot relation triplet extraction.
• To address the data scarcity issue in low-resource scenarios for RTE.

Main Methods:

• KBPT integrates external ontology knowledge with language model prompts to synthesize structured training examples.
• A multiple triplets decoding (MTD) algorithm was developed for extracting multiple relation triplets.
• A collective training method jointly optimizes embedding representations, bridging knowledge and text.

Main Results:

• KBPT demonstrated significant improvements, achieving up to 14.65% and 24.19% F1 score enhancements on Wiki-ZSL and TACRED-Revisit datasets, respectively.
• The model outperformed current state-of-the-art (SOTA) models in F1 score, P-R curves, and AUC.
• The framework is model-agnostic and applicable to various pre-trained language models (PLMs).

Conclusions:

• KBPT offers an effective solution for zero-shot relation triplet extraction, particularly in low-resource settings.
• The integration of external knowledge and prompt-based synthesis enhances extraction performance.
• The proposed method advances the field of information extraction in challenging, data-scarce scenarios.