Kbpt: Knowledge-Based Prompt Tuning for Zero-Shot Relation Triplet Extraction

Abstract

Knowledge representation is becoming an effective way of information extraction. However, considerable studies ignored its application scenarios in the zero-shot setting. In this paper, we propose a novel framework for prompt language models based on external ontology knowledge called Knowledge-Based Prompt Tuning for the Zero-shot Relation Triplet Extraction (KBPT), which encourages further investigation in low-resource regimes to address the data scarcity problem in Relation Triplet Extraction (RTE). The major task of relation triplet extraction in zero-shot learning is to extract multiple triplets that are consisted of head entities, tail entities, and relation labels from an input sentence where the extracted relation labels do not exist in the training set. The fundamental idea of prompt tuning is to construct a prompt template, then append it behind the input text as the pre-trained language models (PLMs) input, thus, transforming the classification task into a masked language model prediction. However, our proposed model does not involve the mask language model prediction but a well-designed prompt template in a structured text format to generate synthetic training data containing the unseen relation category. Concretely, we utilize the relation labels and incorporate virtual tokens sources from the relation semantics to construct a structured prompt template for generating synthetic training instances. Moreover, to further enrich and supplement prior knowledge, we draw on ontology schema based on external knowledge bases to enhance the capability of semantic representation in the prompt template. To address the problem of knowledge heterogeneity, we synergistically optimize these embedding representations by way of collective training. In addition, we carefully design a Multiple Triplets Decoding (MTD) algorithm to break through the limitation of extracting multiple relation triplets in a sentence, and our proposed model is model-agnostic and can be orthogonal to any kind of PLMs. Extensive experimental results on four publicly available datasets with low-resource and zero-shot settings demonstrate the effectiveness of our proposed method. Moreover, our model achieves better performance compared with the current State-Of-The-Art (SOTA) model in terms of ${\rm F_{1}}$ score, precision-recall (P–R) curves and AUC.