IPED: An Implicit Perspective for Relational Triple Extraction based on Diffusion Model

TL;DR

IPED addresses relational triple extraction by replacing explicit element tagging with an implicit block-covered table filling strategy and a block-denoising diffusion model. The approach uses a 3D table space of blocks, a PBES decoding scheme, and a Blk-DDM that iteratively refines block edges and levels to recover all triples efficiently. Empirical results on NYT and WebNLG show state-of-the-art F1 scores and substantially faster inference, particularly on large-relations datasets, validating the method's robustness to SEO, EPO, and SOO patterns. The work introduces a scalable, classifier-free paradigm for table filling that reduces redundant tagging and decoding errors, with potential extensions to broader information extraction tasks.

Abstract

Relational triple extraction is a fundamental task in the field of information extraction, and a promising framework based on table filling has recently gained attention as a potential baseline for entity relation extraction. However, inherent shortcomings such as redundant information and incomplete triple recognition remain problematic. To address these challenges, we propose an Implicit Perspective for relational triple Extraction based on Diffusion model (IPED), an innovative approach for extracting relational triples. Our classifier-free solution adopts an implicit strategy using block coverage to complete the tables, avoiding the limitations of explicit tagging methods. Additionally, we introduce a generative model structure, the block-denoising diffusion model, to collaborate with our implicit perspective and effectively circumvent redundant information disruptions. Experimental results on two popular datasets demonstrate that IPED achieves state-of-the-art performance while gaining superior inference speed and low computational complexity. To support future research, we have made our source code publicly available online.

Figures (4)

• Figure 1: Figure (a) depicts our table-filling strategy along with triple demonstration. For the convenience of illustration, we simplify our three-dimensional tables (as in Figure (b)) into the form of a two-dimensional table in Figure (a), containing nine blocks in total that represent nine triples. Here, dashed rectangles denote the four edges of the blocks, and different colors indicate the levels of the blocks. Figure (b) illustrates the overall diffusion process.
• Figure 2: The overview model structure of IPED. To enhance the illustration of the workflow, we utilize three different colors to denote various feature representations: Pink for level information, Yellow for sentence information, and Red for edge information. Ⓔ represents the encoding of $\mathbf{R}_{\textit{E}}$. $\otimes$ denotes the maxpooling operation. To simplify the illustration, the four Biaffine modules are integrated into one in this overview. To better display the reverse process as in Figure \ref{['figure1']}, a reverse-flow arrow is used to symbolize progressive denoising.
• Figure 3: Performance of IPED with different number of denoising blocks D in terms of F1-score on NYT.
• Figure 4: Performance of IPED with different number of denoising blocks D in terms of inference time on WebNLG. Note that the batch size is 8 during inference.