A Cascade Dual-Decoder Model for Joint Entity and Relation Extraction

TL;DR

This work introduces a cascade dual-decoder for joint entity and relation extraction that treats relations as mapping functions and first detects text-level relations (TR) to guide entity pair extraction (HE and TE) in a cascade. By modeling $p_{ heta}(r|x)$, $p_{ heta}(h|r,x)$, and $p_{ heta}(t|r,h,x)$ within a KL-divergence framework, the method reduces error propagation and naturally handles overlapping triples. Empirical results on NYT, WebNLG, and a real open-pit mining dataset show state-of-the-art or competitive performance, with notable gains in exact-match metrics and relation-element extraction, especially in challenging overlapping scenarios. The approach demonstrates strong practical potential for complex information extraction tasks in knowledge graph construction and domain-specific corpora.

Abstract

In knowledge graph construction, a challenging issue is how to extract complex (e.g., overlapping) entities and relationships from a small amount of unstructured historical data. The traditional pipeline methods are to divide the extraction into two separate subtasks, which misses the potential interaction between the two subtasks and may lead to error propagation. In this work, we propose an effective cascade dual-decoder method to extract overlapping relational triples, which includes a text-specific relation decoder and a relation-corresponded entity decoder. Our approach is straightforward and it includes a text-specific relation decoder and a relation-corresponded entity decoder. The text-specific relation decoder detects relations from a sentence at the text level. That is, it does this according to the semantic information of the whole sentence. For each extracted relation, which is with trainable embedding, the relation-corresponded entity decoder detects the corresponding head and tail entities using a span-based tagging scheme. In this way, the overlapping triple problem can be tackled naturally. We conducted experiments on a real-world open-pit mine dataset and two public datasets to verify the method's generalizability. The experimental results demonstrate the effectiveness and competitiveness of our proposed method and achieve better F1 scores under strict evaluation metrics. Our implementation is available at https://github.com/prastunlp/DualDec.

Figures (7)

• Figure 1: An example to show the proposed dual-decoder model on the sentence " Maria was born in Leipzig, Germany, and has been living here", which contains three overlapping triples.
• Figure 2: Schematic diagram of the overlapping triple problem. The overlapping triple problem can be tackled naturally by first detecting the relations and then extracting the entity pairs corresponding to each extracted relation.
• Figure 3: The framework of the cascade dual-decoder model. In this example, there are two candidate relations detected at the TR decoder, i.e., $Located\_in$ and $Capital\_of$. For each candidate relation ${v}_{r}^{j}$, the RE decoder extracts the corresponding head entities and tail entities sequentially. At different relations' iterations ($j$=1, 2), head entities ${v}_{head|r}$ and tail entities ${tail}_{head,r}$ at the RE decoder will change accordingly. A triple is extracted as follows: for instance, for the second relation $j$=2, two candidate head entities, i.e., $Washington$ ($l$=1) and $U.S.$ ($l$=2), are detected. Then, for $l=1$, a tail entity $U.S.$ is detected, reflecting the triple ($Washington,Capital\_of,U.S$) led by the second candidate relation $Capital\_of$ and the corresponding candidate head entity $Washington$.
• Figure 4: Structure of the Head-entity extractor.
• Figure 5: The matching process of an entity's start position and end position. $\mathbf{E}_i^{head}$ denotes the start position of the ith entity, i$\in$[1,c], c denotes the number of entities in a sentence. * denotes the possible end position of the i-th entity. $\mathbf{p}_i^{se}$ denotes the relative distance between the start position and the end position of the entity.