A Multi-granularity Interaction Network for Relational Triple Extraction

Xiaocheng Luo; Yanping Chen; Zaifu Zhan

cs.CLOct 18, 2026

A Multi-granularity Interaction Network for Relational Triple Extraction

Authors: Xiaocheng Luo, Yanping Chen, Zaifu Zhan

Organizations: School of Computer Science and Technology, Soochow University, Suzhou, 215006, Jiangsu, PR China · State Key Laboratory of Public Big Data, College of Computer Science and Technology, Guizhou University, Guiyang, 550025, Guizhou, PR China · Department of Electrical and Computer Engineering, University of Minnesota, Minneapolis, 55455, MN, USA

Abstract

Relational triple extraction identifies entity relations from raw sentences without depending on entity type information. The performance of traditional methods is seriously degenerated by unseen'' entities, because these methods can't build contextual dependencies relevant unseen'' entities. This paper addresses this challenge by proposing a multi-granularity interaction network that leverages word-word, entity-relation, and head-tail interactions to capture deep semantic dependencies. By incorporating multiple iterations, the multi-granularity interaction network is effective to refine the representations corresponding to head and tail entities. It enhances the global semantic features associated with relational triples, aiding in the remediation of triples that may have been overlooked or omitted in earlier steps. Strong and consistent performance is achieved by our method on two rearranged datasets, with a significant improvement in identifying relational triples, including those involving ``unseen'' named entities. For example, our model improves F1-score by 3.1%/20.5% over previous approaches on the NYT-R/WebNLG-R datasets in the Unseen metric. Code: https://github.com/xiaocen-luo/Interaction-Network.

Figures & tables

Dataset	Entire	Partial	Unseen
$\mathrm{WebNLG}$	756 ( ${\color[rgb]{1,0,0}\mathbf{85.23}\%}$ )	110 (12.40 $\%$ )	21 (2.37 $\%$ )
$\mathrm{NYT}$	2672 ( ${\color[rgb]{1,0,0}\mathbf{77.07\%}}$ )	663 (19.12 $\%$ )	132 (3.81 $\%$ )
$\mathrm{WebNLG}$ - $\mathrm{R}$	439 (44.17 $\%$ )	338 (34.00 $\%$ )	217 (21.83 $\%$ )
$\mathrm{NYT}$ - $\mathrm{R}$	823(13.81 $\%$ )	3985 (66.89 $\%$ )	1150 (19.30 $\%$ )

Table 1: The number (ratio) of triple types in the testing datasets.

Model	Dataset
Model	$\mathrm{WebNLG}$	$\mathrm{WebNLG}$ - $\mathrm{R}$	$\mathrm{NYT}$	$\mathrm{NYT}$ - $\mathrm{R}$
CasRel	88.3	68.6( $\mathbf{{\color[rgb]{1,0,0}\downarrow 19.7}}$ )	90.1	62.9( $\mathbf{{\color[rgb]{1,0,0}\downarrow 27.2}}$ )
TPLinker	89.0	69.1( $\mathbf{{\color[rgb]{1,0,0}\downarrow 19.9}}$ )	92.4	64.7( $\mathbf{{\color[rgb]{1,0,0}\downarrow 27.7}}$ )
PRGC	88.0	61.8( $\mathbf{{\color[rgb]{1,0,0}\downarrow 26.2}}$ )	89.1	62.6( $\mathbf{{\color[rgb]{1,0,0}\downarrow 26.5}}$ )

Table 2: The F1 performance achieved by the current best-performing model across four datasets ( Lee et al., 2022 ) .

Figure 1: An example contains two types of relational triples.

Figure 3: The details of Entity-Relation Interaction.

Figure 4: An Example of entity-pair tagging.

Category	Details of the dataset				Type of triple
Category	Train	Valid	Test	Relation	Entire	Partial	Unseen	Others
NYT - R	56196	5000	5000	24	348	3307	886	459
WebNLG - R	5019	500	703	216	155	178	174	196
NYT	56195	5000	5000	24	4292	473	88	147
WebNLG	5019	500	703	171	580	42	17	64

Table 3: Statistics of the datasets.

Dataset	Overlapping Pattern
Dataset	Normal	SEO	EPO	SOO	$N=1$	$N=2$	$N=3$	$N=4$	$N\geq 5$
NYT - R	3965	709	408	110	3958	733	192	50	48
WebNLG - R	389	293	9	84	411	143	81	58	10
NYT	3266	1297	978	45	3,244	1,045	312	291	108
WebNLG	245	457	26	84	266	171	131	90	45

Table 4: Overlapping patterns in datasets.

N

denotes the total count of relational triples.

Model	$\mathrm{NYT}$ - $\mathrm{R}$						$\mathrm{WebNLG}$ - $\mathrm{R}$
Model	Prec.	Rec.	F1	Entire	Partial	Unseen	Prec.	Rec.	F1	Entire	Partial	Unseen
$\mathrm{CasRel}^{\dagger}$ ( Wei et al., 2020 )	65.9	60.1	62.9	85.8	65.0	42.3	73.6	64.2	68.6	89.6	52.3	41.5
$\mathrm{TPlinker}^{\dagger}$ ( Wang et al., 2020 )	69.0	60.8	64.7	83.3	66.7	46.8	75.1	63.9	69.1	88.5	52.7	42.9
$\mathrm{PRGC}^{\dagger}$ ( Zheng et al., 2021 )	63.5	61.6	62.6	81.6	64.2	45.1	61.6	62.0	61.8	79.2	47.2	28.3
$\mathrm{BiRTE}^{\lx@sectionsign}$ ( Ren et al., 2022 )	64.5	65.4	64.9	86.1	67.0	45.2	71.0	64.6	67.6	89.2	53.9	36.7
$\mathrm{OneRel}^{\lx@sectionsign}$ ( Shang et al., 2022 )	$\underline{70.8}$	58.2	63.9	82.5	66.5	44.2	75.9	66.5	70.9	$\underline{90.8}$	$\underline{57.2}$	39.2

Table 5: Main results.

Model	$\mathrm{NYT}$ - $\mathrm{R}$						$\mathrm{WebNLG}$ - $\mathrm{R}$
Model	Prec.	Rec.	F1	Entire	Partial	Unseen	Prec.	Rec.	F1	Entire	Partial	Unseen
Fully Model	70.2	$\mathbf{67.1}$	$\mathbf{68.6}$	89.1	$\mathbf{70.8}$	$\mathbf{52.6}$	$\mathbf{78.1}$	$\mathbf{67.8}$	$\mathbf{72.6}$	$\mathbf{90.8}$	$\mathbf{57.2}$	$\mathbf{58.7}$
${w/o\ \mathrm{WWI}}$	70.3	65.0	67.6	88.4	69.5	50.7	77.1	63.5	69.7	87.0	55.9	50.0
${w/o\ \mathrm{ERI}}$	69.4	65.0	67.1	$\mathbf{89.3}$	69.2	$\underline{51.7}$	76.8	67.0	71.6	90.0	54.7	55.8
${w/o\ \mathrm{HTI}}$	70.6	64.0	67.2	87.9	69.4	50.4	75.8	$\mathbf{67.8}$	71.6	90.5	55.9	56.4
${w/o\ \mathrm{ALL}}$	$\underline{74.2}$	57.1	64.6	81.2	66.8	44.8	68.1	52.8	59.5	75.7	43.5	43.5

Table 6: Ablation study.

Dataset	Model	$\mathrm{Normal}$	$\mathrm{EPO}$	$\mathrm{SEO}$	$\mathrm{SOO}$	$\mathrm{N}=1$	$\mathrm{~N}=2$	$\mathrm{~N}=3$	$\mathrm{~N}=4$	$\mathrm{~N}\geq 5$
$\mathrm{NYT}$	$\mathrm{CasRel}$	87.3	92.0	91.4	77.0	88.2	90.3	91.9	94.2	83.7
	$\mathrm{TPLinker}$	90.1	94.0	93.4	90.1	90.0	92.8	93.1	96.1	90.0
	$\mathrm{SPN}$	90.8	94.1	94.0	81.8	90.9	93.4	94.2	95.5	90.6
	$\mathrm{PRGC}$	$91.0$	94.5	94.0	90.8	91.1	93.0	93.5	95.5	93.0
	$\mathrm{DirectRel}$	$\mathbf{91.7}$	94.8	90.0	90.0	$\mathbf{91.7}$	$\mathbf{94.1}$	93.5	96.3	92.7
	$\mathrm{BiRTE}$	91.4	94.2	94.7	-	91.5	93.7	93.9	95.8	92.1

Table 7: F1-score (%) on sentences with different overlapping patterns and different numbers of triples for the NYT and WebNLG datasets.

N

represents the number of triples within a sentence.

Figure 5: The performance of tensor-based operations with different vector spaces {3, 4, 5, 6, 8, 10, 12}.

Model	$\mathrm{NYT}$ - $\mathrm{R}$			$\mathrm{WebNLG}$ - $\mathrm{R}$
Model	Prec.	Rec.	F1	Prec.	Rec.	F1
$CCA_{Our}(\times 1)$	70.2	67.1	68.6	78.1	67.8	72.6
$CCA_{Our}(\times 2)$	69.7	61.3	65.2	79.7	66.0	72.2
$CCA_{Ori}(\times 1)$	73.2	58.4	65.0	77.7	65.3	71.0
$CCA_{Ori}(\times 2)$	75.1	57.7	65.3	78.9	65.8	71.8

Table 8: The performance of CCA with different configurations on

\mathrm{NYT}

\mathrm{R}

and

\mathrm{WebNLG}

\mathrm{R}

Model	$\mathrm{NYT}$ - $\mathrm{R}$			$\mathrm{WebNLG}$ - $\mathrm{R}$
Model	Prec.	Rec.	F1	Prec.	Rec.	F1
$[\mathrm{MLP},\mathrm{CA}]$	71.5	59.7	65.1	78.5	64.0	70.6
$[\mathrm{MHA},\mathrm{MLP}]$	73.8	57.4	64.6	79.5	66.4	72.4
$[\mathrm{MLP},\mathrm{MLP}]$	73.9	57.0	64.3	78.1	67.8	72.6
$[\mathrm{MHA},\mathrm{CA}]$	70.2	67.1	68.6	79.7	66.0	72.2

Table 9: The performance of Self Interaction and Multi-Object Interaction with different configurations on

\mathrm{NYT}

\mathrm{R}

and

\mathrm{WebNLG}

\mathrm{R}

Model	$\mathrm{NYT}$ - $\mathrm{R}$			$\mathrm{WebNLG}$ - $\mathrm{R}$
Model	Prec.	Rec.	F1	Prec.	Rec.	F1
$I=1$	70.2	67.1	68.6	78.1	67.8	72.6
$I=2$	69.2	61.9	65.4	70.5	37.3	48.8
$I=3$	64.6	64.1	64.4	69.4	36.6	47.9

Table 10: The performance of interaction network with different number of iterations of on

\mathrm{NYT}

\mathrm{R}

and

\mathrm{WebNLG}

\mathrm{R}

Model	$\mathrm{NYT}$			$\mathrm{WebNLG}$
Model	Prec.	Rec.	F1	Prec.	Rec.	F1
BiRTE	92.2	$\underline{93.8}$	93.0	93.2	94.0	93.6
OneRel	92.8	92.9	92.8	94.1	94.4	94.3
BitCoin	92.9	92.8	93.1	$\underline{94.4}$	$\underline{94.5}$	94.4
SOIRP	$\underline{93.3}$	92.9	93.1	94.3	94.1	94.2
RTF	93.2	93.4	93.3	94.0	$\mathbf{95.3}$	$\mathbf{94.6}$

Table 11: Performance on

\mathrm{NYT}

and

\mathrm{WebNLG}

Model	$\mathrm{NYT}$ - $\mathrm{R}$
Model	Training Time	GPU Memory	F1
$\mathrm{SOIRP}$	6738 seconds	4278 MB	67.5
${Ours}$	7113 seconds	5180 MB	68.6
${Ours}_{w/o\mathrm{WWI}}$	6474 seconds	4612 MB	67.6
${Ours}_{w/o\mathrm{ERI}}$	5724 seconds	4338 MB	67.1
${Ours}_{w/o\mathrm{HTI}}$	6418 seconds	5042 MB	67.2

Table 12: Comparison of model efficiency.

Figure 6: Examples of sentences of the Entire, Partial, and Unseen three triple types from OneRel and our models and the corresponding predictions. The orange mark indicates an error in the extracted relational triples. The blue mark indicates that the extracted relational triple is missing.

Figure 7: Visualization of different model components for two sentences.

Relation Extraction

Relation extraction identifies and classifies relationships between entities or events in text and multimodal data.

See all 82 papers

GLiNER-Relex: A Unified Framework for Joint Named Entity Recognition and Relation Extraction
May 11, 2026Ihor Stepanov, Oleksandr Lukashov, Mykhailo Shtopko +1Relation Extraction KG Construction
Knowledgator Engineering, Kyiv, Ukraine · Baldor Technologies Pvt. Ltd. (IDfy), Mumbai, India
Calibrated Typed Decisions for Entity, Relation, and Event Extraction
Oct 12, 2026Yang Wang, Hanlin Xiao, Yanan Ma +2Event Extraction Document Information Extraction
University of Manchester · Mayo Clinic · Shanghai University of Finance and Economics
LLM-based Atomic Propositions help weak extractors: Evaluation of a Propositioner for triplet extraction
Apr 3, 2026Luc Pommeret, Thomas Gerald, Patrick Paroubek +3KG Construction Relation Extraction
STL · Université Paris-Saclay, CNRS, LISN, 91400, Orsay, France · LISN +3

May 11, 2026J/K move · Enter open · S save

GLiNER-Relex: A Unified Framework for Joint Named Entity Recognition and Relation Extraction

Ihor Stepanov, Oleksandr Lukashov, Mykhailo Shtopko, Vivek Kalyanarangan

Knowledgator Engineering, Kyiv, Ukraine · Baldor Technologies Pvt. Ltd. (IDfy), Mumbai, India

Relation Extraction KG Construction Named Entity Recognition Document Information Extraction

Open paper AskPDF arXiv

Joint named entity recognition (NER) and relation extraction (RE) is a fundamental task in natural language processing for constructing knowledge graphs from unstructured text. While recent approaches treat NER and RE as separate tasks requiring distinct models, we introduce GLiNER-Relex, a unified architecture that extends the GLiNER framework to perform both entity recognition and relation extraction in a single model. Our approach leverages a shared bidirectional transformer encoder to jointly represent text, entity type labels, and relation type labels, enabling zero-shot extraction of arbitrary entity and relation types specified at inference time. GLiNER-Relex constructs entity pair representations from recognized spans and scores them against relation type embeddings using a dedicated relation scoring module. We evaluate our model on four standard relation extraction benchmarks: CoNLL04, DocRED, FewRel, and CrossRE, and demonstrate competitive performance against both specialized relation extraction models and large language models, while maintaining the computational efficiency characteristic of the GLiNER family. The model is released as an open-source Python package with a simple inference API that allows users to specify arbitrary entity and relation type labels at inference time and obtain both entities and relation triplets in a single call. All models and code are publicly available.

Abstract

Relation Extraction

Explore similar work

GLiNER-Relex: A Unified Framework for Joint Named Entity Recognition and Relation Extraction

Calibrated Typed Decisions for Entity, Relation, and Event Extraction

LLM-based Atomic Propositions help weak extractors: Evaluation of a Propositioner for triplet extraction