Joint Entity-Relation Extraction

Updated 8 July 2026

Joint Entity-Relation Extraction is a unified task that detects entities and their semantic relationships from text, documents, and multimodal inputs.
It leverages methodologies such as table filling, span-based classification, and generative decoding to address challenges like cross-sentence reasoning and noisy supervision.
The approach minimizes error propagation through bidirectional interaction between entity recognition and relation prediction, enhancing extraction accuracy.

Joint entity-relation extraction (JERE) is the task of identifying entities and extracting the semantic relations that hold among them within a unified information-extraction framework. In its standard sentence-level form, the input is a sentence $X=(x_1,\dots,x_N)$ and the output is a set of relational structures such as $Y=\{((s_i,e_i),(s_j,e_j),r)\}$ , where $s_i,s_j$ are spans, $e_i,e_j$ are entity types, and $r$ is a relation type; in document-level formulations, the predicted objects become mentions, coreference clusters, cluster types, and relations between entity clusters; in multimodal variants, the output can be a set of quintuples over text-image pairs (Kong et al., 2023, Eberts et al., 2021, Yuan et al., 2024). Across these settings, the central research problem is not only how to avoid pipeline-style error propagation, but how to encode the mutual dependence between entity recognition and relation prediction under overlap, long-context reasoning, noisy supervision, and domain-specific schema constraints (Wang et al., 2023, Zhang et al., 2020).

1. Task scope and formal variants

Most JERE work has been organized around sentence-level extraction, but the output structure is not uniform. CARE formulates the problem over all possible continuous spans in a sentence and predicts entity labels for spans together with relation labels for pairs of predicted entities, returning relational triples in a classic sentence-level setting (Kong et al., 2023). In contrast, JEREX treats extraction as a document-level entity-centric problem: given a document, it predicts mention spans, coreference clusters, entity types for clusters, and relations between entity clusters, with relations defined at the entity-cluster level rather than the mention level (Eberts et al., 2021). KECPM extends the task further to joint multimodal entity-relation extraction, where each example is a text-image pair and the output is a set of quintuples $(e_1,t_1,e_2,t_2,r)$ (Yuan et al., 2024).

This variation in output space has methodological consequences. Sentence-level models can optimize over token pairs, span pairs, or generated sequences; document-level models must additionally solve cross-sentence aggregation and coreference; multimodal models must align textual entity-relation structure with visual evidence. A plausible implication is that “JERE” is best understood as a family of coupled structured-prediction problems rather than a single canonical architecture.

Document-level work also changes what counts as an entity. In JEREX, an entity is a cluster of co-referring mentions, and relation extraction is explicitly performed over entity clusters rather than individual mentions (Eberts et al., 2021). DocRED-FE pushes this further by arguing that realistic JERE also requires fine-grained entity typing, not merely document scope: DocRED already has 96 relation types but only 6 coarse entity types, and the paper treats this mismatch as a weakness for joint modeling (Wang et al., 2023).

2. Architectural paradigms and decompositions

The literature contains several distinct design patterns for JERE, differing mainly in how they represent triples, how tightly they couple entity and relation decisions, and where they place structural bias.

Paradigm	Representative papers	Core formulation
Decomposition and tagging	ETL-Span (Yu et al., 2019), BERT-based multi-head selection (Huang et al., 2019)	Head-conditioned sequence labeling or CRF-based NER plus multi-head relation selection
Table filling and token-pair classification	CARE (Kong et al., 2023), OneRel (Shang et al., 2022), PFN (Yan et al., 2021)	Word-pair or token-pair tables, often with relation-specific labels
Span-based classification	Two-phase paradigm (Ji et al., 2022)	Enumerate spans and span pairs, then classify existence and types
Generative or structured decoding	Seq2UMTree (Zhang et al., 2020), AERJE (Huang et al., 2023), KECPM (Yuan et al., 2024)	Generate triplets or verbalized structures as text
Document-level entity-centric modeling	JEREX (Eberts et al., 2021), similarity-based memory enhancement (Kosciukiewicz et al., 2023)	Mention detection, coreference, entity typing, and entity-level relations

A recurrent theme is that decomposition is not equivalent to pipelining. ETL-Span decomposes triplet extraction as

$p(h,r,t\mid S)=p(h\mid S)\,p(r,t\mid h,S),$

first extracting head entities and then, conditioned on each head, extracting tail entities and relation types (Yu et al., 2019). The two-phase span-based paradigm instead separates existence classification from fine-grained typing: candidate spans are first classified as Entity vs. Not-Entity and candidate ordered pairs as Relation vs. Not-Relation, then the surviving spans and pairs are typed in a second phase (Ji et al., 2022). Seq2UMTree attacks a different structural issue: it argues that flattened seq2seq generation imposes an unnecessary order on an unordered set of triplets, and therefore replaces sequence generation with an unordered multi-tree whose effective autoregressive depth is bounded by three decoding steps per triplet (Zhang et al., 2020).

OneRel represents the opposite extreme. It describes JERE as “one module in one step,” casts the task as fine-grained triple classification, and predicts a tensor $\mathbf{M}^{L\times K\times L}$ over token pairs and relation types using relation-specific horns tagging (Shang et al., 2022). AERJE and KECPM move the task into sequence generation, but do so under different conditions: AERJE uses dynamic prompts and a structured extraction language for API text (Huang et al., 2023), whereas KECPM verbalizes multimodal quintuples and augments few-shot extraction with retrieved prompts, auxiliary knowledge, and self-reflection (Yuan et al., 2024).

3. Interaction mechanisms inside joint models

A major line of work treats JERE as an interaction problem: how should entity-side and relation-side representations communicate without collapsing into feature confusion? CARE makes this explicit. It uses a parallel encoding strategy with two distinct MLPs for NER and RE, then a co-attention module that updates the two streams bidirectionally: $g_i^{e} = h_i^{e} + \sum_{j=1}^{n} \alpha_{ij} h_j^{r}, \qquad g_i^{r} = h_i^{r} + \sum_{j=1}^{n} \beta_{ij} h_j^{e}.$ This is implemented within a table-filling architecture and is motivated by the claim that entity information helps relation extraction while relation information also helps entity recognition and boundary disambiguation (Kong et al., 2023).

PFN makes a related but more granular claim: sequential encoders produce one-way interaction, while parallel encoders produce largely independent task features. Its solution is a partition filter network that splits encoder information into entity-specific, relation-specific, and shared partitions, then filters them into task memories such as

$\mu_e=\rho_e+\rho_s,\qquad \mu_r=\rho_r+\rho_s.$

The shared partition is evenly exposed to both tasks, while task-private partitions are shielded from the other task, and the paper argues that this provides “proper two-way interaction” (Yan et al., 2021).

TriMF extends this interaction view by introducing category memories for entity types and relation types together with multi-level memory flow attention. The model reads from both memories at the subword and word levels and combines them by

$Y=\{((s_i,e_i),(s_j,e_j),r)\}$ 0

then uses a trigger sensor to induce relation-trigger information without trigger annotations (Shen et al., 2021). DArtER pushes the granularity one step further by decoupling the encoder into subject, relation, and object streams, then applying inter-aggregation and intra-aggregation so that triple structure is modeled as interaction among subject-specific, relation-specific, and object-specific features rather than only between NER and RE as coarse tasks (Wang et al., 2024).

The document-level similarity-based memory-enhanced framework makes an analogous point in the JEREX setting. Built on top of JEREX, it writes category information from entity and relation classifiers into memory and reads that information back into token and span representations, thereby creating a feedback loop from later semantic decisions to earlier subtasks such as mention extraction (Kosciukiewicz et al., 2023). Taken together, these papers shift the meaning of “joint” from shared loss alone to explicit representation-level reciprocity.

4. Document-level and fine-grained JERE

Document-level JERE introduces a different set of bottlenecks: mention detection, coreference, entity typing across mentions, and cross-sentence relation evidence. JEREX addresses this by jointly modeling mention localization, coreference resolution, entity classification, and entity-level relation extraction, using multi-instance learning over mention pairs so that entity-level relation labels can be learned without mention-level relation annotations (Eberts et al., 2021). On DocRED, its Multi-Instance Relation Classifier reaches Ign F1 58.44 and F1 60.40 in the standard relation-classification setting, while on the paper’s end-to-end split the joint model reports 92.99 F1 for mention localization, 82.79 for coreference resolution, 80.10 for entity classification, and 40.38 for relation classification (Eberts et al., 2021).

DocRED-FE modifies this landscape by re-annotating part of DocRED with a two-level entity hierarchy of 11 coarse-grained and 119 fine-grained types while keeping DocRED’s 96 relation types (Wang et al., 2023). The dataset contains 2,596 documents, 516k tokens, 50,549 entities, 32,366 relations, and 96 relation types, and its annotation policy is contextual single-label typing rather than unrestricted multi-label ultra-fine typing (Wang et al., 2023). The paper argues that document-level JERE has been studied mostly with sentence-level or coarse-grained assumptions, and its experiments quantify the difficulty of lifting those assumptions. For JEREX, strict RE F1 drops from 40.26 on DocRED to 31.52 on DocRED-FE, relaxed RE F1 from 40.62 to 39.02, and NER F1 from 80.25 to 67.74; for REBEL, strict RE F1 drops from 45.29 to 37.51 while relaxed RE F1 changes only from 46.45 to 45.94 (Wang et al., 2023). The contrast between strict and relaxed evaluation shows that entity typing becomes a principal failure mode once the type ontology becomes realistic.

The same paper also shows that fine-grained entity types improve relation classification across CNN, LSTM, BiLSTM, GAIN, JEREX, and SSAN, and reports an information gain of $Y=\{((s_i,e_i),(s_j,e_j),r)\}$ 1 and information gain ratio $Y=\{((s_i,e_i),(s_j,e_j),r)\}$ 2 relative to original DocRED (Wang et al., 2023). This suggests that document-level JERE is not only a context-length problem but also a semantic-granularity problem: cross-sentence reasoning, coreference, and type specificity become entangled.

The similarity-based memory-enhanced document-level model confirms that the challenge is not purely architectural. On BioCreative V CDR it improves JEREX from 42.88 to 43.75, achieving state of the art, but on Re-DocRED it underperforms the baseline in some settings, and the paper attributes part of this instability to the memory warm-up proportion (Kosciukiewicz et al., 2023). Document-level JERE therefore remains sensitive both to representation design and to training dynamics.

5. Low-resource, noisy, and knowledge-augmented settings

A separate strand of work studies JERE when supervision is scarce or unreliable. DENRL targets distantly supervised JERE, where both entity labels and relation labels are noisy because they are projected from a knowledge base. Its solution combines a GPT-2-based joint tagging backbone with Bag-of-Word Regularization for relation-pattern supervision, Ontology-Based Logic Fusion for entity-relation consistency, and Self-Adaptive Learning for iterative instance filtering (Li et al., 2023). On NYT it reports 68.60 F1 in triplet evaluation and 67.87 in quadruplet evaluation; on Wiki-KBP it reports 46.29 and 45.65, respectively (Li et al., 2023). The central point is not merely the final score, but the explicit treatment of both relation noise and entity noise inside a unified JERE model.

SSDAU attacks robustness through data augmentation rather than denoising. It segments text according to entity-relation structure, performs entity semantic restructuring with hybrid similarity and contextualized embeddings, and filters augmented samples with BERTTopic-based consistency checking (He et al., 22 May 2026). On clean data it improves multiple JERE backbones, and under semantic ambiguity it reports an average F1 decrease of only 8.26% versus 31.91% for baselines (He et al., 22 May 2026). Because it is model-agnostic and explicitly preserves triple-level semantics, it reframes augmentation for JERE as structured recombination rather than general perturbation.

Few-shot settings have produced both multimodal and symbolic responses. KECPM studies few-shot JMERE and constructs few-shot train/dev sets from the original JMERE benchmark while preserving relation-category distribution. Its two-stage framework retrieves semantically similar prompts, asks GPT-3.5-turbo to generate auxiliary knowledge with self-reflection, selects the most relevant response, and feeds the resulting text into a seq2seq model (Yuan et al., 2024). In the main results it reaches 39.96 micro-F1 and 21.29 macro-F1, outperforming SpERT at 37.69 and 19.60 and EEGA at 36.44 and 17.43 (Yuan et al., 2024). The paper’s broader claim is that in few-shot multimodal extraction, image captions plus external knowledge can be more useful than heavy visual-feature fusion.

The LLM+ASP workflow takes a different low-resource route. It fine-tunes GPT-4o on only 10% of the original training data, converts predicted entities and relations into ASP atoms, and filters invalid relations through entity-existence and type-compatibility constraints (Tran et al., 18 Aug 2025). On SciERC it reports ER macro 35.37 versus 16.06 for ASPER, and the paper highlights this as a “2.5 times (35% over 15%) improvement” in the relation extraction task (Tran et al., 18 Aug 2025). The ASP stage mainly reduces false positive relations and is explicitly presented as elaboration tolerant: additional domain knowledge can be added as type_def(...) facts without changing the core ASP program (Tran et al., 18 Aug 2025).

Domain-specific generative JERE also appears in software text. AERJE reformulates API entity and relation extraction as prompt-conditioned sequence generation over Stack Overflow sentences, uses a BERT classifier to choose top- $Y=\{((s_i,e_i),(s_j,e_j),r)\}$ 3 candidate relation types for the prompt, and generates a structured extraction language with T5-v1.1-large (Huang et al., 2023). On its final test set it reports entity F1 96.51 and relation F1 81.20, and the dynamic prompt generator materially improves relation extraction over both static prompts and separated subtasks (Huang et al., 2023).

6. Evaluation practice, recurring misconceptions, and open questions

Evaluation in JERE is notably heterogeneous. CARE reports Micro-F1 for NER and RE, with partial matching on NYT and WebNLG because only the last word of an entity is annotated, and strict full-span matching on SciERC (Kong et al., 2023). DocRED-FE distinguishes Strict RE F1, which requires mention span, coreference resolution, entity type, and relation type to be correct, from Relaxed RE F1, which ignores entity type correctness (Wang et al., 2023). KECPM emphasizes both micro-F1 and macro-F1 because of strong relation-label imbalance in JMERE, and defines macro-F1 as

$Y=\{((s_i,e_i),(s_j,e_j),r)\}$ 4

over relation types (Yuan et al., 2024). These differences mean that leaderboard comparisons across papers are rarely interchangeable.

One recurrent misconception is that “joint” necessarily implies a single decoder or one-step extraction. The literature does not support that restriction. JEREX is explicitly joint because all four subtasks share a single BERT encoder and are trained together through a multi-task loss, even though relation extraction is downstream of predicted mentions and clusters (Eberts et al., 2021). The two-phase paradigm is likewise a joint model despite separating binary existence decisions from fine-grained typing (Ji et al., 2022). Conversely, OneRel’s “one module in one step” is only one endpoint in the design space, not the definition of jointness (Shang et al., 2022).

A second misconception is that relation extraction is the only subtask that benefits from joint modeling. PFN argues the opposite explicitly: its auxiliary experiments suggest that relation prediction is contributory to named entity prediction in a non-negligible way, and its robustness analysis reports a smaller average F1 decline than several relation-free NER baselines under perturbation (Yan et al., 2021). CARE, TriMF, and DArtER are built on the same premise, though with different interaction mechanisms.

A third misconception is that sentence-level coarse-grained benchmarks are adequate proxies for realistic JERE. DocRED-FE argues that most existing work still centers on sentence-level coarse-grained extraction and shows that moving to document-level fine-grained entity typing substantially degrades current baselines while making relation distributions more semantically interpretable (Wang et al., 2023). This suggests that future progress depends on tighter coupling of mention detection, typing, cross-sentence aggregation, and relation prediction.

The open questions named by the papers are correspondingly structural. CARE notes that overlap-specific analyses such as SEO and EPO are not reported (Kong et al., 2023). Seq2UMTree shows that local within-triplet order still matters even after removing order among triplets (Zhang et al., 2020). DocRED-FE leaves open what a genuinely strong document-level fine-grained joint model should look like (Wang et al., 2023). The similarity-based document-level memory model reports sensitivity to memory warm-up (Kosciukiewicz et al., 2023). The LLM+ASP workflow currently validates rather than repairs predictions, and its logical layer is limited mainly to entity-existence and type-compatibility constraints (Tran et al., 18 Aug 2025). Across these limitations, the field’s trajectory is clear: stronger JERE systems are being pushed toward finer ontologies, richer bidirectional interaction, more realistic document and multimodal settings, and explicit handling of noise, data scarcity, and schema knowledge.