Document-level relation extraction means reading a whole article and deciding, for every pair of entities it mentions, which relation (if any) holds between them. It is expensive to annotate. The number of entity pairs grows with the square of the entity count, over 96% of pairs have no relation at all, and about 40% of true facts require linking information across sentences (DocRED). With only 3% of DocRED's training documents labeled, a fine-tuned BERT-base reaches 10.44 F1. The best low-resource fix in our file, a calibration method, lifts that to 37.02 (PRiSM). Meanwhile, graph-based label propagation copies labels from a few annotated examples to similar unlabeled ones by diffusing them along graph edges. It works well on citation networks, but in document relation extraction it has only been tried on flat similarity graphs over sentence-level candidates.
We propose to propagate relation labels over a corpus-wide graph whose nodes are entity pairs, using two kinds of edges. Semantic edges connect similar pairs across documents and carry labels from annotated documents to unannotated ones. Structural edges come from the internal layout of each document, such as shared entities, shared sentences, and coreference, and enforce consistency among the pairs of a single document. The central difficulty is sparsity. The inter-sentence pairs that most need help are exactly the ones that sit structurally isolated inside their document. We repair this with edge induction that must be corroborated by structure, and we stop the dominant "no relation" class from washing out the rare classes. Semantic similarity carries labels between documents; document structure decides which of those labels to trust.
Core Claim
In low-resource document relation extraction, the errors of cross-document similarity propagation are mostly structurally inconsistent. A pair gets a relation label that contradicts what its document's other pairs imply. A second channel of propagation along intra-document structure, with induced edges admitted only when document structure corroborates them, should produce pseudo-labels that are more precise, especially for inter-sentence pairs. Training a standard extractor on these pseudo-labels should beat both calibration alone and flat similarity propagation.
The claim also addresses a tension in the literature. Scale-Free Graph-Language Models (SFGL) builds a k-nearest-neighbor graph from text cosine similarity. It found that at ultra-low label rates this induced graph beats the real citation graph, because real graphs have isolated components that trap labels. HGDL argues the opposite. It keeps an edge only when topology and feature similarity agree ("intersection sparsity"), to suppress noise. We predict that each view is right in a different place. Aggressive induction helps where structure is absent, which in our graph means inter-sentence pairs. Intersection filtering helps where structure exists. The experiments are designed to show this split directly.
Research Program
Graph construction. The graph is built from a document-graph encoder in the style of DHG, which separates a word/sentence structure layer from an entity/mention reasoning layer. Here we use only its entity-pair representations, not its fully supervised training. Every candidate pair in every training document, labeled or not, becomes a node.
Cross-document semantic edges link each pair to its $k$ nearest pairs by cosine similarity. We use cosine rather than Euclidean distance following SFGL's finding that cosine is what produces usefully connected text graphs. The edges are further filtered by agreement between the two pairs' entity-type signatures. This is the only channel through which labels leave annotated documents, as in Jointprop.
Intra-document structural edges link pairs that share an entity and co-occur in a sentence or coreference chain. This follows the observation of Sahu et al. and Christopoulou et al. that sentence and coreference hubs create short paths between distant mentions.
Edge induction for isolated pairs. Pairs whose entities never share a sentence often have no structural edges at all. For these we borrow the method of Improving Word Usage Graphs with Edge Induction. That paper predicts missing edges in sparse lexical graphs from a "log-triangle" feature, which counts length-two paths between the two endpoints, combined with contextual embedding similarity, and showed that the enriched graphs cluster better. Our analogue is a small logistic model trained on labeled documents. It predicts whether two same-document pairs have linked labels, for example (A,B) and (B,C) against (A,C), from triangle counts through shared entities and from contextual similarity.
Induced edges are admitted only above a precision threshold measured on held-out labeled documents. That paper also reports that filling edges iteratively cascades errors, so we induce once and never iterate.
Propagation and training. Following On the Equivalence of Decoupled Graph Convolution Network and Label Propagation, which shows that decoupled graph convolution amounts to "propagate labels, then train a classifier on the soft labels," we precompute the soft labels in closed form:
$$F = (1-\alpha)\,(I - \alpha\,[\lambda S_{\text{sem}} + (1-\lambda) S_{\text{str}}])^{-1} Y.$$
In this equation, $Y$ holds the one-hot gold labels of annotated pairs (zeros elsewhere). $S_{\text{sem}}$ and $S_{\text{str}}$ are the symmetrically normalized semantic and structural adjacency matrices. $\lambda$ weights the two channels, $\alpha$ controls how far labels spread, and $F$ holds the resulting soft labels.
To handle the no-relation class, we treat it as an absorbing sink rather than a diffusing label, so that it cannot spread across the graph. We then rescale each relation column of $F$ so that its total mass matches the class frequency in the labeled documents. Unconstrained propagation otherwise lets the dominant class wash out the rare ones, a known failure mode of label propagation on imbalanced tasks.
The extractor, a BERT-base pair classifier with PRiSM's calibration head built from the PRiSM codebase, is trained on gold labels plus pseudo-labels above a confidence threshold. An optional second round re-encodes pairs with the trained model and rebuilds the semantic edges. This is a lightweight version of GLEM's alternation (GLEM), used only if the first round gains.
Baselines
The naive reference point is BERT-base fine-tuned on the 3% labeled split. PRiSM reports 10.44 F1 for it on DocRED. We rerun it on our splits.
The strongest existing method is PRiSM itself, which reports 37.02 F1 on 3% DocRED with BERT-base. It is the right competitor because it attacks exactly the low-resource imbalance problem, it uses the same encoder, and its code is public. We run it with its released hyperparameters and then give it the same tuning budget as our method, measured in trials on the dev set. Beating it means a gain over 37.02 F1 (or over our reproduced number, if that differs) that holds across at least three random 3% splits. We will also report Re-DocRED and DWIE, where no reported result was found on Lacuna for the low-resource setting, so our reproduced PRiSM run is the reference there.
A strong graph competitor is Jointprop-style propagation. Jointprop runs random-walk propagation over a cosine k-nearest-neighbor graph of span and span-pair candidates and reports +15.89 RE F1 on SciERC at 5% labels. Its result on document-level extraction is untested (no reported result found on Lacuna for DocRED). We reimplement it with our pair encoder, our $k$, and the same pseudo-label threshold, so the only difference is the absence of structural edges.
Two further graph baselines isolate the induction question. The SFGL-style variant adds unconstrained cosine edges everywhere, with no corroboration. The HGDL-style variant keeps only edges where structure and similarity agree, with no induction for isolated pairs.
Ablations remove, one at a time, the structural channel ($\lambda=1$), the induced edges, the entity-type filter, and the no-relation sink with class-mass rescaling. The key ingredient is structure-corroborated induction. Its ablation, induced edges removed, must lose most on inter-sentence pairs.
We also report fully supervised DHG (65.9 F1 on CDR, 83.1 F1 on GDA, from DHG) as a ceiling, not a competitor.
Why not just…?
Why not just use PRiSM with ordinary confidence-based self-training? This is the cheapest serious alternative. PRiSM is already well calibrated (it reduces expected calibration error from 0.359 to 0.010 on 3% DocRED), and calibrated confidences might select pseudo-labels as well as a graph does. It is a real threat. We include PRiSM plus self-training at matched thresholds and matched pseudo-label volume. If our pseudo-labels are not more precise on inter-sentence pairs at equal recall, the graph machinery is not justified and we will say so.
Why not just propagate over a flat k-nearest-neighbor graph of pairs, as Jointprop does? Flat propagation transfers labels between similar pairs but cannot notice when a pair's label contradicts its own document. Our hypothesis is that exactly these errors dominate. The Jointprop reimplementation is the direct test. We will count how many of its false-positive pseudo-labels are overturned by the structural channel. If that fraction is small, the structural edges add little.
Why not just add as many similarity edges as possible, as SFGL does? SFGL's evidence comes from homogeneous citation graphs where every node is a whole document. Pair nodes inside documents are denser and noisier, and DHG reports that over-connected document graphs over-smooth entity distinctions. The SFGL-style baseline, stratified by whether a pair is structurally connected, settles where unconstrained induction helps and where it hurts.
Why not just train an LM and a heterogeneous GNN jointly by EM, like GLEM? GLEM exchanges pseudo-labels between a language model and a graph network on homogeneous graphs. Extending it to multi-level document graphs is heavier, and it does nothing specific about isolated pairs or the dominant no-relation class. We keep a single re-encoding round as an optional extension. The core question is answerable without it.
Proposed Experiments
The main benchmark is DocRED at a 3% labeled rate (about 100 documents), over three random splits, with the remaining training documents unlabeled. We add Re-DocRED, whose false-negative corrections matter for pseudo-label precision, and DWIE for longer documents. For domain transfer we use CDR and GDA with 5% labels. Metrics are relation F1 and ECE. For the pseudo-labels themselves we measure per-class precision and recall against the hidden gold labels, plus a no-relation leakage rate: the fraction of rare-class gold pairs whose pseudo-label becomes "no relation."
The first decisive check is a stratified analysis. We partition unlabeled pairs into intra-sentence pairs, structurally connected inter-sentence pairs, and structurally isolated inter-sentence pairs. The claim predicts three things. Induction gains should concentrate in the isolated stratum. Intersection filtering should gain in the connected stratum. Flat propagation should be worst on inter-sentence pairs overall. If gains are uniform across strata, structure is not doing the work we claim.
The second check sweeps the induction precision threshold and $\lambda$. It traces how pseudo-label precision trades against coverage and reports the regime in which the SFGL view or the HGDL view wins.
The third check varies the label rate from 1% to 10%. The gain over PRiSM with self-training should shrink as labels grow, because the graph should matter most when labels are scarcest. If the gain does not shrink, we would suspect the improvement comes from extra training signal rather than from structure.
References
- DocRED: A Large-Scale Document-Level Relation Extraction Dataset
- PRiSM: Enhancing Low-Resource Document-Level Relation Extraction with Relation-Aware Score Calibration
- Scale-Free Graph-Language Models
- HGDL: Heterogeneous Graph Label Distribution Learning
- Document-level Relation Extraction with Dual-tier Heterogeneous Graph
- Jointprop: Joint Semi-supervised Learning for Entity and Relation Extraction with Heterogeneous Graph-based Propagation
- Inter-sentence Relation Extraction with Document-level Graph Convolutional Neural Network
- Connecting the Dots: Document-level Neural Relation Extraction with Edge-oriented Graphs
- Improving Word Usage Graphs with Edge Induction
- On the Equivalence of Decoupled Graph Convolution Network and Label Propagation
- Learning on Large-scale Text-attributed Graphs via Variational Inference