Low Resource Relation Extraction via Indirect Supervision

97 concepts89 papers

Supervised relation extraction (RE) models—systems designed to identify structured semantic connections between entities in text, such as founded_by(Apple, Steve Jobs)—are notoriously data-hungry. To train a model capable of accurately extracting these relationships across diverse domains, engineers traditionally require thousands of manually annotated sentences for every target relation. This dependency creates a severe bottleneck: human annotation is expensive, slow, and completely scales out of reach when moving to niche domains like biomedicine or attempting to capture the long tail of rare real-world relationships.

To overcome this data scarcity, researchers have shifted away from relying on brute-force manual labeling. Instead, the field has converged on a powerful insight: we can bypass the need for explicit human annotations by leveraging indirect supervision. By reframing relation extraction as other well-solved tasks, filtering noisy automated labels, engineering synthetic data diversity, and anchoring models in structural priors, we can train highly accurate relation extractors with minimal to zero human effort.


Task Reformulation: Borrowing Latent Priors

Instead of teaching a model a brand-new relation classification task from scratch, what if we map that relation to a task the model already knows how to solve? This is the core intuition behind task reformulation.

Pre-trained language models already possess massive, structured representations of syntax and semantics. By changing the interface of relation extraction to match the native pre-training objectives of these models, we can tap into these pre-existing priors.

For instance, Omer Levy et al. demonstrated in Zero-Shot Relation Extraction via Reading Comprehension that an abstract relation can be framed as a natural language question. If a model can answer the question "Who co-founded $x$?" given a text snippet, it has effectively extracted the founded_by relation without ever seeing an explicit relation extraction training example.

Standard RE:  [Steve Jobs] and Wozniak built [Apple]. -> Classify: "founded_by"
Reframed RC:  Context: "Steve Jobs and Wozniak built Apple." 
              Question: "Who co-founded Apple?" -> Answer: "Steve Jobs"

This paradigm extends to other indirect tasks. The authors of Label Verbalization and Entailment for Effective Zero and Few-Shot Relation Extraction transformed relation extraction into a textual entailment (Natural Language Inference) problem by converting abstract labels into simple sentences. Similarly, Keming Lu et al. introduced SURE in Summarization as Indirect Supervision for Relation Extraction, reframing relation extraction as a constrained summarization task.

Why it works

A model trained on millions of documents already understands how to answer questions, determine entailment, or summarize text. Mapping an unfamiliar, abstract relation label to a natural language prompt acts as a bridge, unlocking this latent knowledge and enabling zero-shot or few-shot generalization without requiring task-specific datasets.


Distant Supervision: Anchoring the Optimization Path

If human labels are unavailable, we can use existing knowledge bases (KBs) like Wikidata or Freebase to automatically label raw text—a technique known as distant supervision. If a KB states that (Steve Jobs, founded, Apple), we can find all sentences in a massive corpus containing both "Steve Jobs" and "Apple" and automatically label them as examples of the founded relation.

However, this introduces a massive amount of noise. As Jinjian Zhang et al. note in Combining MIML and Distant Supervision for KBP Slot Filling, a sentence might simply mention both entities without actually expressing the target relation (e.g., "Steve Jobs passed away years before Apple released the iPhone"). Left unchecked, this noise causes semantic drift, where the model begins to misclassify unrelated sentences because it was trained on false positives.

To combat this, researchers use dual-learning and reinforcement learning loops to filter out the noise. Hongtao Lin et al. developed DualRE in Learning Dual Retrieval Module for Semi-supervised Relation Extraction, which treats relation prediction and sentence retrieval as dual tasks that mutually reinforce and validate each other, discarding instances that fail the dual check.

To keep the model strictly grounded, Xuming Hu et al. proposed GradLRE in Gradient Imitation Reinforcement Learning for Low Resource Relation Extraction. This framework uses reinforcement learning to select pseudo-labels from unlabeled data by forcing the model's optimization path to strictly "imitate" the gradient path of a small pool of clean, gold-standard human labels.

Unlabeled Data ---> [Pseudo-Label Selection via RL] ---> Gradient Update
                                 |
       Must imitate the learning path of the clean data
                                 v
Clean Gold Data --> [       Gold-Standard Gradient       ] ---> Anchor Path

Why it works

Instead of blindly trusting noisy, automatically generated labels, these frameworks use a tiny pool of high-confidence data as an anchor. By enforcing mutual consistency or gradient imitation, the model can dynamically assess whether an automatically labeled sentence actually expresses the target relation, filtering out the false positives that cause semantic drift.


Generative Augmentation: Engineering Structural Diversity

With the advent of Large Language Models (LLMs), another intuitive path opened: why not just ask an LLM to generate synthetic training data for our target relations?

The catch is that LLMs suffer from a repetitive generator problem. When prompted to generate sentences for a specific relation, they tend to fall into high-probability "modes," outputting structurally identical templates. If you ask an LLM to generate 100 examples of a company acquisition, it might repeatedly output variations of "Company X acquired Company Y for Z million dollars." If a downstream model is trained on this data, it overfits to the template rather than learning the underlying semantic relationship.

To solve this, low-resource relation extraction frameworks focus heavily on diversity engineering. Yijie Zhong 0001 et al. introduced ODDA in ODDA: An OODA-Driven Diverse Data Augmentation Framework for Low-Resource Relation Extraction, which uses Monte Carlo Tree Search and explicit attribute constraints to maximize syntactic and inter-relation variability during generation.

Similarly, Zexuan Li et al. demonstrated in Generating Diverse Training Samples for Relation Extraction with Large Language Models that combining iterative In-Context Learning with Direct Preference Optimization (DPO) can force LLMs to generate highly diverse, non-repetitive training samples that significantly boost downstream model performance.

Why it works

A model's ability to generalize depends on seeing a robust decision boundary. By forcing synthetic data generators to vary their syntax, sentence length, and vocabulary, we ensure the downstream model learns the core semantic relationship rather than memorizing a shallow surface pattern.


Ontological and Structural Priors: The Safety Net

Finally, relations do not exist in a vacuum; they belong to broader semantic structures and hierarchies. When data is scarce, we can use these pre-existing structures as a form of regularization or smoothing to guide the model.

Congle Zhang et al. demonstrated this in Ontological Smoothing for Relation Extraction with Minimal Supervision with VELVET. Instead of treating each relation independently, the system maps a few labeled examples to complex queries over a knowledge base ontology, effectively "smoothing" the sparse data across related concepts in the hierarchy.

In a similar vein, Zhiyang Xu et al. introduced STF in Learning from a Friend: Improving Event Extraction via Self-Training with Feedback from Abstract Meaning Representation. During self-training on unlabeled data, the model uses Abstract Meaning Representation (AMR) graphs as a structural guide, providing a corrective feedback signal that prevents the model from becoming overconfident in its own mistakes.

Why it works

Language is highly variable, but the underlying logical structure of a relation (e.g., who did what to whom) remains constant. Grounding a low-resource model in a graph or an ontology provides a safety net where data is thin, preventing erratic predictions on unseen text by ensuring that predictions conform to known structural rules.

Key Foundational Work

Combining Lexical, Syntactic, and Semantic Features with Maximum Entropy Models for Information Extraction (2004) — contributed 2 concept elements to this direction.
Semi-supervised Relation Extraction with Label Propagation (2006) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_data_scarcity_in_relation_extraction_f5b4b127,
  title = {data scarcity in relation extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-scarcity-in-relation-extraction/rdr_f5b4b12770f948b8810b7e4a547c51ee},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_data_scarcity_in_relation_extraction_f5b4b127_v1,
  title = {data scarcity in relation extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-scarcity-in-relation-extraction/rdr_f5b4b12770f948b8810b7e4a547c51ee/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake