Relation Extraction (RE) is the backbone of knowledge graph construction, turning unstructured text into structured "triplets" like (Elon Musk, CEO of, SpaceX). Traditionally, this requires thousands of human-annotated examples for every specific relationship type. This "data hunger" creates a massive bottleneck: while we have plenty of text, we lack the expert labels needed to train models for niche domains, rare relations, or emerging topics.
To bypass this, researchers are developing techniques to extract relations with minimal or even zero human supervision. The core intuition is to bridge the gap between raw text and structured knowledge by "hallucinating" training data, reformulating the task into more common NLP problems, or using models to teach themselves.
Hallucinating Training Data
If human labels are expensive, why not let machines generate them? One of the most effective strategies for low-resource RE is Generative Data Augmentation (GDA).
Instead of relying on a small set of "seed" sentences, researchers like Xuming Hu have proposed frameworks like GDA that use generative models to restructure sentences and approximate patterns. Similarly, Benfeng Xu and colleagues introduced S2ynRE, which uses Large Language Models (LLMs) to synthesize domain-specific training data.
The intuition here is that an LLM already "knows" what a relationship looks like from its pre-training. By prompting it to generate diverse examples of a specific relation, we can create a massive synthetic corpus to fine-tune smaller, more efficient extractors. To ensure this synthetic data isn't just noise, frameworks like ODDA by Yijie Zhong use strategic loops to select only the most diverse and "model-appropriate" examples.
Reformulating the Task
Another powerful idea is to stop treating Relation Extraction as a unique classification problem and instead treat it as a task the model already understands, such as summarization or logical entailment.
- RE as Summarization: Keming Lu and Wenxuan Zhou proposed SURE, which frames extraction as a summarization task. Since summarization models are trained to distill the "essence" of a sentence, they are naturally good at identifying the core relationship between two entities.
- RE as Entailment: Oscar Sainz and his team demonstrated that you can turn a relation into a hypothesis. To find if "A was born in B," you ask the model if the sentence entails the statement "The birth place of A is B." This allows for zero-shot extraction because the model uses its general linguistic understanding of "entailment" rather than specific relation labels.
Self-Training and Semantic Drift
When we have a small amount of labeled data and a mountain of unlabeled text, we can use Semi-supervised Learning. The model predicts labels for the unlabeled data (pseudo-labels) and then trains on its own predictions.
The danger here is "semantic drift"—if the model makes a mistake, it learns from that mistake, leading to a feedback loop of noise. To solve this, Hongtao Lin and Xiang Ren developed DualRE, which treats extraction and retrieval as dual tasks that check each other. If the extraction module says "A is the CEO of B," the retrieval module must be able to find other sentences that support that relationship.
Other approaches, like MetaSRE by Xuming Hu, use meta-learning to train a "label generator" that specifically learns how to produce pseudo-labels that are most helpful for the final classifier, effectively filtering out the noise before it can cause drift.
Injecting Knowledge via Prompts
Finally, researchers are moving away from treating relation labels as abstract numbers (e.g., "Relation 42") and instead using their semantic meaning. KnowPrompt, developed by Ningyu Zhang and colleagues, uses "prompt-tuning" to inject knowledge about entities and relations directly into the model's input.
By using virtual words and "rationales"—the specific trigger words that signal a relation—models like RARE by Zhenyu Zhang can learn much faster. Instead of seeing a thousand examples of "born in," the model is told to look for "clues" like "native of" or "hometown," allowing it to reach high performance with a fraction of the data.