Zero-Shot Relational Learning for Knowledge Graphs

This direction is merged into another direction: unseen relation generalization. Its last version stays readable and citable here.

Version 1 of 150 papers52 conceptsfrozen 2026-10-04

Traditional information extraction systems operate under a "closed-world" assumption: they are trained to recognize a fixed set of relationships, such as born_in or employed_by. However, the real world is dynamic. New types of interactions—from novel drug-protein bindings to emerging geopolitical alliances—appear constantly. When a system encounters a relationship type it hasn't seen during training, it typically fails because it lacks a mathematical representation (an embedding) for that specific label.

Zero-Shot Relational Learning aims to bridge this gap. Instead of treating relations as arbitrary ID numbers, this research direction treats them as semantic concepts that can be understood through natural language, logical frames, or structural analogies.

Reframing Extraction as Reading Comprehension

The most intuitive breakthrough in this field came from reframing the problem. If a model cannot classify a sentence into a pre-defined category like per:founder_of, perhaps it can simply "read" the sentence to find the answer to a question.

Omer Levy and colleagues pioneered this approach by reducing relation extraction to a Reading Comprehension (RC) task. By mapping a relation to a natural language question (e.g., "Who founded $x$?"), the model uses its general linguistic understanding to extract the "tail" entity. This allows the system to extract entirely new relations at test-time simply by providing a new question. This "machine comprehension" framing was later extended to multilingual settings by Mostafa Abdou and Isabelle Augenstein, who demonstrated that training on questions in one language helps the model generalize to unseen relations in others.

Synthesizing Embeddings from Descriptions

In Knowledge Graph (KG) completion, the challenge is more technical: how do you perform vector math on a relation that has no vector? Standard models like TransE or RotatE require a learned embedding for every predicate.

To solve this "cold start" problem, Pengda Qin and Xin Eric Wang proposed using Generative Adversarial Networks (GANs) to synthesize relation embeddings from raw text descriptions. If the model knows what "subsidiary of" means in English, the GAN can generate a synthetic vector that behaves like a real embedding in the KG's latent space.

More recently, researchers have turned to Large Language Models (LLMs) to perform this synthesis. Bo Xiong and Zifeng Ding introduced zrLLM, which uses LLMs to generate representations for unseen relations in Temporal Knowledge Graphs. By leveraging the semantic richness of the LLM, the system can forecast future events (e.g., a new type of diplomatic sanction) even if that specific relation never appeared in the historical training data.

Semantic Scaffolding and Frame Reasoning

While LLMs are powerful, they can struggle with the rigid structure required for Knowledge Graphs. Recent work focuses on providing "scaffolds" to guide their reasoning. Zehan Li and Fu Zhang developed FrameRTE, which uses Frame Semantics to help models identify relationship triplets. By breaking a relation down into its linguistic "frame" (the roles and participants involved), the model can generalize its understanding of how entities interact, even for novel categories.

Similarly, Manqing Dong and Chunguang Pan proposed MapRE, which uses "label-aware" mapping. Instead of treating a label as an opaque string, the model explicitly compares the semantic meaning of the relation name to the context of the sentence, creating a bridge that allows for zero-shot categorization.

Zero-Shot Generalization in Medicine

The stakes for zero-shot learning are highest in biology and healthcare, where we often need to predict interactions for experimental drugs or rare diseases that have no prior data. Ruth Johnson and Marinka Zitnik have shown that Graph AI can enable zero-shot predictions by analyzing the proximity of unknown entities to known ones in a massive biological knowledge graph.

For instance, the PlaNet framework, developed by Maria Brbic and Jure Leskovec, predicts the efficacy of novel drugs by reasoning over clinical knowledge graphs. By leveraging chemical similarities and protein-target networks, these models can generalize to experimental compounds that were entirely absent during the training phase, facilitating early-stage risk assessment and drug discovery.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_zero_shot_relational_learning_for_knowle_4ee2f0f1,
  title = {Zero-Shot Relational Learning for Knowledge Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/zero-shot-relational-learning-for-knowledge-graphs/rdr_4ee2f0f1164942ce9c8d27e3a611d770},
  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_zero_shot_relational_learning_for_knowle_4ee2f0f1_v1,
  title = {Zero-Shot Relational Learning for Knowledge Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/zero-shot-relational-learning-for-knowledge-graphs/rdr_4ee2f0f1164942ce9c8d27e3a611d770/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}