Learning Universal Schemas for Open Relation Extraction

This direction is merged into another direction: closed-world relation extraction. Its last version stays readable and citable here.

Version 1 of 118 papers18 conceptsfrozen 2026-10-04

Traditional Information Extraction (IE) operates on a "closed-world" assumption. To extract knowledge, a researcher must first define a fixed schema—a set of specific relationship types like PlaceOfBirth or Employer. While this works for structured databases, it fails to capture the "long tail" of human expression found in the wild. If a relationship isn't in your predefined list, the model is effectively blind to it.

A growing body of research is moving toward Open Relation Extraction, where the goal is to discover and represent relationships dynamically, without being tethered to a rigid, human-defined schema.

The Matrix Completion Intuition

The most significant hurdle in open extraction is that the same underlying relationship can be expressed in infinite ways. "X was born in Y," "X’s hometown is Y," and "X is a native of Y" all point to the same semantic fact.

In Latent Relation Representations for Universal Schemas, Sebastian Riedel and Andrew McCallum proposed treating this as a recommendation problem, similar to how Netflix predicts movie ratings. Imagine a massive matrix where rows are entity pairs (e.g., "Michael Bay, Transformers") and columns are potential relations—both structured ones from databases like Freebase and raw text patterns like "directed by."

By using matrix factorization, the model learns latent representations for these relations. If the model sees that entity pairs sharing the structured relation FilmDirector also frequently share the text pattern "is the director of," it learns that these columns are semantically similar. This allows the system to generalize to new, unseen text patterns by placing them in the same latent space as known relations.

Bridging Supervised and Unsupervised Knowledge

While universal schemas help align text to structured data, we often encounter entirely new relations that have no counterpart in any existing database. How do we identify a relationship we've never seen before?

Maosong Sun and Yuan Yao addressed this in Open Relation Extraction: Relational Knowledge Transfer from Supervised Data to Unsupervised Data. They introduced Relational Siamese Networks (RSNs), which learn a similarity metric from labeled data. Instead of learning to classify a relation as "Type A," the model learns what it means for two relations to be the same. Once trained, the model can look at two unlabeled text snippets and determine if they describe the same relationship, effectively clustering new, "open" relations into meaningful categories without human intervention.

Beyond Sentence Boundaries

Early IE systems were often "myopic," looking only at single sentences. However, complex relationships—especially in scientific domains—are often spread across multiple sentences or even entire documents.

In Simultaneously Self-Attending to All Mentions for Full-Abstract Biological Relation Extraction, Emma Strubell and Patrick Verga demonstrated that Transformer-based architectures (BRANs) can score all potential entity pairs in an abstract simultaneously. This global view is essential for open extraction because it allows the model to synthesize "view learning" concepts—a term coined by Jesse Davis and David Page in View Learning for Statistical Relational Learning—where complex features are discovered by looking at the relational structure across multiple records.

The Generative Shift

The most recent evolution in this field involves moving away from discriminative models (which pick from a list) to generative models (which describe what they see).

As noted in the survey by Pai Liu and colleagues, A Survey on Open Information Extraction from Rule-based Model to Large Language Model, Large Language Models (LLMs) are blurring the lines between open and standard IE. Rather than requiring a candidate list of relations in a prompt, newer frameworks like the one proposed by Wanli Li in Generative Meta-Learning for Zero-Shot Relation Triplet Extraction use meta-learning to master the task of extraction itself.

By treating data-fitting as a low-level task and generalization as a high-level goal, these models can extract "triplets" (Subject, Relation, Object) for relationships they have never encountered in training. This represents the ultimate goal of the field: a system that doesn't just fill in the blanks of a predefined form, but one that can read a text and autonomously map out the web of relationships it contains.

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_learning_universal_schemas_for_open_rela_ea9f52a2,
  title = {Learning Universal Schemas for Open Relation Extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-universal-schemas-for-open-relation-extraction/rdr_ea9f52a2acf44ab18608746aba719208},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_learning_universal_schemas_for_open_rela_ea9f52a2_v1,
  title = {Learning Universal Schemas for Open Relation Extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-universal-schemas-for-open-relation-extraction/rdr_ea9f52a2acf44ab18608746aba719208/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}