On the Representation and Embedding of Knowledge Bases beyond Binary Relations

Jianfeng Wen, Jianxin Li, Yongyi Mao, Shini Chen, Richong Zhang

TL;DR

Most knowledge base embedding models assume all relationships are binary (pairs), but real-world data is often multi-fold (n-ary). This paper demonstrates that forcing n-ary data into binary triples—a common practice in datasets like FB15K—causes significant information loss. The authors propose m-TransH, a model that embeds multi-fold relations directly, achieving a massive performance leap (over 80% Hit@10) compared to traditional triple-based methods.

What Problem Does This Paper Address?

Knowledge Base (KB) embedding involves representing entities and relations as vectors in a continuous space to perform tasks like link prediction or question answering. However, almost all existing models (like TransE or TransH) are built on the "binary assumption": the idea that every fact involves exactly two entities (e.g., Beijing is the capital of China).

In reality, a huge portion of knowledge is non-binary. For example, the fact "Benedict Cumberbatch played Alan Turing in the movie The Imitation Game" involves three entities: an actor, a character, and a movie. Current methods handle this by breaking n-ary relations into multiple binary triples (a process called "Star-to-Clique" conversion). The authors argue this approach is fundamentally flawed because it is irreversible, loses structural context, and introduces "noise" that hinders the learning process.

Methods

The authors advocate for a direct modelling framework rather than a "decomposition" framework.

  1. Canonical Representation: They define an "instance representation" for multi-fold relations where each instance is a function mapping specific roles (like ACTOR or MOVIE) to entities.
  2. m-TransH: This is a generalization of the popular TransH model. While the original TransH calculates the distance between two projected entity vectors, m-TransH calculates a cost function based on a weighted sum of projections for all entities involved in the n-ary relation.
  3. It projects entities onto a relation-specific hyperplane.
  4. It uses a weight vector ($a_r$) to determine the contribution of each role.
  5. The goal is for the weighted sum of these projected entities to align with a relation-specific translation vector ($b_r$).
  6. Efficiency: By modeling the n-ary relation directly, the model avoids the combinatorial explosion of triples that occurs when you try to turn a 5-fold relation into all possible pairs.

Datasets

  • FB15K: The authors analyze this popular benchmark and point out its flaws—specifically that it is the result of a lossy "Star-to-Clique" conversion of Freebase data.
  • JF17K (New): The authors introduce a new dataset derived from Freebase that preserves the original multi-fold structure. It contains 17,629 entities and 181 multi-fold relation types, filtered to ensure entities have enough connectivity for meaningful embedding.

Key Insights and Ideas

  • The "Binary Trap": Over 1/3 of entities in Freebase participate in non-binary relations. Treating these as simple triples is like trying to describe a 3D object using only 2D shadows; you lose the "depth" of the relationship.
  • Irreversibility: The paper mathematically notes that once you convert an n-ary relation to a set of triples (S2C conversion), you cannot perfectly reconstruct the original n-ary relation. This proves that triple-based datasets like FB15K are inherently degraded.
  • Direct vs. Decomposition: Existing models on triple-based data are actually just "decomposing" the cost function. By switching to "direct" modeling, you capture the joint interaction of all entities at once.
  • Fact-Level Info: The authors found that including "Fact IDs" (treating the specific instance itself as an entity) can further improve accuracy, suggesting that knowing which specific instances belong to the same event is valuable.

Key Results

The performance gains are "astonishing," according to the authors:

  • Accuracy (Hit@10): On the JF17K dataset, m-TransH achieved a Hit@10 of over 80%. For comparison, the standard TransH model (using the triple-based approach) hovered around 50%.
  • Mean Rank: m-TransH significantly reduced the Mean Rank (the average position of the correct entity in a list of guesses), showing much higher confidence and precision.
  • Complexity: m-TransH is computationally more efficient. Because it doesn't need to process the inflated number of triples generated by S2C conversion, it trains faster and uses fewer parameters to achieve better results.

Limitations and Future Work

  • Fact ID Bottleneck: While adding Fact IDs (m-TransH:ID) improved Hit@10, it actually worsened the Mean Rank when the embedding dimension (DIM) was low. This is because the model had to embed tens of thousands of unique Fact IDs, exhausting the "degrees of freedom" in low-dimensional spaces.
  • Dataset Scale: While JF17K is a step forward, it is still a filtered subset. Testing these direct modeling approaches on the full, massive scale of Freebase or other KBs remains a challenge.
  • Generalization: The authors suggest that their direct modeling framework could be applied to other "Trans-series" models (like TransR or PTransE), which could lead to even further state-of-the-art improvements.

Summary updated

Abstract

The models developed to date for knowledge base embedding are all based on the assumption that the relations contained in knowledge bases are binary. For the training and testing of these embedding models, multi-fold (or n-ary) relational data are converted to triples (e.g., in FB15K dataset) and interpreted as instances of binary relations. This paper presents a canonical representation of knowledge bases containing multi-fold relations. We show that the existing embedding models on the popular FB15K datasets correspond to a suboptimal modelling framework, resulting in a loss of structural information. We advocate a novel modelling framework, which models multi-fold relations directly using this canonical representation. Using this framework, the existing TransH model is generalized to a new model, m-TransH. We demonstrate experimentally that m-TransH outperforms TransH by a large margin, thereby establishing a new state of the art.
BibTeX
@inproceedings{wen2016representation,
  author = {Jianfeng Wen and Jianxin Li and Yongyi Mao and Shini Chen and Richong Zhang},
  title = {{On the Representation and Embedding of Knowledge Bases beyond Binary Relations}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
  url = {https://openreview.net/forum?id=ryEgQ4fuZH},
}

Loading related content...

Figure