Introduction
Unsupervised Text-Graph Alignment (UTGA) aims to bridge the gap between structured knowledge graphs and natural language without relying on parallel datasets. This bidirectional task—converting graphs to text and vice versa—is fundamental for knowledge-grounded reasoning. However, a significant hurdle remains: Information Asymmetry.
In real-world scenarios, a text and its corresponding graph rarely contain identical information. For instance, a sentence might include descriptive "scene information" (e.g., "on a nice Sunday afternoon") that is absent from a logical graph. Conversely, a graph might encode complex logical rules (e.g., the definition of a grandmother) that are implicit in the text. Standard UTGA models often fail because they attempt to force a symmetric, one-to-one mapping between these asymmetric sources.
The ConstrainedBT Framework
To address this, researchers introduced ConstrainedBT, a framework that explicitly models asymmetric information using latent variable decomposition. The core idea is to categorize information into three distinct types of latent variables:
- Transferable Variable (TV): Information shared by both text and graph.
- Dependent Variable (DV): Non-transferable information that depends on the shared content (e.g., linguistic style).
- Independent Variable (IV): Non-transferable information independent of the shared content (e.g., abstract logical constraints).
Latent Constraints and Disentanglement
ConstrainedBT employs three Variational Auto-Encoder (VAE) based regularizations to manage these variables. The Transferable Constraint (TC) aligns shared info in a common latent space. The Dependent Constraint (DC) uses a Conditional VAE to model format-specific details. Finally, the Independent Constraint (IC) ensures that unique logical rules are preserved even when they aren't present in the text.
By enforcing a "Disentangling Condition," the model can effectively separate shared semantics ($z$) from format-specific nuances ($z_t, z_g$), allowing the generator to "imagine" missing details during the back-translation process.
Experimental Results
The model was evaluated on three benchmarks: Logic2Text (symmetric), LogicNLI (asymmetric logic), and CLUTRR (kinship reasoning). ConstrainedBT significantly outperformed existing unsupervised baselines like CycleGT and CycleCVAE, particularly on asymmetric tasks:
- LogicNLI: Achieved a Node F1 score of 65.9, a 12-point improvement over the best baseline.
- CLUTRR: BLEU-4 scores for text generation jumped from 17.2 to 23.9, demonstrating the model's ability to recover missing scene information.
- Downstream Impact: Using the learned latent variables as prompts for Natural Language Understanding (NLU) tasks improved LogicNLI accuracy from 51.3% to 74.4%.
Conclusion
ConstrainedBT demonstrates that information asymmetry is not a barrier to be removed, but a feature to be modeled. By decomposing latent spaces into transferable and non-transferable components, the framework achieves state-of-the-art performance in unsupervised alignment. Future work will focus on automating hyperparameter tuning for these constraints and integrating explicit symbolic rules directly into the latent space.
Summary updated