Bridging Structure and Semantics: Knowledge Graph Verification via Triple-to-Text Alignment

Ziang CuiarXiv 2026 · arXiv

TL;DR

GenVer is a new framework for verifying Knowledge Graph (KG) triples by "verbalizing" them into natural language. By training a generator to turn structured data into sentences and then using a dual-branch BERT classifier to score them, GenVer achieves state-of-the-art accuracy, even when external evidence is completely missing.

What Problem Does This Paper Address?

Knowledge Graphs (KGs) like Freebase and Wikidata are essential for modern AI, but they are often riddled with errors. Verifying whether a triple (subject, relation, object) is actually true is a major challenge.

Existing methods usually fall into two camps:

  1. Graph-based scoring: These look at the internal structure of the graph (e.g., TransE, ComplEx) but ignore the rich semantic knowledge stored in text.
  2. Evidence retrieval: These search the web or Wikipedia for sentences that confirm a triple. However, as the authors show in Figure 1(a), many triples simply don't have a corresponding sentence in existing corpora, leaving these methods "blind" to a large portion of the graph.
Figure 1: (a) Number of candidate NLRs retrieved per FB15k triple under the restricted search procedure. (b) GenVer uses generated text as a feature for triple classification.

Methods

The authors propose GenVer, a two-stage framework that bridges the gap between graph structure and natural language semantics.

1. Triple-to-Text Alignment & Generation

To train a model to "speak" KG triples, you first need a high-quality dataset of triples paired with sentences. The authors developed an alignment procedure that ranks candidate Wikipedia sentences based on four features:

  • Mention Coverage (MS): Does the sentence contain both the subject and object?
  • Sentence Length (SL): Shorter, more concise sentences are preferred.
  • Relation Relevance (RR): Does the sentence text match the relation label?
  • Pattern Typicality (PT): Does the sentence follow common linguistic patterns for that relation?

They then train a UniLM generator on these pairs. Once trained, this generator can produce a Natural Language Representation (NLR) for any triple, effectively creating "synthetic evidence" where none existed before.

2. Triple Verification Module

The verification step uses a BERT-based joint encoder that processes both the serialized triple and the generated NLR. It features two specialized extraction branches:

  • Structured Feature Extractor: Uses an energy-based gate (inspired by the TransE graph embedding method) to weight the [CLS] token representation.
  • Unstructured Feature Extractor: Uses a CNN over the generated NLR tokens to capture local semantic patterns.
Figure 2: GenVer overview. Triple–NLR alignment supplies training pairs for the UniLM generator. At inference time, BERT jointly encodes each triple and its generated NLR; the two feature extractors feed a classifier. The stages are trained separately.

Datasets

The models were evaluated on FactFB, a dataset derived from the well-known FB15k benchmark. The authors created four variants (FactFB2, 4, 8, and 16) based on relation frequency. Negative examples were generated by randomly perturbing entities or relations within positive triples.

Key Insights and Ideas

  • Generated Text as a Feature, Not Evidence: The authors are careful to note that the generated NLR is a "model-produced feature." It doesn't provide independent proof of truth, but rather allows the classifier to tap into the latent factual knowledge encoded within Pre-trained Language Models (PLMs).
  • The Power of Alignment: Learning the parameters for the triple-text alignment (using a small set of manually labeled pairs) significantly outperformed simple prior-based rules.
  • Semantic Drift: The case studies revealed that the generator can sometimes "hallucinate" or drift away from the triple's meaning (e.g., turning a "nomination" into a "win"). Interestingly, the classifier sometimes assigned high scores even to these unfaithful generations, highlighting a potential vulnerability.

Key Results

GenVer consistently outperformed a wide array of baselines, including pure graph embedding methods (TransE, ComplEx) and text-based KG representation models (SimKGC).

  • Accuracy: GenVer achieved classification accuracies between 88.56% and 91.52% across the FactFB variants.
  • AUROC: It reached a mean AUROC of 0.941, significantly higher than graph-only methods like TransD (0.562) or TransH (0.771).
Figure 4: Mean AUROC across the four FactFB variants, for the evaluated model configurations.

Limitations and Future Work

  • Faithfulness: The "semantic drift" mentioned above is a major hurdle. Future work could involve adding constraints to ensure the generator stays true to the input triple.
  • Negative Sampling: The current evaluation uses random perturbations. In the real world, KG errors might be more subtle, and the model's performance under more realistic error distributions remains an open question.
  • Open-World Assumption: A missing edge in a KG isn't necessarily a "false" assertion in reality. The current binary classification setup simplifies this nuance.

Summary updated

Abstract

Semantic accuracy is an important dimension of knowledge graph quality. Graph-based scoring methods exploit relational structure, while textual representations allow classifiers to draw on knowledge encoded by pre-trained language models (PLMs). We present GenVer, a generative verifier that scores triples using both their serialized form and generated natural language representations (NLRs). A triple-text alignment procedure selects training pairs for a sequence-to-sequence generator; a BERT-based classifier then combines features from each triple and its generated NLR. On four FactFB variants, constructed from FB15k using perturbed negative triples, GenVer obtains the highest classification accuracy among the evaluated baselines, ranging from 88.56% to 91.52%. The experiments suggest that generated representations can help triple classification, while the case analysis also exposes semantic drift and high-confidence errors.
BibTeX
@misc{cui2026bridging,
  author = {Ziang Cui},
  title = {{Bridging Structure and Semantics: Knowledge Graph Verification via Triple-to-Text Alignment}},
  year = {2026},
  eprint = {2611.07780},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.07780},
}
Figure