Bridging Structure and Semantics: Knowledge Graph Verification via Triple-to-Text Alignment
Organizations: Southeast University
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.
Figures & tables
| Corpus | ||||
|---|---|---|---|---|
| 86.12 | 88.28 | 86.55 | 88.35 | |
| 91.52 | 90.67 | 88.56 | 90.51 |
| Methods | ||||
|---|---|---|---|---|
| GenVer | 91.52 | 90.67 | 88.56 | 90.51 |
| DistMult | 81.63 | 79.24 | 81.17 | 83.33 |
| ComplEx | 84.46 | 89.68 | 87.20 | 89.83 |
| TransE | 87.39 | 86.20 | 85.27 | 85.63 |
| TransH | 88.73 | 85.24 | 83.80 | 83.31 |
| TransR | 84.43 | 69.98 | 62.28 | 59.80 |
| Subject | Object | Generated NLR | Target | Score | Case |
|---|---|---|---|---|---|
| Remember the Titans | United States of America | Remember the Titans opened strongly at the United States of America box office | 1 | 0.9999993 | TP |
| Bruce Cohen | United States of America | Bruce Cohen was ranked second in the United States of America | 0 | 0.0000003 | TN |
| The Grudge | Buenos Aires | “I finally found someone" by Buenos Aires and Bryan Adams | 0 | 0.9997579 | FP |
| Wild at Heart | France | Wild at Heart released the film in France | 1 | 0.0203743 | FN |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Property | ||||
|---|---|---|---|---|
| Entities | 4085 | 5680 | 7230 | 7621 |
| Relations | 2 | 4 | 8 | 16 |
| Triples | 142088 | 260640 | 460464 | 752164 |
| Statistic | ||||
|---|---|---|---|---|
| 0.56701 | 18.18729 | 0.08487 | 0.32198 | |
| 0.44968 | 11.36014 | 0.18546 | 0.26298 |
| Triple | Generated NLR (Score) | Generated NLR (ML) |
|---|---|---|
| S: George Lucas P: Award Nominee O: Djimon Hounsou | Djimon Hounsou was nominated for the George Lucas | Djimon Hounsou was nominated for George Lucas |
| S: Anjelica Huston P: Film Region Date O: Spain | Anjelica Huston said, “a huge line from Spain to Spain can be found in the us" | Anjelica Huston opened in Spain |
| S: Hugo Award P: Award Category O: Michael Swanwick | Michael Swanwick has received three writers guild of america awards, including nominations for an Hugo Award for Best Novel | Michael Swanwick has the most nominations for one golden globe and Hugo Award for Best Novel |
| S: Scarlett Johansson P: Award Nomination O: Best Female Lead | Scarlett Johansson appeared in the biopic “pollock" in which she won an Independent Spirit Award for Best Female Lead | Scarlett Johansson won the Independent Spirit Award for Best Female Lead |
| S: Leighton Meester P: Person Profession O: Singer-songwriter | Leighton Meester released the single “suddenly i feel love" with singer-songwriter | Leighton Meester attended the Professional Children’s School and began working as a singer-songwriter |
| S: Actor P: People Profession O: Bono | Bono is the second actor | Bono was such a lousy actor |
Knowledge Graphs
Knowledge graphs represent entities and relationships as structured, queryable networks for integrating data and supporting reasoning.
See all 251 papers