Traceable Context-Enriched Knowledge Graph Generation with Induced Schemas
Authors: Mohammad Sadeq Abolhasani, Yang Ba, Yixuan He, Rong Pan
Organizations: School of Computing and Augmented Intelligence, Arizona State University, Tempe, USA · School of Mathematical and Natural Sciences, Arizona State University, Tempe, USA.
Knowledge graph generation typically relies either on predefined ontologies or on schema-free extraction. Ontology-driven pipelines enforce consistent typing but require costly schema design and maintenance, whereas schema-free methods often produce fragmented graphs with weak global organization, especially in long technical documents with dense, context-dependent information. We propose TRACE-KG (Text-dRiven schemA for Context-Enriched Knowledge Graphs), a framework that jointly constructs a context-enriched knowledge graph and an induced schema without assuming a predefined ontology. TRACE-KG captures conditional relations through structured qualifiers and organizes entities and relations using a data-driven schema that serves as a reusable semantic scaffold while preserving full traceability to the source evidence. Experiments show that TRACE-KG produces structurally coherent, traceable knowledge graphs and offers a practical alternative to both ontology-driven and schema-free construction pipelines. The code, experiments and prompts are available at https://github.com/mosadddeq01/TRACE-KG.
Figures & tables
Figure 1: TRACE-KG pipeline. (1–2) Processing/Chunking : documents are textualized and chunked (scattered, unorganized shapes). (3) EntRec : local entity mentions are extracted but remain unresolved (e.g., duplicate pink rectangles). (4) EntRes : duplicate mentions are consolidated into resolved entities (e.g., one pink rectangle). (5) EntClsRec : resolved entities are grouped into entity classes (e.g., rectangles grouped together). (6) EntClsRes : entity classes are refined into broader class groups (e.g., rectangles and squares under a broader quadrilateral group). (7) RelRec : grounded relation instances are extracted with qualifiers (e.g., mixed solid single-headed, solid double-headed, and dashed single-headed arrows). (8) RelRes : relation labels are canonicalized and organized into relation classes and class groups (e.g., arrow types separated into solid single-headed, solid double-headed, and dashed single-headed arrows, then grouped into broader solid and dashed relation groups). (9) KG : resolved entities and relation instances form the context-enriched KG. (10) Schema : induced entity and relation hierarchies form the schema. Stages (4), (6), and (8) use semantic clustering, structured LLM actions, and deterministic execution.
Method
Ret.Acc
EGU ↑
RWA ↑
SCI ↑
Leak ↓
TriCR →1
Conn. ↑
AvgRank ↓
Schema-free extraction / post-hoc consolidation
OpenIE
56.0%
40.5%
41.5%
0.024
2.3%
3.201
74.0%
3.56
KGGen
63.6%
30.1%
30.1%
0.005
0.0%
0.494
46.1%
4.00
Retrieval-oriented graph baseline
GraphRAG
48.4%
45.1%
45.1%
0.146
0.0%
0.394
91.5%
2.67
Dynamic schema induction baseline
Table 1: Knowledge retention results on MINE-1, grouped by construction setting.
Figure 2: Failure modes behind raw retrieval accuracy on MINE-1. Left: Ret.Acc retained after discounting by connectivity and leakage. Middle: deviation of TriCR from the ideal value 1. Right: Ret.Acc versus EGU.
Table 4
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
Method
∣V∣
∣E∣
AvgEW
AvgDeg ↑
Conn. ↑
Clust. ↑
TRACE-KG
63
90
2.4
1.35
88.5%
0.199
GraphRAG
11
11
1.6
0.98
91.5%
0.150
AutoSchemaKG
107
105
6.6
0.98
61.5%
0.066
OpenIE
170
253
3.2
1.26
74.0%
0.027
KGGen
84
65
1.2
0.75
46.1%
0.011
Appendix
Table 4: Additional structural diagnostics complementary to Table 1 . ∣V∣ and ∣A∣ denote average node and edge counts; AvgEW is average words per entity. AvgDeg is ∣E∣/∣V∣ [ 31 ] ; Conn. is largest weakly connected component fraction [ 30 ] ; Clust. is average clustering coefficient [ 32 , 31 ] . Higher is better only for AvgDeg, Conn., and Clust.
Figure 3: Normalized multi-metric profile on MINE-1. Leak is inverted, and compression is expressed as proximity to TriCR = 1 so that higher values are preferable on all axes.
Figure 4: Effective Graph Utilization (EGU) on MINE-1. Error bars indicate variability across benchmark instances.
#
Ontology
#Types
#Relations
#Sentences
1
University
15
46
156
2
Music
15
35
290
3
Airport
14
39
306
4
Building
14
38
275
5
Athlete
17
37
293
6
Politician
19
40
319
Appendix
Table 5: DBpedia-WebNLG statistics for Experiment 2.
Level
Relations
Concepts
Finest level
64.2%
50.8%
Middle level
28.7%
35.4%
Coarsest level
7.1%
13.8%
Appendix
Table 6: Best-matching TRACE level among accepted alignments (macro-averaged, Combined).
Metric
Separate
Combined
Reduction
Pages / chunks
83 / 339
83 / 339
–
Entity mentions
638
638
–
Raw relation candidates
882
847
4.0%
Resolved entities
403
285
29.3%
Entity classes
124
64
48.4%
Canonical relations
229
156
31.9%
Appendix
Table 7: Cross-document consolidation on the ISO 55000-family corpus. Processing the three standards separately (summing outputs) versus jointly as one corpus. Reduction reflects vocabulary compression from cross-document consolidation.
Axis
Outcome
Count
Rate
Precision
Correct (valid merge / keep / modify)
222/249
89.2%
False merge (distinct entities merged)
11/249
4.4%
Incorrect modify (attribute distortion)
16/249
6.4%
Recall
Missed duplicates (should have merged)
9/120
7.5%
Appendix
Table 8: Entity resolution audit on the ISO 55000-family corpus (manual inspection). Precision is measured over audited resolution actions; recall is measured over sampled resolved entities.
Setting
Ret.Acc
Component ablation
Full TRACE-KG
91.4%
w/o qualifier extraction
82.4%
w/o iterative entity res.
84.3%
w/o relation canonicalization
86.0%
w/o schema induction
88.2%
Appendix
Table 9: Component ablation and backbone sensitivity on 25 MINE-1 articles (375 fact checks).
Group
Element
Grounded
Over-gen.
Distorted
Halluc.
Success
Head
30/32
2/32
0/32
0/32
Relation
28/32
2/32
2/32
0/32
Tail
30/32
2/32
0/32
0/32
Subtotal
88/96
6/96
2/96
0/96
Failure
Head
27/32
3/32
1/32
1/32
Relation
23/32
3/32
4/32
2/32
Appendix
Table 10: Faithfulness audit on 25 MINE-1 articles (GPT-5 backbone), by manual inspection. We sample 64 relation instances—32 from successful and 32 from failed fact checks—and judge each triplet element (head, relation, tail) against its cited provenance chunk.
Metric
Raw (0.61M chars)
Normalized (per 1M chars)
LLM calls
876
1,436
Input tokens
3.60M
5.90M
Output tokens
0.42M
0.69M
Wall-clock time
2.8 h
4.6 h
Appendix
Table 11: End-to-end computational cost of TRACE-KG on the combined ISO multi-document experiment (full pipeline: recognition, schema induction, resolution). Normalized figures are per 1M source characters.
Setting
Entities
Relations
Non-text facts
Text only
214
336
0
Text + VLM
241
389
47
Appendix
Table 12: Effect of multimodal ingestion on the SEMI case-study subset.
Figure 5: Overview of the constructed context-enriched knowledge graph for the case-study document.
Figure 6: Induced entity and relation schema hierarchy for the case-study document.
Figure 7: Traceability example (figure grounding): KG node linked to source diagram.
Figure 8: Traceability example (equation grounding): KG node linked to source equation.
Figure 9: Example resolved entity with schema, confidence, and provenance.
Figure 10: Example context-enriched relation with qualifiers and provenance.
Knowledge graph construction turns text, tables, images, and databases into structured entities and relationships that machines can query and reason over.