Scientific data selection is typically driven by relevance or utility scores, providing limited insight into why a document is retained. We introduce SciGraphSelect, a reference-grounded framework for auditable scientific data selection. SciGraphSelect organizes domain-specific knowledge, reusable reasoning operations, and reference-attested relations into a typed knowledge--reasoning graph, then scores candidate documents using concept coverage, relation and path co-support, and directional alignment. To avoid over-interpreting document embeddings, rich relational structure is preserved on the reference side while candidate-side evidence is treated conservatively without assuming relation entailment. By exposing the reference-grounded evidence behind each selection, SciGraphSelect brings traceability to the upstream construction of scientific training corpora and complements downstream scientific verification.
Figures & tables
Figure 1: Overview of SciGraphSelect . Reference-derived knowledge–reasoning structure is matched against candidate documents to compute structure-aware selection scores and expose the evidence supporting each ranked document.
Takeaway 1. Reasoning becomes scientifically informative only when grounded in domain-specific knowledge; K→R explicitly represents this grounding.
Table 2
Takeaway 2. SciGraphSelect preserves relations supported by scientific references, while making only conservative claims from candidate-document evidence.
Table 3
Heterogeneous Hard-Negative Validation
PeS2o Scientific-Paper Evaluation
Method
P@1%
P@5%
P@10%
AUPRC
AUROC
P@1%
P@5%
P@10%
AUPRC
AUROC
Chemistry
Embedding similarity
.7681
.8197
.7621
.727481
.956629
.8229
.7139
.5833
.601362
.921411
Concept only
.7965
.8666
.7997
.755068
.957812
.8921
.7722
.6129
.647915
.923127
+ Typed relation co-activation
.7992
.8629
.7904
.744855
.953926
.9046
.7796
.6172
.655911
.923901
+ Reasoning/path
.7951
.8442
.7657
.719709
.948361
.9000
.7749
.6159
.653231
.924460
Table 1: Scientific document selection under heterogeneous hard negatives and the 15.1M PeS2o scientific-paper pool. Chemistry uses the largest directional margin, while Biology averages the three largest margins. Bold indicates the best result within each domain, setting, and metric.
Transition
Retained
Entered TP ↑ /FP ↓
Exited TP ↓ /FP ↑
Δ TP ↑
Turnover
Chemistry
Embedding Similarity → Concept
93,352
51,259 / 6,513
40,800 / 16,972
+10,459
38.23%
Concept → Relation co-activation
124,607
23,670 / 2,847
21,783 / 4,734
+1,887
17.55%
Relation co-activation → Reasoning/path
129,516
18,202 / 3,406
18,897 / 2,711
− 695
14.30%
Structure-aware → Final direction-aware
102,992
46,732 / 1,400
37,137 / 10,995
+9,595
31.85%
Biology
Table 2: Top-1% selection transitions over the common PeS2o pool. TP/FP are defined using PeS2o source-domain proxy labels. Arrows indicate the preferred direction when applicable; turnover measures set change rather than quality.
Figure 2: Structured scoring changes both where and what is selected. (a) Local graph evidence for a Chemistry document selected only by the structure-aware score. (b) Joint UMAP of embedding-only, shared, and structure-aware-only selections with active knowledge and reasoning nodes. UMAP is used only for visualization; selection and relation similarities are computed in the original embedding space.
Domain
Embedding Similarity-only ↑
SciGraphSelect-only ↑
Δ↑
Chemistry
.821
.853
+.032
Biology
.915
.963
+.049
Table 3: Human relevance of documents uniquely selected by embedding similarity or SciGraphSelect. All metrics are higher-is-better.
Domain
Rel. ↑
Concept ↑
Reason. ↑
Relation ↑
κ↑
Chemistry
.837
.524
.521
.426
.691
Biology
.939
.890
.789
.503
.719
Table 4: Human evaluation of SciGraphSelect attribution correctness. All metrics are higher-is-better.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Biology
Chemistry
References
38,217
146,871
Knowledge nodes
35,822
197,462
Reasoning nodes
20
23
Backbone edges
5,716
53,686
Attested paths
6,320
56,711
Appendix
Table 5: Statistics of the domain-matched reference graphs.
Graph element
Count
Knowledge nodes
197,462
Reasoning nodes
23
Backbone edges
53,686
K→K edges
12
K→R edges
53,650
R→R edges
24
Appendix
Table 6: Composition of the Chemistry typed graph.
Biology
Chemistry
Knowledge aggregation
family MPool 8
top-16 mean
Reasoning aggregation
gated MPool 3
top-5 mean
Edge aggregation
target MPool 8
top-16 weighted mean
Path aggregation
reliability + MPool 8
top-16 mean
Directional Top- k
3
1
Directional weight η
0.30
0.10
Appendix
Table 7: Domain-specific graph-score aggregation and directional-fusion settings used by SciGraphSelect.
Target domain
Positive
Negative
Prevalence
Biology
5,193,143
9,919,169
34.36%
Chemistry
1,482,270
13,630,042
9.81%
Appendix
Table 8: Composition of the full PeS2o paper evaluation.
Figure 3: Blinded human-evaluation interface. Annotators first read the candidate document and assess its relevance to the target scientific domain before seeing any SciGraphSelect attribution. They then evaluate the support for attributed knowledge concepts, reasoning operations, and relations. Relevance judgments are fixed before attribution labels are revealed.
Domain
Embedding-only
SciGraphSelect-only
Both
Chemistry
25
25
10
Biology
25
25
10
Appendix
Table 9: Human relevance sample per domain.
Domain
SciGraphSelect-only
Both
Chemistry
20
10
Biology
20
10
Appendix
Table 10: Human attribution sample per domain.
Figure 4: Biology selection geometry and local graph evidence. (a) Local evidence for a Biology document selected only by the structure-aware score. Supporting knowledge and reasoning nodes are connected to the document, and the solid arrow marks a reference-attested knowledge-to-reasoning path. (b) Joint UMAP of documents selected only by embedding similarity, by both methods, or only by the structure-aware scorer, together with activated knowledge and reasoning nodes. The circled point corresponds to the example in (a). UMAP is used only for visualization; document selection and graph matching are performed in the original embedding space.
Domain
Comparison
Δ AUPRC
Biology
True topology – rewired edges
+0.00361
Biology
True topology – shuffled paths
+0.00134
Chemistry
True topology – rewired edges
+0.00050
Chemistry
True topology – shuffled paths
+0.00014
Appendix
Table 11: Topology-placebo results. Values denote the AUPRC of the true reference topology minus that of the corresponding placebo.
Removal
n
ΔSK
ΔSR
ΔSgraph
ΔSfinal
Selectivity
Chemistry
Shared control
50
−0.03162
−0.04094
−0.03460
−0.000152
–
Attributed concept
50
+0.01000
−0.03190
+0.01859
+0.000046
+0.000198
Attributed reasoning
50
+0.01164
−0.02518
+0.01633
+0.000044
+0.000196
Biology
Shared control
50
−0.00428
−0.03216
−0.00630
+0.000287
–
Appendix
Table 12: Counterfactual sentence removal under the final domain scorers. Selectivity is the difference in final-score drop between attributed-sentence deletion and the shared Shared control deletion.