File size: 5,334 Bytes
c2a7352
 
 
 
 
 
 
e2f1b37
c2a7352
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
---
language:
- en
- de
license: cc-by-4.0
task_categories:
- text-classification
- other
tags:
- entity-resolution
- record-linkage
- cross-system-matching
- enterprise-data
- benchmark
- rag
- information-extraction
- multilingual
- neurips-2026
pretty_name: CrossER
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
    - split: train
      path: data/splits/train.json
    - split: validation
      path: data/splits/val.json
    - split: test
      path: data/splits/test.json
---

# CrossER: A Benchmark for Context-Dependent Cross-System Entity Resolution

[![License: CC BY 4.0](https://img.shields.io/badge/Data-CC%20BY%204.0-green.svg)](https://creativecommons.org/licenses/by/4.0/)
[![NeurIPS 2026 E&D](https://img.shields.io/badge/NeurIPS%202026-E%26D%20Track-red.svg)](https://neurips.cc/Conferences/2026/CallForEvaluationsDatasets)

**CrossER** is a benchmark for context-dependent cross-system entity resolution where surface features are deliberately misleading. Match pairs average only **0.29 string similarity** (names look unrelated), while non-match pairs average **0.94 similarity** (names look identical).

In real enterprises, matching `Product 4418` to `Maltodextrin DE20 Grade A` requires consulting migration runbooks, classification guides, and Slack threads — not string similarity. CrossER measures the "context gap" across three evaluation modes.

## Dataset Summary

| Metric | Value |
|--------|-------|
| Total Entities | 688 |
| Total Pairs | 1,800 |
| Match / No-Match / Ambiguous | 800 / 800 / 200 |
| Source Systems | 5 |
| Entity Types | 4 |
| Languages | English, German |
| Signal Documents | 8 |
| Noise Documents | 110 |
| Oracle Context Records | 875 |

## Headline Results

| Method | CrossER-Easy | CrossER-Full | CrossER-Hard |
|--------|-------------|-------------|-------------|
| String Matching | 0.741 | 0.363 | 0.000 |
| Fuzzy Matching | 0.771 | 0.455 | 0.000 |
| Embedding Matching | 0.964 | 0.559 | 0.000 |
| Attribute Matching | **1.000** | 0.729 | 0.000 |
| SBERT (multilingual) | 0.843 | 0.604 | 0.222 |
| LLM Zero-Shot | -- | 0.090 | 0.000 |
| LLM + RAG (BM25) | 0.848 | 0.632 | 0.200 |
| LLM + Oracle | **1.000** | **1.000** | **1.000** |

No-context methods score **0.00 F1** on hard pairs. Oracle context closes the gap completely. RAG partially bridges it — retrieval quality is the bottleneck.

## Evaluation Modes

| Mode | Description |
|------|-------------|
| **No Context** | Entity pairs only — what's possible from attributes alone |
| **Raw Context** | 118 enterprise documents (8 signal + 110 noise) — realistic RAG |
| **Oracle Context** | 875 structured migration records — upper bound |

## Named Subsets

| Subset | Pairs | Description |
|--------|-------|-------------|
| **CrossER-Easy** | 257 | Easy matches + obvious negatives; F1 ceiling = 1.000 |
| **CrossER-Medium** | 262 | Medium-difficulty pairs; F1 ceiling = 0.776 |
| **CrossER-Hard** | 203 | Hard matches + adversarial negatives + ambiguous; F1 ceiling = 0.000 (no-context) |
| **CrossER-Full** | 722 | All test pairs |

## Source Systems

| System | Role | Naming Style |
|--------|------|-------------|
| SAP_TC2 | Primary ERP (NA HQ) | Formal English |
| SAP_CFIN | Financial consolidation | Internal codes / abbreviations |
| SAP_APAC | APAC regional ERP | Abbreviated with region prefix |
| LEGACY_ERP | Decommissioned (2019) | Cryptic category codes |
| SHAREPOINT | Tax/compliance reference | Authoritative long names |

## Dataset Structure

```
data/
├── entities.json              # 688 entities across 5 systems
├── pairs.json                 # 1,800 pairs with difficulty tiers
├── splits/                    # train (40%) / val (20%) / test (40%)
├── subsets/                   # CrossER-Easy, -Medium, -Hard, -Full
└── context/
    ├── raw/documents/         # 8 signal documents
    ├── raw/noise/             # 110 noise documents
    └── structured/            # oracle_context.json (875 records)
```

## Quick Start

```python
from datasets import load_dataset

# Load train/val/test splits
ds = load_dataset("smurthy5/CrossER")

# Load a named subset
import json, requests
easy = json.loads(requests.get(
    "https://huggingface.co/datasets/smurthy5/CrossER/resolve/main/data/subsets/crosser_easy.json"
).text)
```

## Prediction Format

```json
[
  {"pair_id": "pair_0001", "predicted_label": "match"},
  {"pair_id": "pair_0002", "predicted_label": "no_match"}
]
```

Valid labels: `match`, `no_match`, `ambiguous`.

## Reproducibility

The dataset is fully reproducible:

```bash
git clone https://github.com/nihalgunu/CrossER
pip install -r requirements.txt
python -m generate.generate_all --seed 42
```

## Citation

```bibtex
@inproceedings{crosser2026,
  author    = {Gunukula, Nihal and Murthy, Sameer},
  title     = {{CrossER: A Benchmark for Context-Dependent Cross-System Entity Resolution}},
  booktitle = {NeurIPS 2026 Evaluations \& Datasets Track},
  year      = {2026},
  url       = {https://huggingface.co/datasets/smurthy5/CrossER}
}
```

## License

- **Code**: Apache 2.0
- **Data**: CC BY 4.0

---

[Phyvant](https://phyvant.com) · [GitHub](https://github.com/nihalgunu/CrossER) · [Paper (NeurIPS 2026)](https://neurips.cc/Conferences/2026/CallForEvaluationsDatasets)