README.md

metadata

language:
  - vi
license: mit
pretty_name: Vietnamese Business Entity Resolution — Realistic Hybrid Benchmark
tags:
  - entity-resolution
  - record-linkage
  - splink
  - synthetic
  - vietnamese
configs:
  - config_name: observed_records
    default: true
    data_files:
      - split: train
        path: observed/observed_records/observed_records_train.parquet
      - split: validation
        path: observed/observed_records/observed_records_validation.parquet
      - split: test
        path: observed/observed_records/observed_records_test.parquet
  - config_name: business_registry
    data_files:
      - split: train
        path: observed/business_registry/business_registry_train.parquet
      - split: validation
        path: observed/business_registry/business_registry_validation.parquet
      - split: test
        path: observed/business_registry/business_registry_test.parquet
  - config_name: tax_registry
    data_files:
      - split: train
        path: observed/tax_registry/tax_registry_train.parquet
      - split: validation
        path: observed/tax_registry/tax_registry_validation.parquet
      - split: test
        path: observed/tax_registry/tax_registry_test.parquet
  - config_name: inspection_record
    data_files:
      - split: train
        path: observed/inspection_record/inspection_record_train.parquet
      - split: validation
        path: observed/inspection_record/inspection_record_validation.parquet
      - split: test
        path: observed/inspection_record/inspection_record_test.parquet
  - config_name: record_split_map
    data_files:
      - split: train
        path: control/record_split_map/record_split_map_train.parquet
      - split: validation
        path: control/record_split_map/record_split_map_validation.parquet
      - split: test
        path: control/record_split_map/record_split_map_test.parquet
  - config_name: ground_truth_entity_map
    data_files:
      - split: train
        path: >-
          evaluation/ground_truth_entity_map/ground_truth_entity_map_train.parquet
      - split: validation
        path: >-
          evaluation/ground_truth_entity_map/ground_truth_entity_map_validation.parquet
      - split: test
        path: >-
          evaluation/ground_truth_entity_map/ground_truth_entity_map_test.parquet
  - config_name: all_positive_pairs
    data_files:
      - split: train
        path: evaluation/all_positive_pairs/all_positive_pairs_train.parquet
      - split: validation
        path: evaluation/all_positive_pairs/all_positive_pairs_validation.parquet
      - split: test
        path: evaluation/all_positive_pairs/all_positive_pairs_test.parquet
  - config_name: hard_negative_pairs
    data_files:
      - split: train
        path: evaluation/hard_negative_pairs/hard_negative_pairs_train.parquet
      - split: validation
        path: evaluation/hard_negative_pairs/hard_negative_pairs_validation.parquet
      - split: test
        path: evaluation/hard_negative_pairs/hard_negative_pairs_test.parquet
  - config_name: record_generation_metadata
    data_files:
      - split: train
        path: >-
          evaluation/record_generation_metadata/record_generation_metadata_train.parquet
      - split: validation
        path: >-
          evaluation/record_generation_metadata/record_generation_metadata_validation.parquet
      - split: test
        path: >-
          evaluation/record_generation_metadata/record_generation_metadata_test.parquet
  - config_name: canonical_entities
    data_files:
      - split: train
        path: canonical/canonical_entities/canonical_entities_train.parquet
      - split: validation
        path: canonical/canonical_entities/canonical_entities_validation.parquet
      - split: test
        path: canonical/canonical_entities/canonical_entities_test.parquet

Vietnamese Business Entity Resolution Dataset — Embedding Focused

Synthetic benchmark cho bài toán Entity Resolution trên dữ liệu doanh nghiệp Việt Nam từ ba nguồn:

• business_registry
• tax_registry
• inspection_record

Dataset design

• Entities: 12,000
• Records: 34,708
• Entity-level split: 70% train, 15% validation, 15% test
• Mọi record của cùng một entity luôn nằm trong cùng một split
• Phần lớn entity có đủ 3 nguồn; một tỷ lệ nhỏ thiếu hoặc có thêm record
• Positive pairs: lexical, semantic-only và mixed/hard

Subsets

Subset	Nội dung
observed_records	Toàn bộ bản ghi quan sát
business_registry	Bản ghi đăng ký doanh nghiệp
tax_registry	Bản ghi cơ quan thuế
inspection_record	Bản ghi kiểm tra
record_split_map	Ánh xạ record sang split
ground_truth_entity_map	Ground-truth entity
all_positive_pairs	Các cặp match thật
hard_negative_pairs	Các cặp non-match khó
record_generation_metadata	Metadata sinh dữ liệu
canonical_entities	Entity chuẩn

Split policy

Việc chia dữ liệu được thực hiện ở cấp entity, không phải cấp record.

Do đó, một doanh nghiệp và toàn bộ record tương ứng của nó chỉ xuất hiện trong đúng một trong ba split: train, validation hoặc test.

Điều này ngăn entity leakage giữa các tập dữ liệu.