BN_VNNews1072

BN_VNNews1072 is a Vietnamese news corpus for single-path hierarchical text classification. It accompanies the manuscript Semantic Path-HMLC: Frozen Label Semantics and Taxonomy Graph Fusion for Single-Path Hierarchical Text Classification, intended for Neurocomputing.

Dataset description

The corpus contains 434,478 usable Vietnamese news articles collected from VnExpress. Each document has exactly one complete gold taxonomy path. The taxonomy contains 14 top-level categories, 1,072 canonical nodes, 981 valid complete paths, and a maximum depth of five. A valid path may terminate at an internal node that also has children.

The collection records cover January 1, 2019 through August 13, 2026. The corpus builds on the BN-VN3S data-construction methodology, but retains native hierarchical labels rather than ten harmonized flat categories. Record-level overlap and a complete mapping between the two releases have not been audited.

Data schema

The internal source file is dataset.parquet. The manuscript identifies content as the article-text field and sublabel as the complete gold path, represented as a list of labels from root to terminal node. The source also contains index, src, tag, day, month, year, source, and author.

This README does not assert that every auxiliary field is complete or that every value has been independently validated. Inspect the actual Parquet schema before using these fields.

Task and taxonomy

This is single-path classification, not multi-label classification. Each example has one gold path and predictions should be one valid complete path. Taxonomy nodes are identified by their full prefixes, because the same surface label may occur under different parents.

The fixed taxonomy contains 981 permissible complete paths, including paths that terminate at internal nodes. The manuscript reports 169 valid terminal paths that also have children.

Experimental protocol

The manuscript reports a primary 80/10/10 train/validation/test split of 347,577 / 43,453 / 43,448 documents, using split seed 2026. Exact-duplicate text groups are kept within one partition. The split is not guaranteed to be near-duplicate-disjoint or temporally disjoint.

These are historical experimental counts. This release does not claim to reproduce the original split unless the original split assignments or manifests are supplied and verified. Do not create a new random split and describe it as the paper's split.

The manuscript reports 948 paths with training support, 663 paths with test support, and 16 test paths absent from training. The controlled class-holdout experiment is a separate protocol and should not be confused with the primary test split.

Evaluation

The primary metrics are exact-path accuracy, macro-F1 over test-supported complete paths, and micro-averaged hierarchical F1 over non-root prefix nodes. Macro-F1 over all 981 taxonomy paths is a separate diagnostic metric. Calibration metrics include categorical NLL, multiclass Brier score, and top-label ECE.

Results in the manuscript are specific to the investigated corpus and experimental configurations. This dataset is not presented as evidence of cross-dataset generalization or unrestricted open-world recognition.

Access and rights

The original article texts are derived from VnExpress. The authors do not claim ownership of publisher content or an unrestricted right to redistribute it. The public BN_VNNews1072 repository is intended to provide metadata and annotations without redistributing original article texts.

This private repository is for authorized storage and research use. Access to the underlying texts is subject to applicable copyright, publisher terms, permissions, and availability. Possession of this file does not confer a license to redistribute publisher-owned content. No unrestricted license is asserted for the article texts.

Researchers seeking access to non-public materials may contact the corresponding author. Access is considered case by case and cannot be guaranteed.

Related code

Semantic Path-HMLC implementation: https://github.com/chaubaonhan/Semantic-Path-HMLC

Citation

The following is a manuscript citation template. Publication details are pending and must be replaced with verified bibliographic information.

@article{chau2026semanticpathhmlc,
  title   = {Semantic Path-HMLC: Frozen Label Semantics and Taxonomy Graph Fusion for Single-Path Hierarchical Text Classification},
  author  = {Chau, Bao Nhan and Tran, Phuoc},
  journal = {Neurocomputing},
  year    = {2026},
  note    = {Manuscript intended for Neurocomputing; publication details pending}
}