PostNauka

Origin and credits

This Russian science communication corpus contains texts from PostNauka, whose authors and editors created the original articles. The historical corpus description says “publications up to February 2017”; individual publication dates have not been verified.

Related research includes N. A. Chirkova and K. V. Vorontsov, Additive Regularization for Hierarchical Multimodal Topic Modeling, Journal of Machine Learning and Data Analysis, 2(2), 187–200 (2016), DOI: 10.21469/22233792.2.2.05, primary PDF. That paper describes an earlier PostNauka corpus of 1,728 documents, rather than the larger populations below.

Related work also includes Irina V. Efimova's 2017 Master's thesis at Moscow Institute of Physics and Technology, Иерархическая мультимодальная тематическая модель коллекции научно-популярных текстов (thesis).

Configurations and text fields

Both configurations retain the three columns id, raw_text, and vw_text and the historical row order. bag-of-words remains the default.

Configuration	Rows	raw_text	vw_text
bag-of-words	3,404	Title/snippet display text	Multimodal BOW features and counts
natural-order-of-words	3,446	Full text from the original cleaned document with the same NOOW ID	Prepared tokens in natural order

Each NOOW raw_text is the corresponding original cleaned file decoded once as UTF-8. Headings, snippets, separators, line breaks and other whitespace are retained. There is no further normalization or inferred character-offset alignment between raw_text and prepared vw_text.

Reconstructed document correspondence

data/document-correspondence.json contains 3,403 one-to-one BOW-to-NOOW pairs reconstructed with the existing PostNauka adapter. For each BOW document, every NOOW document sharing at least one @word token is a candidate. The matcher compares token-set Jaccard scores, accepts a unique maximum without a threshold, and checks the resulting assignments for collisions. There are zero collisions. This is a reproducible token-based reconstruction, not exact human ground truth.

BOW document 8.txt remains unresolved: NOOW documents 7810.txt and 9419.txt tie at score 0.2. The JSON also lists all 43 unmapped NOOW IDs. The two configurations therefore have different populations, and a BOW ID must not be used directly as a NOOW ID.

The files data/noow/topics.json and data/noow/topic_documents.json describe the historical reference topic model with 19 topics plus one background topic and use NOOW IDs; any transfer to BOW must pass through the reconstructed mapping and retain the unresolved case.

For local, offline use, run this example from the downloaded dataset repository (requires datasets):

import json
from pathlib import Path
from datasets import load_dataset

root = Path(".")
bow = load_dataset(str(root.resolve()), "bag-of-words", split="train")
noow = load_dataset(str(root.resolve()), "natural-order-of-words", split="train")
correspondence = json.loads((root / "data/document-correspondence.json").read_text(encoding="utf-8"))
mapping = correspondence["bow_to_noow"]
noow_by_id = {row["id"]: row for row in noow}
for bow_row in bow:
    noow_id = mapping.get(bow_row["id"])
    if noow_id is None:
        continue  # Unresolved BOW document; no match is guessed.
    full_text = noow_by_id[noow_id]["raw_text"]
    # Use full_text alongside bow_row's unchanged multimodal features.

Preparation and integrity

This package combines the historical BOW and NOOW representations with the cleaned source documents. NOOW raw_text is decoded directly from those documents; vw_text contains the prepared features. The token-based correspondence above links the two populations where a unique match is available.

Run sha256sum -c SHA256SUMS from this directory to check the files listed in SHA256SUMS.

Rights

The TopicNet license concerns the TopicNet distribution and does not establish permission to redistribute the underlying PostNauka articles. The PostNauka about page states that all rights are reserved. Permission to redistribute the full articles remains unresolved.

Historical BOW statistics

The following tables describe the 3,404-document BOW representation and are retained from the original dataset card without recalculation. They do not describe the 3,446-document NOOW population.

Some measurable characteristics of the BOW dataset:

• D — number of documents
• <modality name> W — modality dictionary size (number of unique tokens)
• <modality name> len D — average document length in modality tokens (number of tokens)
• <modality name> len D uniq — average document length in unique modality tokens (number of unique tokens)

	D	@title W	@title len D	@title len D uniq	@2gramm W	@2gramm len D	@2gramm len D uniq	@3gramm W	@3gramm len D	@3gramm len D uniq	@snippet W	@snippet len D	@snippet len D uniq	@word W	@word len D	@word len D uniq	@post_tag W	@post_tag len D	@post_tag len D uniq	@author W	@author len D	@author len D uniq
value	3404	13546	3.97944	3.98002	163289	47.9697	37.0029	5848	1.71798	1.30288	46033	13.5232	13.5256	1.43373e+06	421.189	254.58	18031	5.297	5.297	3685	1.08255	1.08255

The @snippet modality is a piece of raw text designed to display top documents.

Information about document lengths in modality tokens:

	len_total@title	len_total@2gramm	len_total@3gramm	len_total@snippet	len_total@word	len_total@post_tag	len_total@author	len_uniq@title	len_uniq@2gramm	len_uniq@3gramm	len_uniq@snippet	len_uniq@word	len_uniq@post_tag	len_uniq@author
mean	3.98002	47.9697	1.71798	13.5244	421.189	5.297	1.08255	3.98002	37.0029	1.30288	13.5256	254.58	5.297	1.08255
std	2.21452	49.9748	3.07432	3.0559	482.422	2.01979	0.413483	2.21452	35.4885	1.81957	3.05377	249.288	2.01979	0.413483
min	1	0	0	5	3	1	0	1	0	0	5	3	1	0
25%	2	15	0	12	106	4	1	2	13	0	12	80	4	1
50%	3.5	33	1	13	294	5	1	3.5	27	1	13	197	5	1
75%	5	67	2	15	563	6	1	5	52	2	15	343	6	1
max	22	778	69	28	4892	18	4	22	444	27	28	2277	18	4

There exists a reference "gold-standard" topic model for 19 + 1 (background) topics. Information about the topics is available for the natural word order version of the dataset.

Citations

@article{chirkova2016hierarchical,
  author = {Chirkova, N. A. and Vorontsov, K. V.},
  title = {Additive Regularization for Hierarchical Multimodal Topic Modeling},
  journal = {Journal of Machine Learning and Data Analysis},
  year = {2016},
  volume = {2},
  number = {2},
  pages = {187--200},
  doi = {10.21469/22233792.2.2.05},
  url = {https://publications.hse.ru/pubs/share/direct/210878218.pdf}
}

@mastersthesis{efimova2017hierarchical,
  author = {Efimova, Irina V.},
  title = {Иерархическая мультимодальная тематическая модель коллекции научно-популярных текстов},
  school = {Moscow Institute of Physics and Technology},
  year = {2017},
  language = {Russian},
  url = {http://www.machinelearning.ru/wiki/images/2/24/Efimova_Diploma.pdf}
}