README.md

---
license: cc-by-4.0
task_categories:
- text-classification
language:
- de
- en
- fr
- it
- ru
- la
- es
- ar
- pl
- tr
tags:
- Subject_Indexing
- Cataloging
---

# Title

"Basisklassifikation" (BK) Training Dataset for Automatic Subject Indexing: Titles and Subjects from the K10plus Library Catalogue

# Description

This is a training dataset for automatic subject indexing containing more than 6 million titles and their corresponding subjects (classes) from the "[Basisklassifikation](https://wiki.k10plus.de/spaces/K10PLUS/pages/437452809/Basisklassifikation+Stand+06+2019)" (BK). Initially introduced in the 1980s, today the Basisklassifikation constitutes the most widely used classification system for subject indexing within the Berlin State Library. As of August 2024, around 11% of all of the works (~8.7 Mio. titles) included in the [K10plus](https://www.bszgbv.de/services/k10plus/) catalogue - the union catalogue of the German library networks GBV and SWB - have already received BK notations. The dataset consists of files in TSV format, intended to be used together with the Annif tool for automatic subject indexing and a vocabulary file of the BK (see [BK Download](https://api.dante.gbv.de/export/download/bk/default/)). In this way, Annif models for the prediction of BK classes have been trained which can be accessed via the [Staatsbibliothek zu Berlin – Preußischer Kulturbesitz community at HuggingFace](https://huggingface.co/SBB).

The dataset was created by the team of the "Mensch.Maschine.Kultur – Künstliche Intelligenz für das Digitale Kulturelle Erbe" research project at Berlin State Library (SBB) which was funded by the Federal Government Commissioner for Culture and the Media (BKM), project grant no. 2522DIG002. More specifically, it was created in the context of sub-project 3 "AI-supported content analysis and subject indexing".


## Homepage

Not applicable.

## Publisher

Staatsbibliothek zu Berlin – Berlin State Library

## Dataset Curators

The dataset was curated and published by two members of the research project "Mensch.Maschine.Kultur" ("Human.Machine.Culture"):

Sophie Schneider, project collaborator in the "Human.Machine.Culture" project at Staatsbibliothek zu Berlin – Berlin State Library, [email protected],  ORCID: [0000-0002-8303-1798](https://orcid.org/0000-0002-8303-1798). Sophie Schneider has studied library and information science and works in the research project "Mensch.Maschine.Kultur". She was responsible for the dataset creation and contributed to the datasheet.

Dr. Jörg Lehmann, project collaborator in the "Human.Machine.Culture" project at Staatsbibliothek zu Berlin – Berlin State Library, [email protected], ORCID: [0000-0003-1334-9693](https://orcid.org/0000-0003-1334-9693). Jörg Lehmann has studied history and comparative literature. He contributed to the datasheet and was responsible for data publication.

Both curators can be contacted with regard to an update or feedback to the datasheet and regarding technical issues. The curators are prepared to incorporate responses and comments into a new version of the datasheet if this deems sensible.


## Other Contributors

This dataset comprises extensive indexing work by librarians from various K10plus institutions and thus describes a resource which evolved steadily over time.

## Point of Contact

Clemens Neudecker, Staatsbibliothek zu Berlin – Berlin State Library, [[email protected]](mailto:[email protected]) 

## Papers and/or Other References

Not applicable.

## Supported Tasks and Shared Tasks

### AI Category 

Text Classification

### Type of Cultural Heritage Application

Cataloging, Text Categorization

### (Cultural Heritage) Application Example

Subject Indexing of (historical/contemporary) Text

# Distribution

## Data Access URL

[https://doi.org/10.5281/zenodo.15690227](https://doi.org/10.5281/zenodo.15690227)

## Licensing Information

[Creative Commons Attribution 4.0 International – CC BY 4.0](https://creativecommons.org/licenses/by/4.0/legalcode)

## File Format

text/csv

## Citation Information
```bibtex
@dataset{schneider_2025_15690227,
  author       = {Schneider, Sophie and
                   Lehmann, Jörg},
  title        = {"Basisklassifikation" (BK) Training Dataset for
                   Automatic Subject Indexing
                  },
  month        = sep,
  year         = 2025,
  publisher    = {Staatsbibliothek zu Berlin – Berlin State Library},
  version      = 1,
  doi          = {10.5281/zenodo.15690227},
  url          = {https://doi.org/10.5281/zenodo.15690227},
}
```

# Composition

## Data Category 

metadata

## Media Category

text

## Object Type

titles, subject indexes

## Dataset Structure

### Data Instances

The dataset constitutes a list of titles and their BK subject notations, one title per line. Each instance consists of the title text and a number of notations, all separated by tab spaces in the form `A title <TAB> keyword_1 <TAB> keyword_n`. As an example, it resembles the following line:

`Modern Coding Theory <http://uri.gbv.de/terminology/bk/31.80> <http://uri.gbv.de/terminology/bk/53.71>` 

### Data Fields

Due to the simplicity in structure, the dataset solely consists of tab-separated strings. The subject notations contain URIs which lead to the specific subclass entry in the BK terminology as provided by the Common Library Network (GBV): [http://uri.gbv.de/terminology/bk/](http://uri.gbv.de/terminology/bk/).

### Compliance with Standard(s)

The "[Basisklassifikation](https://wiki.k10plus.de/spaces/K10PLUS/pages/437452809/Basisklassifikation+Stand+06+2019)" (BK) is a standardized vocabulary.

### Data Splits

A random 80/10/10 split was performed in order to create subsets for training, validation and testing: `train.tsv` contains 4.891.960 data instances and can be used for training Annif models, while `val.tsv` and `test.tsv` both contain 611.495 data instances and are intended to be used for hyperparameter optimization and final testing.

## Language

This is a multilingual dataset, as the size of the overall dataset rather than a specific language was considered in the dataset creation. The 10 most frequent languages are: 

* ger (de): 2.778.958 instances
* eng (en): 1.716.282
* fre (fr): 306.357 
* ita (it): 151.723
* rus (ru): 151.415
* lat (latin): 138.558
* spa (sp): 123.749
* ara (ar): 82.414
* pol (pl): 59.212
* tur (tr): 45.774
	
This analysis was carried out on the basis of the Pica+ `010@` (Pica3: [`1500`](https://swbtools.bsz-bw.de/cgi-bin/k10plushelp.pl?cmd=kat&val=1500)) field, subfield a.

## Descriptive Statistics

In its compressed form, this dataset has a size of 294 MB (308.342.676 Bytes). Besides its diversity in languages, it does not cover a specific time period. However, a strong increase in BK subject assignments can be found for titles published from 1980 onwards (at the time of the classification’s introduction).

# Data Collection Process

## Curation Rationale

The main motivation behind the creation of this dataset was to develop models for automatic subject indexing with the BK classification system. More generally speaking, machine-based support of labor intensive intellectual subject assignment as carried out by librarians was of interest. Since the Annif tool was chosen for training the models, the dataset was created and curated in such a way that direct usage with Annif is enabled. 

## Source Data

### Initial Data Collection

The initial data collection included the following steps:

* downloading the `kxp-subjects.tsv.gz` of: Voß, J., & Verbundzentrale des GBV. (2024). Normalized subject indexing data of K10plus library union catalog (2024-02-26) [Data set]. VZG. [https://doi.org/10.5281/zenodo.10933926](https://doi.org/10.5281/zenodo.10933926) 
* extracting all PPNs assigned with BK notations and the notations themselves
* querying the corresponding BK titles (Pica3 field [`4000`](https://swbtools.bsz-bw.de/cgi-bin/k10plushelp.pl?cmd=kat&val=4000), subfields a and d) for all PPNs from previous step via [unapi](http://unapi.k10plus.de/)
* merging the subject and title data based on PPN, filtering out duplicates (identical titles)
* this led to a dataset of overall 6.114.950 entries, split into 80% train and 10% for test and validation subsets 

### Source Data Producers

As stated above, in terms of classification, the source data was produced by librarians from various institutions contributing to K10plus. Also, the titles included in this dataset were recorded by librarians unless supplied in advance by the publishers.

### Digitisation Pipeline

Not applicable.

## Preprocessing and Cleaning

Not applicable.

## Annotations

Not applicable.

### Annotation Process

Not applicable.

### Annotators

Not applicable.

### Crowd Labour

Not applicable.

## Data Provenance

The BK classification is available as public domain ([CC0](https://creativecommons.org/publicdomain/zero/1.0/)) data. 

## Use of Linked Open Data, Controlled Vocabulary, Multilingual Ontologies/Taxonomies

The BK vocabulary can be accessed online at [http://uri.gbv.de/terminology/bk](http://uri.gbv.de/terminology/bk) and is available in several data formats suitable for Linked Data. 

## Version Information

This is version 1.0 of the dataset.

### Release Date

2025-09-01

### Date of Modification

Not applicable.

### Checksums

**MD5 and SHA256 hashes of the bk_annif_stdc.zip:**

MD5: 4878e017d10fe9b0e1427d7d2fedf153

SHA256: ec60f9fcf2c7f1eb70b4a1cd7ee34f64ea5aaa239ab73369e6048a1d93af8031

## Maintenance Plan

### Maintenance Level

The maintenance of this dataset is limited. The data will not be updated, but any technical issues will be addressed during the lifetime of the research project "Human.Machine.Culture". This project ends in October 2025, and the dataset will be maintained at least until then.

### Update Periodicity

No updates are foreseen.

# Examples and Considerations for Using the Data

## Ethical Considerations

### Personal and Other Sensitive Information

Not applicable.

### Discussion of Biases

The BK was introduced in the 1980s and therefore its structure partly represents outdated ways of thinking. In itself, the BK is biased, for some examples see [https://verbundkonferenz.gbv.de/wp-content/uploads/2024/09/2024-08-20_VK_beckmann_Kunst-oder-Krempel-Potenziale-der-Basisklassifikation.pdf](https://verbundkonferenz.gbv.de/wp-content/uploads/2024/09/2024-08-20_VK_beckmann_Kunst-oder-Krempel-Potenziale-der-Basisklassifikation.pdf), slide 17. Hence, typical biases are e.g. the binary understanding of gender roles, outmoded use of geographical designators or of social movements. However, in order to keep it up to date, the classification system is under constant revision. As of now, the classes suggested for an input text might not be suitable for today’s understanding and might not conform to contemporary values.

### Potential Societal Impact of Using the Dataset

There is no societal impact to be expected from the publication of this dataset.       	

## Examples of Datasets, Publications and Models that (re-)use the Dataset 

On Hugging Face, in the dedicated space of the [Staatsbibliothek zu Berlin (SBB)](https://huggingface.co/SBB/models), there are models that use this dataset.    	

## Known Non-Ethical Limitations

Previously collected information, e.g. on the year or language of the publications, was discarded in order to create a dataset enabling direct usage together with the Annif tool (as the data must be delivered in a specific [data format](https://github.com/NatLibFi/Annif/wiki/Corpus-formats)). This limits the subsequent use, since filtering a specific language, time period, etc. is no longer possible in this version of the dataset.

## Unanticipated Uses made of this Dataset

There are no known unanticipated uses made of this dataset. Users are invited to report the uses they made of this dataset back to the curators, which would enable an update of the datasheet.

Datasheet as of September 1st, 2025