and-effect/mdk_gov_data_titles_clf

Dataset Card for MDK

This dataset was created as part of the Bertelsmann Foundation's Musterdatenkatalog (MDK) project. The MDK provides an overview of Open Data in municipalities in Germany. It is intended to help municipalities in Germany, as well as data analysts and journalists, to get an overview of the topics and the extent to which cities have already published data sets.

Dataset Description

Dataset Summary

The dataset is an annotated corpus of 1258 records based on the metadata of the datasets from GOVDATA. GovData is a data portal that aims to make cities' data available in a standardized way.

The annotation maps the titles of the datasets to a taxonomy containing categories such as 'Verkehr - KFZ - Messung' or 'Abfallwirtschaft - Abfallkalender'. Through the assignment the names of the data sets can be normalized and grouped. In total, the taxonomy consists 250 categories. Each category is divided into two levels:

• Level 1: "Thema" (topic)

• Level 2: "Bezeichnung" (label).

The first dash divides the levels. For example:

You can find an interactive view of the taxonomy with all labels here.

The repository contains a small and a large version of the data. The small version is for testing purposes only. The large data set contains all 1258 entries. The large and small datasets are split into a training and a testing dataset. In addition, the large dataset folder contains of a validation dataset that has been annotated separately. The validation dataset is an additional dataset that we created for the evaluation of the algorithm. It also consists of data from GOVDATA and has the same structure as the test and training data set.

Languages

The language data is German.

Dataset Structure

Data Fields

dataset	size
small/train	18.96 KB
small/test	6.13 KB
large/train	517.77 KB
large/test	118.66 KB

An example of looks as follows:

{
    "doc_id": "a063d3b7-4c09-421e-9849-073dc8939e76",
    "title": "Dienstleistungen Alphabetisch sortiert April 2019",
    "description": "CSV-Datei mit allen Dienstleistungen der Kreisverwaltung Kleve. Sortiert nach AlphabetStand 01.04.2019",
    "labels_name": "Sonstiges - Sonstiges",
    "labels": 166
}

The data fields are the same among all splits:

• doc_id (uuid): identifier for each document
• title (str): dataset title from GOVDATA
• description (str): description of the dataset
• labels_name (str): annotation with labels from taxonomy
• labels (int): labels indexed from 0 to 250

Data Splits

dataset_name	dataset_splits	train_size	test_size	validation_size
dataset_large	train, test, validation	1009	249	101
dataset_small	train, test	37	13	None

Dataset Creation

The dataset was created through multiple manual annotation rounds.

Source Data

The data comes from GOVDATA, an open data portal of Germany. It aims to provide central access to administrative data from the federal, state and local governments. Their aim is to make data available in one place and thus easier to use. The data available is structured in 13 categories ranging from finance, to international topics, health, education and science and technology. GOVDATA offers a CKAN API to make requests and provides metadata for each data entry.

Initial Data Collection and Normalization

Several sources were used for the annotation process. A sample was collected from GOVDATA with actual datasets. For the sample, 50 records were drawn for each group. Additional samples are from the previous version of the MDK that contain older data from GOVDATA. Some of the datasets from the old MDK already contained an annotation, but since the taxonomy is not the same, the data were re-annotated. A sample was drawn from each source (randomly and by manual selection), resulting in a total of 1258 titles.

Annotations

Annotation process

The data was annotated in four rounds and one additional test round. In each round a percentage of the data was allocated to all annotators to caluculate the inter-annotator agreement using Cohens Kappa. The following table shows the results of the of the annotations:

	Cohens Kappa	Number of Annotators	Number of Documents
Test Round	.77	6	50
Round 1	.41	2	120
Round 2	.76	4	480
Round 3	.71	3	420
Round 4	.87	2	416
Validation set	-	1	177

In addition, a validation set was generated by the dataset curators.

Who are the annotators?

Annotators are all employees from &effect data solutions GmbH. The taxonomy as well as rules and problems in the assignment of datasets were discussed and debated in advance of the development of the taxonomy and the annotation in two workshops with experts and representatives of the open data community and local governments as well as with the project members of the Musterdatenkatalog from the Bertelsmann Foundation. On this basis, the &effect employees were instructed in the annotation by the curators of the datasets.

Considerations for Using the Data

The dataset for the annotation process was generated by sampling from GOVDATA and data previously collected from GOVDATA. The data on GOVDATA is continuously updated and data can get deleted. Thus, there is no guarantee that data entries included here will still be available.

Social Impact of Dataset

Since 2017, the German government has been promoting systematic and free access to public administration data with first laws on open data in municipalities. In this way, a contribution is aimed at the development of a [knowledge society] (https://www.verwaltung-innovativ.de/DE/Startseite/startseite_node.html). The categorization of open data of cities in a standardized and detailed taxonomy supports this process of making data of municipalities freely, openly and structured accessible.

Discussion of Biases (non-ethical)

The data was mainly sampled at random from the categories available on GOVDATA. Although all categories were sampled there is still some imbalance in the data. For example: entries for the concept 'Raumordnung, Raumplanung und Raumentwicklung - Bebauungsplan' make up the majority class. Although manual selection of data was also used for not all previous concepts data entries was found. However, for 95% of concepts at least one data entry is available.

Additional Information

Dataset Curators

Friederike Bauer

Rahkakavee Baskaran

Licensing Information

CC BY 4.0