The developed algorithm includes the following steps:

• Identification of relevant websites: First, websites with biographical information about the candidates are identified using the Microsoft Bing Websearch API.


• Text extraction from web pages: The text is then automatically extracted from the identified web pages using an automated version of Selenium Chrome Driver.


• Heuristic relevance check: An automated relevance check checks whether the website contains relevant information about the respective candidate. This check is based on the criteria name, party and region.


• Data extraction: Biographical data such as professional activities, memberships, training details and personal data (year of birth, place of birth, marital status, children, etc.) are extracted from all websites that pass the relevance check. Large language models are often unreliable in free text language generation, but at the same time they are much more reliable in correctly extracting data with appropriate prompt instructions. Large language models such as Meta’s Llama3.1 70BN, OpenAI’s GPT-4-o and Claude 3.5 are used for this process.



To improve data quality, the same data points from up to six different websites are aggregated. The AI ​​is also used to translate websites that are in English into German and to convert free data points into standardized categories for job, training and memberships. To provide information that can be presented chronologically, an annual matrix is ​​created for each candidate that shows all active years (in training, work or politics). Education, occupation and political activities are given equal weight, and professional activities carried out at the same time are given equal weight within the occupation category. Data without complete time information (start and end date) on at least one website will be excluded from further evaluation. The aggregated annual data is evaluated using pivot tables to calculate average years of experience in politics and work as well as in specific job categories. These analyzes also include the percentage distribution by occupational categories and educational backgrounds.

We carry out automated quality controls. The algorithm carries out an automated data quality check, which was selectively validated by manual checks during development. Automated checks include identifying candidates without reliable data, statistical frequency of websites used, and classifying data gaps. The websites used are often Wikipedia, parliamentwatch.de, party websites, websites of parliament or parliamentary groups, personal websites of the candidates and websites of various media providers. As an indication from pilot analyses, we know that on average 1.6 internet sources are found for each data point. The extraction and aggregation of all data is confirmed by at least two language models to minimize the likelihood of hallucinations. For each candidate, we list all websites from which data was extracted to enable manual verification of the information.



We carry out automated quality controls. The algorithm carries out an automated data quality check, which was selectively validated by manual checks during development. Automated checks include identifying candidates without reliable data, statistical frequency of websites used, and classifying data gaps. The websites used are often Wikipedia, parliamentwatch.de, party websites, websites of parliament or parliamentary groups, personal websites of the candidates and websites of various media providers. As an indication from pilot analyses, we know that on average 1.6 internet sources are found for each data point. The extraction and aggregation of all data is confirmed by at least two language models to minimize the likelihood of hallucinations. For each candidate, we list all websites from which data was extracted to enable manual verification of the information.




An example of the prompts used is available here.