This paper is concerned with the detection and classification of occupational titles in German texts, with a focus on the linguistic and structural challenges that are unique to the German language. The study utilizes an extensive dataset on job title variants to assess rule-based and language model-based methodologies across diverse corpora, encompassing historical documents and parliamentary proceedings. The findings indicate that rule-based methods demonstrate robust performance, particularly in structured texts, while large language models exhibit complementary strengths in recognizing complex terms. The work makes a significant contribution to the field by providing valuable annotated data and methodological insights.
In addressing the second research question, the focus is directed towards two distinct datasets. The ifrst dataset comprises German parliamentary debates. The second dataset encompasses GDR job descriptions (Berufsbilder). The principal contribution of this paper is a substantial collection of novel training data and a systematic analysis of the challenges associated with the detection of German job titles according to KldB.
The present document is organized into five sections.
The integration of diverse labor market data sources is widely acknowledged as a multifaceted
undertaking [
We follow our study of automated classifications presented in [
A portion of the research has focused on the classification of OJAs according to the O*NET framework
[
A limited number of publications have been published on the subject of German job titles, with a
particular emphasis on the German KldB. For instance, a technical report based on OJAs [
2See https://statistik.arbeitsagentur.de/DE/Navigation/Grundlagen/Klassifikationen/Klassifikation-der-Berufe/ KldB2010-Fassung2020/KldB2010-Fassung2020-Nav.html.
3.1. Data To identify job titles, an extensive dataset was utilized, encompassing 526,535 synonyms and variants of male, female, and neutral job titles. This dataset was provided by the German Federal Employment Agency (BA)3. For illustrative purposes, the following examples are given: “Meister – Maßschneiderei” and “Herrenschneidermeisterin”, both of which link to KldB 28293-901. However, it should be noted that the first five digits are utilized exclusively, and a link to 28293 is established. The presence of numerous duplicates results in a non-unique linkage. The most prominent example is the term “Meister” (Master), which is linked to nearly all crafts. Furthermore, the dataset under consideration contains terms associated with occupations. For instance, the terms “Kohle”, “Naturwerkstein”, and “Anlagenführung” are linked to 21212-129. Consequently, this dataset can also be utilized to identify any implicit relations in a text that link to occupations. Nevertheless, for the aforementioned approach, a blacklist of these terms was created. While these terms are classified under the gender-neutral section, the removal of these terms would also result in the removal of all gender-neutral occupational titles.
The initial test dataset comprises job descriptions (Berufsbilder) utilized in the former German
Democratic Republic (GDR) for the purpose of vocational guidance, see [
The Corpus of Plenary Proceedings of the German Bundestag (CPP-BT) is another dataset that will
be utilized in this study, see [
und Natursteinen. Durch ein Fachschulstudium können Sie sich zum Ingenieur für Baustofftechnologie, Ingenieur der Bauelementeproduktion oder Ingenieur für Verfahrenstechnik
FilterList
LLMs PhraseMatcher
Occupation title Data Portal of the German Bundestag and the Documentation and Information System for Parliamentary Materials (DIP) up to the respective cut-of date.
While the initial dataset does not include any individual names, the subsequent dataset highlights the complexity of identifying individuals by their family names when these names are associated with specific job titles, as previously discussed. Examples include: • Carsten Schneider spricht jetzt für die SPD-Fraktion. (Carsten Schneider now speaks on behalf of the SPD parliamentary group.) • Kollege Schneider, ich glaube, die Bürger haben richtig entschieden • Da wird Herr Weber nicht begeistert sein! (Mr. Schneider, I believe that the citizens have made the right decision!) • Herr Koch war sowieso dagegen. (Mr. Koch was against it anyway.) • Stefan Müller [Erlangen] [CDU/CSU] • Zimmermann, Sabine DIE LINKE
The workflow delineates two sequential steps: initial identification of job titles, followed by subsequent determination of whether said titles refer to an individual or pertain to a previously mentioned occupation. To identify all relevant candidates, a PhraseMatcher is employed, leveraging the data set comprising job titles and synonyms. The present study utilizes Python 3.11.2 and the Spacy library. For a visual representation of the complete workflow, refer to Figure 2.
The initial step in this process is to either substantiate or refute the candidates’ claims, as certain terms may not be applicable to a specific position. The approach employed involves the utilization of a filter list, which comprises a manually curated list, or the execution of queries against diverse LLMs to ascertain the nature of the term in question, specifically determining whether it pertains to an occupation or not. In the evaluation, the initial approach is designated as "Rule based." • If the previous word is a well-known given name (list retrieved from Wikidata), return true. • If the previous word is a particular title (Kollege, Kollegin, Herr, Frau, or Dr.) or a party name (SPD, CDU/CSU, FDP, ...) return true.
• If the previous or next word is a title or begins with a bracket, return true.
The subsequent task is to ascertain whether the term in question is a family name. The initial approach is a rudimentary rule-based strategy that utilizes a concise sequence of other words:
The second approach once again delegates this question to diferent LLMS. In the evaluation, the initial approach is again designated as "Rule based."
As illustrated in Table 1, the following metrics are presented on GDR job descriptions. The rule-based approach is the most efective method for identifying job title candidates. However, it is subject to certain limitations when it comes to terms such as “Aufbereitung”.
It has been observed that certain models appear to encounter dificulties with extended terms, such as “Ingenieur für Verfahrenstechnik”. This particular term was identified by both the rule-based approach and llama3.2, yet it did not elicit any results from other models. Additional challenges were identified, including the term “Ausbildung” (training), which was identified by the smaller models. Nevertheless, the rule-based approach was the only method that correctly identified more exotic terms such as “Fähnrich” and even “Berufskraftfahrer”.
It is not surprising that the rule-based approach is the most efective method for name detection. As this text category does not include names, the application of all rules is not guaranteed. Nevertheless, it is noteworthy that qwen3 also achieved this optimal performance.
As illustrated in Table 2, the result metrics for an annotated subset of 20 plenary discussions of the German Bundestag are presented. Once more, the rule-based approach to identifying job titles evidently surpasses the capabilities of LLM approaches. In particular, the qwen3 model has been observed to detect a considerable number of generic terms, such as “Beratung” (consulting), as occupational titles. However, both the phi4 and qwen3 models have demonstrated challenges in identifying terms such as “Wohnungsbau” and “Recht”. However, the performance of phi4 appears to be satisfactory, as evidenced by its satisfactory performance in both macro and weighted 1-score metrics. In contrast, the performance of all other models is clearly biased towards macro weights.
With regard to the identification of family names, the precision and recall metrics demonstrate eficacy across all models for weighted scores. However, the eficacy of both phi4 and the rule-based approach has been demonstrated. The annotated data set exhibits significant imbalance, with nearly 94% of all candidates comprising non-job titles and only 0.2% consisting of family names. Consequently, the macro values are not readily discernible. However, preliminary findings suggest the potential eficacy of LLM approaches over the proposed rule-based approach.
This study investigated the complex task of detecting and classifying occupational titles in German texts, highlighting the linguistic and contextual challenges that render this problem particularly challenging. The research utilizes a comprehensive dataset comprising 526,535 job title variants and synonyms, and it analyzes two distinct textual domains: GDR vocational descriptions, and plenary proceedings of the German Bundestag. This approach ofers a comprehensive perspective on the nuances involved in occupational detection.
The experimental results underscore the continued efectiveness of rule-based approaches, particularly in structured texts such as GDR job descriptions. The eficacy of these methods was demonstrated by their high precision and recall in identifying occupational titles and family names, with several cases exhibiting superior performance in comparison to large language models (LLMs). However, LLMs such as llama3.2 demonstrated proficiency in detecting more complex, multi-word occupational expressions that were occasionally overlooked by rule-based systems. Despite these advancements, the process of mapping occupations to fine-grained classification codes, such as the 5-digit level of the German Klassifikation der Berufe (KldB), persists as a challenging endeavor. The task is further complicated by the presence of ambiguities caused by overlapping uses of names and job titles, as well as the presence of gender-specific, gender-neutral, and historically derived terms.
In light of these findings, this work suggests several promising areas for future research. First, the
usage of other AI-approaches like BERT, see [
Furthermore, the implementation of cross-lingual mapping and alignment with international classification systems such as ISCO and ESCO would expand the practical relevance of these methods beyond national boundaries. As researchers increasingly experiment with LLMs for occupational coding, the development of standardized benchmarks and evaluation frameworks will be essential to ensure comparability and reproducibility of results.
In sum, the present study furnishes a substantial dataset and a series of methodological insights that establish the foundation for more precise and extensible systems in occupational text analysis. The ifndings are of particular pertinence for applications in labor market research, policy analysis, and computational social science, where the automated identification of occupational information persists as a salient concern.
During the preparation of this work, the authors used DeepL in order to: Grammar and spelling check. After using these tool(s)/service(s), the authors reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.