RESEARCH · METHODOLOGICAL EXPERIMENT
Discourse, Power, Network
Examining Kádár-era discourses on the Hungarian National Theatre
How can a large language model–based annotation tool be integrated into the workflow of late-twentieth-century institutional history? This research codes the political and professional discourse on the Hungarian National Theatre between 1957 and 1988 along symbolic and pragmatic argumentative categories — at once a digital source-edition experiment and a methodological reflection on AI-assisted qualitative coding.
01
THE HISTORICAL QUESTION
The paradox of contemporary history and the symbolic stake of the National Theatre
Researchers of the second half of the twentieth century are not faced with a scarcity of sources but, on the contrary, with an almost unmanageable abundance of documents. Even surveying the written legacy of Kádár-era bureaucracy is a substantial task; systematic, quantitative analysis is virtually impossible with traditional methods. This paradox set the starting point of the research.
The actual stake of the debate
A central dimension of the political and professional discourse on the Hungarian National Theatre between 1957 and 1988 was the question of the building’s location. The debate, however, was never simply about an architectural decision: the symbolic stake was what the National Theatre is — national heritage, an instrument of socialist cultural policy, a prestige object, or merely a building-management task. These interpretations alternated, intertwined and competed across three decades.
Hypothesis: symbolic and pragmatic argumentation
The central hypothesis is that in the decision-making about the location of the National Theatre, symbolic arguments — the National Theatre as an embodiment of national culture, state representation and Hungarian identity — and pragmatic arguments — economic, architectural and operational considerations — appeared in different proportions across the various phases of the period. The coding scheme is organised around this duality: 16 sub-categories under the SY (symbolic) and PR (pragmatic) main categories.
Periods of the era
The corpus can be divided into four well-defined periods structured by two key turning points around the National Theatre — the 1963–64 demolition decision and the 1978 change of management.
- 1957 — 1962 Postwar reorganisation Symbolic re-appropriation of the National Theatre within socialist cultural policy; dominance of political discourse.
- 1963 — 1964 Turning point: demolition of the building The decision on the demolition of the Blaha Lujza Square building — a temporary rise of pragmatic argumentation.
- 1965 — 1977 The ‘theatreless’ years Long years of provisional arrangements; the symbolic register returns in the architectural-competition debates.
- 1978 — 1988 Turning point: change of management and new plans Restructuring of the discourse, reinforcement of institutional considerations — pragmatic argumentation regains weight.
The basis for reconstructing the discourse is a published source collection that makes the emblematic documents on the era’s National Theatre accessible. The investigation builds on a category-based, systematic coding of these sources.
02
THE ANNOTATION TOOL
A browser-based annotator built through iterative cowork
The annotation tool itself is a methodological innovation of the research: it was not built through traditional software development but through prompt-based, iterative dialogue with Anthropic’s Claude Opus 4.6 model. The application is a single, standalone HTML file — UI, logic and export module all live in the same browser environment.
Methodology of the development
The workflow ran in cowork mode: the researcher’s intent was formulated as natural-language prompts, the AI-generated code was tested, the results were fed back, and the next iteration responded to the issues that had emerged. Development was organised around the research question — every new module served an analytical need directly. The principles of prompt design: develop one module at a time, keep changes testable, and ensure that the source code remains readable and editable throughout.
Functions of the tool
The blocks below introduce each function of the tool along three dimensions: what is its purpose, how does it work, and how was it built — from the perspective of prompt iteration and key challenges.
Document list and metadata
Purpose
Manage and search 50 documents in one place, with source reference, date, document type and length recorded.
Mechanism
Each document occupies one row in a left-hand list — date, title, coding count. Filterable by free-text search and document type (political or theatrical discourse).
Build
The first module. The challenge was maintaining metadata-schema consistency — the coding prompt and the rendering layer expected the same fields.
Annotation interface — five entity types
Purpose
Mark mentions of persons, organisations, works, places and dates in the text — proper-name forms can later be linked into a network.
Mechanism
Mouse selection on the text, entity-type chooser in the right-hand panel. Mentions are gathered into a list, exportable, and shown colour-coded in the text view.
Build
Handling overlapping annotations and multi-word mentions required several iterations. The key to refining the prompt: separate the data model (start/end positions) from the rendering logic.
Text labels — the coding scheme
Purpose
Hierarchical coding of argumentative units: SY (symbolic) and PR (pragmatic) main categories, each with eight sub-categories.
Mechanism
The researcher assigns a label to a selected text fragment; after AI pre-coding, only acceptance or modification is required. Labels are navigable in the hierarchy.
Build
Flexibility of the category tree was the main requirement — the scheme evolved during the research, so previously coded data could not be lost when the scheme was restructured.
HuSpaCy named-entity recognition
Purpose
Speed up automatic proper-name recognition and provide a baseline for manual annotation.
Mechanism
Optional Python backend (HuSpaCy Hungarian model). NER results are imported by the browser interface; the researcher validates and corrects them.
Build
The data-exchange format between backend and frontend required several rounds of refinement — the JSON schema needed to be both human-readable and machine-processable.
Network graph
Purpose
Visualisation of the relational network of co-occurring entities (persons, organisations) across the entire corpus.
Mechanism
Force-directed graph — edges weighted by co-occurrence count. Filterable by entity type, period, document type.
Build
Tuning the graph-simulation parameters (repulsion, damping, node sizing) generated the longest iteration trail — we sought a balance between visual clarity and data fidelity.
Statistical analysis module
Purpose
Distribution of codings per document, period, and code category; measurement of the agreement between AI pre-coding and researcher validation.
Mechanism
Multiple views: per-document code distribution, corpus-level SY/PR ratio, time-series curves, agreement-metric tables.
Build
The set of statistics was shaped by the research questions — every new analytical question prompted a new view, in incremental expansion.
TEI XML, CSV and JSON export
Purpose
Results can be transferred to other research environments and publication platforms (especially the TEI-based source edition).
Mechanism
One click exports the entire corpus or a single document — the TEI schema preserves both proper-name annotations and text labels.
Build
Selecting and adapting the TEI schema to the document structure required prior consultation; lighter formats (CSV, JSON) were added later to improve usability.
03
USING THE TOOL
The workflow step by step
The methodological procedure consists of seven sequential steps — from digitising the source to quantitative analysis. For each step we indicate what happens in the tool, why the step matters, and what methodological safeguards are needed for reliable application.
-
Structuring the source corpus
Archival documents are organised into a unified table — each record describes one document (identifier, date, archive reference, length, document type). The full text is stored by the tool.
Why it matters: consistency of metadata provides the scaffold on which coding and network analysis are built.
-
Finalising the coding scheme
Sixteen sub-categories arranged under SY/PR main categories, each with a coding manual (what it covers, what it excludes, examples).
Why it matters: AI pre-coding is reliable only when category boundaries are precisely circumscribed. The scheme was iteratively refined through pilot codings.
-
Running AI pre-coding
For each document, the Claude model receives the full scheme description plus the document text and proposes codeable text units. Results appear in the annotation interface for the researcher.
Why it matters: the model does not decide; it prepares — the actual coding decision always lies with the researcher. AI directs attention rather than replacing it.
-
Researcher validation and correction
Each suggestion can be examined individually: accept, modify (category change), reject, or extend (adjusting the boundaries of the text unit).
Why it matters: agreement rates can be measured here — the divergence between AI suggestions and researcher decisions directly signals where the scheme or the prompt needs further refinement.
-
Named-entity recognition and validation
The HuSpaCy Hungarian model is run on the full corpus. Results are reviewed in every case; misidentified or unrecognised mentions are corrected.
Why it matters: network analysis is accurate only when proper names are normalised (different spellings of the same person mapped to a single node).
-
Statistical and network analysis
The visualisation module operates on the coded and validated data. Analyses can be run from several angles: SY/PR distribution per period, ratio by document type, code environments of key actors.
Why it matters: qualitative coding becomes quantitatively interpretable here. Visualisation is not only presentation — it also generates new questions.
-
Export and source edition
The full coded corpus is exported to TEI XML and transferred to the publication platform — the digital source edition becomes a standalone, citable publication.
Why it matters: the work does not end with analysis — the source edition ensures that researcher decisions can be verified and the corpus extended.
04
RESULTS
The structure and temporal dynamics of the discourse
The fifty-document corpus contains 472 approved codings. The SY/PR distribution supports the basic claim of the hypothesis: in the discourse around the National Theatre, symbolic argumentation is dominant but not exclusive.
Dynamics of the discourse
The distribution across the four periods is uneven: at the two key turning points — around the 1963–64 demolition decision and the 1978 change of management — a temporary increase in pragmatic argumentation can be observed. The differences between periods are statistically detectable, yet even in the most ‘pragmatic’ phase, symbolic argumentation retains its leading role.
Discourse type and rhetorical register
Documents in the corpus fall into two main types: political discourse (party- and government-level documents, ministerial memos) and theatrical discourse (directors’ notes, programme materials, press items). The two types show different SY/PR ratios: in theatrical discourse symbolic argumentation is more pronounced, while pragmatic considerations gain proportionally more weight in political documents — especially in the turning-point years.
Key actors and code environments
The connection between named-entity recognition and coding operates on two levels. At the first level, we examine the proportion in which each individual’s name appears in symbolic versus pragmatic code environments — the actor’s individual argumentative profile. At the second level, the relationships among actors (co-occurrences) are weighted by the type of code environment: it becomes visible not only who appeared together, but in what argumentative space they were placed alongside one another.
Performance of AI pre-coding
Agreement between coding suggestions generated by Claude Opus 4.6 and the final coding accepted by the researcher was measured per document. In this study the agreement metrics signal not only the model’s technical performance but also the clarity of the coding scheme and the quality of prompt design — every divergence pointed either to an edge case in the scheme or to an imprecisely formulated prompt element. After iterative refinement, agreement reached a consistent level across the entire corpus.
Methodological lessons
The research confirms: a large language model does not replace, but complements, qualitative historical work. AI pre-coding works reliably when (1) the coding scheme is well articulated, (2) prompts undergo iterative refinement, and (3) every suggestion passes through researcher validation. The digital source edition and the interactive analysis interface together ensure that the method does not remain opaque: the reader can trace the sources, the codings and the analytical results alike.