Historical Taxonomy Construction

• Historical taxonomy construction is the systematic creation of hierarchical representations of historical artifacts and events using data-driven clustering and expert curation.
• It employs iterative batch coding, inter-coder reconciliation, and algorithmic filtering to yield transparent and reproducible taxonomies for fields like digital humanities and economic history.
• Modern approaches integrate machine learning, phylogenetic methods, and LLM-driven semantic analysis to enable quantitative cross-analysis and predictive modeling of historical trends.

Historical taxonomy construction refers to the principled development of structured, hierarchical representations—typically tree-based—of artifact classes, event types, products, or other entities across historical domains. This field synthesizes empirical data-driven encoding, algorithmic clustering, domain-expert curation, and, increasingly, machine learning to create transparent, reproducible, and analyzable concept structures. These taxonomies serve as the backbone for comparative analysis and evolution studies across science, culture, economics, and the digital humanities.

1. Foundations and Motivations

The taxonomy of historical objects has transitioned from ad hoc, resemblance-based groupings (“phenetics”) to rigorous, multivariate, and often phylogenetic formalisms. Early traditions, such as Aristotelian grouping by observable traits for practical classification, yielded hundreds of incompatible schemes. The 18th-century innovations of Adanson and Linnaeus—full multivariate character use and binomial nomenclature—established objectivity and reproducibility. Modern taxonomy construction extends these principles to non-biological and historical domains, employing branching structures to capture both similarity and genealogical descent or transmission-with-modification (Fraix-Burnet, 2016). In digital humanities and economic history, formal taxonomies enable quantification, cross-comparison, and predictive modeling.

2. Empirical, Human-in-the-Loop Taxonomy Construction

A prominent empirical framework is illustrated by the VisTaxa method for historical visualizations (Zhang et al., 3 May 2025). The process is staged as iterative, coder-driven, batched labeling across a sampled corpus, with explicit mechanisms for consensus building and documentation. Key steps include:

a) Corpus Preparation: A manageable but representative subset of artifacts (e.g., images of pre-1950 visualizations) is curated from a larger archival set—such as 13,000 examples from OldVisOnline.

b) Batch Coding: Batches (e.g., 4 × 100 images) are annotated in four steps:

• S1: Individual taxonomy creation and memo writing.
• S2: Collective structural reconciliation, supported by a visual taxonomy tree-comparison interface.
• S3: Relabeling per the harmonized tree.
• S4: Image–label reconciliation, with dissensus recorded.

c) Data Structures: The taxonomy is a rooted tree $T$ with taxa-vectors $V_i$ assigned to each data instance $d_i$; a codebook tracks mappings, and a glossary ensures terminological coherence.

d) Decision Protocols: Policies specify multi-labeling, incidentality thresholds, and exclusion criteria (e.g., non-visual artifacts). Memoization tracks boundary decisions, term definitions, and interpretive disagreements.

e) Inter-Coder Agreement Metrics: Exact label match, image-wise Jaccard, and node-set IoU trace convergence, with exact match ratios rising from 0–10% (S1) to ≥95% by S4 in exemplars.

f) Taxonomy Synthesis: Final trees retain nodes present in a majority of coders’ structures; categories are split or merged on clear subclusters or semantic overlap. The resulting hierarchy may feature multiple levels, with top-level classes (maps, charts, diagrams) and specialized subtypes.

The genericity of this method allows adaptation across non-visual artifact domains, applying the same batch coding, memoization, and reconciliation architecture (Zhang et al., 3 May 2025).

3. Algorithmic and Machine-Assisted Construction

Automated and semi-automated approaches, particularly for large-scale, dynamically evolving corpora, have emerged. One exemplary model is the taxonomy network approach in economic history, which reconstructs the directed product development space (Zaccaria et al., 2014):

• Bipartite Projection: Countries and products are encoded as binary matrices $M_{cp}$ (with values determined by metrics such as Revealed Comparative Advantage), from which projections onto the product space yield initial co-activation scores.
• Normalization and Filtering: Co-occurrence weights $B_{pp'}$ are normalized against product ubiquity and country diversification, and a “maximum-picking” filter retains only the most significant outgoing edge per node, forming a sparse, directed graph.
• Temporal Validation: Directionality and causal interpretation are supported by activation tensors and enabling matrices, measuring whether the presence of product $p$ in year $y{-}1$ predicts the emergence of $p'$ in year $y$.
• Case Study Alignment: Empirical activation sequences (e.g., the stepwise industrial trajectory of South Korean electronics) conform closely to the inferred taxonomy’s paths, validating the network’s explanatory power.

Such frameworks yield actionable structures: sparse, interpretable product taxonomies that empirically reflect developmental sequencing and can inform policy or identify developmental “stepping stones” (Zaccaria et al., 2014).

4. Multi-Agent and LLM-Enhanced Historical Taxonomy Construction

LLMs and multi-agent systems have been leveraged for scalable historical taxonomy creation in text-rich or complex domains. CHisAgent (Tang et al., 9 Jan 2026) demonstrates a modular approach for event taxonomy induction from classical Chinese chronicles, utilizing three specialized agents:

• Inducer: Bottom-up clustering and merging of extracted event types into coarse-to-fine hierarchies, using text embeddings (e.g., cosine similarity of text-embedding-3-small vectors) and iterative concept merging.
• Expander: Top-down structural refinement, introducing intermediate or missing sibling concepts, deduplicating semantically overlapping nodes, and ensuring hierarchical completeness through guided LLM prompting.
• Enricher: Completion and faithfulness, integrating frequent corpus events, topic-modeled concepts, and external ontologies (e.g., CBDB) while deduplicating via semantic similarity thresholds.

This pipeline iterates classification, semantic similarity computation, deduplication, and evidence-based insertion, ultimately producing a domain-specific taxonomy indexed across multiple cultural spheres (e.g., politics, military, economy). Evaluation employs both reference-based (node recall, novelty, coverage rate) and reference-free metrics (path granularity, structural/content scatter, coverage on held-out corpora), providing a comprehensive framework for cross-cultural and structural assessment (Tang et al., 9 Jan 2026).

5. Formal Algorithms and Representational Choices

Historical taxonomy construction integrates diverse algorithmic techniques:

• Clustering: Employed both in unsupervised data-driven contexts (e.g., Adanson’s multivariate clustering, KMeans on image features, graph clustering) and as a stage in LLM-guided pipelines.
• Phylogenetic Methods: In evolutionary contexts, character-based (maximum parsimony, cladistics) and distance-based (minimum spanning tree, neighbor-joining) methods reconstruct transmission-with-modification structures, as seen in both biological and cultural studies (Fraix-Burnet, 2016).
• Network Filtering: Maximum-picking, significance-testing, and projection-based sparsification create interpretable, manageable graphs from dense co-occurrence or affinity matrices (Zaccaria et al., 2014).
• LLM-based Semantic Embedding: Modern pipelines leverage high-dimensional textual embeddings for semantic similarity estimation, node deduplication, and granular coverage validation (Tang et al., 9 Jan 2026).
• Decision Rules and Human Input: Multi-coder memoization, glossary use, and manual label reconciliation continue to anchor structural decisions in expert judgment, particularly for ambiguous or highly polysemic artifacts (Zhang et al., 3 May 2025).

6. Practical Considerations and Limitations

Common challenges in historical taxonomy construction include:

• Ambiguity, Anachronism, and Coverage Gaps: Classical texts and artifacts may resist automated parsing, necessitating human-in-the-loop disambiguation or domain-specific extractor fine-tuning (Tang et al., 9 Jan 2026).
• Inter-Coder Disagreement: Systematic dissensus and differences in interpretive “lenses” can persist; reserved dissensus and curated memo-writing are recommended to transparently record open issues (Zhang et al., 3 May 2025).
• Machine Bias and Semantics: Clustering or LLM-based expansion can introduce bias or spurious splits; rigorous evaluation and cautious threshold selection are necessary.
• Tree vs. Network Structures: Although tree-based taxonomies dominate, many historical processes (hybridization, horizontal transfers) necessitate more general network models—outer-planar split networks, inferences of reticulation, or community detection extensions (Fraix-Burnet, 2016).
• Computational Feasibility: For large spaces (e.g., 1000+ nodes, decades of temporal data), complexity scaling must be addressed, though current algorithms remain tractable for typical historical/archival corpora (Zaccaria et al., 2014).

7. Applications and Cross-Domain Generalization

Historical taxonomies support a wide array of research activities:

• Design Space Exploration: Clarification of available historical forms (e.g., visualization types) (Zhang et al., 3 May 2025).
• Evolutionary Analysis: Tracing descent, diffusion, or innovation paths in biology, linguistics, material culture, or technology (Fraix-Burnet, 2016).
• Economic Development Mapping: Empirical forecasting of developmental trajectories (“stepping stones”) and policy design based on product activation orderings (Zaccaria et al., 2014).
• Cultural and Event Taxonomy Construction: Integrating structured knowledge for digital humanities, cross-cultural studies, and ontology-driven research (Tang et al., 9 Jan 2026).

Further, the genericity of modern protocols—including iterative batch coding, expert reconciliation, embedding-driven similarity analysis, and multi-agent architectures—supports generalization to any appropriately digitized and feature-rich class of historical artifacts; only the extractor modules, domain glossaries, and high-level domains require adaptation (Zhang et al., 3 May 2025, Tang et al., 9 Jan 2026).