💬 NLP

Curation of a Palaeohispanic Dataset for Machine Learning

This paper addresses the scarcity of suitable resources for studying Palaeohispanic languages by constructing a structured dataset specifically designed to enable the application of Machine Learning techniques to this under-researched field.

Original authors: Gonzalo Martínez-Fernández, Jose F Quesada, Agustín Riscos-Núñez, Francisco José Salguero-Lamillar

Published 2026-04-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Gonzalo Martínez-Fernández, Jose F Quesada, Agustín Riscos-Núñez, Francisco José Salguero-Lamillar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, dusty attic filled with thousands of ancient letters written in a language that no one speaks anymore. These letters are scattered across Spain and France, carved into stone, painted on pottery, or scratched into metal. They tell stories of people who lived over 2,000 years ago, but the letters are messy: some are broken, some are written backwards, and the ink is faded.

This is the world of Palaeohispanic languages—the ancient tongues spoken in the Iberian Peninsula before the Romans arrived. For decades, historians and linguists have been the "detectives" trying to solve the mystery of these letters. They've done a great job, but they've mostly been working with paper notes and spreadsheets that are hard for computers to read.

The Problem: A Library in a Foreign Language
Think of the existing data like a library where every book is written in a different handwriting, some pages are torn out, and the books are stacked in a chaotic pile. If you wanted to use a super-smart robot (an Artificial Intelligence) to read these books and find patterns, the robot would be confused. It can't understand "handwritten notes" or "messy piles." It needs a neat, organized list of numbers and clear categories.

The Solution: The Great Organizing Project
This paper is about a team of researchers who decided to build a digital bridge between these ancient ruins and modern computer science. They didn't just scan the letters; they cleaned, sorted, and translated the information into a format that Machine Learning (AI) can actually use.

Here is how they did it, using some everyday analogies:

1. Gathering the Clues (Data Collection)

The team started with the "Hesperia Data Bank," which is like a giant, central museum archive. It had thousands of records, but they were messy.

The Filter: They had to ignore "fake" letters (forgeries) and focus only on the real ones.
The Goal: They wanted to turn the museum's dusty catalog into a clean, digital spreadsheet.

2. Cleaning the Mess (Data Transformation)

This is the hardest part. Imagine you have a list of addresses that says "The big house near the river in the blue town." A computer doesn't know what that means. The researchers had to translate this into GPS coordinates (latitude and longitude).

Location: Instead of writing "Valencia," they converted it into numbers like 39.48 and -0.38. This allows the AI to understand that two towns are close to each other, just like a map app does.
Time: The dates were written in phrases like "Beginning of the 3rd Century BC." The team turned these into mathematical ranges (e.g., "Between -300 and -200") so the computer could calculate time gaps.
The Text: The ancient inscriptions had a lot of "noise"—comments from scholars, missing letters marked with brackets, and weird symbols. The team created a "cleaning robot" that stripped away the scholar's notes and fixed the broken letters, leaving just the raw ancient text.

3. The Result: A New Tool for Time Travelers

The end result is a dataset (a giant digital table) with 1,751 entries.

Before: A historian had to read a PDF, guess the date, and manually type the location into a map.
Now: A computer can instantly look at 1,751 entries, see patterns in how words are written, guess where a text came from based on its location, or even try to translate a word it has never seen before.

Why Does This Matter?

Think of this dataset as giving a pair of glasses to a blind robot.

Before: The robot was blind to these ancient languages. It couldn't learn from them because the data was too messy.
Now: The robot can "see" patterns. It might discover that a specific type of pottery always has a specific type of writing, or that a certain word appears more often in the north than the south.

This doesn't mean the languages are fully solved yet. It's like finding a new, perfectly organized map for an explorer. The explorer (the AI) still has to figure out the territory, but now they have the best tools possible to do it.

In short: The authors took a chaotic pile of ancient, broken letters and turned them into a neat, computer-friendly instruction manual. This opens the door for AI to help us finally understand the voices of people who lived 2,000 years ago.

1. Problem Statement

The study addresses the lack of structured, machine-readable data for Palaeohispanic languages (languages spoken in the Iberian Peninsula and Southern France prior to the Roman arrival in 218 BC).

Context: These languages (including Iberian, Celtiberian, Lusitanian, and Tartessian) are largely undeciphered or only partially understood. Research has historically relied on manual, linguistic, and archaeological methods.
The Gap: While digital archives exist (notably the Hesperia Data Bank), the data is presented in unstructured formats (mostly strings with irregularities, annotations, and natural language descriptions). This format is unsuitable for Machine Learning (ML) and Natural Language Processing (NLP) techniques, which require numerical or categorical inputs.
Challenge: The complexity of the data involves:
- Scriptio Continua: Words are often not separated, requiring segmentation.
- Mixed Scripts: Use of semi-syllabaries, Greek-Iberian alphabets, and Latin alphabets.
- Unstructured Metadata: Dating and location data are provided as free-text strings rather than coordinates or intervals.

2. Methodology

The authors developed a pipeline to transform the raw Hesperia Data Bank entries into a structured CSV dataset suitable for ML models. The process involved two main phases: Data Collection and Data Transformation.

A. Data Collection

Source: The primary source was the Hesperia Data Bank (Complutense University of Madrid), focusing specifically on epigraphic (inscription) data, as it offers the most attributes for computational analysis.
Scope: The dataset includes Celtiberian and Iberian inscriptions. Suspicious or forged entries were retained in the raw data but flagged for potential filtering.
Initial Attributes: 14 raw attributes were extracted, including archaeological site, MLH/Hesperia references, text content, location (town/province), material, medium, writing direction, technique, sign inventory, dual system status, word separators, and approximate dating.

B. Data Transformation

The authors applied specific transformation rules to convert string-based attributes into numerical or categorical features:

Location (Geospatial Encoding):
- Problem: "Archaeological site" was excluded due to lack of geospatial data.
- Solution: Municipality and Province strings were mapped to latitude and longitude coordinates using external government datasets (Spanish National Geographical Institute and French Government data). This preserves spatial distance relationships useful for clustering or regression.
Chronology (Interval Extraction):
- Problem: Dates were written in natural text (e.g., "Beginning of the 3rd century BC").
- Solution: A systematic parsing algorithm converted text into numerical intervals $[lower\_bound, upper\_bound]$ $[l o w er_b o u n d, u pp er_b o u n d]$ .
  - Centuries were converted to year ranges (e.g., 3rd C. BC $\rightarrow$ [-300, -201]).
  - Modifiers (e.g., "Comienzos" for "Beginning") adjusted the intervals (e.g., restricting to the first 20% of the century).
  - The final representation was converted to $[centerpoint, amplitude]$ (mean and width) for easier model ingestion.
Text Processing (Cleaning & Normalization):
- Problem: Raw texts contained epigraphic commentaries, bibliographic citations, and Leiden Convention symbols (e.g., [---] for missing characters, () for restored abbreviations) that act as noise for ML models.
- Solution: Two versions of the text were created:
  - Original: Retained for reference.
  - Cleaned: Stripped of annotations, citations, and non-linguistic symbols.
  - Specific Rules: Missing characters were standardized to [---]; parentheses/brackets indicating restoration were removed; vowel redundancies (common in syllabaries) were stripped; and line-break markers were removed to join text into single lines.
Categorical Encoding:
- Attributes like Material, Medium, Writing Direction, Technique, Sign Inventory, and Separators were mapped to integer codes.
- Rare values or spelling variations were grouped into "miscellaneous" categories to reduce dimensionality.
- Null values were preserved rather than imputed or dropped to maintain data integrity for specific experiments.

3. Key Contributions

First Structured ML Dataset: The creation of the first curated dataset specifically designed for applying Machine Learning to Palaeohispanic languages.
Feature Engineering: The development of robust preprocessing pipelines for geospatial coordinates, historical date intervals, and noisy epigraphic text.
Dataset Specifications:
- Size: 1,751 instances (inscriptions).
- Features: 36 columns.
- Format: CSV file containing both raw strings (for traceability) and encoded numerical/categorical features (for immediate ML use).
Open Source Tools: The release of Python generator scripts on GitHub, allowing the community to update the dataset as new linguistic research emerges.

4. Results

The paper does not present a specific ML model performance (e.g., accuracy scores) but rather delivers the infrastructure required for such experiments.

Data Quality: The resulting dataset successfully converts unstructured archaeological records into a format compatible with standard ML libraries (e.g., Scikit-learn, PyTorch).
Feature Diversity: The dataset supports diverse analytical tasks:
- Supervised Learning: Classification of script types, dating prediction, or language family attribution.
- Unsupervised Learning: Clustering inscriptions by geographic or morphological similarity.
- NLP Tasks: Word segmentation (due to scriptio continua), morphological analysis, and cognate detection.

5. Significance

Bridging Disciplines: This work acts as a critical bridge between Historical Linguistics/Archaeology and Computer Science. It moves the field from purely manual analysis to data-driven discovery.
Scalability: By automating the data curation process, the authors enable the application of advanced NLP techniques (like Transformers or Large Language Models) to ancient, low-resource languages.
Future Research: The dataset opens new avenues for:
- Deciphering undeciphered scripts (like the South-Western script) by finding patterns in undeciphered texts.
- Analyzing the evolution of writing systems and their relationship to neighboring cultures.
- Improving the understanding of the morphosyntax of language isolates like Iberian.
Dynamic Nature: The authors emphasize that the dataset is a "living" resource; as linguistic interpretations change, the provided scripts allow for easy updates, ensuring the data remains relevant.

In summary, this paper provides the essential data foundation required to apply modern AI techniques to the study of ancient Iberian languages, potentially accelerating the decipherment and understanding of these lost cultures.