1 of 9

CHALLENGES IN THE DETECTION OF DIALECT FOR HISTORICAL LANGUAGES

THE CASE OF OLD IRISH TEXT RESOURCES

ADRIAN DOYLE

ADRIAN.ODUBHGHAILL@MU.IE

2 of 9

OLD IRISH

Archaic Irish

(400 – 600)

Old Irish

(600 – 900)

Middle Irish

(900 – 1200)

Early Modern Irish

(1200 – 1600)

Modern Dialectal Irish

(1600 – Present)

Modern Standard Irish

(1958 – Present)

Scottish Gaelic

Manx Gaelic

3 of 9

OLD IRISH AND DIALECTS

Early Old Irish

(600 – 700)

Mid. Old Irish

(700 – 800)

Late Old Irish

(800 – 900)

RIA MS 23 E 25 (Leabhar na hUidhre)

(isos.dias.ie)

St. Gallen, Stiftsbibliothek, Cod. Sang. 904, f. 194

(www.e-codices.ch)

Thomas F. O’Rahilly. 1932. Irish Dialects Past and Present. Dublin Institute for Advanced Studies, Dublin: 16.

4 of 9

COMPUTATIONAL APPROACHES

The Problem

No Old Irish text data labelled for dialect

Neural Models and LLMs

Can’t train models
Can’t test models’ accuracy

Stylometric Approach

Cannot quantify accuracy
Clustering allows visual assessment

Maged S. Al-Shaibani and Moataz Ahmed. 2025. The Arabic AI Fingerprint: Stylometric Analysis and Detection of Large Language Models Text.
Jack Grieve.2023. Register Variation Explains Stylometric Authorship Analysis. Corpus Linguistics and Linguistic Theory, 19(1): 47–77.
M.Z. Lahjouji-Seppälä, A. Rabus, and R. von Waldenfels. 2022. Ukrainian Standard Variants in the 20th Century: Stylometry to the Rescue. Russian Linguistics, 46(3): 217–232.

Aim, to use established dialect detection techniques to see if they can add to the discussion of Old Irish dialects

Traditional approaches have focussed heavily on lexical level
Computational approaches can also take into account:

orthography
morphology
syntax
sub-word character sequences

Many computational approaches have been applied successfully to identify dialects in various other languages

LLMs
neural networks
even more rudimentary classifier models

***

Such approaches cannot be applied to Old Irish data

Models rely on labelled data for pre-training, training, fine-tuning, and testing
No such data exists for Old Irish

No texts can be clearly assigned to any known dialect

***

Stylometric approaches

Demonstrated to be effective in areas like
Identifying diachronic development
and Register Variation
Separating AI generated text from human
incl. dialectal social media posts

A stylometric method engineered to separate texts based on dialectal features

Best available option for Old Irish

Limitation: can’t quantify accuracy

However:

Clustering allows for visual assessment by experts

Determine what the real cause is for separation of clusters

5 of 9

METHODOLOGY

Corpus PalaeoHibernicum (CorPH)
12 texts in total

Texts all drawn from one source:

Linked to named geographical location
Linked to named author
Large corpus of glosses

Text selected because:

Each sentence labelled by source text

Texts split into sentences

6 of 9

METHODOLOGY

Pre-trained model: paraphrase-multilingual-mpnet-base-v2
768-dimensional vector space
Cosine Similarity

Sentence Embeddings

HDBSCAN
Min. cluster size = 15
Min. samples = 5

Clustering

UMAP

Dimensionality Reduction

Sentence embedding carried out using a pre-trained multilingual model from HuggingFace

Multilingual model because:

Hope: to avail of model’s transfer learning capabilities

Apply general understanding of semantic and syntactic structure when encoding Old Irish

Sentence Embedding model because:

Allows syntactic structure to be taken into account

Similarity measured using cosine distance

Unsupervised Clustering

Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN)
HDBSCAN used because:

Not necessary to predefine the number of clusters

Necessary with k-means clustering, for e.g.
Undesirable because the number of potential dialects is unknown

Instances not belonging to any dense region are labelled as noise by HDBSCAN

Not assigned to any cluster
Removes thousands of likely irrelevant sentences
Remaining clusters more meaningful

Min cluster size = 15
Min samples = 5

Ensure lower number of clusters
More likely for clusters to represent something on the scale of dialect

Not authorship of common turns of phrase

Dimensionality Reduction using UMAP

Embeddings projected in 2D

7 of 9

RESULTS

.i. adǽ

.i. adæ /�.i. adae

adæ

adǽ /

ádǽ /

adé /

ádæ

Looking at the Results:

Experiment impacted by a lack of data
Clusters formed from only 4/12 texts after noisy data removed
Only one gloss from Milan Priscian

Looking at the clusters in the plot:

Considerable overlap between content from Milan glosses and Sg. Priscian glosses

Might affirm arguments that little dialectal variation can be found
Cannot result from similarity of topic (biblical texts and grammatical text)

Annals of Ulster form their own cluster to the centre-right of the plot.

Might reflect stylistic distinction between glosses and annalistic entries
Might also reflect some other form of stylistic distinction

***

Several smaller clusters around edge of plot from Milan glosses alone

Various forms address to God in the vocative
Might be topic-specific rather than dialectal

Can’t be tested without more biblical text data

Requires further investigation

8 of 9

FUTURE WORK

Application to modern Gaelic languages

Reattempt on Old Irish with more data

9 of 9

GO RAIBH MAITH AGAIBH!��THANK YOU!