Lombard, an underresourced language variety spoken by approximately 3.8 million people in Northern Italy and Southern Switzerland, lacks a unified orthographic standard. Multiple orthographic systems exist, creating challenges for NLP resource development and model training. This paper presents the first study of automatic Lombard orthography classification and LombardoGraphia, a curated corpus of 11,186 Lombard Wikipedia samples tagged across 9 orthographic variants, and models for automatic orthography classification. We curate the dataset, processing and filtering raw Wikipedia content to ensure text suitable for orthographic analysis. We train 24 traditional and neural classification models with various features and encoding levels. Our best models achieve 96.06% and 85.78% overall and average class accuracy, though performance on minority classes remains challenging due to data imbalance. Our work provides crucial infrastructure for building variety-aware NLP resources for Lombard.
Figures & tables
Orthography
Sample Text
MILCLASS
Sera settada in terra col coo in man, e i gombed sui genœucc: me ziffolava el vent in di cavij: demaneman che vegneva on quaj bôff, el me portava comè ona vôs che vegneva de lontan; […] "I was seating on the ground with the head in the hands, and the elbows on the knees: as it came in whiffs, it brought me as a voice from afar; […]"
LOCC
Sera setada in tera cul coo in man, e i gumbed süi genögg: me zifulava el vent in di cavii: demaneman che vegneva un quai buf, el me purtava cumè una vus che vegneva de luntan; […]
SL
S’era setada in terra col coo in man, e i gombeds sui jenœgg: me cifolava el vent ind i cavei: demaneman qe vegneva un quai bof, al me portava comè una vox qe vegneva de lontan; […]
NOL
S’era setada in terra col coo in man, e i gombet sui sgenoeugg: me scifolava el vent in di cavej: demaneman che vegneva un quaj bof, el me portava comè una vos che vegneva de lontan; […]
LSI
Cula paròla Ticines a sa inteend i dialett che i è parlaa in Tesin (vün dai 26 cantón svizzer) e in Mesulcina e Calanca (dó vall dal Canton Grison). […] "With the word Ticines , we mean the dialects spoken in Ticino (one of the 26 Swiss cantons) and in Mesolcina and Calanca (two valleys of Grisons)"
LORUNIF
L’ortografìa orientàl ünificàda l’è ’n tentatìf de dàga ’na rispòsta al bizògn de ’n sistéma de scritüra bù per töte le varietà piö parlàde del Lombàrt Orientàl […] "The unified oriental orthography is an attempt to answer the need for a writing system for all the most spoken varieties of Eastern Lombard."
Table 1: Summary of the Lombard orthographies in the dataset. The first column give their abbreviated name, while the second shows a text sample for each one. When available (MILCLASS, LOCC, SL, NOL) we used the same text. 19 19 19 From La Fuggitiva ("The Fugitive") by Tommaso Grossi (1791-1853). Retrieved from https://lmo.wikipedia.org/wiki/Dialett_milanes . Own translation.
Class
Type
Total Lines
Train
Valid
Test
Dataset Total
Removed
N
%
N
%
N
%
N
%
N
%
MILCLASS
Local
79,196
3,606
40.29
471
42.13
446
39.89
4,523
40.43
74,673
94.29
LOCC
Macro-dialectal
34,794
2,907
32.48
345
30.86
380
33.99
3,632
32.47
31,162
89.56
LORUNIF
Macro-dialectal
76,455
1,901
21.24
240
21.47
229
20.48
2,370
21.19
74,085
96.90
SL
Pan-Lombard
3,761
174
1.94
22
1.97
16
1.43
212
1.90
3,549
94.36
NOL
Pan-Lombard
2,349
109
1.22
9
0.81
8
0.72
126
1.13
2,223
94.64
Table 2: Distribution of orthographic classes across train, validation, and test sets. Total Lines shows the number of lines in the Wikipedia corpus before filtering. Dataset Total reports the number of lines in the cleaned corpus. Removed shows lines filtered out due to quality criteria (N) and the percentage removed relative to Total Lines (%). The no_tag category contains lines without orthography tags that were excluded from the dataset.
Model
MILCLASS
LOCC
LORUNIF
SL
NOL
CRES
BREMOD
BERGDUC
Overall
Avg Class
Traditional Models - Logistic Regression
Log. byte
97.31
92.89
89.08
100.0
75.00
94.74
53.85
83.33
93.38
85.78
Log. byte+char+word
97.31
94.74
94.76
100.0
75.00
100.0
46.15
66.67
95.08
84.33
Log. char
95.07
93.42
84.28
100.0
75.00
94.74
53.85
83.33
91.67
84.96
Log. word
94.62
94.21
88.65
93.75
37.50
94.74
69.23
66.67
92.39
79.92
Traditional Models - Support Vector Machine
Table 3: Complete per-class accuracy (%) across all 24 models, with overall and average class accuracy. Bold indicates the best performance per metric. Classes are ordered by sample size. Overall accuracy measures performance on all test samples, while high average class accuracy shows consistent performance across classes.
Figure 1: Boxplot of the results for all models. Traditional models are in purple, and neural models are in blue.
Several of the world's languages are still under-resourced in terms of Natural Language Processing (NLP) tools. This is mostly due to the lack of high-quality datasets to train, develop, and evaluate systems and models for several tasks, such as Machine Translation (MT). We conduct a manual audit of the parallel and monolingual corpora available for Lombard, an under-resourced language continuum from Italy. Our analysis reveals that the perceived abundance of web-scraped data is an illusion, with massive datasets plagued by severe language misidentification, boilerplate text, and non-linguistic noise. Furthermore, we analyze the orthographic composition of the valid Lombard portions across web-scraped datasets, curated corpora, and benchmarks. Our findings show conflicting orthographical systems and severe representational bias across all corpora: high-quality data is heavily skewed towards Western Lombard varieties, with Eastern ones left on the margins. This underscores the need for variety-aware, community-driven data curation rather than purely quantity-driven scraping.
We present a crowdsourced dataset for Piedmontese, an endangered Romance language of northwestern Italy. The dataset comprises 145 Italian-Piedmontese parallel sentences derived from Flores+, with translations produced by speakers writing in their natural orthographic style rather than adhering to standardized conventions, along with manual word alignment. We use this resource to benchmark several large language models on tokenization parity, topic classification, and machine translation. Our analysis reveals that Piedmontese incurs a tokenization penalty relative to higher-resource Romance languages, yet LLMs achieve classification performance approaching that of Italian, French, and English. Machine translation results are asymmetric: models translate adequately from Piedmontese into high-resource languages, but generation into Piedmontese remains challenging. The dataset and code are publicly released.
Gianluca Vico, Jindřich Libovický
Charles University, Faculty of Mathematics and Physics, Institute of Formal and Applied Linguistics, Malostranské náměstí 25, 118 00 Praha, Czech Republic
Lemmatization -- the task of mapping an inflected word form to its dictionary form -- is a crucial component of many NLP applications. In this paper, we present RUMLEM, a lemmatizer that covers the five main varieties of Romansh as well as the supra-regional standard variety Rumantsch Grischun. It is based on comprehensive, community-driven morphological databases for Romansh, enabling RUMLEM to cover 77-84% of the words in a typical Romansh text. Since there is a dedicated database for each Romansh variety, an additional application of RUMLEM is variety-aware language classification. Evaluation on 30'000 Romansh texts of varying lengths shows that RUMLEM correctly identifies the variety in 95% of cases. In addition, a proof of concept demonstrates the feasibility of Romansh vs. non-Romansh language classification based on the lemmatizer.
Dominic P. Fischer, Zachary Hopton, Jannis Vamvas
Department of Computational Linguistics, University of Zurich