RUMLEM: A Dictionary-Based Lemmatizer for Romansh

Dominic P. Fischer; Zachary Hopton; Jannis Vamvas

cs.CLApr 13, 2026

RUMLEM: A Dictionary-Based Lemmatizer for Romansh

Authors: Dominic P. Fischer, Zachary Hopton, Jannis Vamvas

Organizations: Department of Computational Linguistics, University of Zurich

Abstract

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.

Figures & tables

Figure 1: The main functionalities of rumlem .

	Unique Entries	Single Words	DE Translations	POS Tags	Gender	Infl. Verbs
Sursilvan	147,977	93,211	147,971	67,201	72,460	5,031
Sutsilvan	58,584	39,191	58,581	42,817	30,876	3,021
Surmiran	74,986	44,365	74,986	39,224	36,956	2,947
Puter	89,908	32,084	89,807	13,712	36,918	3,383
Vallader	106,690	35,322	106,438	10,854	48,435	3,779
RG	249,169	94,291	249,165	98,046	161,942	3,867

Table 1: Description of the Pledari Grond dictionary for each Romansh idiom as well as Rumantsch Grischun. A single ‘|’ means ‘thereof’: Unique entries, thereof X Single Words, German Translations, etc. Infl. Verbs refers to the number of unique verbs for which inflected forms are provided.

	Vocab	Mapped Forms	Lemmas	Noun	Adj	Verb	Other
Sursilvan	223,826	222,860	36,505	23,206	4,977	5,858	2,464
Sutsilvan	129,519	87,902	19,326	12,467	2,671	3,033	1,155
Surmiran	149,078	84,481	22,838	15,145	3,107	3,204	1,382
Puter	180,361	107,758	26,201	15,534	3,122	3,102	4,443
Vallader	165,354	109,090	30,479	19,841	4,821	3,625	2,192
RG	180,690	112,914	43,118	31,200	6,099	4,049	1,770

Table 2: rumlem ’s data coverage, single words only. Mapped Forms describes the amount of entries linked to a lemma. A single ‘|’ means ‘thereof’: Vocab, thereof X Mapped Forms, thereof X Lemmas, etc.

Variety	Form + features	Gloss
rm-surmiran	fomanto [PoS=ADJ; Gender=FEM; Number=SG]	hungrig
rm-surmiran	fomantar [PoS=V; VerbForm=PTCP; Tense=PST; Gender=FEM; Number=SG]	aushungern
rm-vallader	fomantà [PoS=ADJ; Gender=FEM; Number=SG]	ausgehungert
rm-vallader	fomantà [PoS=ADJ; Gender=FEM; Number=SG]	hungrig
rm-vallader	fomantada [PoS=N; Gender=FEM; Number=SG]	Ausgehungerte
rm-vallader	fomantada [PoS=N; Gender=FEM; Number=SG]	Hungrige

Table 3: rumlem ’s output given the token ‘fomantada’, with potential lemmas and morphological annotations of the form itself returned, as well as the German translation (variants).

Length	2–10	10–50	50–300	300–800	800+	All
Sursilvan	0.85	0.85	0.87	1.00	1.00	0.86
Sutsilvan	1.0	0.99	1.0	1.00	1.00	1.00
Surmiran	0.92	0.94	0.99	1.00	1.00	0.95
Puter	0.97	0.98	0.99	1.00	1.00	0.98
Vallader	0.94	0.91	0.93	1.00	1.00	0.92
RG	0.89	1.00	1.00	1.00	1.00	1.00

Table 4: Classification accuracy by Rumantsch variety across length buckets (number of tokens).

Figure 2: Distributions of Romansh (turquoise) and other Romance languages (rust) according to the highest Romansh variety score assigned to each text sample.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Length	2–10	10–50	50–300	300–800	800+	All
Surs.	0.76	0.81	0.84	0.83	0.78	0.84
Suts.	0.71	0.77	0.77	0.73	0.73	0.77
Surm.	0.73	0.80	0.84	0.82	0.79	0.82
Puter	0.79	0.83	0.84	0.84	0.81	0.84
Vall.	0.66	0.77	0.80	0.83	0.81	0.80
RG	0.93	0.79	0.79	0.80	0.78	0.79

Appendix

Table 5: Coverage ratios by text length and variety. Values for each variety and bucket are averaged across samples. ‘All’ shows the total number of lemmatizable tokens divided by the total tokens in the bucket/variety, not an average.

Length	2–10	10–50	50–300	300–800	800+	Tot
Surs.	68	2647	4173	7	5	6900
Suts.	6	1190	1795	5	7	3003
Surm.	1113	3751	2204	6	6	7080
Puter	660	2783	2468	5	7	5923
Vall.	277	2209	3202	5	7	5700
RG	9	1218	3052	6	6	4291

Appendix

Table 6: #samples by Romansh variety across buckets.

Length	50–300	300–800	800–2000	Tot
French	2517	1551	693	4761
Italian	2493	1620	671	4784
Romanian	2128	1513	922	4563
Catalan	2595	1575	593	4763
Romansh	4242	661	75	4978
Tot	13975	6920	2954	23849

Appendix

Table 7: #samples by Romance language across buckets.

Figure 3: Romansh (turqoise) and other Romance languages (rust) winning variety score distributions in the bucket with token length 300–800.

Figure 4: Romansh (turqoise) and other Romance languages (rust) winning variety score distributions in the bucket with token length 800–2000.

Low-Resource Language Processing

Low-resource language processing builds and evaluates language technologies for languages with limited data, tools, or digital representation.

See all 565 papers

Robust Language Identification for Romansh Varieties
Mar 16, 2026Charlotte Model, Sina Ahmadi, Jannis VamvasLanguage Identification Low-Resource Language Processing
University of Zurich
LombardoGraphia: Automatic Classification of Lombard Orthography Variants
Mar 30, 2026Edoardo Signoroni, Pavel RychlýLow-Resource Language Processing
Faculty of Informatics, Masaryk University, Botanická 68a, 602 00 Brno, Czechia
ROMEVA: Geometry-Preserving Vocabulary Expansion for Roman Urdu Language Models
Jun 21, 2026Mahnoor Khan, Afsheen Asif, Milhan Afzal Khan +2Multilingual Language Models Word Embeddings
School of Electrical Engineering and Computer Science (SEECS), National University of Sciences and Technology (NUST), Islamabad, Pakistan · Department of Computer Science, University of Agriculture Faisalabad, Faisalabad, Pakistan

Mar 16, 2026J/K move · Enter open · S save

Robust Language Identification for Romansh Varieties

Charlotte Model, Sina Ahmadi, Jannis Vamvas

University of Zurich

Language Identification Low-Resource Language Processing

Open paper AskPDF arXiv

The Romansh language has several regional varieties, called idioms, which sometimes have limited mutual intelligibility. Despite this linguistic diversity, there has been a lack of documented efforts to build a language identification (LID) system that can distinguish between these idioms. Since Romansh LID should also be able to recognize Rumantsch Grischun, a supra-regional variety that combines elements of several idioms, this makes for a novel and interesting classification problem. In this paper, we present a LID system for Romansh idioms based on an SVM approach. We evaluate our model on a newly curated benchmark across two domains and find that it reaches an average in-domain accuracy of 97%, enabling applications such as idiom-aware spell checking or machine translation. Our classifier is publicly available.

Abstract

Low-Resource Language Processing

Explore similar work

Robust Language Identification for Romansh Varieties

LombardoGraphia: Automatic Classification of Lombard Orthography Variants

ROMEVA: Geometry-Preserving Vocabulary Expansion for Roman Urdu Language Models