pith. sign in

arxiv: 2604.11233 · v1 · submitted 2026-04-13 · 💻 cs.CL

RUMLEM: A Dictionary-Based Lemmatizer for Romansh

Dominic P. Fischer , Zachary Hopton , Jannis Vamvas This is my paper

Pith reviewed 2026-05-10 15:21 UTC · model grok-4.3

classification 💻 cs.CL
keywords lemmatizationRomanshmorphological databasesvariety classificationlow-resource languagesdictionary-based NLPlanguage identification
0
0 comments X

The pith

A dictionary-based lemmatizer for Romansh covers 77-84% of typical text words across its five varieties and identifies the variety in 95% of cases.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces RUMLEM, a lemmatizer that relies on separate morphological databases for each of the five main Romansh varieties plus the standard Rumantsch Grischun. This design maps inflected forms to lemmas and simultaneously reveals which variety a text comes from. A reader would care because Romansh remains a low-resource language with dialectal splits, so a working tool could unlock basic text processing without needing massive new datasets. The reported coverage and classification rates show that community-curated dictionaries can deliver usable results today.

Core claim

RUMLEM leverages dedicated community-driven morphological databases for each Romansh variety and the supra-regional standard to perform lemmatization, reaching 77-84% word coverage on typical texts. Testing on 30,000 texts of varying lengths yields 95% accuracy in identifying the correct variety, and a proof-of-concept shows the same mechanism can separate Romansh from non-Romansh language.

What carries the argument

Variety-specific morphological databases that allow lookup of inflected forms to return the lemma while also flagging the source variety through database match patterns.

If this is right

  • Basic NLP pipelines for Romansh can now include lemmatization without training new models from scratch.
  • Variety classification becomes a byproduct of lemmatization rather than a separate task.
  • The same database approach can support language identification between Romansh and other languages.
  • Higher-level applications such as search or summarization gain a practical entry point for this language.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Community-maintained linguistic databases may offer a scalable route for other languages that lack large annotated corpora.
  • The method could extend to additional Romansh varieties or related minority languages with similar internal variation.
  • Combining the dictionary output with lightweight statistical models might raise coverage further while preserving the variety signal.

Load-bearing premise

The community-driven morphological databases are complete and accurate enough for all varieties, and the 30,000 evaluation texts represent real-world usage.

What would settle it

A fresh collection of Romansh texts from underrepresented regions or speakers that drops word coverage below 70% or variety identification below 90% accuracy would undermine the reported effectiveness.

Figures

Figures reproduced from arXiv: 2604.11233 by Dominic P. Fischer, Jannis Vamvas, Zachary Hopton.

Figure 1
Figure 1. Figure 1: ). Together with Model (2025), our system is among the very few systems to reliably perform such a classification; with the additional benefit lavuraiva Lemma: ‘lavurar’ or ‘lavurer’ Idiom: Vallader or Puter Morph: Impf. Tense, 1./3. Sg [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distributions of Romansh (turquoise) and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Romansh (turqoise) and other Romance lan [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Romansh (turqoise) and other Romance lan [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper presents RUMLEM, a dictionary-based lemmatizer for the five main varieties of Romansh plus the standard Rumantsch Grischun, built on community-driven morphological databases. It claims that RUMLEM covers 77-84% of words in a typical Romansh text and, when applied to 30,000 texts of varying lengths, correctly identifies the variety in 95% of cases. A brief proof-of-concept for Romansh vs. non-Romansh language classification is also included.

Significance. If the empirical results hold under proper validation, the work supplies a practical, immediately usable resource for a genuinely low-resource language with dialectal variation. The dictionary-driven design is transparent and leverages existing community assets rather than requiring new annotated corpora, which is a pragmatic strength for Romansh NLP. The variety-identification application is a natural and potentially useful byproduct.

major comments (3)
  1. [Abstract / Evaluation] Abstract and Evaluation section: The headline figures (77-84% coverage; 95% variety identification on 30,000 texts) are presented without any description of the test-set construction, the procedure for handling out-of-vocabulary tokens, the definition of 'typical text,' or any error analysis. Because coverage is bounded by database completeness, the absence of these details makes it impossible to judge whether the reported numbers are robust or merely reflect the particular evaluation sample.
  2. [Method / Database description] Database and Method sections: The central claim that the five variety-specific plus standard databases are 'comprehensive' is load-bearing for both coverage and the 95% variety-classification result, yet no quantitative assessment of database size, coverage gaps, inter-variety consistency, or labeling accuracy is supplied. Any systematic incompleteness or cross-variety leakage would directly cap performance and bias the highest-match-rate classifier.
  3. [Evaluation] Evaluation section: No baseline lemmatizer (rule-based, statistical, or neural) is reported, nor is there any comparison against a simple string-matching or frequency-based variety classifier. Without these controls it is unclear whether the 95% figure represents a genuine advance or could be matched by far simpler methods.
minor comments (2)
  1. [Abstract] Abstract: '30'000' uses a locale-specific thousands separator; replace with '30,000' for international readability.
  2. [Proof-of-concept subsection] The proof-of-concept language-classification experiment is mentioned only in the abstract and receives no quantitative results or experimental details in the body; either expand or remove the claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive critique. We address each major comment below with clarifications drawn from our evaluation process and commit to revisions that add the requested transparency without altering the core claims.

read point-by-point responses

  1. Referee: [Abstract / Evaluation] Abstract and Evaluation section: The headline figures (77-84% coverage; 95% variety identification on 30,000 texts) are presented without any description of the test-set construction, the procedure for handling out-of-vocabulary tokens, the definition of 'typical text,' or any error analysis. Because coverage is bounded by database completeness, the absence of these details makes it impossible to judge whether the reported numbers are robust or merely reflect the particular evaluation sample.

    Authors: The 30,000 texts were sampled from publicly available Romansh corpora maintained by community organizations, stratified across varieties and lengths to represent real-world usage; 'typical text' denotes documents of median length (~300 words) in that collection. OOV tokens were left unlemmatized and excluded from match-rate calculations for variety identification. We will expand the Evaluation section with explicit test-set methodology, OOV handling rules, and a concise error analysis of the 5% misclassifications. revision: yes

  2. Referee: [Method / Database description] Database and Method sections: The central claim that the five variety-specific plus standard databases are 'comprehensive' is load-bearing for both coverage and the 95% variety-classification result, yet no quantitative assessment of database size, coverage gaps, inter-variety consistency, or labeling accuracy is supplied. Any systematic incompleteness or cross-variety leakage would directly cap performance and bias the highest-match-rate classifier.

    Authors: The databases are the official, expert-curated morphological resources released by the primary Romansh language institutions; we will insert summary statistics (lemma counts per variety, overlap rates) and a brief discussion of curation standards and inter-variety consistency in the revised Method section. This directly addresses potential leakage concerns while preserving the transparency of the dictionary-driven design. revision: yes

  3. Referee: [Evaluation] Evaluation section: No baseline lemmatizer (rule-based, statistical, or neural) is reported, nor is there any comparison against a simple string-matching or frequency-based variety classifier. Without these controls it is unclear whether the 95% figure represents a genuine advance or could be matched by far simpler methods.

    Authors: The paper emphasizes a pragmatic, zero-training-data approach suited to low-resource settings. We will add a simple frequency-based variety classifier and a string-overlap baseline in the revised Evaluation section to provide context; neural baselines remain infeasible given the absence of large annotated Romansh corpora. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical results from direct lookup on held-out texts

full rationale

The paper presents RUMLEM as a straightforward dictionary lookup system built on pre-existing community morphological databases for each Romansh variety. Coverage percentages (77-84%) and variety identification accuracy (95%) are obtained by applying the lemmatizer to 30,000 separate evaluation texts and counting matches, with no mathematical derivations, fitted parameters, or self-referential definitions. No equations or self-citations reduce any claimed result to quantities defined by the same data or inputs. The method is self-contained against external benchmarks (the held-out texts), satisfying the default expectation of no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The system depends on pre-existing community morphological databases but introduces no new free parameters, mathematical axioms, or postulated entities; all performance claims rest on empirical lookup and counting.

pith-pipeline@v0.9.0 · 5456 in / 1201 out tokens · 40891 ms · 2026-05-10T15:21:14.780814+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

  1. [1]

    In4th Workshop on African Natural Language Processing

    Lexicon and rule-based word lemmatization approach for the somali language. In4th Workshop on African Natural Language Processing. Marco Passarotti, Marco Budassi, Eleonora Litta, and Paolo Ruffolo. 2017. The lemlat 3.0 package for morphological analysis of Latin. InProceedings of the NoDaLiDa 2017 Workshop on Processing Histor- ical Language, pages 24–31...

    work page 2017

  2. [2]

    Cla Rauch ha orientà davart l’Archiv Cul- tural d’Engiadina Bassa (fotografia: Bene- dict Stecher).[...] Haben Sie noch kein Konto? Registrieren Sie sich hier [...]

    SwissBERT: The multilingual language model for Switzerland. InProceedings of the 8th edition of the Swiss Text Analytics Conference, pages 54–69, Neuchatel, Switzerland. Association for Computa- tional Linguistics. A Preprocessing Tests The example below illustrates a preprocessing test case. The first line shows the raw dictionary data, while the lines f...

    work page 2000