Pith. sign in

REVIEW 5 major objections 7 minor 1 cited by

A single multilingual model can jointly predict morpheme boundaries and interlinear glosses, and an interleaved output format forces the two to align.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 10:09 UTC pith:HBEX3RNH

load-bearing objection A genuinely useful joint segmentation-and-glossing model with an honest but confounded SOTA claim; worth engaging, but the GlossLM comparison and the alignment metric need tightening before the headline result is bulletproof. the 5 major comments →

arxiv 2601.10925 v3 pith:HBEX3RNH submitted 2026-01-16 cs.CL

Massively Multilingual Joint Segmentation and Glossing

Michael Ginn , Lindia Tjuatja , Enora Rice , Ali Marashian , Maria Valentini , Jasmine Xu , Graham Neubig , Alexis Palmer This is my paper
classification cs.CL
keywords interlinear glossingmorphological segmentationjoint predictionlanguage documentationsequence-to-sequence modelalignment metriclow-rank adaptationlow-resource languages
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Automated interlinear glossing has been accurate in the abstract, but linguists in documentation settings found the outputs untrustworthy because models assigned glosses to whole words without revealing the morpheme boundaries. This paper proposes PolyGloss, a family of multilingual sequence-to-sequence models pretrained on an expanded, cleaned corpus of 353,266 examples, that predicts morphological segmentation and glosses together in one forward pass. Its best variant uses an interleaved format in which every gloss label is followed by its morpheme in parentheses, so any well-formed output is by construction aligned between segmentation and glosses. On nine evaluation languages, the paper reports lower morpheme error than the prior GlossLM model and much lower error than open-source LLMs up to 8B parameters, along with near-perfect alignment, and it shows that per-language perplexity predicts glossing accuracy with r-squared 0.951. It also demonstrates rapid adaptation to an unseen language with low-rank adapters trained in about twelve minutes.

Core claim

PolyGloss is the first neural model family the authors are aware of that jointly predicts interlinear glosses and their morphological segmentation from raw text. The core finding is that the two tasks are not in tension: training a single sequence-to-sequence model on both, with glosses and segments interleaved as GLOSS(morpheme) pairs, improves glossing (average morpheme error rate 0.234) and segmentation (morpheme F1 0.862) while achieving a perfect alignment score of 1.000 on every test language, because the format guarantees that every gloss is attached to a specific morpheme. Joint training outperforms separate glossing and segmentation models (which fall to 0.915 alignment on average)

What carries the argument

The central mechanism is the interleaved prediction format: the model is trained to output strings in which each gloss label is immediately followed by its corresponding morpheme in parentheses, e.g., 'INTERJ(o) you.know(wol_e)-ZERO(0)=ART(n)', turning joint segmentation and glossing into a single sequence-generation task with a hard per-word coupling. Because the causal decoder attends to the morpheme while generating its gloss, misalignment between the two predictions becomes structurally impossible for well-formed output. The paper defines a novel alignment score to quantify this coupling: predicted gloss and segmentation strings are abstracted to 'x' sequences (each morpheme becomes x, b

Load-bearing premise

The paper's headline results rest on a comparison against GLOSSLM that the authors concede is not a clean head-to-head, and on an alignment score that any well-formed interleaved output trivially maxes out — so both the glossing gains and the perfect alignment need controlled re-checks before the state-of-the-art claim can be taken at face value.

What would settle it

Re-train GLOSSLM on the expanded PolyGloss corpus with PolyGloss's learning rate, batch size, and epoch count; if its morpheme error rate on Arapaho, Tsez, and Gitksan matches PolyGloss's, the joint objective is not the source of the improvement. Separately, ask two linguists to judge, on a random 100-sentence sample, whether each PolyGloss gloss-morpheme pair attaches to the semantically correct morpheme; if disagreement with the 1.000 alignment score is frequent, the alignment metric overstates real alignment.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Documentation software can present linguists with a single suggested annotation that includes both morpheme boundaries and glosses, removing the extra human step of re-segmenting words and making automated suggestions interpretable.
  • A curation system can compute per-language perplexity on its own unlabeled text and decide in advance whether to show PolyGloss's suggestions, fall back to a trivial gloss predictor, or show nothing for that language.
  • Field linguists with a few hundred glossed examples and a single GPU can adapt PolyGloss to an unseen language in about twelve minutes via low-rank adaptation, rather than training or serving a monolingual model.
  • Because the interleaved format guarantees alignment, downstream consumers of interlinear glossed text (paradigm learners, taggers, machine translation pivots) receive annotations whose gloss and segmentation lines cannot disagree in count or order.
  • The alignment score, usable as an RLVR reward, offers a path to improving glossing for new languages that have gold glosses but no gold morphological segmentations.

Where Pith is reading between the lines

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

  • Editorial inference: the perfect 1.000 alignment score is partly tautological — any well-formed interleaved output trivially scores 1.000 on the abstracted 'x' edit-distance metric — so the metric measures format compliance, not whether the gloss is semantically attached to the correct morpheme; human judgment of attachment correctness would be a stronger test.
  • Editorial inference: because the GlossLM baseline was not trained on the full cleaned corpus, the reported glossing gains on overlapping languages may be partly attributable to more data, better formatting, and new hyperparameters rather than to the joint interleaved objective; a controlled re-training of GlossLM on the PolyGloss corpus would isolate the contribution.
  • Editorial inference: the strong perplexity–MER correlation was measured on only nine languages; a natural test is whether it extrapolates to typologically distant or under-represented languages, and if so it becomes a cheap monitoring tool for any documentation workflow.
  • Editorial inference: the interleaved format could be extended to constrained decoding against a user's preferred glossing convention, or to multi-word constructions, suggesting a family of hard-alignment sequence formats beyond the single-morpheme-in-parentheses scheme.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 7 minor

Summary. The paper presents POLYGLOSS, a family of ByT5-based seq2seq models that jointly predict morphological segmentation and interlinear glosses from raw text, trained on an extended and cleaned version of the GLOSSLM corpus (341k examples, 2,077 languages). Three task formats are compared (multitask, concatenated, interleaved), with the interleaved format—where each gloss is immediately followed by its morpheme in parentheses—reported as best. The authors claim SOTA glossing performance, outperforming GLOSSLM on arp/ddo/git and beating ICL baselines (Qwen 3 0.6B, Gemma 3 4B, Aya Expanse 8B) on glossing, segmentation, and alignment. Additional contributions include a per-language perplexity heuristic for predicting glossing error, a LoRA adaptation demonstration on Vamale, and a small-scale GRPO experiment using the alignment score as a reward. The central empirical claim is the joint model's superiority on all three metrics, with the interleaved format achieving perfect alignment by construction.

Significance. If the claims hold, this is a useful step toward making automated glossing tools practical for documentary linguists: it is the first model to jointly output morpheme boundaries and aligned glosses, addresses user concerns about interpretability (Rice et al., 2025), and provides a concrete adaptation recipe (LoRA) plus a heuristic for deciding when to trust model predictions. The extended corpus and open release are valuable resources. The ablation comparing multitask, concatenated, and interleaved formats is a meaningful design study. However, the paper's headline claims are weakened by two load-bearing issues: the GLOSSLM comparison is confounded by corpus and hyperparameter changes, and the alignment score is trivially 1.000 for any well-formed interleaved output, making the alignment 'win' a format artifact rather than evidence about genuine segmentation–gloss alignment. The perplexity–MER correlation is also evaluated only on the nine training/evaluation languages, so its generalizability is unestablished.

major comments (5)
  1. [§5.1, Table 3] The claim 'outperforms GLOSSLM on glossing' (Abstract) rests on a comparison that is not head-to-head. The text concedes 'the original GLOSSLM model did not include all of the segmented training data, so this is not a perfect head-to-head comparison.' The POLYGLOSS gains over GLOSSLM on arp/ddo/git (MER 0.161→0.152, 0.095→0.072, 0.870→0.597) co-vary with a 91,416-example corpus expansion (§2) and changed hyperparameters (15 epochs, lr 5e-5, batch 64; §4.3). §6 even attributes the gains to 'better dataset preprocessing and improved training hyperparameters.' Without an ablation that trains a GLOSSLM-style model on the new corpus with the new hyperparameters (or at least a controlled comparison that isolates the joint objective), the superiority over GLOSSLM is not established. This is the central SOTA claim and needs a controlled experiment.
  2. [§3.3, Table 5] The alignment score is 1.000 for any well-formed interleaved output by construction. The abstracting procedure replaces every morpheme in both segmentation and gloss sequences with 'x' and keeps boundaries; in the interleaved format, the gloss and morpheme are co-located, so the two abstracted strings are identical if the format is respected. Thus Table 5's perfect 1.000 for POLYGLOSS is not evidence of correctly aligned glosses—it only indicates the model adhered to the output template. The claimed 'beats ... on alignment' is therefore a format artifact. The metric needs to be redefined to evaluate whether the segmented units correspond to the glossed units in a way that is not trivially satisfied (e.g., by checking word boundaries, or by comparing against gold alignments, or by penalizing empty/extra segments). The current footnote in §3.3 acknowledges the score can be high without cor
  3. [§6.1, Figure 2] The claim that 'per-language perplexity can roughly predict glossing performance for any language' is based on a correlation computed on the same nine evaluation languages that were used for training and validation. The validation-set perplexity and test-set MER are both measured for languages in the training corpus; this is an in-sample fit, not a prediction for unseen languages. The paper even calls it a 'practical heuristic' but provides no evidence on held-out languages. A small leave-one-language-out experiment or evaluation on a few truly unseen languages (e.g., Vamale or languages in the Fieldwork dataset not in the test set) is needed to support the generalization claim.
  4. [Tables 3–5] None of the main tables report error bars, confidence intervals, or significance tests, despite the text in §7.1 noting 'error bars are large because of the variance across languages, and should not be used to determine significance.' With only nine languages and substantial variance (e.g., git MER 0.870 vs 0.597, which drives much of the average improvement), the aggregate differences, and even some per-language differences (arp 0.152 vs 0.161), may be within noise. The paper should provide per-language confidence intervals or a paired significance test across languages (e.g., paired bootstrap or Wilcoxon) for the primary GLOSSLM comparison and the task-format ablations. Without this, the 'state-of-the-art' language is not statistically grounded.
  5. [§8, Figure 6] The LoRA adaptation experiment compares POLYGLOSS with a raw ByT5 base that has never seen glossing data; it is unsurprising that raw ByT5 'never improves.' This is a strawman baseline that does not isolate the benefit of continued pretraining. The demonstration that POLYGLOSS adapts with 380 examples is still informative, but the figure's message would be stronger if the baseline were a GLOSSLM model or a ByT5 model trained on the multitask format. As it stands, the adaptation claim ('quickly adapted to a new dataset') is only weakly supported, though this is a secondary contribution.
minor comments (7)
  1. [§3.1] 'c.f.' should be 'cf.'; also consider defining 'morpheme error rate' abbreviation before first use in Table 3 caption.
  2. [§6.1] Figure 2 caption uses 'r2 = 0.951' while the text says 'r² = 0.951'; use a consistent notation (R² is standard for the coefficient of determination).
  3. [Appendix B, Table 8] The Qwen 0.6B (ICL) row appears to be missing entries for Lezgi, Natugu, Nyangbo, and Ruuli (only 7 numbers are shown instead of 10). Please correct the table.
  4. [§7.1] The sentence 'its alignment score is significantly worse' uses 'significantly' without a statistical test. Rephrase to 'numerically worse' or add a test.
  5. [§9] The GRPO experiment reports 'all three metrics improve slightly' but no error bars or significance; this is a feasibility demonstration and should be labeled as such.
  6. [References] The reference 'Aycock et al.' is incomplete (no year, no venue) and appears in the Related Work section without a full citation; the duplicate 'Girrbach (2023a)/(2023b)' entries should be merged.
  7. [Table 4] The 'Pipeline' row in Table 8 repeats the 'Finetuned ByT5 (separate)' segmentation numbers exactly for several languages; verify whether this is an error or by construction (e.g., because the pipeline's first stage is the separate segmentation model).

Circularity Check

1 steps flagged

Alignment metric is definitionally forced by the interleaved format; main glossing/segmentation results are empirical, though the GLOSSLM comparison is admittedly not head-to-head.

specific steps
  1. self definitional [§3.3 (Alignment), §4.1 (Interleaved Prediction), §7.3 (Effect of Task Format), Table 5]
    "each morpheme sequence is converted to a single “x” character, and morpheme boundaries (“-” and “=”) are left unchanged (§3.3); “We hypothesized that this setting would have the best alignment, as any well-formed output should be perfectly aligned” (§4.1); “The interleaved format results in perfect alignment thanks to its explicit constraint” (§7.3); Table 5: POLYGLOSS 1.000 for all languages."

    The alignment metric abstracts both segmentation and gloss sequences to x's plus boundary markers. In the interleaved format each gloss is emitted with its morpheme in parentheses, so for any well-formed output the two abstracted sequences are identical by construction, giving an edit distance of zero and a score of 1.000. PolyGloss's perfect alignment score in Table 5 is therefore an artifact of the output format, not evidence of learned alignment to the intended gloss–morpheme structure; the paper itself notes “a model could achieve a perfect alignment score while predicting incorrect glosses and segmentation” (fn. 2, §3.3). The advertised advantage over LLMs on alignment reduces to format conformance.

full rationale

Most of PolyGloss's derivation chain is empirical and not circular: glossing and segmentation claims rest on held-out test sets (Tables 3–4), monolingual/multilingual ablations (§7), and external baselines (GLOSSLM, in-context LLMs). The corpus extension and LoRA adaptation are separate empirical contributions. Self-citations to GlossLM and Rice et al. are background/motivation and comparison points; no load-bearing argument reduces to an unverified self-citation or to an imported uniqueness/ansatz. The one definitional circularity is the alignment score: §3.3 abstracts sequences so that the interleaved format trivially scores 1.000, and §7.3 credits the format for “perfect alignment.” Thus the claimed superiority over LLMs on alignment is a format-conformance artifact. Separately, the GLOSSLM comparison is admittedly not head-to-head (§5.1: “the original GLOSSLM model did not include all of the segmented training data”), so the SOTA glossing claim is confounded by corpus and hyperparameter changes; this is a validity concern rather than a circularity, and it keeps the score moderate instead of making the glossing result circular. Overall, partial evaluation-design circularity on the alignment component, with the other central claims intact.

Axiom & Free-Parameter Ledger

10 free parameters · 5 axioms · 0 invented entities

The central empirical claims rest on several domain assumptions about data quality, metric validity, and language representativeness. No new physical entities are introduced. The listed hyperparameters are hand-chosen and not fitted by optimization over held-out data.

free parameters (10)
  • learning_rate = 5e-5
    Chosen manually for ByT5 and Qwen PolyGloss training (§4.3, Table 6).
  • batch_size = 64 (ByT5), 18 (Qwen)
    Hand-selected batch sizes (§4.3, Table 6).
  • training_epochs = 15
    Fixed for all PolyGloss runs (§4.3).
  • lora_rank = 8
    Rank-8 adapters for Vamale adaptation (§8).
  • lora_epochs = 25
    LoRA training epochs for Vamale (§8).
  • grpo_beta = 0.1
    KL penalty coefficient for GRPO tuning (§9).
  • grpo_lr = 5e-5
    Learning rate for GRPO tuning (§9).
  • grpo_batch_and_group_size = 8
    Batch and group sizes for GRPO (§9).
  • grpo_epochs = 50
    Number of GRPO epochs (§9).
  • sampling_temperature = 0.6
    Sampling temperature for GRPO (§9); top-p 0.9 and repetition penalty 1.05 also set by hand.
axioms (5)
  • domain assumption IGT corpus annotations are consistent enough after cleaning for joint training; the 34,894 misaligned examples can be safely placed in training rather than removed.
    §2: 'we ensure the offending examples are within the training split, so as not to affect evaluation.' If misaligned examples teach bad alignments, the model could be harmed.
  • domain assumption Byte-level tokenization (ByT5) is suitable for morphological segmentation and rare languages.
    §4.2: 'By using byte-level tokenization, ByT5 avoids the issues that arise for rare languages with subword tokenizers.' This is taken from prior work (He et al., 2023), not re-derived.
  • domain assumption The nine evaluation languages and the train/eval/test splits are representative and free of leakage from continued pretraining.
    §2, §6: The SOTA claim is averaged over these nine languages; if splits leaked, the MER numbers would be inflated.
  • domain assumption The proposed alignment score, computed after converting morpheme sequences to 'x' strings, is a meaningful measure of alignment quality.
    §3.3: The score is novel and not compared to human judgments; for the interleaved format it is 1.000 by construction, so its validity rests on the assumption that format-conformance equals alignment.
  • domain assumption Validation-set per-language perplexity reliably predicts test morpheme error rate on unseen languages.
    §6.1: The r²=0.951 is computed on the same nine languages; the paper does not test on a held-out language, so the generalization is assumed.

pith-pipeline@v1.3.0-alltime-deepseek · 14941 in / 15106 out tokens · 142869 ms · 2026-08-03T10:09:10.232556+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Massively Multilingual Joint Segmentation and Glossing." pith.science (2026). https://pith.science/paper/HBEX3RNH

@misc{pith2026260110925,
  author       = {Pith},
  title        = {Pith review of: Massively Multilingual Joint Segmentation and Glossing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HBEX3RNH}},
  note         = {Machine review of arXiv:2601.10925}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Automated interlinear gloss prediction with neural networks is a promising approach to accelerate language documentation efforts. However, while state-of-the-art models like GlossLM achieve high scores on glossing benchmarks, user studies with linguists have found critical barriers to the usefulness of such models in real-world scenarios. In particular, existing models typically generate morpheme-level glosses but assign them to whole words without predicting the actual morpheme boundaries, making the predictions less interpretable and thus untrustworthy to human annotators. We conduct the first study on neural models that jointly predict interlinear glosses and the corresponding morphological segmentation from raw text. We run experiments to determine the optimal way to train models that balance segmentation and glossing accuracy, as well as the alignment between the two tasks. We extend the training corpus of GlossLM and pretrain PolyGloss, a family of seq2seq multilingual models for joint segmentation and glossing that outperforms GlossLM on glossing and beats various open-source LLMs on segmentation, glossing, and alignment. In addition, we demonstrate that PolyGloss can be quickly adapted to a new dataset via low-rank adaptation.

Figures

Figures reproduced from arXiv: 2601.10925 by Alexis Palmer, Ali Marashian, Enora Rice, Graham Neubig, Jasmine Xu, Lindia Tjuatja, Maria Valentini, Michael Ginn.

Figure 1
Figure 1. Figure 1: An interlinear glossed text example, showing [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Relationship between validation set perplex [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Scores for POLYGLOSS multilingual models using three different data formats. Scores are averaged across nine languages and reported with standard error. 7.3 Effect of Task Format Results for the three ByT5-based POLYGLOSS models are very similar regardless of task format ( [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation comparing monolingual joint mod [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Morpheme error rate for Vamale when training [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Scores before and after GRPO tuning for Gitksan. ing that all three metrics improve slightly. This approach could be scaled to the full training dataset as an additional post-training step. Furthermore, RL could be used when adapting the model to a new language that lacks gold-labeled morphologi￾cal segmentations, since the alignment score will force the segmentation to (minimally) be aligned with the gold… view at source ↗
Figure 8
Figure 8. Figure 8: Scores for POLYGLOSS multilingual models using the Interleaved format and different base models. Scores are averaged across nine languages and reported with standard error. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. GlossAssist -- A Tool to Simplify Corpus Creation and Study the Effect of NLP Models in Low-Resource Documentation Settings

    cs.CL 2026-06 unverdicted novelty 5.0

    GlossAssist presents an interactive glossing tool built on CWoMP that incorporates annotator corrections as active learning to expand a mutable morpheme lexicon and improve predictions.

Reference graph

Works this paper leans on

44 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

    Seyede Faezeh Hosseini Alast and Sasan Baleghizadeh. 2024. https://doi.org/10.1177/13621688211020423 The interplay of glossing with text difficulty and comprehension levels . Language Teaching Research, 28(3):1201--1230

    work page doi:10.1177/13621688211020423 2024

  2. [2]

    Can llms really learn to translate a low-resource language from one grammar book? In The Thirteenth International Conference on Learning Representations

    Seth Aycock, David Stap, Di Wu, Christof Monz, and Khalil Sima'an. Can llms really learn to translate a low-resource language from one grammar book? In The Thirteenth International Conference on Learning Representations

  3. [3]

    Khuyagbaatar Batsuren, Omer Goldman, Salam Khalifa, Nizar Habash, Witold Kiera \'s , G \'a bor Bella, Brian Leonard, Garrett Nicolai, Kyle Gorman, Yustinus Ghanggo Ate, Maria Ryskina, Sabrina Mielke, Elena Budianskaya, Charbel El-Khaissi, Tiago Pimentel, Michael Gasser, William Abbott Lane, Mohit Raj, Matt Coler, and 76 others. 2022. https://aclanthology....

    2022

  4. [4]

    Bender, Michael Wayne Goodman, Joshua Crowgey, and Fei Xia

    Emily M. Bender, Michael Wayne Goodman, Joshua Crowgey, and Fei Xia. 2013. https://aclanthology.org/W13-2710/ Towards creating precision grammars from interlinear glossed text: Inferring large-scale typological properties . In Proceedings of the 7th Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities, pages 74--83, Sofia...

    2013

  5. [5]

    Rogier Blokland, Niko Partanen, Michael Rie ler, and Joshua Wilbur. 2019. Using computational approaches to integrate endangered language legacy data into documentation corpora: Past experiences and challenges ahead. In Workshop on Computational Methods for Endangered Languages, Honolulu, Hawai’i, USA, volume 2, pages 24--30

    2019

  6. [6]

    Camilo Andrés Bonilla Carvajal. 2025. https://doi.org/10.17533/udea.ikala.356253 Interlinear translations reduce cognitive load on efl vocabulary acquisition . Íkala, Revista de Lenguaje y Cultura, 30(1)

    work page doi:10.17533/udea.ikala.356253 2025

  7. [7]

    Aditi Chaudhary, Arun Sampath, Ashwin Sheshadri, Antonios Anastasopoulos, and Graham Neubig. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.246 Teacher perception of automatically extracted grammar concepts for L 2 language learning . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3776--3793, Singapore. Association ...

    work page doi:10.18653/v1/2023.findings-emnlp.246 2023

  8. [8]

    John Dang, Shivalika Singh, Daniel D'souza, Arash Ahmadian, Alejandro Salamanca, Madeline Smith, Aidan Peppin, Sungjin Hong, Manoj Govindassamy, Terrence Zhao, Sandra Kublik, Meor Amer, Viraat Aryabumi, Jon Ander Campos, Yi-Chern Tan, Tom Kocmi, Florian Strub, Nathan Grinsztajn, Yannis Flet-Berliac, and 26 others. 2024. https://arxiv.org/abs/2412.04261 Ay...

    Pith/arXiv arXiv 2024

  9. [9]

    DeepSeek-AI and alia. 2025. https://arxiv.org/abs/2501.12948 Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning . Preprint, arXiv:2501.12948

    Pith/arXiv arXiv 2025

  10. [10]

    Micha Elsner and David Liu. 2025. https://doi.org/10.18653/v1/2025.sigmorphon-main.1 Prompt and circumstance'':'' a word-by-word LLM prompting approach to interlinear glossing for low-resource languages . In Proceedings of the 22nd SIGMORPHON workshop on Computational Morphology, Phonology, and Phonetics, pages 1--14, Albuquerque, New Mexico, USA. Associa...

    work page doi:10.18653/v1/2025.sigmorphon-main.1 2025

  11. [11]

    Ryan Alden Georgi. 2016. From Aari to Zulu : massively multilingual creation of language tools using interlinear glossed text . Ph.D. thesis

    2016

  12. [12]

    Michael Ginn, Mans Hulden, and Alexis Palmer. 2024 a . https://doi.org/10.18653/v1/2024.findings-emnlp.337 Can we teach language models to gloss endangered languages? In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5861--5876, Miami, Florida, USA. Association for Computational Linguistics

    doi:10.18653/v1/2024.findings-emnlp.337 2024

  13. [13]

    Michael Ginn, Sarah Moeller, Alexis Palmer, Anna Stacey, Garrett Nicolai, Mans Hulden, and Miikka Silfverberg. 2023. https://doi.org/10.18653/v1/2023.sigmorphon-1.20 Findings of the SIGMORPHON 2023 shared task on interlinear glossing . In Proceedings of the 20th SIGMORPHON workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 1...

    doi:10.18653/v1/2023.sigmorphon-1.20 2023

  14. [14]

    Michael Ginn, Lindia Tjuatja, Taiqi He, Enora Rice, Graham Neubig, Alexis Palmer, and Lori Levin. 2024 b . https://doi.org/10.18653/v1/2024.emnlp-main.683 G loss LM : A massively multilingual corpus and pretrained model for interlinear glossed text . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 12267--12...

    work page doi:10.18653/v1/2024.emnlp-main.683 2024

  15. [15]

    Leander Girrbach. 2023 a . https://doi.org/10.18653/v1/2023.sigmorphon-1.17 T \"u - CL at SIGMORPHON 2023: Straight-through gradient estimation for hard attention . In Proceedings of the 20th SIGMORPHON workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 151--165, Toronto, Canada. Association for Computational Linguistics

    work page doi:10.18653/v1/2023.sigmorphon-1.17 2023

  16. [16]

    Leander Girrbach. 2023 b . https://doi.org/10.18653/v1/2023.sigmorphon-1.19 T \"u - CL at SIGMORPHON 2023: Straight-through gradient estimation for hard attention . In Proceedings of the 20th SIGMORPHON workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 171--185, Toronto, Canada. Association for Computational Linguistics

    doi:10.18653/v1/2023.sigmorphon-1.19 2023

  17. [17]

    Taiqi He, Kwanghee Choi, Lindia Tjuatja, Nathaniel Robinson, Jiatong Shi, Shinji Watanabe, Graham Neubig, David Mortensen, and Lori Levin. 2024. https://doi.org/10.18653/v1/2024.acl-long.34 W av2 G loss: Generating interlinear glossed text from speech . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

    work page doi:10.18653/v1/2024.acl-long.34 2024

  18. [18]

    Mortensen, Graham Neubig, and Lori Levin

    Taiqi He, Lindia Tjuatja, Nathaniel Robinson, Shinji Watanabe, David R. Mortensen, Graham Neubig, and Lori Levin. 2023. https://doi.org/10.18653/v1/2023.sigmorphon-1.22 S ig M ore F un submission to the SIGMORPHON shared task on interlinear glossing . In Proceedings of the 20th SIGMORPHON workshop on Computational Research in Phonetics, Phonology, and Mor...

    work page doi:10.18653/v1/2023.sigmorphon-1.22 2023

  19. [19]

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  20. [20]

    Manuel Mager, \"O zlem C etino g lu, and Katharina Kann. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.423 Tackling the low-resource challenge for canonical segmentation . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5237--5250, Online. Association for Computational Linguistics

    work page doi:10.18653/v1/2020.emnlp-main.423 2020

  21. [21]

    Angelina McMillan-Major. 2020 a . https://aclanthology.org/2020.scil-1.42/ Automating gloss generation in interlinear glossed text . In Proceedings of the Society for Computation in Linguistics 2020, pages 355--366, New York, New York. Association for Computational Linguistics

    2020

  22. [22]

    Angelina McMillan-Major. 2020 b . https://doi.org/10.7275/TSMK-SA32 Automating Gloss Generation in Interlinear Glossed Text . Proceedings of the Society for Computation in Linguistics, 3(1):338--349. Publisher: University of Mass Amherst

    work page doi:10.7275/tsmk-sa32 2020

  23. [23]

    Sarah Moeller and Mans Hulden. 2018. https://aclanthology.org/W18-4809 Automatic Glossing in a Low - Resource Setting for Language Documentation . In Proceedings of the Workshop on Computational Modeling of Polysynthetic Languages , pages 84--93, Santa Fe, New Mexico, USA. Association for Computational Linguistics

    2018

  24. [24]

    Sarah Moeller, Ling Liu, Changbing Yang, Katharina Kann, and Mans Hulden. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.424 IGT 2 P : From interlinear glossed texts to paradigms . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5251--5262, Online. Association for Computational Linguistics

    work page doi:10.18653/v1/2020.emnlp-main.424 2020

  25. [25]

    Sebastian Nordhoff and Thomas Kr \"a mer. 2022. https://aclanthology.org/2022.ldl-1.3/ IMTV ault: Extracting and enriching low-resource language interlinear glossed text from grammatical descriptions and typological survey articles . In Proceedings of the 8th Workshop on Linked Data in Linguistics within the 13th Language Resources and Evaluation Conferen...

    2022

  26. [26]

    Alexis Palmer, Taesun Moon, and Jason Baldridge. 2009. Evaluating automation strategies in language documentation. In Proceedings of the NAACL HLT 2009 Workshop on Active Learning for Natural Language Processing, pages 36--44

    2009

  27. [27]

    Alexis Palmer, Taesun Moon, Jason Baldridge, Katrin Erk, Eric Campbell, and Telma Can. 2010. https://doi.org/10.33011/lilt.v3i.1217 Computational strategies for reducing annotation effort in language documentation: A case study in creating interlinear texts for Uspanteko . Linguistic Issues in Language Technology, 3

    work page doi:10.33011/lilt.v3i.1217 2010

  28. [28]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485--5551

    2020

  29. [29]

    Rita Ramos, Everlyn Asiko Chimoto, Maartje Ter Hoeve, and Natalie Schluter. 2025. https://doi.org/10.18653/v1/2025.acl-long.1447 G ramma MT : Improving machine translation with grammar-informed in-context learning . In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 29920--29940, Vienn...

    work page doi:10.18653/v1/2025.acl-long.1447 2025

  30. [30]

    Enora Rice, Katharina von der Wense, and Alexis Palmer. 2025. https://doi.org/10.18653/v1/2025.emnlp-main.568 Interdisciplinary research in conversation: A case study in computational morphology for language documentation . In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 11284--11296, Suzhou, China. Associa...

    doi:10.18653/v1/2025.emnlp-main.568 2025

  31. [31]

    Shruti Rijhwani, Daisy Rosenblum, Michayla King, Antonios Anastasopoulos, and Graham Neubig. 2023. https://aclanthology.org/2023.computel-1.4/ User-centric evaluation of OCR systems for kwak ' wala . In Proceedings of the Sixth Workshop on the Use of Computational Methods in the Study of Endangered Languages, pages 19--29, Remote. Association for Computat...

    2023

  32. [32]

    Erich Round, Mark Ellison, Jayden Macklin-Cordes, and Sacha Beniamine. 2020. https://aclanthology.org/2020.lrec-1.351/ Automated parsing of interlinear glossed text from page images of grammatical descriptions . In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 2878--2883, Marseille, France. European Language Resources Association

    2020

  33. [33]

    Bhargav Shandilya and Alexis Palmer. 2025. https://aclanthology.org/2025.coling-main.499/ Boosting the capabilities of compact models in low-data contexts with large language models and retrieval-augmented generation . In Proceedings of the 31st International Conference on Computational Linguistics, pages 7470--7483, Abu Dhabi, UAE. Association for Comput...

    2025

  34. [34]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. https://arxiv.org/abs/2402.03300 Deepseekmath: Pushing the limits of mathematical reasoning in open language models . Preprint, arXiv:2402.03300

    Pith/arXiv arXiv 2024

  35. [35]

    Gemma Team. 2025. https://goo.gle/Gemma3Report Gemma 3

    2025

  36. [36]

    Heli Uibo, Jack Rueter, and Sulev Iva. 2017. Building and using language resources and infrastructure to develop e-learning programs for a minority language. Proceedings of the Joint 6th Workshop on NLP for Computer Assisted Language Learning and 2nd Workshop on NLP for Research on Language Acquisition at NoDaLiDa , 134:61--67

    2017

  37. [37]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, and alia. 2025 a . https://arxiv.org/abs/2505.09388 Qwen3 technical report . Preprint, arXiv:2505.09388

    Pith/arXiv arXiv 2025

  38. [38]

    Changbing Yang, Franklin Ma, Freda Shi, and Jian Zhu. 2025 b . https://doi.org/10.18653/v1/2025.emnlp-main.69 L ing G ym: How far are LLM s from thinking like field linguists? In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 1314--1340, Suzhou, China. Association for Computational Linguistics

    work page doi:10.18653/v1/2025.emnlp-main.69 2025

  39. [39]

    Changbing Yang, Garrett Nicolai, and Miikka Silfverberg. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.261 Multiple sources are better than one: Incorporating external knowledge in low-resource glossing . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4537--4552, Miami, Florida, USA. Association for Co...

    work page doi:10.18653/v1/2024.emnlp-main.261 2024

  40. [40]

    Olga Zamaraeva. 2016. https://doi.org/10.18653/v1/W16-2021 Inferring morphotactics from interlinear glossed text: Combining clustering and precision grammars . In Proceedings of the 14th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology , pages 141--150, Berlin, Germany. Association for Computational Linguistics

    work page doi:10.18653/v1/w16-2021 2016

  41. [41]

    Xingyuan Zhao, Satoru Ozaki, Antonios Anastasopoulos, Graham Neubig, and Lori Levin. 2020. https://doi.org/10.18653/v1/2020.coling-main.471 Automatic interlinear glossing for under-resourced languages leveraging translations . In Proceedings of the 28th International Conference on Computational Linguistics, pages 5397--5408, Barcelona, Spain (Online). Int...

    doi:10.18653/v1/2020.coling-main.471 2020

  42. [42]

    Mortensen, and Alex Waibel

    Zhong Zhou, Lori Levin, David R. Mortensen, and Alex Waibel. 2020. https://arxiv.org/abs/1911.02709 Using interlinear glosses as pivot in low-resource multilingual machine translation . Preprint, arXiv:1911.02709

    Pith/arXiv arXiv 2020

  43. [43]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  44. [44]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...