{"id":"0d6ef3e8-9b10-418e-abc3-3a3a9e75752a","arxiv_id":"2502.02722","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A thesis presenting T-Projection and constrained decoding, which improve cross-lingual sequence labeling for low-resource languages, plus the Medical mT5 model and open resources.","lead":"This doctoral thesis develops and evaluates cross-lingual transfer methods for low-resource NLP: T-Projection, a two-step annotation projection method, a constrained decoding algorithm for zero-shot sequence labeling with text-to-text models, and Medical mT5, a multilingual medical text-to-text model. It reports state-of-the-art results on African NER and other tasks, and releases code, datasets, and models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Intrinsic projection-quality comparisons may be inflated because T-Projection is fine-tuned on the same English source sentences whose translations are later used as the manual-projection gold standard.","rationale":"The reader correctly identifies mT5 candidate recall as a load-bearing assumption: if the correct target span is not in the top-k candidates, the NMT-based selection step cannot recover it, and Table 4.3 quantifies this ceiling at 98.0 F1. However, the more fundamental issue is that this ceiling and the intrinsic comparison in Table 4.2 may be optimistically biased by the experimental design. Section 4.5 explicitly trains T-Projection on the full English datasets used to create the manual-projection evaluation sets, including the labeled English Europarl test data for NER. This means the model has seen the exact English source content before being asked to identify the corresponding spans in the parallel target sentences. The reported 8.6-point margin over SimAlign could therefore reflect memorization of source labels plus cross-lingual pattern matching rather than a general projection capability on unseen source text. This does not invalidate the downstream extrinsic result on MasakhaNER2.0, where the final NER model is evaluated on held-out target test data, but it does weaken the generality of the intrinsic 'state-of-the-art projection' claim and the interpretation of the 98.0 upper bound. The correct response is to keep the CONDITIONAL verdict and require a held-out-source intrinsic evaluation before accepting the wide-margin claim as established for arbitrary parallel data. The reader's weakest-assumption analysis is useful but should be extended to question the validity of the recall measurement itself, hence 'partial' agreement.","tokens_in":52318,"tokens_out":11388,"duration_ms":112281,"concrete_test":"Retrain T-Projection on only the standard English training splits (CoNLL 2003 train, ABSA train, AbstRCT Neoplasm train) and recompute the intrinsic projection F1 and the upper-bound recall on manual projections for held-out source sentences (e.g., CoNLL 2003 test, ABSA test, AbstRCT test). If the upper-bound F1 drops materially below the reported 98.0, or if the margin over SimAlign shrinks by more than a few points, then the headline projection-quality claim is contingent on fine-tuning on the evaluation source side. Report both configurations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that T-Projection outperforms prior annotation projection methods by 8.6 F1 points rests primarily on the intrinsic evaluation in Table 4.2, with the mT5 candidate-recall ceiling in Table 4.3 (98.0 average F1) presented as the method's upper bound. A closer reading of Section 4.5 shows that T-Projection is trained on the very source sentences used to construct the evaluation gold standard. For NER, the text states: 'T-Projection and XLM-RoBERTa are trained using the full English CoNLL 2003 dataset together with the labeled English Europarl test data' and then evaluated on the manual projections for the Europarl-based NER test set. Similarly, for OTE and AM, T-Projection is trained on 'the full English ABSA 2016 dataset' and 'the full Neoplasm data from the AbstRCT dataset,' respectively, before being evaluated on manual projections of those same datasets. Because the English side of the evaluation pairs is in the training set, mT5 can memorize source spans and use cross-lingual correspondence at inference, inflating both the candidate-recall upper bound and the final projection F1. This is not the same as measuring projection quality on source sentences never seen during fine-tuning. The downstream MasakhaNER2.0 evaluation is less affected because the final NER model is evaluated on held-out human-annotated target test sets, but the headline intrinsic margin and the '98.0 upper bound' are not held-out estimates. The reader's concern about mT5 recall is thus valid but incomplete: the recall ceiling itself may be optimistically biased by the train/evaluation overlap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This PhD thesis addresses cross-lingual transfer learning for sequence labeling tasks (NER, OTE, Argument Mining) in low-resource languages. It makes three main contributions: (1) a comparative study of data-based versus model-based transfer with an error analysis (Chapter 3); (2) T-Projection, a two-stage annotation projection method that generates candidates with a fine-tuned mT5 model and selects among them using MT-based translation probabilities (Chapter 4); and (3) a constrained decoding algorithm for text-to-text LLMs that enforces the output structure of sequence labeling, evaluated on NER, OTE, and Event Extraction across up to 25 languages (Chapter 5). Chapter 6 applies these ideas to the medical domain by releasing Multilingual Medical Corpus and Medical mT5. The headline claims are that T-Projection outperforms prior projection methods by 8.6 F1 points on average in intrinsic evaluation and that constrained decoding lets mT0-XL beat encoder-only models in zero-shot cross-lingual sequence labeling.","tokens_in":52544,"tokens_out":9046,"duration_ms":79300,"significance":"If the results hold, this is a practically significant body of work for low-resource NLP. The thesis releases open-source software, datasets, and models (T-Projection, Medical mT5, Easy Label Projection), which is a genuine strength for reproducibility and community adoption. The two-step projection design, with separate analysis of candidate generation and selection, is a well-structured contribution that improves over alignment-based methods, especially for long span projections (Argument Mining). The constrained decoding algorithm addresses a real failure mode of text-to-text models (hallucinations and invalid output structure) and shows consistent gains over unconstrained decoding. The comparison between data and model transfer in Chapter 3 and the application to African languages in Chapters 4 and 5 are useful empirical anchors for practitioners.","major_comments":[{"comment":"The statement that 'the constrained beam search will always yield an output that is at least as good as, if not superior to, unconstrained beam search' is logically incorrect. Constraining the search space to valid sequences can exclude the highest-probability decoding, and an unconstrained output that is invalid may still contain correct entity spans that yield a higher F1 after parsing than any valid constrained output. What is guaranteed is that constrained decoding produces a well-formed output and removes hallucinated tokens outside the input. The empirical results in Tables 5.4-5.8 are not in question, but this theoretical claim should be removed or revised accordingly.","section":"Section 5.3.2, final paragraph"}],"minor_comments":[{"comment":"These central tables report single-run F1 scores without variance estimates, whereas Chapter 3 reports five-seed averages with standard deviations. This makes it difficult to assess small per-language differences, e.g., Igbo 71.4 vs 70.9 in Table 4.5.","section":"Tables 4.2, 4.3, 4.5, 5.4-5.8"},{"comment":"Chapter 5 (constrained decoding) is not listed in the scientific contributions included in the thesis; the provenance of this work should be clarified, whether it is an unpublished chapter or it has appeared elsewhere.","section":"Section 1.4.1"},{"comment":"The claim that 'these are the best zero-shot results achieved for MasakhaNER2.0' should be qualified, since the evaluation covers 8 of the 20 languages and a specific training setup (English CoNLL plus NLLB translations); as written it could be misread as a claim over the full benchmark.","section":"Section 4.6.2"},{"comment":"The 'upper bound' defined by selecting the correct candidate when present and otherwise the most probable candidate mixes candidate recall and selection accuracy; please define it explicitly as a practical ceiling rather than a pure recall bound.","section":"Section 4.5.2"},{"comment":"Several typos and small errors: Chapter 6 title 'Domain-Spacific' should be 'Domain-Specific'; Figure 5.5 caption says 'unconstrained and unconstrained beam search' instead of 'constrained and unconstrained'; Section 2.2 has 'Comercial' for 'Commercial'; Figure 2.15 caption has 'coss-lingual'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The thesis is largely a compilation of three published papers (EMNLP 2022, EMNLP 2023, LREC-COLING 2024), with Chapter 5 not appearing in the contributions list; this should be clarified editorially. The most significant scientific concern is the intrinsic evaluation leakage in Chapter 4; while the transductive setup matches the practical use case of projecting labels from one's own labeled source data, the abstract and conclusions overstate the result as a general improvement over prior projection methods. Requiring a held-out intrinsic evaluation or a clear reframing would strengthen the manuscript. The extrinsic MasakhaNER2.0 evaluation and the constrained decoding experiments are the most convincing parts of the empirical contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the thesis. The two new methods are real: T-Projection's two-step design (mT5 generates candidate spans, an NMT model ranks them) is not in the cited literature, and the constrained decoding algorithm for zero-shot sequence labeling is a clean, low-overhead extension of beam search. The ablations are informative: both steps matter, and the n-gram and most-probable-candidate baselines show it. The code, models, and datasets are public, which counts for a lot.\n\nThe soft spot is the intrinsic evaluation in Chapter 4. T-Projection is trained on the same English sentences that were translated and manually projected to build the evaluation gold standard. For NER, it is even trained on the labeled Europarl test data that the manual projections come from. That means the 8.6-point F1 margin over SimAlign and the 98.0 candidate-recall upper bound are not held-out estimates; they are optimistic by an unknown amount. This is a load-bearing part of the paper's headline claim. The reader's worry about mT5 recall is real, but the recall ceiling itself may be inflated for the same reason.\n\nThe extrinsic evaluation on MasakhaNER2.0 is the stronger evidence and it survives the leak: the downstream model is tested on human-annotated target test sets, so the reported gains are credible, even if only 8 of 20 languages are covered (the languages mT5 supports). Still, the \"best to date\" phrasing should be scoped to the compared baselines and languages.\n\nChapter 4 tables also omit error bars and run counts, which is a step down from Chapter 3's reporting. That is fixable.\n\nWho gets value: anyone working on data-based cross-lingual transfer for sequence labeling, especially for low-resource languages. The thesis deserves a serious referee: the methods are plausible and useful, but the intrinsic numbers need to be recomputed on disjoint English source data before the main claim is safe.\n\nMy recommendation: engage with it, send it to review, and insist on the held-out intrinsic evaluation. If the margin shrinks, it will shrink to something still respectable.","headline":"Real methods and useful open resources, but the intrinsic evaluation is compromised by training/evaluation overlap and the headline margin needs to be recomputed.","tokens_in":53159,"tokens_out":3627,"would_cite":true,"duration_ms":34329,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This thesis establishes a generate-then-rank annotation projection method and a constrained decoding algorithm that together improve cross-lingual sequence labeling for low-resource languages.","keywords":["cross-lingual transfer","annotation projection","sequence labeling","low-resource languages","constrained decoding","text-to-text language models","named entity recognition","multilingual medical NLP"],"falsifier":"Run T-Projection on a language pair or domain not covered by the thesis, with human-projected gold labels, and measure how often the correct target span appears in the generator's top-$k$ candidate lists. If that recall is well below the 98.0 average $F_1$ reported in the thesis, the selection step will not be able to compensate, and the claimed margin over SimAlign should shrink proportionally.","tokens_in":52037,"feed_emoji":"🌍","tokens_out":14768,"duration_ms":109526,"temperature":0.7,"pith_summary":"This thesis tries to establish two mechanisms for cross-lingual transfer in low-resource sequence labeling. T-Projection is an annotation projection method that generates candidate target-language spans with a fine-tuned multilingual text-to-text model and ranks them by machine-translation probability; it outperforms previous projection methods by 8.6 average $F_1$ points across Named Entity Recognition, Opinion Target Extraction, and Argument Mining. The second mechanism is a constrained decoding algorithm that forces text-to-text language models to copy the input sentence and emit well-formed labels, enabling mT0-XL to beat encoder-only models in zero-shot cross-lingual sequence labeling. The same methods are applied to build the first open multilingual text-to-text medical model, showing that the approach reaches real-world low-resource domains.","feed_headline":"New projection method beats word alignments by 8.6 F1","feed_subtitle":"Generate label candidates with a multilingual model, rank them by translation score, train low-resource systems.","key_machinery":"The load-bearing machinery is T-Projection's two-step pipeline: candidate generation, in which a fine-tuned multilingual text-to-text model (mT5) sees the unlabeled target sentence with category tags taken from the source sentence and generates many candidate spans; and candidate selection, in which a machine-translation model scores each candidate by normalized, symmetrized translation probability and chooses the best non-overlapping assignment. The second mechanism is the constrained-decoding finite-state automaton, which restricts each decoding step to tokens that copy the next input word or open or close a label tag, implemented as constrained beam search so the output is always the input sentence with valid tags. The first mechanism removes reliance on word alignments; the second removes reliance on a model's free-form output discipline.","core_discovery":"On the paper's own terms, the central discovery is that cross-lingual sequence labeling improves when the two standard tools are replaced: word-alignment-based annotation projection gives way to a generate-then-rank method, and free-form generation by text-to-text models gives way to constrained decoding. T-Projection fine-tunes a multilingual text-to-text model to propose candidate target-language spans for each labeled category in the parallel sentence, then scores those candidates with machine-translation probabilities from M2M100 or NLLB200, symmetrized across translation directions, and selects the best non-overlapping spans. In intrinsic evaluation against manually projected gold labels across OTE, NER, and Argument Mining in five languages, T-Projection averages 93.9 $F_1$ and outperforms the second-best system, SimAlign, by 8.6 points; in the extrinsic African-languages evaluation it beats zero-shot encoder transfer by 15.6 average $F_1$ points. The constrained-decoding contribution defines a finite-state automaton that forces a text-to-text model to copy the input sentence word-by-word and insert or close only valid HTML-style tags; applied to mT0-XL, this eliminates hallucinated words and unparseable outputs and outperforms both unconstrained beam search and encoder-only baselines in zero-shot cross-lingual sequence labeling, most strongly for languages morphologically distant from English. The medical case study extends the same data- and model-transfer methods to build the first open multilingual text-to-text model for the medical domain.","pith_inferences":["The paper's own upper-bound result implies that future gains should come from improving candidate recall: at 98.0 average $F_1$ the selection step is already near-ceiling, so the bottleneck is the generator's ability to put the right span in its top candidates.","A testable extension is to carry the generate-then-rank scheme to other span-level tasks, such as event extraction or coreference, where word alignments are known to be fragile.","Constrained decoding might be pushed further by training the generator to expect the constraint, rather than imposing it only at inference, which could reduce the language-mixing errors that still appear in zero-shot outputs.","Because T-Projection still needs labeled data in a source language and a good translation system into the target, its practical reach is bounded by those resources; the thesis's African-languages results suggest the boundary is wider than word-alignment methods allowed, but not infinite."],"forward_implications":["Data-based transfer becomes the recommended route when no large multilingual encoder has strong proficiency in the target language: T-Projection's projected data beat zero-shot encoder transfer in six of eight African languages, by 15.6 average $F_1$ points.","Annotation projection for long, structured spans no longer needs word alignments; on Argument Mining projection T-Projection reaches 96.0 $F_1$, 9.4 points above the next-best system.","Text-to-text models become viable for zero-shot cross-lingual sequence labeling once decoding is constrained to copy the input words and produce well-formed tags; mT0-XL with constrained beam search beats unconstrained decoding and encoder-only baselines, especially for languages distant from English.","The same transfer toolkit carries into new domains: a multilingual text-to-text medical model, trained on compiled medical corpora and evaluated on sequence labeling and question answering, shows that low-resource medical NLP can use the same data- and model-transfer routes."],"supporting_citations":[{"why":"Supplies the mT5 text-to-text model that T-Projection fine-tunes for candidate generation.","marker":"Xue et al. 2021"},{"why":"Provides M2M100, the machine-translation model used to score candidate translation probabilities.","marker":"Fan et al. 2021"},{"why":"Provides NLLB200, used for low-resource African language pairs and for scoring when M2M100 lacks coverage.","marker":"Costa-jussà et al. 2022"},{"why":"Defines the normalized and symmetrized translation-probability scoring that ranks the projection candidates.","marker":"Vamvas and Sennrich 2022"},{"why":"SimAlign is the second-best system in the intrinsic evaluation that T-Projection beats by 8.6 average $F_1$.","marker":"Jalili Sabet et al. 2020"},{"why":"AWESOME is the leading word-alignment baseline and the comparison point for extrinsic projection quality.","marker":"Dou and Neubig 2021"},{"why":"EasyProject is the marker-based joint translation-and-projection baseline in the African-languages extrinsic evaluation.","marker":"Chen et al. 2023"},{"why":"CODEC is the two-stage marker-based projection baseline that T-Projection outperforms in seven of eight languages.","marker":"Le et al. 2024"},{"why":"Supplies mT0-XL, the text-to-text model whose constrained decoding beats unconstrained decoding and encoder-only baselines.","marker":"Muennighoff et al. 2023"},{"why":"Provides XLM-R, the encoder-only model family whose zero-shot cross-lingual behavior is the baseline the thesis improves on.","marker":"Conneau et al. 2020"}],"fun_headline_variants":["T-Projection beats word alignments by 8.6 F1","Constrained decoding improves zero-shot cross-lingual labeling","Generate-then-rank projection for low-resource NLP","Zero-shot labeling with constrained decoding","Cross-lingual transfer via generate-and-rank"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"T-Projection's gains depend on the fine-tuned multilingual generator placing the correct translation of each labeled source span among its top candidates for the parallel sentence; if candidate recall fails, the machine-translation ranking step cannot recover, and the paper's own upper-bound experiment puts that recall ceiling at 98.0 average $F_1$.","fun_headline_variants_meta":{"raw":{"variants":["T-Projection beats word alignments by 8.6 F1","Constrained decoding improves zero-shot cross-lingual labeling","Generate-then-rank projection for low-resource NLP","Zero-shot labeling with constrained decoding","Cross-lingual transfer via generate-and-rank"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001037,"raw_usage":{"total_tokens":4467,"prompt_tokens":1152,"completion_tokens":3315,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":768,"completion_tokens_details":{"reasoning_tokens":3251}},"tokens_in":768,"tokens_out":3315,"duration_ms":23528,"temperature":1.0,"reasoning_tokens":3251,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T11:21:23.481561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run T-Projection on a language pair or domain not covered by the thesis, with human-projected gold labels, and measure how often the correct target span appears in the generator's top-$k$ candidate lists. If that recall is well below the 98.0 average $F_1$ reported in the thesis, the selection step will not be able to compensate, and the claimed margin over SimAlign should shrink proportionally.","supporting_citations":[],"review_version":1}