Gloss-Free Representation Learning for Cross-Dataset Sign Spotting

# Gloss-Free Representation Learning for Cross-Dataset Sign Spotting
Source: [https://arxiv.org/html/2608.11332](https://arxiv.org/html/2608.11332)
Ezgi Ekin[https://orcid.org/0009-0006-6008-4308](https://orcid.org/0009-0006-6008-4308)Affiliation:Hacettepe University, Ankara, TürkiyeAffiliation:E\-mail[\{oguztufekcioglu,ezgiekin21,kaancevik21,hacerkeles\}@hacettepe\.edu\.tr](mailto:)Mustafa Kaan Çevik[https://orcid.org/0009-0000-0092-2660](https://orcid.org/0009-0000-0092-2660)Affiliation:Hacettepe University, Ankara, TürkiyeHacer Yalim Keles[https://orcid.org/0000-0002-1671-4126](https://orcid.org/0000-0002-1671-4126)Affiliation:Hacettepe University, Ankara, Türkiye

###### Abstract

Sign\-language research for resource\-constrained languages is often limited by the cost of dense linguistic labels, including glosses, temporal boundaries, and sign order\. Broadcast news provides a practical alternative by pairing continuous signing with spoken\-language transcripts, but this supervision is weak because text and signing are only loosely aligned\. Morphologically rich languages such as Turkish add a further difficulty, since the same lexical meaning can appear in many inflected forms, while some derived forms should remain separate\. We study whether weak transcript\-based supervision can pretrain a reusable sign encoder in this morphologically rich setting, where inadequate text normalization can fragment pseudo\-gloss targets and weaken representation learning\. Unlike prior pseudo\-gloss pipelines designed mainly to improve translation, we test whether the pretrained encoder transfers as a reusable representation for cross\-dataset sign spotting\. We pretrain on TSL\-News, a Turkish broadcast corpus collected for this work, using pseudo\-gloss labels derived from transcripts rather than manually annotated glosses\. We compare two pseudo\-gloss construction strategies, namely rule\-based morphological lemmatization and constrained LLM\-assisted normalization over a fixed vocabulary\. We evaluate the learned representations through cross\-dataset sign spotting on a new TSL Spotting Benchmark built from the TSL Dictionary corpus\. The LLM\-assisted encoder raises top\-5 temporal localization mean IoU from 0\.235 with raw spatial features to 0\.465, with 56\.2% of examples reaching an IoU of at least 0\.50\. A frequency analysis further suggests that localization quality is not mainly driven by memorization of frequent pseudo\-gloss labels\. In a downstream translation check, the same pretraining improves BLEU\-4 from 9\.60 to 11\.04 and ROUGE from 23\.48 to 27\.43\. These results show that loosely aligned broadcast data can provide effective weak supervision for learning sign representations that capture both lexical content and temporal structure\.

###### Keywords:

Turkish Sign Language Sign spotting Translation Gloss\-free Weak supervision Pseudo\-glosses Temporal localization

## 1Introduction

Sign language translation \(SLT\)\[[3](https://arxiv.org/html/2608.11332#bib.bib19),[17](https://arxiv.org/html/2608.11332#bib.bib20),[2](https://arxiv.org/html/2608.11332#bib.bib18)\], sign language recognition\[[7](https://arxiv.org/html/2608.11332#bib.bib17)\], and sign spotting\[[2](https://arxiv.org/html/2608.11332#bib.bib18)\]all rely on visual representations that preserve both lexical content and temporal structure\. However, the annotations commonly used to train such models are costly to obtain\. Gloss labels require linguistic expertise, temporal boundaries require frame\-level or segment\-level annotation, and sign order may differ from the corresponding spoken\-language sentence\. These requirements are especially restrictive for resource\-limited sign languages, where parallel corpora and expert gloss annotations are scarce\.

Broadcast news data offer a practical source of continuous signing paired with spoken or written language translations\. However, this supervision is weak: the text and signing are only loosely aligned, and they may differ in lexical choice, word order, and level of detail\. These issues are more difficult in a morphologically rich spoken language such as Turkish, where many surface forms can refer to the same lexical unit, while some derived forms should remain distinct\. The question addressed in this paper is therefore not whether a new spotting or translation architecture can be designed, but whether weak text supervision can be used to pretrain a reusable sign encoder\. We study this question in the pretraining stage of a Turkish Sign2GPT\-style pipeline\[[20](https://arxiv.org/html/2608.11332#bib.bib5)\], treating it as a gloss\-free representation learning problem\. The goal is to test whether sentence\-level pseudo\-gloss labels derived from Turkish text can shape a video encoder into a representation that transfers beyond the original broadcast data\. We evaluate this transfer through sign spotting: given an isolated dictionary sign as a visual template, the model must localize the corresponding sign inside a sentence\-level example video\. Rather than evaluating only on the pseudo\-gloss labels used during training, we test the learned representation on the TSL Dictionary corpus \(TSLD\), using isolated signs as templates and sentence videos as localization targets\.

The evaluation uses normalized cross\-correlation \(NCC\) between hidden\-state sequences\. Given an isolated TSLD sign video and a continuous TSLD sentence video, the isolated\-sign representation is slid over the continuous representation, and the highest\-scoring temporal windows are compared with human temporal annotations\. This setup directly tests the representation\-learning question\. If the pretrained encoder has learned reusable sign structure, an isolated dictionary sign should match the corresponding sign inside a sentence video, although neither video comes from the broadcast corpus used for pretraining\.

The results support this hypothesis\. The encoder pretrained with the constrained LLM\-assisted pseudo\-gloss strategy improves NCC\-based temporal localization over the unadapted spatial\-feature baseline\. Mean IoU increases from 0\.124 to 0\.271 at top\-1, from 0\.192 to 0\.403 at top\-3, and from 0\.235 to 0\.465 at top\-5 on TSL Spotting Benchmark \(TSL\-SB\)\. The same strategy also improves both vocabulary coverage and localization accuracy compared with the rule\-based morphological lemmatization baseline\. These gains are achieved under weak supervision\. During pretraining, the model receives only sentence\-level pseudo\-gloss labels derived from Turkish text, without exact sign labels or temporal boundary annotations\.

##### Contributions\.

Four contributions are made\. First, we repurpose a Sign2GPT\-style pseudo\-gloss pretraining pipeline\[[20](https://arxiv.org/html/2608.11332#bib.bib5)\]– originally proposed to supply weak supervision for gloss\-free translation – to ask a question this line of work has not addressed: whether the resulting representation is itself strong enough to support cross\-dataset sign spotting, a task central to both sign language recognition and translation\. This reframes an existing weak\-supervision pretraining paradigm as a representation\-learning question, tested outside both its original translation objective and its training corpus\. Second, we design two Turkish\-specific pseudo\-gloss construction strategies, moving from rule\-based morphological lemmatization to a constrained LLM\-assisted lexical normalization, to address the lexical variation introduced by Turkish’s agglutinative morphology\. Third, we document TSL\-News and TSL\-SB as broadcast pretraining and cross\-dataset spotting resources for this setting\. Fourth, we introduce a multi\-angle evaluation protocol – combining isolated\-template NCC matching, temporal IoU, direct pseudo\-gloss score localization, and all\-vocabulary local NCC – that isolates representation quality from confounds such as vocabulary coverage or classifier bias, complemented by a downstream translation check\.

## 2Related Work

##### Sign spotting and dictionary\-based localization\.

Sign spotting asks whether and where a target sign appears in continuous, co\-articulated signing\. Early modern spotting systems emphasize the same practical difficulty faced here: isolated dictionary productions are useful queries, but they differ from continuous broadcast signing in speed, context, and signer behavior\. Watch, Read and Lookup learns spotting embeddings from sparse localized labels, subtitles, and isolated dictionary examples using a multiple\-instance contrastive objective, explicitly treating subtitles and dictionaries as weak supervisors rather than dense gloss annotations\[[11](https://arxiv.org/html/2608.11332#bib.bib1)\]\. Large\-vocabulary continuous sign language recognition \(CSLR\) from spoken\-language supervision extends this direction by using retrieval\-style objectives, pseudo\-label cleaning, synonym aggregation, and dense evaluation with temporal sign intervals on the BBC\-Oxford British Sign Language \(BOBSL\) broadcast dataset\[[16](https://arxiv.org/html/2608.11332#bib.bib2)\]\. Recent segmentation work instead predicts temporal boundaries directly with begin\-inside\-outside \(BIO\) labels, transformer temporal modeling, hand\-pose features, and connectionist temporal classification \(CTC\) sequence constraints\[[5](https://arxiv.org/html/2608.11332#bib.bib3)\]\. These methods show that spotting quality depends not only on a classifier, but also on the representation used for cross\-domain matching and on the temporal proposal mechanism\. The present work follows the dictionary\-query view of spotting, but uses it as an evaluation protocol: isolated TSLD signs are templates, continuous TSLD examples are search videos, and NCC over hidden states tests whether weak broadcast pretraining has learned reusable lexical\-temporal structure\.

##### Gloss\-free and weakly supervised sign representations\.

Gloss annotations remain expensive because they require linguistic expertise, sign\-order decisions, and often temporal alignment\. Gloss\-free SLT therefore tries to avoid a manually annotated gloss bottleneck\. Gloss\-Free Sign Language Translation with Visual\-Language Pretraining \(GFSLT\-VLP\) aligns visual and textual representations through visual\-language pretraining and masked language modeling without gloss supervision\[[23](https://arxiv.org/html/2608.11332#bib.bib4)\], while Sign2GPT uses automatically extracted pseudo\-glosses to pretrain a visual encoder before connecting it to a large language model\[[20](https://arxiv.org/html/2608.11332#bib.bib5)\]\. Other representation\-learning work is also relevant: learnt contrastive concept \(LCC\) embeddings use spoken\-language word embeddings to regularize sign embeddings and enable automatic temporal localization\[[19](https://arxiv.org/html/2608.11332#bib.bib6)\], and SignRep learns sign\-specific self\-supervised red\-green\-blue \(RGB\) representations that transfer to recognition, dictionary retrieval, and translation\[[21](https://arxiv.org/html/2608.11332#bib.bib7)\]\. These studies support the broader premise that useful sign representations can be learned without dense manual gloss boundaries\. Our setting is narrower and more diagnostic: the pseudo\-gloss labels are Turkish text\-derived sentence\-level bags, and success is measured by cross\-dataset temporal localization rather than by translation scores or isolated recognition accuracy\.

##### Temporal alignment and post\-processing\.

Weak supervision in continuous signing also requires careful temporal handling\. CSLR models commonly use sequence losses or alignment constraints to learn from video\-level gloss order without frame boundaries, and motion\-aware encoders such as CorrNet show that local hand and face trajectories are central to continuous sign representations\[[6](https://arxiv.org/html/2608.11332#bib.bib8)\]\. After visual scoring, post\-processing often determines whether weak predictions become usable annotations: subtitle windows are padded or shifted, still frames are trimmed, repeated high\-confidence detections are merged, score thresholds suppress isolated spikes, and synonym or subtitle alignment is used to reduce lexical noise\. Gloss Alignment uses spoken\-language embeddings to reassign spotted glosses to neighboring subtitles when subtitle timing and signing order are misaligned\[[18](https://arxiv.org/html/2608.11332#bib.bib9)\]; large\-language\-model disambiguation similarly reranks dictionary\-based spotting candidates with sentence\-level linguistic context after visual matching\[[4](https://arxiv.org/html/2608.11332#bib.bib10)\]\. In this paper, the post\-processing is deliberately simple–fixed trimming, NCC peak selection, and top\-kktemporal IoU–so that the experiment primarily probes representation transfer\. The failure cases therefore connect directly to the literature: stronger proposal generation, context\-aware reranking, and text\-gloss realignment are natural next steps once the encoder has shown cross\-dataset spotting signal\.

## 3Method

### 3\.1Gloss\-Free Weak Pseudo\-Gloss Pretraining

LetV=\{xt\}t=1TV=\\\{x\_\{t\}\\\}\_\{t=1\}^\{T\}be a sign\-language video and letG⁡\(V\)⊆𝒢G\(V\)\\subseteq\\mathcal\{G\}be the set of pseudo\-gloss labels extracted from the paired Turkish text\. These labels are sentence\-level bags of words rather than sign\-language glosses with temporal boundaries; no manually annotated sign glosses are used during pretraining\. The stage\-1 model maps the video into temporal hidden states

H=Eθ​\(V\)∈ℝT′×d\.H=E\_\{\\theta\}\(V\)\\in\\mathbb\{R\}^\{T^\{\\prime\}\\times d\}\.\(1\)Here,xtx\_\{t\}is the input frame at timett,TTis the input video length,𝒢\\mathcal\{G\}is the pseudo\-gloss vocabulary,EθE\_\{\\theta\}is the trainable video encoder with parametersθ\\theta,HHis the sequence of hidden states,T′T^\{\\prime\}is the output sequence length after temporal downsampling, andddis the hidden feature dimension\. Note thatttindexes the input video at frame resolution \(t=1,…,Tt=1,\\dots,T\), whereas the downsampled hidden\-state sequenceHHis indexed separately below\. In the current model setting,EθE\_\{\\theta\}uses DINOv2\[[14](https://arxiv.org/html/2608.11332#bib.bib11)\], a self\-supervised vision\-transformer backbone, for frame features with lightweight adaptation, followed by a MetaFormer\[[22](https://arxiv.org/html/2608.11332#bib.bib12)\]temporal encoder for token mixing with local attention and downsampling\. The output hidden states are passed to a prototype head whose class prototypes are initialized from Turkish fastText\[[1](https://arxiv.org/html/2608.11332#bib.bib13)\]subword word embeddings\. For classggat positiont′∈\{1,…,T′\}t^\{\\prime\}\\in\\\{1,\\dots,T^\{\\prime\}\\\}in the downsampled sequence, the head computes a cosine similarityst′,gs\_\{t^\{\\prime\},g\}between a projected hidden state and the word prototype\. The temporal class score combines class and time normalization,

pg=∑t′softmaxg⁡\(st′,g/τc\)​softmaxt′⁡\(st′,g/τt\),p\_\{g\}=\\sum\_\{t^\{\\prime\}\}\\operatorname\{softmax\}\_\{g\}\\left\(s\_\{t^\{\\prime\},g\}/\\tau\_\{c\}\\right\)\\operatorname\{softmax\}\_\{t^\{\\prime\}\}\\left\(s\_\{t^\{\\prime\},g\}/\\tau\_\{t\}\\right\),\(2\)Here,pgp\_\{g\}is the sentence\-level score for pseudo\-glossgg,softmaxg\\operatorname\{softmax\}\_\{g\}normalizes similarities across pseudo\-gloss classes, andsoftmaxt′\\operatorname\{softmax\}\_\{t^\{\\prime\}\}normalizes across theT′T^\{\\prime\}downsampled time steps\. The trainable temperaturesτc\\tau\_\{c\}andτt\\tau\_\{t\}control the sharpness of the class and time distributions in the TSL\-News pretraining setting\. The model is trained with binary cross entropy \(BCE\) over the sentence\-level pseudo\-gloss set:

ℒpg=BCE\(\{pg\}g∈𝒢,\{𝟙\[g∈G\(V\)\]\}g∈𝒢\)\.\\mathcal\{L\}\_\{\\mathrm\{pg\}\}=\\operatorname\{BCE\}\\left\(\\\{p\_\{g\}\\\}\_\{g\\in\\mathcal\{G\}\},\\\{\\mathbb\{1\}\[g\\in G\(V\)\]\\\}\_\{g\\in\\mathcal\{G\}\}\\right\)\.\(3\)Here,ℒpg\\mathcal\{L\}\_\{\\mathrm\{pg\}\}is the pseudo\-gloss pretraining loss and𝟙\[g∈G\(V\)\]\\mathbb\{1\}\[g\\in G\(V\)\]is an indicator target that equals 1 when pseudo\-glossggappears in the text\-derived label set for videoVV, and 0 otherwise\. Because no temporal target is supplied, any temporal structure inHHmust emerge from video dynamics, the encoder inductive bias, and the pressure to explain the weak pseudo\-gloss set\. Figure[1](https://arxiv.org/html/2608.11332#S3.F1)summarizes the full pipeline, including pseudo\-gloss construction, stage\-1 pretraining, and the downstream translation stage that reuses the pretrained visual encoder\.

![Refer to caption](https://arxiv.org/html/2608.11332v1/figures/pipeline_diagram_updated.png)Figure 1:Overview of the pretraining and translation pipeline\. Turkish text is converted into pseudo\-glosses via lemmatization; video frames pass through a frozen DINOv2 encoder and a trainable sign encoder, whose output feeds both the pseudo\-gloss pretraining head and, for context, a frozen GPT/XGLM decoder \(both LoRA\-adapted\)\.Technical details on Sign2GPT\[[20](https://arxiv.org/html/2608.11332#bib.bib5)\]pretraining are provided in Appendix[0\.A\.1](https://arxiv.org/html/2608.11332#Pt0.A1.SS1)\.

### 3\.2Turkish Pseudo\-Gloss Construction

Turkish Sign Language broadcast pretraining requires converting Turkish sentence translations into weak lexical targets\. This is not a trivial word\-filtering step because Turkish is highly agglutinative: inflectional suffixes can mark case, possession, number, person, tense, and modality, while derivational suffixes can create related but lexically different words\[[12](https://arxiv.org/html/2608.11332#bib.bib21),[13](https://arxiv.org/html/2608.11332#bib.bib22)\]\. A surface\-word vocabulary therefore fragments the same lexical concept across many forms, but overly aggressive normalization can also merge targets that should remain distinct for signing\. We therefore compare two Turkish\-specific pseudo\-gloss construction strategies, both of which produce sentence\-level lexical sets rather than manual sign glosses\.

##### Morphology\-lemma strategy\.

The first strategy uses rule\-based Turkish morphological analysis, following the finite\-state/two\-level morphology tradition for agglutinative Turkish\[[12](https://arxiv.org/html/2608.11332#bib.bib21),[13](https://arxiv.org/html/2608.11332#bib.bib22)\]\. For each Turkish surface token, the analyzer returns possible morphological parses; each parse is decomposed into a structured representation, and the root morpheme of the first inflectional group is treated as a lemma candidate\. Since this stage does not perform sentence\-level morphological disambiguation, multiple analyses of the same token are resolved by selecting the root that appears most often among the returned parses\. Tokens with no root analysis are discarded, one\-character lemmas are removed, and repeated lemmas inside the same sentence are collapsed because the pretraining objective supervises pseudo\-gloss presence rather than token count\. This strategy mainly removes inflectional variation, so forms such as case\-, possessive\-, plural\-, tense\-, or agreement\-marked words contribute to a shared lemma target\. Compared with a raw, unlemmatized surface\-token baseline \(not manually annotated glosses\) of approximately 16K targets, this yields a pseudo\-gloss vocabulary of 4802 classes\. Its limitation is that analyzer ambiguity, named entities, derivational morphology, and broadcast phrasing can still create lexical targets that do not correspond cleanly to signed units\.

##### LLM lexical\-rule strategy\.

The second strategy uses an LLM as a constrained lexical normalizer rather than as a free gloss generator\. The normalization rules select content\-bearing Turkish units – nouns, proper names, verbs, adjectives, adverbs, pronouns, and numbers – while suppressing function words such as conjunctions and question particles\. Inflectional morphology is removed; verbs are normalized toward an active infinitive or base form; spelling, diacritic, and proper\-name variants are canonicalized; and derivational suffixes are preserved when they change the lexical meaning\. Passive and causative verb variants are pruned when a simpler base verb is already available, reducing redundant supervision from closely related verbal forms\.

To keep the output closed\-vocabulary, translation tokens are then matched against the allowed pseudo\-gloss inventory using exact lexical matches, root/prefix matches, and fuzzy string candidates; the LLM may only choose from these candidates or reject the token\. This turns the LLM step into a rule\-constrained disambiguation and normalization stage rather than an unbounded annotation source\. The resulting pseudo\-gloss vocabulary contains 6,539 classes, 1737 more than the morphology\-lemma vocabulary\. Thus, the two alternatives differ not only in size, but also in how they trade off coverage, derivational specificity, and lexical consistency\. The resulting TSLD coverage difference is reported in Table[2](https://arxiv.org/html/2608.11332#S4.T2), where Overlap \(%\) is computed relative to a fixed 2004 pool of TSLD sign\-word entries considered matchable under either method, not relative to each method’s own pseudo\-gloss vocabulary size\.

Additional implementation details for the LLM lexical\-rule pseudo\-gloss construction, including the prompt structure, mapping statistics, and training hyperparameters, are provided in the Supplementary Material\.

## 4Datasets and Cross\-Dataset Evaluation Protocol

### 4\.1Datasets

##### TSL\-News\.

We introduce TSL\-News, a new Turkish news broadcast corpus, to support weak\-supervision\-based techniques in the sign language domain \(Table[1](https://arxiv.org/html/2608.11332#S4.T1)\)\. It contains television \(TV\) news videos from 2021–2023 with Turkish transcripts, but no manual gloss labels, sign\-order labels, or temporal sign boundaries\. The corpus therefore matches the limited\-resource setting of this paper: supervision is available only at the sentence level, while the model must learn temporally useful sign representations from noisy spoken\-language text\.

Table 1:TSL\-News broadcast pretraining corpus\.
##### TSLD\.

The Turkish Sign Language dictionary corpus \(TSLD\) is used for transfer evaluation\. TSLD is based on the publicly available TID Dictionary corpus\[[10](https://arxiv.org/html/2608.11332#bib.bib23)\]; the local crawl contains 2004 dictionary headword entries with isolated sign videos and example or meaning videos\. It provides sign\-word dictionary entries, isolated sign videos, and sentence\-level example videos\. For this work, TSL\-SB is prepared by matching TSLD sign\-word entries or single\-token synonym variants to the TSL\-News pseudo\-gloss vocabularies \(the matching procedure is detailed in Sec\.[4\.2](https://arxiv.org/html/2608.11332#S4.SS2)\), then manually temporally annotating sign occurrences in example videos\. This separates training from evaluation: the encoder is pretrained on broadcast TSL\-News, but spotting is measured on dictionary and example videos from TSLD\.

### 4\.2Benchmark Construction

The evaluation links two resources\. TSL\-News supplies the broadcast training distribution and pseudo\-gloss vocabularies\. TSLD supplies sign\-word dictionary entries, isolated sign videos, and continuous example sentence videos\. An overlap vocabulary is constructed by scanning TSL\-News pseudo\-gloss tokens and matching TSLD sign\-word dictionary entries or single\-token synonym variants, after removing a conservative list of Turkish stopwords\. Multiword phrase matches are intentionally avoided to reduce noisy overlap from frequent function words\.

Table 2:TSLD overlap construction statistics with TSL\-News for the two pseudo\-gloss construction methods\. Vocabulary sizes identify the corresponding weak\-supervision settings\. Overlap \(%\) is normalized by the fixed pool of 2004 matchable TSLD sign\-word entries, not by the pseudo\-gloss vocabulary size\.TSL\-SB contains 596 annotated sign words and 1842 sentence\-level temporal annotations, of which 1817 fall inside the LLM lexical\-rule vocabulary and 1159 inside the morphology\-lemma vocabulary, with a shared intersection of 1137 used for the controlled comparison in Sec\.[5\.1](https://arxiv.org/html/2608.11332#S5.SS1)\(the underlying label export contains 2127 segment\-level intervals, whose role is detailed in Sec\.[5\.3](https://arxiv.org/html/2608.11332#S5.SS3)\)\. Together with the isolated sign videos used to build template representations, the benchmark contains 4363 videos\. Because the model output layer is defined by its pseudo\-gloss dictionary, each evaluation filters examples whose target sign word is absent from the active vocabulary\.

### 4\.3NCC Localization

For an annotated TSLD example, letVsentV\_\{\\mathrm\{sent\}\}be the continuous sentence video andVisoV\_\{\\mathrm\{iso\}\}be the isolated dictionary video\. ThenR=Eθ​\(Vsent\)R=E\_\{\\theta\}\(V\_\{\\mathrm\{sent\}\}\)is the continuous sentence representation andQ=Eθ​\(Viso\)Q=E\_\{\\theta\}\(V\_\{\\mathrm\{iso\}\}\)is the isolated\-sign representation\. The isolated sequenceQQis slid overRR, and normalized cross\-correlation \(NCC\)\[[8](https://arxiv.org/html/2608.11332#bib.bib14)\]is computed at offsetuu:

NCC⁡\(u\)=1\|Q\|​∑i=1\|Q\|⟨Qi−μQσQ,Ru\+i−μR,uσR,u⟩\.\\operatorname\{NCC\}\(u\)=\\frac\{1\}\{\|Q\|\}\\sum\_\{i=1\}^\{\|Q\|\}\\left\\langle\\frac\{Q\_\{i\}\-\\mu\_\{Q\}\}\{\\sigma\_\{Q\}\},\\frac\{R\_\{u\+i\}\-\\mu\_\{R,u\}\}\{\\sigma\_\{R,u\}\}\\right\\rangle\.\(4\)Here,uuis the candidate start offset in the continuous sequence,\|Q\|\|Q\|is the isolated\-template length,QiQ\_\{i\}is theiith isolated hidden state, andRu\+iR\_\{u\+i\}is the aligned continuous hidden state\. The termsμQ\\mu\_\{Q\}andσQ\\sigma\_\{Q\}are the mean and standard deviation of the isolated template, computed per feature dimension over the\|Q\|\|Q\|template steps, whileμR,u\\mu\_\{R,u\}andσR,u\\sigma\_\{R,u\}are the corresponding statistics of the continuous window starting atuu\. The inner product compares the standardized template and continuous features\. The top\-kkNCC peaks define candidate temporal windows of length\|Q\|\|Q\|\. Each window is compared against the human annotation using interval IoU, and the best IoU among the firstkkranked candidates is reported fork∈\{1,3,5\}k\\in\\\{1,3,5\\\}\. All main runs trim 0\.5 seconds from both ends of the continuous search region and 0\.2 seconds from both ends of isolated signs\. This trimming removes non\-signing margins because idle frames can dominate NCC matching if they are left in the search/template sequences\.

## 5Experiments

### 5\.1Main NCC Results

Table[3](https://arxiv.org/html/2608.11332#S5.T3)reports NCC localization performance\. The spatial\-feature baseline uses raw pretrained spatial features before stage\-1 pseudo\-gloss pretraining, evaluated on the same example set as the LLM lexical\-rule vocabulary \(hence the identicalNNand Skip values\)\. The trained LLM lexical\-rule model improves top\-1 mean IoU by 0\.147 absolute over this baseline and nearly doubles top\-5 mean IoU \(0\.465 vs\. 0\.235\)\. It also improves top\-5 IoU@0\.50 from 23\.9% to 56\.2%\.

Table 3:NCC temporal localization\.NNis the number of evaluated examples, Skip is the number filtered out by vocabulary mismatch, and Top\-5@0\.50 is the percentage of examples whose best top\-5 candidate reaches at least 0\.50 IoU\.##### Pseudo\-gloss Construction and Coverage\.

The LLM lexical\-rule strategy increases the matched TSLD sign vocabulary from 1,060 to 1,398 signs\. It also reduces the number of benchmark examples skipped due to vocabulary mismatch from 683 to 25\. In the full coverage\-oriented comparison in Table[3](https://arxiv.org/html/2608.11332#S5.T3), the LLM lexical\-rule encoder is evaluated on 1817 examples, while the morphology\-lemma encoder is evaluated on 1159 examples; therefore, these rows should be read as both a representation and coverage comparison\. To control for the different evaluated example counts, we additionally compute scores on the 1137 annotated example groups covered by both vocabularies\. On this shared subset, the morphology\-lemma encoder reaches 0\.175/0\.297/0\.368 top\-1/top\-3/top\-5 IoU, while the LLM lexical\-rule encoder reaches 0\.271/0\.412/0\.473, corresponding to gains of 0\.096/0\.115/0\.105\.

Figure 2:Top\-5 IoU survival curves for NCC localization\. Each curve uses its own evaluated set: spatial features \(N=1817N\{=\}1817\), morphology\-lemma \(N=1159N\{=\}1159\), and LLM lexical\-rule \(N=1817N\{=\}1817\)\. The dashed line markst=0\.50t\{=\}0\.50, corresponding to the Top\-5@0\.50 column in Table[3](https://arxiv.org/html/2608.11332#S5.T3)\.Figure[2](https://arxiv.org/html/2608.11332#S5.F2)shows the full top\-5 IoU distribution behind the averages in Table[3](https://arxiv.org/html/2608.11332#S5.T3)\. The curves make two effects visible\. Where each curve meets the vertical axis \(t=0t\{=\}0\) shows how many examples were localized at all: the spatial baseline starts near 48%, meaning the sign is completely missed \(top\-5 IoU=0=0\) in about half of the examples, while the LLM lexical\-rule encoder starts near 75%, cutting the complete\-miss rate to about one quarter\. The curves also shift to the right, meaning that when a sign is found, it is found more precisely: the median top\-5 IoU rises from 0\.000 for the spatial baseline and 0\.358 for the morphology\-lemma encoder to 0\.553 for the LLM lexical\-rule encoder\.

Figure[3](https://arxiv.org/html/2608.11332#S5.F3)relates localization quality to the number of TSL\-News train sentences in which each target pseudo\-gloss appears\. The association is weak \(ρ=−0\.069\\rho\{=\}\-0\.069\), and the quartile means remain close even for low\-frequency signs, suggesting that the LLM lexical\-rule encoder is not simply memorizing frequent pseudo\-gloss labels\. Two caveats apply\. By construction, the benchmark contains only signs that appear at least once in the pretraining text, so truly unseen signs are not tested\. Within this matched range, however, localization quality does not depend on how often a sign was seen during pretraining; it appears to be driven more by other factors, such as how visually distinctive the sign is and how much its dictionary production differs from its sentence\-context production\.

Figure 3:TSL\-News train frequency versus TSL\-SB localization for the LLM lexical\-rule encoder\. Each point is one matched sign word \(N=589N\{=\}589\); thexx\-axis islog10⁡\(frequency\+1\)\\log\_\{10\}\(\\mathrm\{frequency\}\{\+\}1\), where frequency counts train sentences whose pseudo\-gloss set contains the target\. Black diamonds show frequency\-quartile means\. Spearmanρ=−0\.069\\rho\{=\}\-0\.069\.

### 5\.2Pseudo\-Gloss Pretraining and Translation Check

Table[4](https://arxiv.org/html/2608.11332#S5.T4)isolates pretraining quality from downstream translation behavior\. The stage\-1 validation result focuses on pseudo\-gloss construction and shows that rule\-based morphological lemmatization is much stronger than a larger raw surface\-token space\. The LLM lexical\-rule vocabulary is flagged separately because validation F1 is not directly comparable across target vocabularies of different sizes\. We include the LLM lexical\-rule row for completeness, but interpret F1 primarily within each target vocabulary rather than as a direct fairness comparison across vocabularies of different sizes\.

Table 4:TSL\-News pseudo\-gloss pretraining validation\. Val F1 is validation F1\-score, the harmonic mean of precision and recall\. The gain is measured against the raw surface\-token baseline\.In the translation check, the Sign2GPT\-style translation setup \(frozen XGLM decoder with LoRA adapters, as in Fig\.[1](https://arxiv.org/html/2608.11332#S3.F1)\) is kept unchanged; only the encoder initialization differs\. We report both LLM lexical\-rule and morphology\-lemma pretrained encoders; the strongest translation check uses the morphology\-lemma pseudo\-gloss construction\. Reusing the pseudo\-gloss pretrained encoder improves all test bilingual evaluation understudy \(BLEU\) and recall\-oriented understudy for gisting evaluation \(ROUGE\) metrics\[[15](https://arxiv.org/html/2608.11332#bib.bib15),[9](https://arxiv.org/html/2608.11332#bib.bib16)\]over training without stage\-1 pretraining\. These translation numbers contextualize the changed pretraining stage rather than introducing a new translation method\.

Table 5:TSL\-News translation check\. The translation setup is fixed; only pseudo\-gloss pretraining of the encoder changes\. B1–B4 denote BLEU\-1–BLEU\-4 and R denotes ROUGE\.
### 5\.3Auxiliary Localization Diagnostics

Table[6](https://arxiv.org/html/2608.11332#S5.T6)compares auxiliary localization diagnostics on the LLM lexical\-rule setup\. All columns report temporal IoU except Top\-5@0\.50, which reports the percentage of examples whose best top\-5 candidate reaches at least 0\.50 IoU\. Table[7](https://arxiv.org/html/2608.11332#S5.T7)separately reports rank\-based retrieval metrics for the all\-vocabulary local NCC setting\.

Figure[4](https://arxiv.org/html/2608.11332#S5.F4)keeps the main qualitative frame evidence by showing frames from the ground\-truth example span and from the isolated videos of the top\-5 retrieved sign\-word candidates\. If high local NCC scores reflect visual similarity in the learned sign representations, the top\-ranked isolated signs should show similar hand and body configurations to the ground\-truth span\.

![Refer to caption](https://arxiv.org/html/2608.11332v1/figures/local_gt_ncc_selected_frames_blurred.png)Figure 4:Qualitative frame comparison for local all\-vocabulary NCC\. The top row shows the ground\-truth*Karışık*\(“complicated”\) span, and the remaining rows show the top\-5 isolated retrieval candidates\. Similar hand and body configurations support that local NCC retrieves visually nearby sign representations\. Faces are blurred to avoid exposing signer identity\.Table 6:Auxiliary localization diagnostics on the LLM lexical\-rule benchmark\.NNis the number of evaluated examples for the first two rows and evaluated segment occurrences for All\-vocabulary local NCC\.Table 7:All\-vocabulary local NCC retrieval metrics\. Recall@kkranks the correct sign word among all local NCC candidates; Avg\. cand\. is the mean number of candidate isolated templates\.Additional qualitative evaluations for the localization diagnostics are provided in Appendix[0\.A\.2](https://arxiv.org/html/2608.11332#Pt0.A1.SS2)\.

## 6Conclusion

This paper studies whether weak text\-derived pseudo\-gloss supervision can pretrain reusable visual representations for limited\-resource Turkish Sign Language\. Using TSL\-News for broadcast pretraining and TSL\-SB for cross\-dataset evaluation, NCC over hidden states shows that the pretrained encoder transfers better than raw spatial features, the morphology\-lemma pseudo\-gloss baseline, and direct pseudo\-gloss score localization\. The main conclusion is therefore about representation learning: even noisy sentence\-level pseudo\-glosses can shape a visual encoder into a lexical\-temporal representation that supports dictionary\-query sign spotting and provides a useful starting point for downstream translation checks\. The evaluation remains limited to representation quality through sign spotting: TSL\-SB relies on Turkish word\-form overlap rather than expert sign glosses, isolated dictionary productions can differ from sentence\-context productions, TSL\-News contains only three signers, and the all\-vocabulary diagnostic still searches within the annotated interval plus a small buffer\. Future work will extend this diagnostic toward open\-ended proposals, sentence\-level reranking, broader TSL\-SB validation, and full translation studies that use translation as a consumer of the learned encoder rather than the central contribution\. The TSL\-News corpus and the TSL\-SB benchmark annotations will be released upon publication to support research on limited\-resource sign languages\.

## Acknowledgements

This work is supported by the Scientific and Technological Research Council of Türkiye \(TÜBİTAK\) under the 1001 Scientific and Technological Research Projects Funding Program \(Project No\. 124E618\)\. We acknowledge the EuroHPC Joint Undertaking for awarding us access to Vega at IZUM, Slovenia, through Development Access allocation 2025D08\-090\.

## References

- \[1\]P\. Bojanowski, E\. Grave, A\. Joulin, and T\. Mikolov\(2017\)Enriching word vectors with subword information\.Transactions of the association for computational linguistics5,pp\. 135–146\.Cited by:[§3\.1](https://arxiv.org/html/2608.11332#S3.SS1.p1.2)\.
- \[2\]D\. Bragg, O\. Koller, M\. Bellard, L\. Berke, P\. Boudreault, A\. Braffort, N\. Caselli, M\. Huenerfauth, H\. Kacorri, T\. Verhoef,et al\.\(2019\)Sign language recognition, generation, and translation: an interdisciplinary perspective\.InProceedings of the 21st international ACM SIGACCESS conference on computers and accessibility,pp\. 16–31\.Cited by:[§1](https://arxiv.org/html/2608.11332#S1.p1.1)\.
- \[3\]M\. De Coster, D\. Shterionov, M\. Van Herreweghe, and J\. Dambre\(2024\)Machine translation from signed to spoken languages: state of the art and challenges\.Universal Access in the Information Society23\(3\),pp\. 1305–1331\.Cited by:[§1](https://arxiv.org/html/2608.11332#S1.p1.1)\.
- \[4\]L\. J\. He, O\. M\. Sincan, and R\. Bowden\(2025\)Sign spotting disambiguation using large language models\.InAdjunct Proceedings of the 25th ACM International Conference on Intelligent Virtual Agents, IVA Adjunct,Vol\.25\.Cited by:[§2](https://arxiv.org/html/2608.11332#S2.SS0.SSS0.Px3.p1.1)\.
- \[5\]L\. J\. He, H\. Walsh, O\. M\. Sincan, and R\. Bowden\(2025\)Hands\-on: segmenting individual signs from continuous sequences\.In2025 IEEE 19th International Conference on Automatic Face and Gesture Recognition \(FG\),pp\. 1–5\.Cited by:[§2](https://arxiv.org/html/2608.11332#S2.SS0.SSS0.Px1.p1.1)\.
- \[6\]L\. Hu, L\. Gao, Z\. Liu, and W\. Feng\(2023\)Continuous sign language recognition with correlation network\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 2529–2539\.Cited by:[§2](https://arxiv.org/html/2608.11332#S2.SS0.SSS0.Px3.p1.1)\.
- \[7\]O\. Koller\(2020\)Quantitative survey of the state of the art in sign language recognition\.arXiv preprint arXiv:2008\.09918\.Cited by:[§1](https://arxiv.org/html/2608.11332#S1.p1.1)\.
- \[8\]J\. P\. Lewis\(1995\)Fast normalized cross\-correlation\.InVision interface,Vol\.10,pp\. 120–123\.Cited by:[§4\.3](https://arxiv.org/html/2608.11332#S4.SS3.p1.1)\.
- \[9\]C\. Lin\(2004\)Rouge: a package for automatic evaluation of summaries\.InText summarization branches out,pp\. 74–81\.Cited by:[§5\.2](https://arxiv.org/html/2608.11332#S5.SS2.p2.1)\.
- \[10\]B\. Makaroğlu and H\. Dikyuva \(Eds\.\)\(2017\)Güncel türk İşaret dili sözlüğü\.Aile ve Sosyal Hizmetler Bakanlığı,Ankara\.Cited by:[§4\.1](https://arxiv.org/html/2608.11332#S4.SS1.SSS0.Px2.p1.1)\.
- \[11\]L\. Momeni, G\. Varol, S\. Albanie, T\. Afouras, and A\. Zisserman\(2020\)Watch, read and lookup: learning to spot signs from multiple supervisors\.InProceedings of the Asian conference on computer vision,Cited by:[§2](https://arxiv.org/html/2608.11332#S2.SS0.SSS0.Px1.p1.1)\.
- \[12\]K\. Oflazer and C\. Guzey\(1994\)Spelling correction in agglutinative languages\.InFourth Conference on Applied Natural Language Processing,pp\. 194–195\.Cited by:[§3\.2](https://arxiv.org/html/2608.11332#S3.SS2.SSS0.Px1.p1.1),[§3\.2](https://arxiv.org/html/2608.11332#S3.SS2.p1.1)\.
- \[13\]K\. Oflazer and I\. Kuruoz\(1994\)Tagging and morphological disambiguation of turkish text\.InFourth Conference on Applied Natural Language Processing,pp\. 144–149\.Cited by:[§3\.2](https://arxiv.org/html/2608.11332#S3.SS2.SSS0.Px1.p1.1),[§3\.2](https://arxiv.org/html/2608.11332#S3.SS2.p1.1)\.
- \[14\]M\. Oquab, T\. Darcet, T\. Moutakanni, H\. Vo, M\. Szafraniec, V\. Khalidov, P\. Fernandez, D\. Haziza, F\. Massa, A\. El\-Nouby,et al\.\(2023\)Dinov2: learning robust visual features without supervision\.arXiv preprint arXiv:2304\.07193\.Cited by:[§3\.1](https://arxiv.org/html/2608.11332#S3.SS1.p1.2)\.
- \[15\]K\. Papineni, S\. Roukos, T\. Ward, and W\. Zhu\(2002\)Bleu: a method for automatic evaluation of machine translation\.InProceedings of the 40th annual meeting of the Association for Computational Linguistics,pp\. 311–318\.Cited by:[§5\.2](https://arxiv.org/html/2608.11332#S5.SS2.p2.1)\.
- \[16\]C\. Raude, K\. Prajwal, L\. Momeni, H\. Bull, S\. Albanie, A\. Zisserman, and G\. Varol\(2024\)A tale of two languages: large\-vocabulary continuous sign language recognition from spoken language supervision\.arXiv preprint arXiv:2405\.10266\.Cited by:[§2](https://arxiv.org/html/2608.11332#S2.SS0.SSS0.Px1.p1.1)\.
- \[17\]N\. Shahin and L\. Ismail\(2024\)From rule\-based models to deep learning transformers architectures for natural language processing and sign language translation systems: survey, taxonomy and performance evaluation\.Artificial Intelligence Review57\(10\),pp\. 271\.Cited by:[§1](https://arxiv.org/html/2608.11332#S1.p1.1)\.
- \[18\]H\. Walsh, O\. M\. Sincan, B\. Saunders, and R\. Bowden\(2023\)Gloss alignment using word embeddings\.In2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops \(ICASSPW\),pp\. 1–5\.Cited by:[§2](https://arxiv.org/html/2608.11332#S2.SS0.SSS0.Px3.p1.1)\.
- \[19\]R\. Wong, N\. C\. Camgoz, and R\. Bowden\(2023\)Learnt contrastive concept embeddings for sign recognition\.InProceedings of the IEEE/CVF International Conference on Computer Vision,pp\. 1945–1954\.Cited by:[§2](https://arxiv.org/html/2608.11332#S2.SS0.SSS0.Px2.p1.1)\.
- \[20\]R\. Wong, N\. C\. Camgoz, and R\. Bowden\(2024\)Sign2GPT: leveraging large language models for gloss\-free sign language translation\.arXiv preprint arXiv:2405\.04164\.Cited by:[§1](https://arxiv.org/html/2608.11332#S1.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2608.11332#S1.p2.1),[§2](https://arxiv.org/html/2608.11332#S2.SS0.SSS0.Px2.p1.1),[§3\.1](https://arxiv.org/html/2608.11332#S3.SS1.p2.1.1)\.
- \[21\]R\. Wong, N\. C\. Camgoz, and R\. Bowden\(2025\)Signrep: enhancing self\-supervised sign representations\.InProceedings of the IEEE/CVF international conference on computer vision,pp\. 22804–22814\.Cited by:[§2](https://arxiv.org/html/2608.11332#S2.SS0.SSS0.Px2.p1.1)\.
- \[22\]W\. Yu, M\. Luo, P\. Zhou, C\. Si, Y\. Zhou, X\. Wang, J\. Feng, and S\. Yan\(2022\)Metaformer is actually what you need for vision\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 10819–10829\.Cited by:[§3\.1](https://arxiv.org/html/2608.11332#S3.SS1.p1.2)\.
- \[23\]B\. Zhou, Z\. Chen, A\. Clapés, J\. Wan, Y\. Liang, S\. Escalera, Z\. Lei, and D\. Zhang\(2023\)Gloss\-free sign language translation: improving from visual\-language pretraining\.InProceedings of the IEEE/CVF International Conference on Computer Vision,pp\. 20871–20881\.Cited by:[§2](https://arxiv.org/html/2608.11332#S2.SS0.SSS0.Px2.p1.1)\.

## Appendix 0\.ASupplementary Material

### 0\.A\.1Sign2GPT Adaptation and Training Configuration

We retain the Sign2GPT macro\-architecture while adapting it for Turkish weak pseudo\-gloss pretraining and a controlled translation check\. Tables[0\.A\.1](https://arxiv.org/html/2608.11332#Pt0.A1.T1)and[0\.A\.2](https://arxiv.org/html/2608.11332#Pt0.A1.T2)specify the frozen and trainable components and the corresponding optimization settings\. fastText initializes the prototype vectors but neither selects lemmas nor consolidates Turkish inflections\. Because no official trained spaCy\-Turkish POS\-tagging and lemmatization pipeline was available, pseudo\-gloss normalization instead uses the compared Turkish morphological analyzer and constrained LLM mapper\.

Table 0\.A\.1:Stage\-1 pseudo\-gloss pretraining configuration\.Table 0\.A\.2:Stage\-2 translation\-check configuration\.
### 0\.A\.2Qualitative NCC Curve Diagnostics

The pseudo\-gloss classifier uses a known class\-score curve, target\-known NCC uses the corresponding isolated template, and all\-vocabulary NCC introduces competition among templates\. The latter searches only the annotated interval with a 0\.25\-second margin and is therefore local retrieval rather than full\-video open\-vocabulary spotting\. ItsNNcounts segment occurrences, whereas the other diagnostics count example groups, so their IoU values are not directly comparable\.

Figures[0\.A\.1](https://arxiv.org/html/2608.11332#Pt0.A1.F1)and[0\.A\.2](https://arxiv.org/html/2608.11332#Pt0.A1.F2)provide qualitative diagnostics for the two NCC settings summarized in the main paper\. The target\-known diagnostic uses only the isolated template of the annotated target sign and shows whether its NCC peak aligns with the human temporal annotation\. The local all\-vocabulary diagnostic uses the same annotated interval plus a small buffer but lets the target compete against all candidate isolated templates, making it visible whether high\-ranking alternatives correspond to similar temporal spans rather than arbitrary background motion\.

![Refer to caption](https://arxiv.org/html/2608.11332v1/figures/target_known_ncc_curves.png)Figure 0\.A\.1:Target\-known NCC curve diagnostic\. The plot visualizes whether the isolated template for the known target sign produces NCC peaks near the human annotated sign interval\. The searched sign word/gloss in this sequence is*Karışık*\(“complicated”\)\. The predicted span \(pink\) largely overlaps the ground\-truth span \(green\); the span plot below the curve separates the top\-kkcandidates for easier inspection\.![Refer to caption](https://arxiv.org/html/2608.11332v1/figures/local_gt_ncc_curves.png)Figure 0\.A\.2:Local all\-vocabulary NCC diagnostic\. The plot visualizes the target sign against competing TSLD isolated templates inside the annotated interval plus the 0\.25 second buffer\. The searched sign word/gloss in this sequence is*Karışık*\(“complicated”\)\. Each candidate’s predicted span is drawn in a distinct color and largely overlaps the ground\-truth span \(green\); the span plot below the curve separates the candidates for easier inspection\.
### 0\.A\.3LLM Lexical\-Rule Pseudo\-Gloss Details

The LLM lexical\-rule pseudo\-gloss pipeline is implemented in two stages\. First, candidate pseudo\-glosses are constructed from a fixed allowed vocabulary using exact, root/prefix, and fuzzy matches\. Second, a local OpenAI\-compatible chat\-completions server selects one candidate or rejects all candidates\. The final experiments use a frozen, manually reviewed merged word\-to\-pseudo\-gloss mapping; this mapping is then applied deterministically to the TSL\-News corpus CSVs and converted to the stage\-1 pseudo\-gloss dictionary used in training\.

#### Worked Pseudo\-Gloss Examples

Table[0\.A\.3](https://arxiv.org/html/2608.11332#Pt0.A1.T3)contrasts representative outputs: the rule\-based analyzer may return shorter roots such as*kur*and*söyle*, whereas the constrained mapper selects fuller or canonical entries such as*kuruluş*and*söylemek*from the fixed vocabulary\.

Table 0\.A\.3:Side\-by\-side pseudo\-gloss construction examples from TSL\-News\. English translations are explanatory only and are not model inputs\.Table 0\.A\.4:LLM lexical\-rule extraction settings\.The word\-level system prompt used by the mapper is:

> Turkce kelimeyi pseudo\_gloss ile eslestir\. Yalniz ADAYLAR listesinden sec; listede olmayan gloss yazma\. Yeni kelime uretme\. Kelime cekimli fiil veya ek almis isim olabilir\. Kok/anlam olarak dusun\. Adaylarda kelimenin kokune benzeyen bir gloss varsa MUTLAKA onu sec\. Ornek: gosterdi\-\>gostermek, gosterecek\-\>gostermek, gosteren\-\>gostermek\. Ornek: geldiniz\-\>gel, artirdik\-\>artir, cikarilacak\-\>cikar, bultenine\-\>bulten\. NONE sadece adaylarin hicbiri kelimeyle anlam veya kok olarak ilgili degilse kullanilir\. Sadece JSON dondur: \{"sonuclar":\[\{"kelime":"\.\.\.","gloss":"\.\.\."\}\]\}

Its English translation is:

> Match the Turkish word to a pseudo\-gloss\. Choose only from the CANDIDATES list; do not write a gloss that is not in the list, and do not invent a new word\. The word may be an inflected verb or a suffixed noun\. Reason by root and meaning\. If one candidate resembles the root of the word, choose it\. Use NONE only when none of the candidates are related to the word by meaning or root\. Return only the requested JSON object\.

The corresponding user prompt has the structure:

> KELIMELER VE ADAYLAR: \{\[ \{"kelime": word, "candidates": \[candidate\_1, \.\.\.\]\}, \.\.\. \]\} Sadece aday listesinden sec\.

Its English translation is:

> WORDS AND CANDIDATES: \{\[ \{"word": word, "candidates": \[candidate\_1, \.\.\.\]\}, \.\.\. \]\} Choose only from the candidate list\.

For sentence\-level mapping, the alternative mapper uses the same local chat\-completions interface with the default model aliasqwen, temperature 0\.0, four workers, and candidate chunks of 250\. Its system prompt is JSON\-only and closed\-vocabulary: select pseudo\-glosses only fromADAYLAR, do not invent new words, and return\{"pseudo\_gloss":\["\.\.\."\]\}\. Repeated identical sentences are cached within a run, so they receive the same pseudo\-gloss set in the generated CSV\. The final reported experiments do not depend on re\-calling the LLM: the merged word mapping is frozen, and rebuilding the 6539\-class TSL\-News pseudo\-gloss corpus from that mapping is deterministic\. A separate independent repeat\-call consistency audit of the original local LLM server was not logged\.