Dialect Gloss
Abstract
Brief Description This variable proposes LLM-Assisted Dialect Gloss Detection (see also Bommarito & Michael, 2025; Liang et al., 2026) and refers to the automated identification, contextual interpretation, and measurement of dialectal, vernacular, or otherwise non-standard linguistic forms in digital texts (Joshi et al., 2025). The procedure combines dictionary-based detection with large language model (LLM)-assisted contextual glossing. A gloss is understood here as a standardized explanation or rendering of a dialectal expression that supports computational analysis while preserving the original linguistic form (Bommarito & Michael, 2025). The variable operationalizes dialect use by identifying vernacular tokens, phrases, or grammatical forms and converting these observations into measurable outputs. One output is dialect intensity, calculated either as a continuous measure, such as the number or proportion of dialect markers per comment, post, or document, or as an ordinal measure, such as low, medium, or high dialect presence. Additional outputs may include glossed text, lists of detected dialect terms, uncertainty scores (provided by the model), and dialect dictionaries developed from corpus-level aggregation. For instance, the comment “cah even scratch my bamcee in peace inno” contains three detected dialect forms: cah, bamcee, and inno. These are contextually glossed as “I can't even scratch my behind in peace, you know,” while the original wording is retained for analysis. Based on the proportion and distribution of detected dialect forms, the comment receives a dialect intensity score of 0.38. LLM-Assisted Dialect Gloss Detection is particularly relevant for automated content analysis in dialect-rich environments where standard preprocessing pipelines may normalize, misclassify, or erase vernacular forms. In such contexts, spelling variation, code-switching, phonetic writing, regional grammar, and culturally specific vocabulary can reduce the validity of downstream analyses such as sentiment analysis, topic modeling, stance detection, or frame analysis. The original text remains unchanged, while the gloss layer provides a machine-readable representation for subsequent automated analysis. While low-resource language glossing is fairly successful if the model is provided with sufficient context, the approach should be combined with a human/machine-in-the-loop approach to avoid reintroducing the language-flattening aspects of automated translation that this approach aims to avoid (Alqazlan et al., 2025; Hong et al., 2022). Field of Application/Theoretical Foundation The theoretical foundation of the variable lies in sociolinguistic work on language variation, indexicality, and the social meaning of dialect (Eckert, 2008; Labov, 1972). Vernacular forms can signal locality, group membership, stance, affect, humor, intimacy, resistance, or cultural belonging (Ming, 2020). At the same time, the variable is grounded in computational communication research and automated content analysis. Dictionary-based procedures are widely used but depend on the quality and coverage of the underlying dictionary. This is problematic in dialect-rich contexts, where relevant terms may be absent from standard dictionaries, may appear in multiple spellings, or may only be interpretable within local context. LLM-assisted glossing can help address these limitations by identifying additional candidate forms and producing contextual metadata. References/Combination with other Methods of Data Collection LLM-Assisted Dialect Gloss Detection can be combined with manual content analysis or serve as a preprocessing step for other automated content analysis variables. For example, sentiment analysis may be conducted on both the original text and the glossed text to examine whether dialect-aware preprocessing changes sentiment scores. Topic modeling or frame analysis may be applied after glossing to reduce the exclusion of dialect-rich comments from interpretable topic clusters. A typical automated procedure consists of the following steps: Corpus preparation: Collect digital texts and retain the original text without normalization or spelling correction. Seed dictionary construction: Develop an initial vernacular dictionary from existing linguistic resources, prior research, expert knowledge, or manually reviewed corpus samples. Dictionary-based detection: Identify known dialect tokens or phrases through rule-based or dictionary-based matching. LLM-assisted glossing: Use constrained prompts to identify additional dialect forms, propose contextual glosses, and generate structured metadata. Human validation: Review a benchmark sample of extracted terms and glosses, ideally with coders or validators familiar with the relevant language variety. Validators assess whether the extracted forms are contextually valid, whether the glosses accurately represent their meaning, and whether cultural meaning, pragmatic force, and the voice of the original comment have been preserved. Disagreements or inaccurate outputs are used to revise the dictionary, glosses, and prompting procedure. External sources, including published online and offline dictionaries are used as external validators. Dictionary refinement: Terms are shown to validators in batches of 25 to confirm Add validated forms to the dictionary and remove false positives. Scoring: Calculate dialect intensity as a raw count, normalized rate, binary indicator, or ordinal category, e.g. Downstream analysis: Use the original and glossed text for further automated or qualitative analysis. Example Studies for Automated Content Analysis Table 1 provides examples of work on dialect detection in low resource languages. Robinson et al. (2024) introduce Kreyòl-MT, a large-scale machine-translation dataset and model infrastructure for Creole languages across Latin America, the Caribbean, and Africa. Smith et al. (2026) develop CariBOT, a Trinidadian English Creole to Standard English architecture for improving chatbot responses. Their Caribe library preprocesses Creole input before it reaches a standard English chatbot. Jackson et al. (2025) assess GPT-4’s communicative competence with Trinidadian English Creole and show the struggle with vocabulary, authenticity, and correct identification. Table 1: Measurement of „dialect detection“ in low resource languages Author(s) Sample Procedure Formal validity check with manual coding as benchmark Code Availability (Bengsch, forthcoming) 30,000 TikTok comments Dictionary lookup to detect known Trinidadian vernacular forms → constrained zero-shot LLM (GPT-4) prompting to identify and contextually gloss unmatched forms → consolidation of dictionary- and LLM-detected forms through Python → manual review of detected forms and glosses Dictionary- and LLM-assisted detections compared with a manually coded benchmark sample; detection and glossing performance assessed separately https://github.com/GeriNZ/caribbeandialect (Robinson et al., 2024) Parallel and monolingual datasets for 41 Creole languages; 107 sources such as anthology websites promoting cultural heritage or language revitalization, educational materials, and government documents Extraction of parallel and monolingual text → text cleaning, normalisation, deduplication and train–development–test splitting → training of multilingual SentencePiece models and fine-tuning of mBART-50 with YANMTT → machine-translation evaluation using held-out test sets Not conducted with manual coding as benchmark. Models were evaluated quantitatively on held-out translation data using chrF and existing machine-translation benchmarks Not available (Jackson et al., 2025) Facebook pages for the Trinidad and Tobago Ministry of Health, Ministry of National Security, Ministry of Education, Ministry of Works and Transport and Loop TT (a Trinidad and Tobago news source) Manual corpus-based identification of TEC features → zero-shot GPT-4 prompting via the OpenAI API, with five completions per prompt → expert comparison of outputs against paired Standard English prompts, authentic corpus usage, dictionaries, and grammatical descriptions. Expert manual evaluation against corpus, dictionary, and grammatical descriptions; Task-specific accuracy percentages reported Not available (Smith et al., 2026) Twitter, Lise Winer’s Dictionary of English Creole of Trinidad and Tobago Collection of Trinidadian Creole terms from social media and lexicographic sources → dictionary-based parsing and replacement of Creole forms with Standard English equivalents using the Caribe Python library → T5-KES transformer-based grammatical correction of the converted text Not conducted with manual coding as benchmark. Automated evaluation and examples are reported, but no systematic comparison with an independently manually coded sample is provided https://pypi.org/project/Caribe/ References Alqazlan, L., Fang, Z., Castelle, M., & Procter, R. (2025). A novel, human-in-the-loop computational grounded theory framework for big social data. Big Data & Society, 12(2), 20539517251347598. Bengsch, G. (forthcoming). Platformed identity and methodological adaptation: Analysing Trinidadian POV culture in the age of AI. In N. Mcleod & A. Obika (Eds.), Repurposing Media and Communication Skills in the Caribbean. Bommarito, I., & Michael, J. (2025). OpenGloss: A Synthetic Encyclopedic Dictionary and Semantic Knowledge Graph. arXiv preprint arXiv:2511.18622. Eckert, P. (2008). Variation and the indexical field 1. Journal of sociolinguistics, 12(4), 453–476. Hong, M. K., Chen, F., Chen, Y.-Y., & Klenk, M. (2022). Supporting qualitative coding with machine-in-the-loop. Proceedings of the 36th Conference on Neural Information Processing Systems (NeurIPS 2022), Jackson, S., Beekhuizen, B., Zhao, Z., & McEwen, R. (2025). GPT-4-Trinis: assessing GPT-4’s communicative competence in the English-speaking majority world