BRIDGE Workshop on Computational Sociolinguistics

In this presentation we report on recent and ongoing work combining historical sociolinguistics and Construction Grammar to study linguistic productivity in English. By discussing specific case studies ranging from the ADJ-ity and V-ment constructions to BE going to V and the way-construction, we show that both morphological and syntactic productivity exhibit similar sociolinguistic variation and that factors such as gender are often but not always significant. Using novel methods, we demonstrate that type frequencies alone do not tell the whole story of change in constructional productivity but that both intra- and extralinguistic factors need to be accounted for. In particular, we analyse gender variation in the semantic expansion of constructions by drawing on techniques from distributional semantics. We find that while men often use constructions more productively than women at the level of type frequencies (e.g. the number of different verbs used in the BE going to V construction), women may lead their semantic expansion (e.g. from concrete to abstract senses). We also explore AI methods for semantic analysis and the enrichment of corpus metadata with sociolinguistic information.

This talk presents a Computational Construction Grammar analysis of register and regional variation across large English corpora, covering two digital registers (tweets and YouTube transcripts) and five inner-circle varieties. We show that both sources of variation produce systematic, largely independent patterns in the constructional network, but differ in where they manifest: register effects are more pervasive and concentrated in abstract, high-level constructions, while regional effects are sparser and surface in lower-level constructions. We account for this asymmetry in terms of a continuum of constructional salience, and discuss implications for integrating sociolinguistic theory into Construction Grammar.

Computational models can detect fine-grained acoustic differences between accents, yet human listeners do not necessarily use the same cues (Strycharczuk et al. 2020). This paper investigates why machine and human accent classification diverge, focusing on two possibilities: that listeners attend to different features from those that are acoustically most diagnostic, and that they have limited access to the acoustic feature space.

Using the Salient Language in Context (SLIC) approach (Montgomery, Walker & Woods 2026), real-time listener attention was captured through time-aligned clicks and accompanying commentaries. Ninety-eight listeners from two regions of England responded to five Northern English accent samples (reported elsewhere in Montgomery, Vriesendorp & Walker 2025). Analyses address four questions: whether attending to more cues improves classification accuracy; whether cue accessibility differs across samples; whether listener-identified cues can classify accents computationally; and which cues drive classification.

Results show that greater attention does not improve accuracy. However, some accents elicit more perceptually accessible cues than others. Crucially, listener-identified cues contain structured regional information and can successfully classify accents in a computational model. At the same time, the ability to articulate a cue predicts successful regional placement.

These findings support a model in which regional distinctiveness operates across three layers: acoustic structure, cue accessibility, and social or regional mapping (cf. Preston 2017). Human perception is constrained at the level of access, helping to explain the divergence between human and machine classification. The paper argues that the key bottleneck in accent recognition is not the signal itself, but listeners’ access to it, with implications for how accents are evaluated and socially interpreted (Preston 2013).

References

Montgomery, Chris, Hielke Vriesendorp & Gareth Walker. 2025. Feature attention and accent recognition: human listeners’ responses to five Northern English accents. Frontiers in Psychology. Frontiers 16. https://doi.org/10.3389/fpsyg.2025.1613018.

Montgomery, Chris, Gareth Walker & Harry Woods. 2026. Salient Language in Context (SLIC): a web app for collecting real-time attention data in response to audio samples. Linguistics Vanguard. De Gruyter Mouton 11(1). 397–406. https://doi.org/10.1515/lingvan-2025-0028.

Preston, Dennis R. 2013. Language with an Attitude. In J. K. Chambers & Natalie Schilling (eds.), The Handbook of Language Variation and Change, 157–182. 2nd edn. Chichester, United States: John Wiley & Sons, Incorporated. (24 September, 2021).

Preston, Dennis R. 2017. The cognitive foundations of language regard. Poznan Studies in Contemporary Linguistics; Berlin. Berlin, Germany, Berlin: Walter de Gruyter GmbH 53(1). 17–42. http://dx.doi.org.sheffield.idm.oclc.org/10.1515/psicl-2017-0002.

Strycharczuk, Patrycja, Manuel López-Ibáñez, Georgina Brown & Adrian Leemann. 2020. General Northern English. Exploring Regional Variation in the North of England With Machine Learning. Frontiers in Artificial Intelligence 3. 48. https://doi.org/10.3389/frai.2020.00048.

Donald Trump’s lifelong navigation of national celebrity in the United States presents an opportunity to bring together research strands exploring (a) lifespan language change and (b) politicians’ selection of variants to present public identities. Trump’s home dialect of New York City English (NYCE) provides indexicalities for Trump to draw upon for image curation--for instance to highlight “New Yorker” identity or project qualities of toughness. But NYCE may also be sociolinguistically marked as incongruous with attainment of high social status, and traditional NYCE features are generally receding among New Yorkers--both factors that could encourage shift away from NYCE. Using a corpus of public speech events from 1980 to 2023, compiled and anaylsed using emergent computational technologies, this project identifies areas of stability and change in Trump's NYCE vowels. I explore implications of these findings for lifespan language change, and for the study of sociophonetic data via new computational methods.

Data annotation remains a significant bottleneck in discourse analysis, particularly for complex, context-sensitive tasks such as the identification of speech acts, metaphors, and stance. In this talk, we present a series of experiments designed to evaluate the capabilities of large language models (LLMs) for automating discourse annotation, with a focus on metaphor. We test a range of LLMs across three methodological approaches to LLM-assisted metaphor identification: prompt engineering, Retrieval-Augmented Generation (RAG), and fine-tuning. Our findings show that state-of-the-art closed-source models can achieve high levels of accuracy, with fine-tuning yielding a median F1 score of 0.79. A comparison of human and model outputs further reveals that most disagreements are systematic, reflecting well-known grey areas and enduring conceptual challenges in metaphor theory.

We complement these empirical findings with a showcase of LinguistAgent, a no-code, user-friendly platform that leverages a reflective multi-model architecture to automate linguistic data annotation. The system implements a dual-agent workflow which comprises an Annotator and an optional Reviewer to simulate a professional peer-review process. The platform processes input texts in real time via the LLM's API and provides real-time token-level evaluation (averaged F1 score) against human gold standards. LinguistAgent supports comparative experiments across the same three methodological approaches employed in our metaphor study.

I present Authorial Language Models (ALMs), a method for authorship attribution in which a base language model is further pretrained on each candidate author's known writings and attribution is decided by perplexity. Central to ALMs is the analysis of comparative negative log-likelihood (CNLL): by tracking how each author model's negative log-likelihood accumulates word by word through a questioned document, the attribution becomes fully transparent, revealing the specific lexical and syntactic choices that distinguish one author from another. This analysis shows that content word classes carry a higher density of authorship information than function word classes—challenging a long-standing assumption in stylometry. I discuss the implications for computational sociolinguistics, arguing that ALMs offer a powerful, interpretable lens for studying idiolectal variation in text. Time permitting, I will also present two variants—reverseALMs (fine-tuned on the questioned text) and distillALMs (fine-tuned on both questioned and candidate texts)—extending the framework to broader forensic and sociolinguistic casework scenarios.

This talk explores whether Large Language Models (LLMs) can identify the English variety used by the author of a text. Traditional profiling methods rely either on linguists’ intuition supported by reference sources or on computational/stylometric methods, each with its own limitations. The use of LLMs for this task offers a third alternative that combines the strength of both methods while mitigating their weaknesses. By testing LLMs with simple one-shot prompts, the talk presents preliminary evaluations of their effectiveness in authorship profiling and the extent of their sociolinguistic knowledge of English varieties.

This talk presents a computational sociolinguistic approach to forensic authorship profiling that focuses on inferring an author’s regional background through systematic analysis of linguistic variation. Using a large corpus of geolocated German social media posts, it moves beyond analyst-selected regionalisms in profiling and models regional language use at scale. Geospatial statistical methods are applied to address data sparsity and improve spatial coverage, which enables the development of an algorithm for predicting a text's most likely origin. The results demonstrate how computational sociolinguistic can enhance the accuracy, scalability, and explainability of regional profiling in forensic contexts.

Although representation learning has transformed semantic modeling in NLP, learned representations of linguistic style remain underexplored — partly due to conflicting definitions of style across NLP and sociolinguistics, and partly due to unclear advantages of separate style representations. This talk presents current approaches to training, evaluating, and applying neural style representations, shows how they can benefit both NLP and sociolinguistic research, and critically examines open challenges.

Large language models are routinely characterised either as stochastic parrots incapable of genuine generalisation or as emergent reasoners approaching general intelligence. In this talk I will argue that both positions are empirically inadequate and propose a unifying alternative: context-directed extrapolation, under which LLMs extrapolate from statistical priors in their training data, with prompt context directing which priors are activated.

I will substantiate this account through evidence from controlled studies of model capabilities (Lu et al., ACL 2024), and demonstrate that construction grammar provides a strong evaluation framework for mapping where extrapolation succeeds and where it characteristically fails (Scivetti et al., 2025). I will then show that the same mechanism carries direct social consequences: because LLMs extrapolate from whatever associative pathways are statistically present in training data, deployment in socially sensitive contexts introduces risks that current safety mechanisms are not designed to address. I will present experimental evidence to demonstrate that conversations on superficially benign topics can expose users to potentially radicalising content (Smith et al., under review).

Context-directed extrapolation thus connects linguistic generalisation, model evaluation, and social consequence within a single coherent framework, with implications for how LLMs should be evaluated, deployed, and regulated.

References:

Tayyar Madabushi, H., Torgbi, M., & Bonial, C. (2025). Neither stochastic parroting nor AGI: LLMs solve tasks through context-directed extrapolation from training data priors. arXiv. https://arxiv.org/abs/2505.23323

Lu, S., Bigoulaeva, I., Sachdeva, R., Tayyar Madabushi, H., & Gurevych, I. (2024). Are emergent abilities in large language models just in-context learning? In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vol. 1, pp. 5098–5139). Association for Computational Linguistics. https://aclanthology.org/2024.acl-long.279/

Smith, L. G. E., Bocheva, D., Brown, O., Lowery, C., Tarpy, H., Torgbi, M., Worajitwannakul, S., & Tayyar Madabushi, H. (under review). Large language models may promote radicalization via exposure to extremist content and politicized communities. PsyArXiv. https://osf.io/preprints/psyarxiv/my3gw_v1

There is an urgent need for shareable, validated, representative corpus data organized by language and register and population. We need this data both (i) to support replicable studies in computational sociolinguistics but also (ii) to support socially-reponsible model training and evaluation in the larger field of computational linguistics. For example, researchers often talk about concepts like out-of-domain performance without ever defining what a domain is or how different two domains are. My goal in this talk is to outline the research challenges, from language and code-switching identification to outlier detection, and move towards a shared framework.

Quantitative research in corpus linguistics is usually based on the relative frequencies of linguistics types, including words, parts of speech, and constructions. For example, stylometry often analyses the relative frequency of function words, while register studies often analyse the relative frequencies of grammatical categories, where the relative frequency of a given type is computed by counting the total number of tokens of that type in the corpus and dividing this value by the total number of tokens of any type in the corpus. Language modelling, however, provides a new basis for quantitative corpus linguistics. Rather than measuring the frequency of each word type, we can measure the unpredictability of each word token in the corpus based on the token-level negative-log-likelihood values returned by a language model. Furthermore, we can also represent each type as a distribution of negative-log-likelihood values. In this talk, I introduce this new paradigm for quantitative corpus linguistics with examples drawn from authorship analysis.