PaperlayerView exact revision on arXiv

Source-linked AI summary

Computational Sociolinguistics: A Survey

Dong Nguyen, A. Seza Doğruöz, Carolyn P. Rosé, Franciska de Jong

arXiv:1508.07544v2cs.CL

TL;DR

Computational sociolinguistics addresses the limited computational treatment of language’s social dimension and the challenges posed by fluid social variables and linguistic variation. This survey synthesizes computational-linguistic research on sociolinguistic themes and examines how computational methods and sociolinguistic perspectives can complement and challenge one another. It concludes that closer synergy between the communities is possible, while multilingual and mixed-language data remain important open challenges.

  • Problem

    Computational linguistics has primarily modeled language’s informational dimension, while social variables are fluid and social-media language contains substantial colloquial and linguistic variation.

  • Method

    The paper surveys computational-linguistic studies of sociolinguistic themes and uses triangulation as a methodological approach for strengthening social-science findings through multiple views.

  • Results

    The survey shows that large-scale data-driven computational methods can complement sociolinguistic studies, while sociolinguistics can inform and challenge computational methods and assumptions.

  • Takeaways & Limitations

    Additional alignment of research agendas and methodological frameworks is needed for genuine synergy between computational linguistics and sociolinguistics.

  • Takeaways & Limitations

    Computational sociolinguistics has focused disproportionately on English and monolingual data, while many NLP tools are not optimized for mixed-language texts.

Abstract

from arXiv · show

Language is a social phenomenon and variation is inherent to its social nature. Recently, there has been a surge of interest within the computational linguistics (CL) community in the social dimension of language. In this article we present a survey of the emerging field of "Computational Sociolinguistics" that reflects this increased interest. We aim to provide a comprehensive overview of CL research on sociolinguistic themes, featuring topics such as the relation between language and social identity, language use in social interaction and multilingual communication. Moreover, we demonstrate the potential for synergy between the research communities involved, by showing how the large-scale data-driven methods that are widely used in CL can complement existing sociolinguistic studies, and how sociolinguistics can inform and challenge the methods and assumptions employed in CL studies. We hope to convey the possible benefits of a closer collaboration between the two communities and conclude with a discussion of open challenges.

1. Introduction

Computational Sociolinguistics examines language’s social dimension by combining sociolinguistic perspectives with computational, data-driven methods. The survey maps this emerging multidisciplinary field, its applications, and the challenges of working with socially variable language data.

  • CL has historically emphasized informational content and language structure, while social media has expanded interest in language’s social dimension.Social media also introduces colloquial language and greater variation than commonly used CL datasets.
  • Because linguistic choices signal identity and support speakers’ social goals, social variables remain fluid and language variation reflects speaker agency.Agency can produce creative departures from conventions that communicate indirect meanings and may contribute to language change.
  • Sociolinguistics traditionally combines qualitative and quantitative analysis with surveys and ethnographic research, often using relatively small controlled datasets.Massive, messy digital data may require computational methods rather than conventional sampling and analysis alone.
  • The survey reviews methods and research on social identity, social interaction, multilingual communication, and computational applications of socially informed language analysis.Its stated scope is selective, primarily determined by computational-linguistics research addressing sociolinguistic topics.
  • Computational Sociolinguistics integrates sociolinguistics and computer science to study the relation between language and society computationally.
  • Accounting for social variation has potential application value, including user profiling and improvements to word prediction, cyberbullying detection, and sentiment analysis.The text notes that the variables most useful for interpretation can also be difficult to detect automatically.

2. Methods for Computational Sociolinguistics

Computational sociolinguistics brings computational modeling, sociolinguistic validation, and social-science methods into dialogue. Its methods must address differences in research values, speaker agency, and the limitations of social-media data.

  • Collaboration: The field aims to stimulate collaboration so sociolinguistics can enrich computational models while computational approaches refine sociolinguistic theory.The article identifies reciprocal contributions rather than a one-way transfer of methods.
  • Validation of Modeling Approaches: Computational linguistics emphasizes creative model development and predictive accuracy, whereas social sciences emphasize conceptual validity and reliability.Reliability concerns reproducibility, while validity concerns isolating issues from confounds and capturing the intended construct.
  • Validation of Modeling Approaches: Triangulation strengthens findings by examining the same object through multiple analytical lenses designed to highlight different details.The article presents triangulation as analogous to ensemble methods such as bagging.
  • Theory versus Empiricism: Computational models should integrate existing theories and use their results to refine or develop new theories, rather than focus only on task performance.The article contrasts this approach with atheoretical empiricism and calls for attention to qualitative social-science concerns.
  • Quantitative versus Qualitative Approaches: Many computational sociolinguistic studies emphasize social structures, but this can overlook speakers’ agency in using language to construct identity.The article notes that some computational work has begun acknowledging agency when interpreting findings.
  • Spotlight on Corpora and Other Data: Social-media data can complement traditional sociolinguistic data, but sampling, demographic uncertainty, and nonrepresentative users complicate interpretation.API access limitations can introduce sampling bias and remove contextual information, especially for secondary analyses.

3. Language and Social Identity

Computational sociolinguistic research examines how language varies with social identity and how text can reveal social variables. It also emphasizes that these variables interact and that automatic prediction remains difficult.

  • Language and Social Identity Construction: Speakers use language to construct and present social identities through choices among languages, dialects, and styles.This identity work involves creating, presenting, and sustaining identities, within constraints on achievable variation.
  • Automatic Social-Variable Inference: Computational studies increasingly infer social variables such as gender, age, location, ethnicity, and social class from text.The task is treated as automatic metadata detection, but labels for some variables are difficult to obtain.
  • Automatic Social-Variable Inference: Automatic age and gender prediction from text alone is nontrivial, although systems have outperformed humans in comparative studies.A crowdsourced system still incorrectly predicted gender for 16% of Twitter users.
  • Interacting Social Variables: Gender, age, and other social variables are interdependent, so failing to control for them can misinterpret language-variation findings.The interaction is complicated further when demographic distributions differ across age ranges in uncontrolled data.
  • Section Scope: The survey organizes computational work on social identity around datasets, gender, age, location, and related NLP tasks.These topics structure the section’s overview of computational studies of language variation.

3.1 Data Sources

Data sources have shifted from formal or controlled speech and writing toward informal social-media text. Building usable datasets requires explicit labeling strategies and sample restrictions that shape who is included.

  • Data Sources: Early studies used formal corpora and controlled conversations, whereas social-media research increasingly uses informal text from uncontrolled settings.Blogs and later Twitter became important sources for studying gender and age variation.
  • Dataset Construction: Dataset construction involves two recurring tasks: labeling users’ social variables and selecting an appropriate sample.These choices are central to studying relations between social variables and language use.
  • Labeling: Researchers obtain labels from user-provided profile information, explicit textual patterns, names, or other external demographic sources.User-provided information can be accurate but is often available for only a small fraction of users; name-based labels can misclassify atypical names.
  • Sample Selection: Sample restrictions often follow the labeling strategy, including selecting users with stated demographics, gender-specific names, or geotagged tweets.Such restrictions determine which people and messages enter the dataset.

3.2 Gender

Computational gender research has concentrated on text classification, but sociolinguistic evidence challenges fixed binary representations of gender. Findings also show that features, genre, interaction, and social networks affect observed patterns and generalization.

  • Modeling Gender: Computational linguistics has primarily modeled gender as a binary biological attribute using supervised text-classification methods.This setup treats gender as fixed and neglects speaker agency, contrasting with sociolinguistic views of gender as socially constructed.
  • Features and Patterns: Studies have emphasized predictive features, including lexical, character-level, grammatical, stylistic, and attribute-based signals.Character-level language models were reported as more robust than token-level models, while LIWC-only models generally underperformed unigram and n-gram models.
  • Features and Patterns: Gender-associated linguistic features can change after controlling for genre or occupation, showing that contextual variables affect observed classification patterns.In one blog-corpus study, previously predictive features were no longer effective after controlling for occupation.
  • Features and Patterns: Models generally perform worse across datasets but remain above random, and stretchy patterns were proposed to improve cross-domain generalizability.Stretchy patterns represent flexible sequences of categories for modeling stylistic variation.
  • Social Interaction: Interactional context can improve gender classification in conversations, whereas Twitter social-network features have generally offered little benefit when sufficient text is available.Gender-specific language was more frequent in same-gender conversations, and partner-dependent classifiers improved performance on telephone corpora.
  • Social Interaction: Gender composition and interactional features also support analyses beyond classification, including power, language change, and interactional style.For interactional-style classification, incorporating features from both speakers improved performance.
  • Interpretation of Findings: Network correlations, conflicting linguistic markers, and human judgments indicate that individual users may diverge from gender-stereotypical language patterns.The survey also cautions that gender is shaped differently across cultures and languages and that within-gender variation persists.

3.3 Age

Computational sociolinguistic research models age through categorical and continuous representations, while examining linguistic features associated with age. Studies also distinguish life-stage interpretations from chronological age and caution that cross-sectional data cannot separate aging from generational change.

  • Modeling Age: Age prediction commonly models chronological age categorically, although researchers use different boundaries and groupings.Researchers distinguish chronological, biological, and social age, but most computational studies focus on chronological age.
  • Modeling Age: Higher performance is obtained when classification uses larger age gaps, such as teens versus people in their 40s or 50s.Experiments found larger separations easier than smaller gaps, such as teens versus people in their 20s or 30s.
  • Modeling Age: Age categories complicate comparisons because meaningful boundaries are difficult to select and vary with dataset age distributions.Continuous regression removes the need to define age categories; latent-variable models can jointly model age and topics.
  • Features and Patterns: Unigrams, part-of-speech features, pronouns, prepositions, determiners, and articles have been used to predict age, with some findings potentially language-specific.Younger people tend to use more singular pronouns, while older people use more plural pronouns, prepositions, determiners, and articles.
  • Features and Patterns: Younger users tend to use more non-standard forms, including lengthening, contractions, Internet acronyms, slang, swear words, capitalization, and stance markers.Words such as “like” are also associated with younger ages.
  • Features and Patterns: Cross-sectional age prediction data cannot distinguish age grading from generational change in observed language differences.A single-time-point dataset may show either individuals changing across life or cohorts differing from one another.

3.4 Location

Computational research on regional variation uses dialect-labeled or location-tagged data to identify dialects, regions, and geographically associated linguistic features. These approaches support large-scale analysis, but dialect boundaries and geographic labels remain difficult to define.

  • Approaches: Dialectometry aggregates linguistic variables to examine linguistic differences between regions, complementing dialectology’s focus on individual variables.Computational linguistics, machine learning, and spatial analysis have influenced this work.
  • Approaches: Regional variation can be studied through supervised dialect identification or unsupervised discovery from location-tagged data.The latter may also be framed as location prediction when locations, rather than dialect labels, are the supervised targets.
  • Features and Patterns: Location prediction and dialect identification frequently use word and character n-grams or unigram word features.Sociolinguistically inspired features have been proposed as potential performance improvements.
  • Features and Patterns: Location-tagged Twitter data has been used to analyze dialect variants and compare computational estimates with traditional sociolinguistic collection methods.Doyle analyzed variants such as double modals using baseline queries to estimate conditional distributions given metadata.
  • Features and Patterns: Linguistic innovations in the United States spread geographically through demographically similar areas, particularly with regard to race.This finding comes from longitudinal analysis of tweets.
  • Limitations: Dialect labeling presumes clear boundaries, but language varieties often form continua and their labels can be influenced by political boundaries.The distinction among languages, dialects, minority languages, and regional varieties remains debated.

3.5 Text Classification Informed by Identity Information

Identity-related linguistic variation can inform NLP tasks beyond predicting speaker attributes. Studies incorporate gender- and age-related language patterns into classification models for cyberbullying, power, subjectivity, polarity, sentiment, and topic classification.

  • Text Classification Informed by Identity Information: Gender- or age-specific linguistic features improve several NLP tasks, whereas treating gender as a binary feature alone is not always effective.Reported applications include cyberbullying detection, power-direction prediction, subjectivity and polarity classification, sentiment analysis, and topic classification.

4. Language and Social Interaction

Computational sociolinguistics studies how language reflects and shapes social interaction in pairs, groups, and communities. Large-scale online data support analyses of relationships, power, audience design, accommodation, and language change, while data bias remains an important constraint.

  • Data sources: Online forums, discussion communities, and organizational email corpora provide large-scale interaction data for studying language in social contexts.Examples include cancer-support communities, review sites, and the Enron email corpus, whose organizational structure can support analyses of workplace relationships and power.
  • Relationships and power: Computational studies use language to analyze social relationships and power, including distinctions between weak and strong ties and relationships embedded in organizational hierarchies.The literature also examines power through Wikipedia Talk Pages, political debates, and Supreme Court arguments.
  • Relationships and power: Lower-status participants accommodated more than higher-status participants in studies of Wikipedia Talk pages and Supreme Court debates.Accommodation measures how speakers’ language aligns with conversational partners and has been used to study power dynamics.
  • Audience design: Audience characteristics influence language choice, sarcasm detection, hashtag use, and the use of non-standard lexical variables on Twitter.These studies connect linguistic choices with audience familiarity, user mentions, and minority-language use.
  • Language change: Community members may first align with established language norms and later contribute to language change, while communities themselves evolve through word innovations and changing usage.Studies track these dynamics through community detection, longitudinal language models, and cross-entropy comparisons.
  • Caveats: Interpretations of computational analyses of language change require careful attention to data bias and incomplete understanding of the data’s dynamics and content.The survey notes that corpus composition, including a bias toward scientific publications, can challenge conclusions about linguistic change.

5. Multilingualism and Social Interaction

This section surveys multilingual communication, including language mixing, computational processing of mixed-language texts, and analyses of language choice and switching. It also highlights the shift from small, manually analyzed data toward large-scale online multilingual data and specialized NLP tools.

  • Multilingual speakers may mix languages or dialects depending on contextual factors such as occasion, while language contact shapes how languages influence one another.
  • Language mixing ranges from occasional word-level switches to structural morphological or syntactic borrowings, with interaction influencing the types of switches.
  • Online multilingual populations provide opportunities for large-scale analysis, although much prior work remains qualitative and online data collection has often been limited in volume and duration.
  • Processing mixed-language text requires tools beyond standard monolingual NLP systems, including word-level language identification and language-specific processing methods.
  • Research has developed approaches for parsing, language modeling, and topic modeling of mixed-language texts, including systems validated on large real-world data.
  • Computational studies analyze language choice through users’ social networks and predict language switches using lexical, syntactic, multi-word-unit, and emoticon features.

6. Research Agenda

The research agenda calls for broader empirical scope, stronger compatibility between sociolinguistic and computational methods, and NLP tools that preserve and represent language variation. It emphasizes multilingual, multimodal, cross-domain, interpretable, and socially informed research.

  • Expanding the Scope of Inquiry: Computational sociolinguistics should expand beyond lexical and stylistic variation to examine a wider range of linguistic and social phenomena.
  • Expanding the Scope of Inquiry: Research should include underrepresented social variables such as social class because richer labels support deeper understanding of variation and more robust analyses.
  • Expanding the Scope of Inquiry: Studies should move beyond English and monolingual data, since multilingual social-media texts can contain multiple languages within a syntactic unit.
  • Expanding the Scope of Inquiry: Multimodal research can combine verbal and nonverbal dimensions of communication through video and audio recordings with speech tracks.
  • Adapting Methodological Frameworks to Increase Compatibility: Effective collaboration requires computational models to engage with sociolinguistic theory and prioritize interpretability rather than treating theory and empiricism as opposing ends.
  • Adapting Methodological Frameworks to Increase Compatibility: Hierarchical or structured modeling can control multiple sources of variation and reduce misattribution to extraneous variables.
  • Adapting Methodological Frameworks to Increase Compatibility: Cross-domain models help distinguish features that behave differently across contexts and improve interpretation of findings.
  • Tuning NLP Tools to Requirements of Sociolinguistics Research: NLP preprocessing should handle informal variation without normalizing away the variation central to sociolinguistic analysis or biasing performance toward particular social groups.

7. Conclusion

The conclusion frames computational sociolinguistics as an emerging field connecting computational linguistics and sociolinguistics. The survey reviews existing work while identifying how each community can complement and challenge the other.

  • Language should be studied not only for propositional content but also as a dynamic social entity involving linguistic variation and agency.
  • The survey defines a research agenda for Computational Sociolinguistics and reviews CL studies addressing sociolinguistic themes.
  • Large-scale data-driven CL methods can complement sociolinguistic studies, while sociolinguistics can inform and challenge CL methods and assumptions.

2012. Detecting influencers in written

The supplied passages are bibliography entries for prior work, including studies of computer-mediated sociolinguistic data and latent Dirichlet allocation.

  • The references include proceedings for a 2012 workshop on Language in Social Media.
  • One cited work presents latent Dirichlet allocation in the Journal of Machine Learning Research.
  • One cited work reviews methodological issues in sociolinguistic research using computer-mediated data.

2005. A quantitative analysis of lexical

The cited work spans computational analyses of language variation involving gender, identity, interaction, multilingual communication, and text classification.

  • Gender and lexical variation: Research examines gender variation in telephone conversations, email discourse, Twitter, weblogs, and conversational speech.The references include studies of gender differences, gender inference, gender attribution, and gender-genre variation across these settings.
  • Identity: Several studies connect language use with identity, including discourse identity, identity work, and gender as a sociocultural construct.The cited works address discourse and identity, verbal identity construction, and gender trouble.
  • Multilingual communication: Multilingual communication is represented through code-switching, Arabic-French corpora, Dutch-Turkish contact, and multilingual Internet research.The references include code-switched language identification, part-of-speech tagging, code-switching prediction, and contact-induced change.
  • Interaction: Computational approaches cover social interaction through politeness, style accommodation, power, and social relationships.References address linguistic accommodation, coordination in dialogue, power differences, politeness, and relationship identification.

2012. Geolocation prediction in social

The cited work includes computational studies of geolocation, subgroup detection, dialectology, online communities, and demographic variation in social-media language.

  • Online social structure: Studies of online interaction examine subgroups, communication networks, collaborative learning, and leadership or social influence.The references include methods for detecting discussion subgroups, exploring communication networks, and analyzing collaborative processes.
  • Demographic variation: Social-media and review-site research links demographic factors such as age, gender, and prestige with linguistic behavior and classification performance.The cited studies address age and gender variation, linguistic adaptation, demographic effects on classification, and review sites as sociolinguistic resources.
  • Geolocation and dialects: Dialect research addresses dialect areas, dialect continua, dialectometry, dialect identification, and challenges in processing social-media dialects.The cited works span theoretical dialectology and computational dialect processing.

2012. Multilevel and Longitudinal Modeling

The cited work covers statistical modeling, authorship and demographic profiling, language variation, interaction, and multilingual processing across varied datasets and tasks.

  • Modeling: Statistical and probabilistic methods include Stata, GLLAMM, approximate inference, hierarchical Bayesian models, and latent attribute detection.These references represent computational tools and models for inference involving latent variables and user attributes.
  • Authorship and profiling: Authorship and profiling research applies probabilistic grammars, idiolects, sociolects, and author-profiling tasks to text analysis.The cited work includes authorship attribution using probabilistic context-free grammars and profiling based on demographic characteristics.
  • Longitudinal and demographic variation: Longitudinal and demographic studies examine language change across online communities and variation associated with age and gender.The references include user lifecycle and linguistic change, age prediction, age and gender lexica, and age or gender effects on blogging.
  • Interaction and multilingual processing: Interaction and multilingual studies address similarity, accommodation, communication, code-switching, language identification, and part-of-speech tagging.The cited work spans computer-mediated communication, speech-style accommodation, code-switching prediction, shared tasks, and tagging.
Loading 1508.07544v2…