Veranstaltungsarchiv

This paper discusses a case of conversational agent use in a telephone inquiry service. Voice agent was implemented to answer customers’ calls and provide them with information on various government and communal services. Recently, there have been new interest in understanding social interaction involving voice user interfaces (Reeves 2017; Porcheron et al. 2018; Kirschthaler et al. 2020). Smart speakers or smart assistants (Hector, Hrncal 2020; Englert, Hoffmann, Waldecker 2022) are typical artificial agents people encounter in their everyday life which come under researchers’ notice. In addition to that, the examined case presents a technology implemented “in the wild” in such a way that people rather are obliged to use it, than voluntarily choose to do so. Consequently, a wide range of people outside of early adopters and technology enthusiasts interact with this automated call operator.

The paper employs conversation analysis (CA) to investigate interactional problems and communicative resources (Suchman 2007) in phone calls involving robot operator. To elucidate this issue, I have analysed audio recordings of over 100 phone calls.

Analysis shows that a telephone call involving conversational agent is a non-linear, dynamic process, not determined by the pre-given characteristics of its participants. People try different techniques and test their assumptions, track machine’s responses to their actions and abandon previously planned lines of action as something goes wrong. The properties of the participants are not just “manifested”, but actively produced in the interaction — they are the result of the joint work of a conversational agent and a human. In the course of interaction, user’s ideas about what the machine’s abilities and limitations are, are constantly developed and changed. Empirical data will be discussed to illustrate situated resources that people employ in order to make sense of interaction and establish a relevant “partner model” (Deppermann, Schmidt 2016).

Studying the interaction of people with the robot Nao, Pelican and Broth argue that despite the difficulties in communicating with the robot, people quickly adapt to their counterpart with its limited capabilities, adjusting their actions and speech to how the technology works (Pelikan, Broth 2016). Their study demonstrates that people are not only skilled interaction participants, but they are also able to quickly become skilled users, competent human–robot interaction participants. In the paper we describe some techniques that make interaction between a person and a conversational agent possible and help overcome barriers and troubles. In its working, the machine is devoid of nuances and is not able to orient to the details of situation in the way people do. The conversational agent has only a set of pre-defined actions and semantic relations available, and it cannot recognize many of the things people express in conversation. Its limited scripts and their outcomes – requests, pauses and repetitions – act as mechanisms that control the conversation. Despite conversational agent’s inflexibility, users gradually reveal its important limitations and capabilities in the interaction and adapt to them.

About ten years ago, smart speakers were seen as bringing AI, Big Data and smart technologies into the domestic sphere (Chacksfield/Pino 2019). While smartphones have become the digital device used in everyday life – and for certain, poorer strata of society, often the only one –, smart speakers, albeit heavy marketing and rave reviews after introduction, have not become as popular. About a third of German households (Statista 2022) and US inhabitants (NPM 2022) have smart speakers in their homes. While the iPhone is still one of the money-makers for Apple, Amazon’s voice and smart speaker branch seems to be losing money (Kim 2022); the commercial drive to turn smart speakers into the smartphone of the home and automated voice recognition and language processing into the next go-to domestic digital interface has been not too successful.

While a number of factors need to be considered, we argue that the low level of interest is has to do with the (counter-intuitive) media practices that are necessary for users to get smart speakers to work. We try to explain the relative non-success of smart speakers based on our multi-method research on smart-speaker use combining sequence analysis, interviews, digital methods and historical research into the history of voice-driven machine-learning applications and its analogue predecessors. In this way, we add a broad and multi-level analysis to the growing literature on smart-speaker use.

Our research suggests that smart speakers do not live up to the advertised promises of comfort and assistance; instead, they often need additional assistance by users to complete even simple tasks. In this sense, users have to adopt linguistic and media practices to a media technology that is based on so-called “Natural Language Processing”. Users have been acquainted to voice-related technologies in call centres and hotlines, to talking machines and to talking to them; this might explain the fast uptake of these devices; their limited use however might have to do with the fact that using a smart speaker has a lot in common with calling an automated hotline. In this sense, they are often used as basis for a smart-home infrastructure and less as an everyday assistant or companion.

In sum, our paper places the advent and limited commercial success of smart-speaker and voice-interface technology in the context of the relative uselessness of these services and the drive by platform companies such as Amazon and Alphabet to train their speech-related algorithms on a free and big data base (Turow 2021).

Chacksfield, Marc, and Nick Pino. 2019. “Amazon Echo (2014) Review”. TechRadar (19 November 2019). URL: https://www.techradar.com/reviews/amazon-echo-2014.
Kim, Eugene. 2022. “Amazon Is Gutting Its Voice Assistant, Alexa. Employees Describe a Division in Crisis and Huge Losses on ‘a Wasted Opportunity’”. Business Insider. Accessed 22 November 2022. URL: https://www.businessinsider.com/amazon-alexa-job-layoffs-rise-and-fall-2022-11.
NPM. “The Smart Audio Report.” 2022. National Public Media. URL: https://www.nationalpublicmedia.com/insights/reports/smart-audio-report/.
Statista. 2022. “Statista-Dossier zu Smartphones”. 3179. URL: https://de.statista.com/statistik/studie/id/3179/dokument/smartphones-statista-dossier/.
Turow, Joseph. 2021. The Voice Catchers. New Haven, CT: Yale University Press. DOI: 10.12987/9780300258738.

As interactive language technologies increasingly become part of our everyday lives, one of the biggest frustrations remains their rigidity: they are designed to avoid turbulence by funneling people into pre-set dialogue flows, cannot gracefully repair breakdowns, and are devoid of social accountability. This has people adapting to tools rather than the other way around. Here we critically assess notions of language and technology that lie at the base of current developments, and propose a redirection inspired by Ivan Illich’s (1973) notion of convivial tools. A general challenge in this area is that while critical and radical deconstruction is necessary, the search is on for constructive ways of applying cumulative insights from decades of empirical work on human interaction. In this contribution, we aim to address this challenge in two ways.

First, we expose the narrow linguistic and cultural roots of much of today’s NLP (Joshi et al. 2020) and contrast it with the wealth of data and knowledge available in cross-linguistic corpora of everyday informal conversation (Dingemanse and Liesenfeld 2022). Looking at everyday language use can bring to light unseen diversity in semiotic practices, and can uncover ways in which interactional resources can serve human empowerment and inspire the design of language technology (Liesenfeld and Buschmeier 2023). Second, we build on new empirical work as well as prior critical and ethnomethodological work into social interaction (McIlvenny 1993; Button et al. 1995) to develop constructive ways of evaluating artificial conversational agents. Current evaluation methods in NLP/AI focus on the putative “humanness” and “fluidity” of “language generation” (Finch and Choi 2020) — all of these unexamined notions in need of technolinguistic scrutiny. Using insights from conversation analysis, we instead make a case for an action-oriented approach to evaluation (cf. Housley, Albert, and Stokoe 2019). We report on a set of heuristics for human evaluation that can help to systematically probe the interactive capabilities of dialogue systems. Such ‘counterfoil research’ (Illich 1973) is of critical importance to arrive at tools that better support human flourishing. This work provides a stepping stone for engineers and conversation designers who seek to ground their work in observable orientations to social action, rather than automated metrics and scoreboards divorced from interactional practices.

Combining critical and practical perspectives will contribute to respecifying taken-for-granted notions of language and technology, will deepen our understanding of the irreducibly social and interactive aspects of human tool use, and will allow the translation of some of these insights into technology design and engineering practices.

Button, Graham, Jeff Coulter, John R. E. Lee, and Wes Sharrock, eds. 1995. Computers, Minds, and Conduct. Cambridge, UK;  Cambridge, MA, USA: Polity Press; Basil Blackwell [distributor].
Dingemanse, Mark, and Andreas Liesenfeld. 2022. “From Text to Talk: Harnessing Conversational Corpora for Humane and Diversity-Aware Language Technology”. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long
Papers), 5614–33. Dublin: Association for Computational Linguistics.
Finch, Sarah E., and Jinho D. Choi. 2020. “Towards Unified Dialogue System Evaluation: A Comprehensive Analysis of Current Evaluation Protocols”. In Proceedings of the 21th Annual Meeting of the Special Interest Group on Discourse and Dialogue, 236–45. 1st virtual meeting: Association for Computational Linguistics.
Housley, William, Saul Albert, and Elizabeth Stokoe. 2019. “Natural Action Processing: Conversation Analysis and Big Interactional Data”. In Proceedings of the Halfway to the Future Symposium 2019, 1–4. Nottingham, UK: ACM.
Illich, Ivan. 1973. Tools for Conviviality. London: Calder and Boyars.
Joshi, Pratik, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. “The State and Fate of Linguistic Diversity and Inclusion in the NLP World”. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 6282–93 . Online: Association for Computational Linguistics.
Liesenfeld, Andreas, and Hendrik Buschmeier. 2023. “Robust and Flexible Interactive Language Technology for Human Empowerment”. Presented at the International Pragmatics Association, Brussels.
McIlvenny, P. B. 1993. “Constructing Societies and Social Machines: Stepping Out of the Turing Test Discourse”. Journal of Intelligent Systems 3(2–4). DOI: 10.1515/JISYS.1993.3.2-4.119.

Synthetic conversational agents populate messenger services, smart devices, and corporate websites today. Some are created as open-domain artificial companions, friends or even romantic partners like Kuki AI and Replika, while others figure as virtual assistants (Siri, Alexa, etc.) or are built as task-oriented chatbots. Notwithstanding the sensationalist coverage of individual “breakthrough” language technologies, such as OpenAI’s GPT-3 and Google’s LaMDA, a great number of present-day chatbots perform rather mundane tasks. In customer service for example, chatbots serve as conversational agents that provide answers to common questions and that automate relatively simple support tasks largely autonomously.

In our presentation, we will ask how such chatbots “come into existence in a coming together of heterogeneous elements” (Vogl 2007:16) as capable conversational agents. We will focus on chatbot frameworks like Rasa, Chatterbot or Microsoft’s BotFramework that provide meta technological resources which “scaffold” (Wimsatt and Griesemer 2007) the creation, deployment, and maintenance of so-called task-oriented chatbots. As such, they constitute the largely invisible infrastructure of semi-autonomous conversational AI.

Drawing on research in critical algorithm, data, software and infrastructure studies (e.g., Mackenzie 2017; Crawford and Joler 2018; Eghbal 2020), we analyze how frameworks modularize and structure the development of chatbots, thus allowing framework-users to orchestrate, adapt and fine-tune general purpose language models for specific tasks like taking pizza orders from customers – a task that has become the Hello World! example in chatbot development tutorials. The frameworks we consider, provide abstract functional resources that developers can use and build upon to create a specific conversational agent and to deploy it on various channels. At the same time, these frameworks take part in shaping their users’ understanding of what chatbots are, which capabilities they can achieve and how conversational encounters can feed into the ongoing training of chatbots.

By paying close attention to the meta-technological resources that enable the creation of chatbots, our aim is to unpack the broad range of technologies and practices that support or ensure the fragile capacities of semi-autonomous agentic media: the specification of conversational intents, the definition of processing pipelines, the fine-tuning of language models, the integration of contextual knowledge bases, and the ongoing monitoring of conversational performances. In light of recent debates about bias built into language models as well as potential harms that might stem from their use (Bender et al. 2021), we thus contend that chatbot frameworks are important sites for critical inquiry into the imaginaries and politics of AI.

Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? &#x1f99c”. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–23. FAccT ’21. New York, NY, USA: Association for Computing Machinery. DOI: 10.1145/3442188.3445922.
Crawford, Kate, and Vladan Joler. 2018. “Anatomy of an AI System: The Amazon Echo As An Anatomical Map of Human Labor, Data and Planetary Resources”- AI Now Institute and Share Lab (blog). 2018. URL: https://anatomyof.ai.
Eghbal, Nadia. 2020. Working in Public: The Making and Maintenance of Open Source Software.
Mackenzie, Adrian. 2017. Machine Learners: Archaeology of a Data Practice. Cambridge, MA: MIT Press.
Vogl, Joseph. 2007. “Becoming-Media: Galileo’s Telescope”. Grey Room 29 (October): 14–25. DOI: 10.1162/grey.2007.1.29.14.
Wimsatt, William C., and James R. Griesemer. 2007. “Reproducing Entrenchments to Scaffold Culture: The Central Role of Development in Cultural Evolution”. In Integrating Evolution and Development: From Theory to Practice, edited by Roger Sansom and Robert N. Brandon, 227–323. Cambridge, MA: MIT Press.

The humanities conceptualize language as an essentially social construction, a symbolic form that is deeply rooted in the history of human culture. This double determination of language (as being socially anchored and culturally grounded) informs scholarly approaches to understanding language and its fundamental importance for human activity in the world. Contrary to this, the technical discussion of language in NLP and Machine Learning is often characterized by a “natural” given of linguistic information as “data”. NLP terminology is full of culturally loaded terms, be it the “knowledge” of language models, their “understanding” of language, or the notion of “intelligence” in technical applications (Bender & Koller 2020). In addition, fundamental concepts like the term “model” itself build on a notion of language as “naturally occurring”. Among other things, this results in the description of NLP resources as models “of language” – instead of models “for language”.

Considering the current hype – both in NLP and its public uptake – around Large Language Models (e.g., GPT-4, LaMDA, LLaMA) and their applications as sources of “information” (e.g., ChatGPT, Bard, Bing) in artificial communication (Esposito 2022), we will discuss the consequences of a culturally-grounded understanding of language for NLP and Machine Learning. Starting from the idea that language is a social, embodied, and highly contextually determined practice, we examine current NLP terminology to highlight some of its consequences for working with language:

A) By ascribing intrinsically cultural values to technical derivations of actual language practice, NLP neglects the manifold socio-cultural prerequisites of automatic language processing in the sense of a naturalization of the object (e.g., the relationship between human perspective and “bias” in NLP).

B) By using metaphors coined for socio-cognitive processes to describe the performance of technical artifacts NLP blurs the crucial difference between language as human practice and its technical reproduction to simulate practice (e.g., in the context of Artificial Intelligence or “sentient” language models; tiku 2022).

Bender, Emily M., and Alexander Koller. 2020. “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data”. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–98. Online: Association for Computational Linguistics. URL: https://aclanthology.org/2020.acl-main.463/
Esposito, Elena. 2022. Artificial communication. How Algorithms Produce Social Intelligence. Cambridge, MA: MIT Press.
Tiku, Nitasha. 2022. “The Google engineer who thinks the company’s AI has come to life”. The Washington Post (11 June 2022). Online. URL: https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/ 

Many of the applications that are used in machine learning and natural language processing (NLP) are written in a computer language called Python. Python has become one of the fastest growing programming languages and knowledge of it can be considered a valuable form of social capital (Bourdieu 1977). The structure of Python, explicitly introduced as a language itself, reinforces a language ideology that sees language as a semantic, referential, and universal system (Jablonski n.d.; Lovrenčic et al. 2009; Danova 2017). This contrasts with the position taken by most linguistic anthropologists that sees language as primarily pragmatic, indexical, and highly localized in its ability to convey meaning (Silverstein 1976; 1979; Gal and Irvine 2019; Nakassis 2018; Eckert 2008).

As part of a fieldwork placement, I have been taking two online courses on Coursera (“Getting Started with Python” and “Natural Language Processing with Python’s NLTK Package”), that contain both explicit and implicit endorsements of this language ideology. Here, language appears exclusively as data encoded in corpora of written text and is visualized with parse trees from generative grammar. Language is said to be “tokenized” and “chunked” by Python code, with words being strictly divided into two categories of either “content words” (in which the analyst should be most interested) and “context words” (which are mostly ignored). Python’s Natural Language Tool Kit (NLTK) is sold as a suite of libraries and programs for “symbolic and statistical natural language processing” (Bird et al. 2019). The use of the word “natural” in this phrase indexes a different stance towards language than is usually found in social sciences departments.

An analysis of course material from these two courses shows that a semantico-referential ideology can be considered doxa (Bourdieu 1977a) in the field of natural language processing, one that marginalizes competing ideologies of language as a phenomenon inextricably tied to other forms of semiosis including body language, prosody, and context. As artificial intelligence applications work their way into the everyday lives of consumers, it is important to investigate the language ideologies that accompany them. Tracing how these ideologies are reproduced through popular programming course material can help reveal what is excluded when the focus is purely on language that is written, and where meaning is assumed to exist in a one-to-one mapping to linguistic signs. In particular, speakers of languages that exist only in an oral form are rendered invisible through NLP analysis (Wilson 2022). Also, contextualization cues that emerge in real-time conversation are ignored in NLP, even though in spoken discourse they are indispensable in helping speakers navigate the semantic ambiguity inherent in indexicals (Gumperz 1982; Schegloff 1996; Moerman 1998). These are some of the real-world consequences of ideologies that circulate unseen in communities of practice.

Bird, Steven, Ewan Klein, and Edward Loper. 2019. Natural Language Processing with Python. Natural Language Toolkit [http://nltk.org/] Version 3.0. Accessed 22 November 2022. URL: https://www.nltk.org/book/.
Bourdieu, Pierre. 1977a (1972). Outline of A Theory of Practice. Translated by Richard Nice. Cambridge, UK: Cambridge University Press.
Bourdieu, Pierre. 1977. “The Economics of Linguistic Exchange”. Social Science Information 16(6): 645-668.
Danova, Ina. 2017. “Computer code as the next universal language”. Pegasus Digital (18 July 2017). Accessed 22 Novemver 2022. URL: https://pegus.digital/computer-code-as-the-next-universal-language/.
Eckert, Penelope. 2008. “Variation and the indexical field”. Journal of Sociolinguistics 12(4): 453–476.
Gal, Susan and Judith Irvine. 2019. Signs of Difference. Cambridge, UK: Cambridge University Press.
Gumperz, John J. 1982. Discourse Strategies. Cambridge, UK: Cambridge University Press.
Jablonski, Joanna. n.d. “Natural Language Processing With Python’s NLTK Package”. Real Python. Accessed 22 November 2022. URL: https://realpython.com/nltk-nlp-python/.
Lovrenčic, Alen, Mario Konecki, Tihomir Orehovački. 2009. “1957-2007: 50 Years of Higher Order Programming Languages”. Journal of Information and Organizational Sciences 33(1): 79-150.
Moerman, Michael. 1988. Talking Culture: Ethnography and Conversation Analysis. Philadephia, PA: University of Pennsylvania Press.
Munk, Anders Kristian, Asger Gehrt Olesen, and Mathieu Jacomy. 2022. “The Thick Machine: Anthropological AI between explanation and explication”. Big Data & Society (Jan-June): 1-14. 
Nakassis, Constantine. 2018. “Indexicality’s Ambivalent Ground”. Signs and Society 6(1).
Schegloff, Emanuel A. 1996. “Confirming Allusions: Toward an Empirical Account of Action”. The American Journal of Sociology 102(1).
Silverstein, Michael. 1976. “Shifters, linguistic categories, and cultural description”. In Meaning in Anthropology, edited by Keith H. Basso and Henry A. Selby. Albuquerque, NM: University of New Mexico Press.
Silverstein, Michael. 1979. “Language Structure and Linguistic Ideology”. In The Elements: A Parasession on Linguistic Units and Levels, edited by Paul R. Clyne, William F. Hanks, and Carol L. Hofbauer, 193-247. Chicago, IL: Chicago Linguistic Society. 
Wilson, Joseph. 2022. “AI Can’t Fully Capture Oral Languages”. SAPIENS. Accessed 22 November 2022. URL: https://www.sapiens.org/language/ai-oral-languages/

Recent advances in statistical language modeling have endowed computers with an impressive range of new linguistic abilities. As these abilities have continued to broaden, they have sparked debates about just what these models understand about language and, by extension, about the nature of language and meaning itself. This talk presents recent work deconstructing the still popular word2vec algorithm, which helped to usher in the past decade of rapid progress, in an effort to understand how the particularities of which words co-occur in text give rise to the elaborate performances of models trained on those texts. This work is then used as an occasion to put contemporary scientific speculations about how these models work into a historical context in which theories of pragmatics and distributional semantics have traded places in the description of why NLP systems work (or fail). This context clarifies how technical developments and linguistic theory come together to shape the scientific imaginary of NLP researchers, and offers a way in to thinking about the role of linguistic theory in the future development of the field.

Large Language Models (LLM) are becoming the new standard in Natural Language Processing (NLP). LLMs such as BERT, RoBERTa, and GPT-x (Chowdhery et al., 2022) are pretrained on very large datasets of language obtained from the web. GPT-3, to date the largest of these models, is trained on 175 billion parameters, and “can generate samples of news articles which human evaluators have difficulty distinguishing from articles written by humans” (Brown et al., 2020), as the authors claim in a research paper, and which was also demonstrated by an article in the Guardian that was written by GPT-3¹. While the latter helps to impress 1 the general public, the NLP community measures performance via quantitative metrics like F1-scores and benchmarks like GLUE,² through which different models can be compared and ranked. Here, LLMs outperform other models on diverse tasks such as language translation, question answering, or sentiment analysis, so that proponents of LLMs have claimed (near) human parity for these tasks (Talmor et al., 2021).

However, the success of LLMs is not undisputed, and research in NLP and Machine Learning outlining failure cases of the models – such as overfitting to data artifacts, lack of generalization, etc. – has emerged (Poliak et al. 2018, Utama et al., 2020, Eger and Benz 2020, Beese et al., 2022). Yet although researchers may be aware that LLMs may fail under certain conditions, the boundaries of what the models are able to do are largely unknown. While there has been important work from cognitive science or linguistics addressing this question (Bender & Koller, 2020; Ettinger, 2020; Marcus, 2020), we discuss what insights can be gained by a sociological perspective: (1) The first is a socio-theoretical grounding of prominent terms that are used in NLP to describe what models can do. E.g., ethnomethodology would suggest that expressions rely on their context to be understood, and their meanings will vary over the course of an interaction (Pütz, 2019b), which seems to present a challenge to computational approaches that treat meanings as static. (2) The second is a methodological reflection about practices of model assessment in NLP. E.g., researchers working with LLMs rely on quantitative model performance measures with little consideration of their limitations. Furthermore, while prediction errors that indicate where models fail are often considered by researchers, they usually do so in an ad-hoc fashion without methodological grounding. (3) Third, we suggest that theoretical grounding and methodological reflection should inspire new forms of interdisciplinary research, and provide an example of bringing together NLP and sociology. Here, we focus on parliamentary debates as a challenging application domain for LLMs that involves dialogue in an institutionalized setting, changes in footing (Goffman, 1979; Pütz, 2019a), and sarcastic statements that require an appreciation of tone and context. Insofar as we fine-tune LLMs for such challenges with the help of appropriate annotations, we should gain additional insights into what LLMs can and cannot do. We also discuss alternative methods for model evaluation, such as error analysis with grounded theory and dynamic dialogue benchmarks.

¹  After being provided with a written prompt and an introduction, GPT-3 generated eight different outputs and the editors of the Guardian decided to not publish just one of them, but to pick what they deemed to be the best paragraphs and arrange them into one text. Otherwise, the editors claimed, “editing GPT-3’s op-ed was no different to editing a human op-ed”. https://www.theguardian.com/commentisfree/2020/sep/08/robot-wrote-this-article-gpt-3. However, researchers have also used generated text to demonstrate in what regards GPT-x models fail to provide not only coherent but also text that is meaningful (Bender & Koller, 2020; Marcus, 2020).
²  URL:https://gluebenchmark.com/.

Representing semantic relationships across a corpus typically involves the construction of large incidence matrices. Ubiquitous even among the latest NLP models, these ‘document–term’ matrices (DTMs) tabulate the number of times particular words appear within each corpus text. Developed through the early 1960s, they are one in a series of what Bernhard Rieder, in his recent book about the algorithmic ordering of information, calls “intermediate forms” (2020, 215): once produced, NLP practitioners use them to model meaningful associations in language data. DTMs are therefore a major brace in the semiotic architecture of NLP. Ultimately, they enable practitioners to dovetail term counts with textual semantics.

But while DTMs may grid the semantics of language models, they themselves do not originate from NLP. Instead, this paper shows that it was research in early information retrieval (IR) that was most responsible for these matrices. Retrievalists’ work on language data paralleled their contemporaries in linguistics and communication studies, but often their research efforts remained independent and only later migrated outside IR. DTMs are particularly illustrative of this dynamic, and in charting their emergence I highlight how retrievalists used these matrices to operationalize a concept of semantic association that significantly differed from the theoretic assumptions of similar fields. Indeed, much as with machine learning now, early IR was largely without theory, being defined instead by an ever growing body of techniques that privileged epistemological consensus over scientific method, the benchmark over proof (see Roberge and Castelle 2021). If not theory, then, by what coordinates did retrievalists audit their techniques? How did forms like the DTM shape what retrievalists understood to be the meaning of language data, and what insights might such understandings hold for the logics that underpin contemporary NLP?

My exploration of these questions will involve canvassing the work of retrievalists like H.P. Luhn and Lauren Doyle to synthesize a more general articulation of semantic association in IR. In doing so, I set this concept against a recurrent pattern among these and other retrievalists: many were highly attuned to experimental psychology and often incorporated this research into their own. In particular, free association provided several research groups in early IR with a key conceptual template against which to rationalize—and thereby justify—their methods. DTMs, I conclude, were a primary structure that retrievalists would use to buttress their semantic models against this foil.

Borko, Harold. 1962. “The Construction of an Empirically Based Mathematically Derived Classification System”. In Proceedings of the May 1-3, 1962, Spring Joint Computer Conference on – AIEE-IRE ’62 (Spring), 279. San Francisco, CA: ACM Press. DOI: 10.1145/1460833.1460865.
Rieder, Bernhard. 2020. Engines of Order: A Mechanology of Algorithmic Techniques. Amsterdam: Amsterdam University Press. DOI: 10.5117/9789462986190.
Roberge, Jonathan, and Michael Castelle, eds. 2021. The Cultural Life of Machine Learning: An Incursion into Critical AI Studies. Cham, Switzerland: Palgrave Macmillan. DOI: 10.1007/978-3-030-56286-1.

If we want to understand the impact of digital technologies on societal phenomena like language, it is necessary to go beyond the individual and study public reconfigurations of communicative practice and their epistemological implications (Hepp et al. 2023). Digital infrastructures and algorithmic culture bring about various specifically distributed human-to-human practices, transnational types of community formation and post-national types of public space. With this, ideas of how language is ordered in territory, how it should ideally look like, and what language is are changing.

In this talk, I draw on language ideology research (e.g., Gal and Irvine 2019) and ask how concepts regarding multilingualism and linguistic diversity are reconfigured through NLP technologies such as multilingual transformer models (Devlin et al. 2019). To this end, I first discuss hegemonic ideas regarding the distribution of diverse language forms from the age of literacy. Based on national epistemologies and cultures of writing, languages are here understood as systems of linear signs that are ordered according to territorial containers (Linell 2005). Multilingualism is conceived as a special case and typically conceptualized as a form of “double monolingualism”, that is, linguistic diversity is understood as appearing in systemically equal, parallel entities (Gramling 2021). Sociolinguistic hierarchies are related to the existence of writing systems, to the materialization of language in books and to the support of governmental institutions. This has the effect of a particularly strong marginalization of language practices that do not conform to a concept of a stable systemic entity (e.g., creole languages, fused forms).

In NLP language technologies, languages are above all approached as data sets. These are drawn from the web, where particularly Wikipedia texts hold a prominent place as ‘gold standard’ representation of language. As national language standards are mostly adhered to in Wikipedia, monolingual ideologies of the age of print literacy – languages as standardized expressions of territorial culture – are reproduced. Yet, those who develop language technologies have little interest in the orders of national community. In the logics of data capitalism, the main goal on sides of those who produce language models is to collect data to ensure advantages in political and economic realms. With a planetary perspective on communication of which as many traces as possible are gathered, boundaries between languages are not seen as based on ethnic or national culture but are found by machine-learning algorithms detecting patterns in data. Given the hegemonic status of English in technology development, tech-sociolinguistic hierarchies entail a dichotomy of ‘English’ vs. ‘Other’ and the marginalization of those practices that neither show up in public online interaction nor are supported by artificial data construction. Yet, many commercial initiatives are feverishly working on providing language technologies for these ‘low-resource languages’ (e.g. NLLB 2022). Non-stable, internally variable, diffuse and multimodal language practices still represent a challenge – but may at the same time open paths of resistance in an age of global surveillance (Zuboff 2019).

Taken together, these observations confirm the posthumanist argument and show that media technology is constitutional of sociolinguistic order and our understanding of what is ‘language’.

Devlin, Jacob, Ming-Wei Chang, Kenton Lee, Kristina Toutanova. 2019. “BERT: Pre-training of deep bidirectional transformers for language understanding”. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1. DOI: 10.18653/v1/N19-1423.
Gal, Susan, and Judith T. Irvine. 2019. Signs of difference. Language and ideology in social life. Cambridge, MA: Cambridge University Press.
Gramling, David. 2021. The invention of multilingualism. Cambridge, MA: Cambridge University Press.
Hepp, Andreas, Wiebke Loosen, Stephan Dreyer, Juliane Jarke, Sigrid Kannengießer, Christian Katzenbach, Rainer Malaka, Michaela Pfadenhauer, Cornelius Puschmann, and Wolfgang Schulz. 2023. “ChatGPT, LaMDA, and the hype around communicative AI: the automation of communication as a field of research in media and communication studies”. Human Machine Communication 6.
Linell, Per. 2005. The written language bias in linguistics. Its nature, origins and transformations. London: Routledge.
NLLB Team et al. 2022. “No Language Left Behind: Scaling Human-Centered Machine Translation”. arXiv, abs/2207.04672.
Zuboff, Shoshana. 2019. The age of surveillance capitalism: the fight for a human future at the new frontier of power. New York, NY: Public Affairs: New York.

The ongoing automation through AI technologies enable and enact research lines in health care that draw on voice data as biomarkers. We examine in particular how ‘stress’ as a biomarker is conceived of, transformed and translated into a set of signals. How these are standardized and re-used to automate mental health assessment tools that measure the physical properties of sound waves in a voice signal, such as frequency, amplitude, and duration, to extract features that are related to stress and to identify patterns that may be associated with certain mental health conditions. Natural language processing techniques are also used to analyze the content of speech and identify words or phrases that are related to stress. For example, certain types of language, such as negative self-talk or catastrophic thinking, are often associated with high levels of stress. Recent advances, as seen through ‘Whisper’, combines the acoustic and linguistic analysis techniques, that can extract a wide range of biomarkers from a person’s voice. The key advantage of Whisper’s approach is the large number of biomarkers that it extracts from a single voice sample. Within this line of research, we question how automation plays a role in the assessment and monitoring of stress, and how it enters prevention programs and workplace matters.

We draw from case studies that answer the emerging infrastructuring of health data through IT aspirations, such as industry collaborations with university research institutions and clinical studies. Health data are often based on administrative infrastructures and biomedical records, at the population levels and often overlape (Bauer, 2014). But what is at stake when voice databases are created in conjunction with cooperations or entirely by the private sector? What kind of datafication, data reuse and infrastructure mining is ongoing in the field of stress research, given the datafication, in an occupational-, social-, and neuroscientific epidemiology of stress?

Bauer, S. 2014. “From Administrative Infrastructure to Biomedical Resource: Danish Population Registries, the ‘Scandinavian Laboratory’, and the ‘Epidemiologist’s Dream’”. Science in Context 27(2): 187-213.

Linguistic methods are always also ideologies, ways of understanding language and its relationship to the social world. Following on the conference theme of cooperation, members of Culture and Computation Lab – a linguistic anthropologist, and two computational linguists – offer an account of the theoretical and ideological questions which emerged in the lab’s project ConMan, an NLP resource aimed at documenting and ultimately enabling the detection of conspiracy theories which is informed by critical digital and linguistic anthropological research. Moving beyond an understanding of language as naturally occurring and “context free” which animates much of NLP, our lab refers to critical sociolinguistic and anthropological understandings of language as a non-neutral medium that cannot be separated from the social world. As media anthropology (Gershon 2017) shows us that digital spaces are not neutral, or rule governed sites of information processing, platform affordances also shape social, affective, ideological spaces; indeed, this very idea of abstract, rules governed processes can create, hide, and legitimate relations of oppression and inequality (Noble 2018).

This is especially important in the world of digital political discourse in which ConMan aims to intervene the proliferation of fascist, far-right ideology online. ConMan combines an anthropological framework and machine learning to analyze conspiracy theories. This project will build on linguistic anthropological frameworks to look at the production, circulation, and consumption of conspiracy narratives, their social and ideological effects, and design an NLP tool to identify them. Initial conversations moved from understanding the “truth” or “correctness” of these narratives to looking at their uptake and consequences. This move, we hope, also brings us beyond “fact-checking” approaches which legitimate powerful, and often exclusionary, institutional discourses and cultural studies approaches which characterize conspiracy as “popular knowledge” and elide its damaging political and social effects.

The specific case of ConMan also raises broader theoretical and methodological questions: What would it mean to have a culturally grounded, socially anchored vision of NLP? What are the interactions, cooperations, and contradictions between the social and the computational? Is it possible to have a critical computational sociolinguistics? Can this collaboration create new forms of NLP, critical machine learning? Or is a critical sociolinguistics limited to critique – to understanding, and countering the ways AI creates, erases, and legitimates inequalities.

Gershon, Ilana. 2017. “Language and the Newness of Media.” Annual Review of Anthropology 46 (1): 15–31. DOI: 10.1146/annurev-anthro-102116-041300.
Noble, Safiya Umoja. 2018. Algorithms of Oppression. How Search Engines Reinforce Racism. New York, NY: New York University Press. URL: https://www.degruyter.com/isbn/9781479833641.

Despite the widespread assumption that NLP is the operational side of a neutral/universal way of representing/experiencing human language, critical approaches increasingly present NLP as the byproduct of a particular language ideology whose genesis can be traced, and that has led to a specific development of the ‘science of language’ and its correlated technologies. As such (as a local story succeeding as a global design, to put it in Walter Mignolo’s terms), it is designing our contemporary algorithmic and datafied communication culture through an ontological bias (different from the more popular ethic biases that make the headlines in critical AI). This biased design is basically unaffected by the deep cultural conceptions of language/communication as intersubjective and proto-aesthetic phenomena, features still to be traced in a more transparent way in ‘minoritised’ communication ideologies. Decades-long efforts for world linguistic diversity maintenance have oscillated between two poles of a continuum: Preservation through (academic) documentation (to accumulate archive knowledge /corpus approach to language material)/ Revitalisation emphasising the centrality of the speaker/speech communities in their environments. Languages with no speakers (documentation) vs speakers at the centre of the (revitalisation) programmes. These terms echo a key issue at the heart of current approaches to language technologies applied to the generality of human natural language: a) The increasing abstracting away of language: LMs only have success in tasks that can be approached by manipulating linguistic form; meaning is inferred via formal distributional properties, etc. b) The abstracting away of speakers themselves, their increasing alienation from the interactional experience and from the environment. Bearing in mind that the notion of the ‘speaker’/ ‘speech community’ very much co-variates with related cultural constructs –person, human (body), voice– and the communication ideologies they are inscribed in.

Linguistic diversity maintenance efforts currently focus on strategies mainly involving language technologies grounded on those structural biases. The diversity of human representations and experiences of language/communication (and associated interactional worlds) may therefore be potentially excluded through the “inclusion” of that plural heritage via the extreme version of a language/communication ideology that increasingly abstracts away the speaker and the process of meaning and world-making.

The UK’s Department of Work and Pension (DWP) is currently trailing AI-powered technology for matchmaking in more than 20 Jobcentres Plus, government-funded employment agencies. The automated system is designed to connect people with local roles and training routes. In my paper, I intend to ethnographically explore both the conditions under which AI powered ‘matchmaking’ is implemented in UK’s jobcentres and the effect of this technology on migrants seeking work in the UK.

Drawing on past work on language and employability and more recent scholarship in critical algorithm studies, automated hiring, and the anthropology of digital media and technology, in my contribution, I argue that AI is not a singular, stable, rigid and non-human, powerful and obscure technological device that enters matchmaking processes, but rather unstable and malleable, ideologically and historically enacted by the multiple institutionally and discursively regulated practices through which people engage with AI and come to think and do matchmaking. Understanding technology in this way means to focus on the situated interpretative processes which frame the choices, decisions and social processes by which people on the ground define and discover problems and identify acceptable solutions, for example about migrant unemployment. This includes the communicative processes through which AI gets to be conceived of for job matchmaking, how AI reconfigures matchmaking, is implemented in jobcentres, adapted to the needs of migrant jobseekers, and comes to legitimize hierarchies of workerhood, processes of selection and exclusion. It also forces us to focus on how the ideologies and worlds of reference of the people engaging with these algorithmic systems come to matter, i.e. how their engagement with AI’s matchmaking recommendations is framed by how they understand the relevance of technology, the world of work, and the value of cultural and linguistic difference.

Through a fine-grained ethnographic analysis of caseworkers and migrant jobseekers AI mediated matchmaking interactions, in this paper I show how AI-powered matchmaking tools are constructed and brought into being in and through communicative processes. This produces new understandings about how algorithmic systems get articulated interactionally with recruitment processes, and are integrated in institutionally scripted process of matchmaking, how they are made to coexist with longer histories of gatekeeping, and with new logics of recruitment and selection. This new knowledge allows a wider understanding of the human side of AI and contributes to our scholarly ability to align AI with projects of social justice and equality.