Band 22, Heft 2 - Special Issue: Artificial Intelligence in Pragmatics and Language Acquisition; Guest Editors: Istvan Kecskes†, Carlotta Sparvoli and Hanh Dinh

Language models are mathematical functions and, as such, induce vector spaces in which input is embedded. Comparing the point clouds of concept vectors across such language models and similar computer vision models, we see surprising similarities. This sheds new light on the Innateness Debate. Much linguistic structure can be induced from extra-linguistic data. Language models are generally thought to be too sample-inefficient to be good models of language acquisition, but what about language models initialized by computer vision models?

This study explores the growing importance of text simplification, particularly in light of increasing global migration and the demand for accessible information. We compare AI-generated and human-produced simplifications of Japanese news articles, focusing on whether GPT-4 can outperform humans in simplification tasks and how their strategies differ. Analyzing 420 original articles and their simplified versions created by human experts and ChatGPT (GPT-4), we employed a readability formula specifically designed for Japanese texts and conducted text feature analysis. Our findings reveal that AI-simplified texts achieve significantly higher readability scores than human-simplified versions while retaining more original content. Specifically, AI-simplified texts exhibit shorter average sentence length and lower complex word usage. We argue that AI, leveraging transformer architectures with self-attention mechanisms, can perform multidimensional simplification more effectively than humans, potentially due to its ability to process information beyond sentence boundaries during training. Furthermore, this study highlights the cognitive challenges humans face in text simplification, offering insights into human language processing, including its limitations and creative potential. Thus, our research confirms the potential of AI in language processing while simultaneously providing a perspective for understanding human cognitive processes in language use. The findings have implications for language education, linguistic research, and the development of accessible information dissemination methods, particularly for initiatives like “Easy Japanese.” We also suggest future research directions, including exploring human-AI collaboration in text simplification.

Can AI reproduce human interaction? It can, but only stereotypically. While it can simulate (and even exaggerate) dialogic engagement, its lexicon is less diverse, and the speech acts it realizes are more repetitive and less varied (we took directives as an example). Most importantly, AI struggles to represent ‘conversational uniqueness’, that is ways to interact that define the specificity of a particular conversation and are not entirely conventional. We discovered this after analyzing dialogic resonance (the re-use of an interlocutor’s construction), recombinant creativity (the creative reformulation of an interlocutor’s construction), relevance acknowledgement (the acknowledgement of what an interlocutor said) and other variables from a sampled section of the CallHome Corpus of Chinese telephone conversations. After feeding ChatGPT with speakers’ demographics and contextual information, we asked it to reproduce telephone interactions among Chinese family members. We then fitted a conditional mixed effects Bayesian regression comparing the two datasets. We found that AI over-generalizes human dialogue. It provides a stereotypical way of conversing but shows scarce flexibility in including ‘atypical’ and unconventional utterances, which are, in turn, constitutive of human interactions that occur in real life.

As large language models (LLMs) increasingly permeate various domains of human life, their ability to accurately comprehend and appropriately respond to irony has become a critical challenge. Irony, as a linguistic phenomenon heavily reliant on contextual, cultural, and cognitive factors, places elevated demands on LLMs’ comprehension. Guided by a systematic theoretical framework, this study integrates qualitative and quantitative methods to construct a comprehensive test set. By analyzing the responses of ChatGPT-4o and comparing them with those of human participants, the study examines the model’s accuracy in understanding and responding to different types of irony. The findings reveal that both human participants and ChatGPT-4o achieved perfect accuracy in comprehension tasks involving situational irony, visual irony, and multimodal irony. However, significant difficulties were observed in the comprehension of verbal irony. Furthermore, the study explores key factors affecting ChatGPT-4o′s performance and identifies the primary mechanisms the model tends to rely on when processing ironic utterances. The results indicate that verbal irony, which requires a more sophisticated grasp of emotional tone and complex cognitive abilities, constitutes the primary factor affecting the performance of LLMs in understanding irony. Meanwhile, Grice’s maxim of quality and inferences about interpersonal relationships are the main mechanisms that LLMs tend to rely on when processing ironic utterances. The findings provide empirical support and developmental pathways for enhancing the capacity of AI systems to handle complex pragmatic phenomena. Furthermore, the study offers important insights into the integration of linguistic theory with artificial intelligence, highlighting new directions for future interdisciplinary research.

Recent advancements in artificial intelligence technology have given impetus to extensive research across a number of disciplines, including semantics and pragmatics, which focus on human-machine linguistic interactions, dialogues in particular, that generate a feeling of almost natural conversation going on. The interest in such interactions has called to rethinking of pragmatic frameworks through which language use between human interlocutors is conceptualized. On the other hand, such research has affirmed that some pragmatic models, normally developed to pertain to human-human conversation, prove applicable and suitable to human-machine interaction, as is the sociocognitive approach (SCA). The concept that stands out in this respect is one of asymmetry in incrementing common ground between speakers that come from different languages. However, the subject of research presented in this paper are the fictional dialogues/conversations between characters in the novel by Kazuo Ishiguro, Klara and the Sun , that was first published four years ago. Following the trail of the question of credibility of fictional characters’ voices that has been illuminated in works on literary pragmatics and drawing on the concept of ‘in-betweenness’ (taken to be crucial for constructing culture by sociologists of culture) I examine those dialogues that Klara, a humanoid artificial friend (AF) enters within at least three different types of communities (K) - the community of other AFs and the communities she forms with humans – children and adults. Taking into account factors such as conceptual and background knowledge, egocentrism and salience, I observe the emergence of common ground between the fictional conversationalists, whom I take to be ‘intercultural speakers.’ The emergent common ground, however, consistently proves, in the majority of such conversations, deficient, cropped and unattainable. I argue, finally, that this failure to increment common ground gives credibility to Klara’s voice, making her a permanent ‘inbetweener.’

With the advance of contemporary internet based artificial intelligence (AI), many interactive scenarios emerge. The interaction between a user and their device centers around the provision and build-up of information that is recorded for further individual or communicative usage. Social media and messaging services are providing AI-based affordances that enable the implementation of adjacency, the provision of background knowledge and recipient design through links and hashtags, and expressions of stance and acknowledgement through likes. Further to that, virtual assistant technologies have become popular and widespread. Through AI, the mobile and home-based assistants are built to directly employ and facilitate pragmatic realities of communication, like turn taking, talk and reaction on the basis of common ground, addresses and ways of responding to them. In this article, I show that modern AI applications for domestic and mobile communication usage, both between humans and for human-machine interaction, are built around the idea of common ground. Taking into account the privacy concerns that are associated with sizable data transmission, I introduce the term übercommon ground for information that is tracked from users’ online activity and used for the benefit of a better browsing experience, but also potentially passed on without the users’ consent.

Conversational AI is advancing rapidly, enabling significant improvements in chatbots’ conversational abilities. Currently, available conversational chatbots (e.g., Snapchat’s MyAI) appear to generate fairly realistic, often human-like output. As collaboration between humans and machines becomes more common, and AI systems are increasingly viewed as more than just tools, understanding human communication in such contexts is crucial. Despite the vast array of applications and the increasing number of human-bot interactions, research on how humans interact with conversational chatbots is scarce. One possible reason for this gap is that studying human-computer communication may require adaptations of existing pragmatic frameworks, due to the unique characteristics of these interactions. A key feature of such conversations is their asymmetrical nature. In this paper, we present evidence that the sociocognitive approach (SCA), which takes into account the asymmetry between interlocutors as regards their possible common grounds, has explanatory potential to describe human-AI-powered chatbot interactions. We collected data from thirty-two L1 Hungarian participants interacting with a conversational chatbot on three consecutive days. The turn-by-turn analysis of the 96 conversations provides insights not only into the nature of common ground humans presuppose with a conversational agent, but also into the processes of building emergent common ground over time. Furthermore, we present linguistic evidence that both egocentrism and cooperation play a role in human-chatbot interaction. While the former is manifested in approaching the chatbot as if it were human, the latter appears to play a role in changing strategies that serve common ground seeking and building.

The study explores the effectiveness of ChatGPT in fostering intercultural pragmatic competence, focusing on Japanese learners of English. Using Kecskes’ sociocognitive framework, the research examines ChatGPT’s role in conversations involving meaning negotiation and repair mechanisms. The study analyzed over 10,000 responses generated by ChatGPT during interactions with 45 students from different nationalities, ranging from upper-beginner to upper-intermediate proficiency levels, in simulated social situations. A comprehensive examination of AI and participant responses and feedback yields a sophisticated understanding of the AI strengths and flaws, along with intercultural communication repair strategies for meaning negotiation. Specifically, findings reveal that while ChatGPT effectively models certain cooperative aspects of communication, it struggles with maintaining relevance and lacks human-like repair strategies. Students often initiate self-repair instead of other-repair, indicating limited interactive feedback from ChatGPT. Furthermore, excessive information. On the positive side, ChatGPT encourages confidence in students through predictions and co-constructed phrases, although its performance is repetitive and lacks deep contextual understanding. The study recommends cautious use of ChatGPT in pragmatic learning and highlights the need for more advanced frameworks to enhance AI’s role in intercultural communication education.

We consider the compatibility of artificial intelligence (AI) chatbots with human pragmatics and whether they have utility as a tool in the understanding of human communications, and language in use. We claim that, precisely because pragmatics reflects a human communication-oriented endeavour, there are a number of cognitive processes to do with processing context, speech acts, common ground, implicatures, and inference, that are potentially difficult for an AI chatbot to accurately handle and mimic. Another claim is that an AI chatbot manages common ground differently to a person. In order to advance our knowledge of the pragmatic behaviour of the AI chatbots, we undertake empirical testing of Microsoft’s Copilot, Google Gemini, and OpenAI’s ChatGPT. We test against: a) the behaviour of these AI chatbots for how they treat local, linguistic, encyclopaedic and contextual knowledge; b) the AI chatbots’ knowledge and understanding of common ground; c) the AI chatbots’ understanding of different kinds of speech acts; d) reference resolution across a discourse; e) implicatures and inferencing. We address the following questions: 1) To what degree does an AI chatbot engage with pragmatic processes? 2) Can an AI chatbot mimic the human cognitive architecture? 3) To what extent can an AI chatbot be trusted with the quality of the knowledge it produces?

Designing chatbots that provide a good user experience, guide users to their goals, and are inclusive and accessible is vital to ensure that key public and commercial services are available to a wide range of users who hold varying social norms and exhibit different patterns of social interaction. However, chatbot design can be undermined by a lack of consideration of user needs and LLMs that are not fine-tuned to account for these differences. Using examples from the testing and development of a medical appointment booking chatbot, this paper showcases how the established methods, approaches and insights of intercultural pragmatics can be used to optimize the dialogue design of chatbots and improve the intent recognition of the language models which drive them. Specifically, this work draws on the CCSARP framework for classification of requests and the GAAFFE framework, which were applied to both naturally occurring and simulated data. The paper also discusses the boundaries and limitations of these approaches.