Introducing Human-Centeredness in AI-Assisted Lexicography

arXiv:2607.11808 · cs.CL, cs.AI · Submitted 2026-07-13 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Introducing Human-Centeredness in AI-Assisted Lexicography".

Jane: The gist This paper proposes a human-centered artificial intelligence (HCAI) framework for AI-assisted lexicography, arguing that AI should augment rather than replace lexicographers by focusing on four interrelated dimensions:

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper called "Introducing Human-Centeredness in AI-Assisted Lexicography." We’re starting by talking about who wrote it and what the title actually signals to us. It’s by Antonio San Martín and Catherine Trekker from the University of Quebec.

Jane: That title sounds very specific, Tom. What does "Human-Centeredness" mean in this context? I'm picturing a system that prioritizes human feelings over machine performance, which is something we need to unpack for our listeners.

Lu: It’s about putting human needs and well-being at the center of AI development itself. It’s challenging the old way of thinking where you just focus on how fast a machine can work, regardless of what that does to the people using it.

Meng: So this paper isn't just about making a faster dictionary; it’s about rethinking the whole relationship between us and these new AI systems in language work. That suggests a deeper kind of consideration than just adding features to an existing tool.

Tom: Exactly, Meng. It means we need to look at the entire process, not just the final product, when we integrate this technology into our jobs. The authors are setting up this whole conceptual structure around how AI should interact with human expertise and language knowledge.

Jane: So if I had to put it in simple terms for someone driving right now, it means instead of a car that just goes as fast as possible, we design a system where the driver is always in control and the car's speed adapts to *their* needs on the road.

Lu: That’s a good way to think about it. It frames AI not as an independent entity trying to solve problems, but as an extension of human capability within our specific professional context.

Tom: Right, and this framework identifies four main areas where we need to examine this relationship: the augmented lexicographer, the sociotechnical context of AI integration, bias in it, and how we design the tools themselves.

Jane: Those four dimensions sound very comprehensive. It covers everything from the person using it to the environment around them and even how we build the actual software. That’s a lot to digest at once for our listeners.

Meng: I think that comprehensive approach is what makes this paper interesting from an engineering side too, because it forces us to look beyond just the code and into the real-world impact on people.

Lu: It’s about creating a set of principles that guide development so that we don't accidentally build something that undermines the value of human linguistic and cultural knowledge.

Tom: So, as we move forward, we need to keep this title in mind—it sets the standard for how AI should be developed in fields like lexicography moving forward.

The paper's summary: Jane: Now that we know what it’s called, let’s get into the actual substance of "Introducing Human-Centeredness in AI-Assisted Lexicography." What is the paper actually summarizing?

Tom: Basically, they argue that generative AI has a huge potential to help lexicographers access information in new ways and be more efficient. But they immediately point out that this power comes with serious risks if we aren't careful about how we use it.

Lu: The core summary is that GenAI chatbots can fulfill user needs by letting people ask complex, contextual questions and get customized answers flexibly, which is a big deal for accessing lexical information quickly.

Meng: But they immediately follow that up by saying these tools are prone to errors and hallucinations, which means the output isn't always reliable or accurate when it comes to precise language.

Jane: So the summary is highlighting this tension: we have powerful new ways to get knowledge, but we have to acknowledge the pitfalls like errors and inconsistency before we even start using them widely.

Tom: And they frame this tension by proposing a whole framework—the HCAI one—to manage it. It’s not just about the tool itself; it’s about how we integrate that tool into our professional lives responsibly.

Jane: So, instead of just saying "AI is good" or "AI is bad," they are proposing a structured way to think about the integration process itself. That seems like a mature way to approach this kind of technology adoption.

Lu: They are essentially mapping out the relationship between AI and lexicography across those four dimensions we talked about, showing how automation should serve augmentation instead of just replacement.

Tom: It really sets up a clear structure for understanding the debate, moving it away from a simple "AI vs. Human" argument toward a more nuanced discussion about collaboration.

Meng: That’s helpful because it moves the conversation out of abstract philosophy and into concrete areas where we can actually build better tools and workflows for real people.

Jane: So, to put it plainly, the paper summarizes that AI is a powerful new assistant for finding words, but we need a careful plan—a human-centered plan—to make sure that assistant actually helps us be better lexicographers.

The paper's improvements: Tom: Okay, so we’ve established what the paper is saying about the framework and the risks. Now let’s look at what they actually propose as solutions, as outlined in "Introducing Human-Centeredness in AI-Assisted Lexicography."

Jane: What are these proposed improvements? Are they just minor tweaks to existing tools, or are they suggesting a complete overhaul of how we think about building these things?

Lu: They suggest concrete improvements centered around high automation and strong human control. The idea is that this balance between machine and human intervention shouldn't be decided for the whole task, but rather at the level of individual subtasks.

Meng: That sounds like a practical way to manage the workload so we don't get overwhelmed by a single massive AI suggestion. It’s about micro-control over the workflow.

Tom: They also push hard on amplification over emulation again—AI should be used to extend what we already do, not try to perfectly copy it. This is where leveraging large datasets for pattern detection comes in handy, as they suggest.

Jane: So instead of asking AI to write a whole entry from scratch, we use it to find patterns in huge text samples and then we apply our human knowledge on top of that foundation. That feels like using the machine’s speed wisely.

Lu: Then there’s the part about designing tools around the needs of lexicographers first, which means the interface has to be built specifically for us, not just for generic AI processing. And again, involving lexicographers in designing those tools is a big point.

Tom: And they also suggest specific design elements to keep control high. They propose using things like confirmation dialogs—that little pause that makes you reassess the AI suggestion before you click anything.

Jane: That intentional friction sounds like a necessary feature to combat that feeling of blindly accepting what the machine spits out, especially when dealing with something as nuanced as language definition.

Meng: And they also talk about things like role-aware permissions and version control, which would let us override AI suggestions if we feel they’re off track for our specific task or version of the work. That gives us back that agency I mentioned earlier.

Lu: Plus, they advocate for using Retrieval-Augmented Generation, or RAG systems where the AI pulls data from corpora that *we* select ourselves, which makes the output traceable and verifiable.

Tom: And to tie in with those design points, they suggest delaying the AI suggestions until after we’ve done our initial analysis to avoid that early anchoring effect on our expert reasoning.

Jane: So, it’s a multi-layered approach: controlling how much automation happens, ensuring the tool fits our needs, and building in checks and balances to keep us in the driver's seat throughout the entire process.

Conclusion: Tom: We’ve gone through the framework, looked at how they identify bias, and discussed all these practical design improvements in "Introducing Human-Centeredness in AI-Assisted Lexicography." So what’s the final message we should take away from this paper?

Jane: The main point is that advancing a human-centered approach requires sustained research. They aren't claiming the problem is solved, but they are laying out exactly where we need to focus our efforts next.

Lu: They emphasize that this isn't just a single fix; it’s a sustained empirical research agenda needed to keep pushing these HCAI principles into practice for language work.

Meng: I think the implication is that the future of AI in this field depends entirely on us, the professionals, actively shaping how these systems are developed and controlled.

Tom: Right. It’s a call to action for everyone involved—the researchers, the engineers, and especially us lexicographers—to keep demanding tools that prioritize our expertise over mere machine speed or cost reduction.

Jane: I think this paper provides a really strong roadmap for how we can move toward a more collaborative future with AI in language resources. It’s about building trust through transparency and control.

Lu: Ultimately, the paper seeks to provide a conceptual starting point that supports the beneficial integration of AI into lexicographic practice by keeping human needs at the very center of everything we build.

Meng: So, when we look at this work, it’s less about how powerful the AI is and more about how thoughtfully we choose to wield that power in our specific domain.

Tom: Exactly. It’s a call to keep building tools that respect the expertise and cultural heritage that make language knowledge valuable in the first place. That’s what this paper on Human-Centeredness in AI-Assisted Lexicography is all about.

University of Quebec in Trois-Rivières

cs.CL, cs.AI

Submitted: 2026-07-13

Updated: 2026-10-08

Importance score: 83/100

The gist: The gist This paper proposes a human-centered artificial intelligence (HCAI) framework for AI-assisted lexicography, arguing that AI should augment rather than replace lexicographers by focusing on

Key concepts

Augmented Lexicographer
This concept views AI as a tool that enhances human expertise rather than replacing it. It involves balancing high automation with strong human control on a task-by-task basis, allowing lexicographers to leverage AI's pattern recognition while retaining their unique world knowledge and linguistic intuition.
Amplification vs. Emulation
Instead of trying to make AI perfectly mimic human intelligence, this principle suggests AI should extend it. Lexicographers use AI for its speed in processing data, but they integrate this with their own embodied experience and deep linguistic competence to create richer results.
Bias in Lexicography
AI systems can introduce several biases, including linguistic bias (favoring certain languages), cultural bias (reflecting training data norms), and epistemic bias (overrepresenting dominant knowledge). Lexicographers must critically check AI output against evidence to ensure accuracy and fairness.
Human-Centered Design in Tools
AI tools must be designed around the needs of lexicographers, involving them actively in the design process. This ensures usability, allows users to control automation levels, and incorporates features like Retrieval-Augmented Generation (RAG) for verifiable sources.

Terminology

Summary

The gist

This paper proposes a human-centered artificial intelligence (HCAI) framework for AI-assisted lexicography, arguing that AI should augment rather than replace lexicographers by focusing on four interrelated dimensions: the augmented lexicographer, the sociotechnical context of AI integration, bias, and the design of AI-powered tools.

The Human-Centered Artificial Intelligence Framework

The paper introduces a conceptual framework for AI-assisted lexicography grounded in HCAI principles to guide the integration of artificial intelligence into this field. This framework places human needs and well-being at the center of AI development, challenging approaches that prioritize machine performance over human values. The framework argues that automation should serve to augment professional expertise rather than displace or marginalize it.

The Augmented Lexicographer

The framework identifies four interrelated dimensions through which the relationship between AI and lexicographic work can be understood and critically examined. The augmented lexicographer is envisioned through two HCAI principles:

  1. High automation and strong human control: This principle holds that high levels of automation and human control are not mutually exclusive but can coexist productively. The balance between machine and human intervention must be determined at the level of individual subtasks rather than for an entire task.

  2. Amplification rather than emulation: This principle holds that AI should not aim to replicate human intelligence. AI systems are viewed as a tool that extends human intelligence because they excel at processing large datasets and detecting patterns at great speed. Lexicographers can draw on world knowledge, embodied experience, introspective linguistic competence, and direct engagement with language communities beyond corpus evidence.

The Sociotechnical Context of AI Integration

AI integration is understood within a broader social context involving stakeholders, institutions, cultures, norms and spaces. A key concern is ensuring that AI augmentation does not dehumanize lexicographers by prioritizing profit extraction or cost reduction. For integration to remain human-centered, lexicographers must retain meaningful agency over how AI is incorporated into their workflows.

Bias in AI-Assisted Lexicography

The framework critically examines bias through several dimensions stemming from training data composition and algorithmic design. These biases include:

** Linguistic bias: This denotes the tendency of LLMs to favor certain languages or linguistic features in their outputs. Crosslinguistic interference may occur when one language influences outputs in other languages. Temporal bias refers to the risk that LLMs generate lexicographic information that is misaligned with the relevant time frame. Cultural bias denotes the tendency of LLMs to reproduce cultural assumptions, values, norms, and perspectives that dominate their training data. Epistemic bias refers to the tendency of LLMs to overrepresent dominant forms of knowledge by amplifying what is most prevalent in their training data. Social bias refers to the reproduction of prejudicial views, stereotypical representations, and discriminatory attitudes directed toward particular social groups. Lexicographers must critically evaluate AI-generated output against corpus evidence to ensure descriptive adequacy. 4. The Design of AI-Powered Lexicographic Tools**

The design of tools must be guided by human-centered design (HCD) principles. First, systems and interfaces must be designed around the needs of lexicographers. Second, lexicographers themselves must be actively involved in the conception and refinement of AI-powered tools. To preserve human control, tools should combine strong usability with features that allow users to selectively enable or disable particular system functions and calibrate automation levels depending on the task or individual preferences. Furthermore, explainability techniques like Retrieval-Augmented Generation (RAG) are valuable because they enable AI outputs to be linked to identifiable and verifiable sources. Design must also mitigate risks like reduced critical engagement through empirically calibrated intentional friction, such as confirmation dialogs that encourage lexicographers to reassess AI suggestions before validation. The paper concludes by stating that Advancing a human-centered approach to AI-assisted lexicography requires a sustained empirical research agenda.

Improvements for AI systems

  1. A system capable of operating in post-editing lexicography by automatically generating drafts while maintaining a clear human control loop, ensuring that human expertise and computational efficiency operate synergistically rather than as a zero-sum relationship.

  2. Implementation of systems designed around lexicographers' needs to prevent reverse adaptation, specifically by incorporating role-aware permissions, version control, and the capacity of lexicographers to override AI suggestions.

  3. Integration of Retrieval-Augmented Generation (RAG) systems that retrieve corpora selected by lexicographers themselves, ensuring that the output is linked to identifiable and verifiable sources rather than relying solely on potentially unreliable web data.

  4. Development of an interface featuring empirically calibrated intentional friction, such as confirmation dialogs, to mitigate the risk of reduced critical engagement by discouraging automation-induced overtrust.

  5. A mechanism that delays AI suggestions until after the lexicographer has completed an initial analysis, preventing the anchoring effect where early proposals narrow expert reasoning.

Sources

Related papers