Natural Language Visualization and the Future of Data Analysis and Presentation

has been like classical art. We used to commission a report from our data analyst—our Michelangelo—and wait patiently. Weeks later, we received an email with a magnificent hand-carved masterpiece: a link to a 50-KPI dashboard or a 20-page report attached. We could admire the meticulous craftsmanship, but we could not change it. What’s more: we could not even ask follow-up questions. Neither the report nor our analyst, since she was already busy with another assignment.

That’s why the future of data analysis does not belong to an ‘analytical equivalent’ of Michelangelo. It is probably closer to the art of Fujiko Nakaya.

Fujiko Nakaya is famous for her fog ‘sculptures’: breathtaking, living clouds of fog. But she doesn’t ‘sculpt’ the fog herself. She has the idea. She designs the concept. The actual, complex work of building the pipe systems and programming the water pressure to produce fog is done by engineers and plumbers.

The paradigm shift of Natural Language Visualization is the same.

Imagine that you need to understand a phenomenon: client churn increasing, sales declining, or delivery times not improving. Because of that, you become the conceptual artist. You provide the idea:

What were our sales in the northeast, and how did that compare to last year?

The system becomes your master technician. It does all the complex painting, sculpting, or, as in Nakaya’s case, plumbing in the background. It builds the query, chooses visualizations, and writes the interpretation. Finally, the answer, like fog in Nakaya’s sculptures, appears right in front of you.

Computer, analyze all sensor logs from the last hour. Correlate for ion fluctuations.

Do you remember the bridge of the Enterprise starship? When Captain Kirk needed to research a historical figure or Commander Spock needed to cross-reference a new energy signature, they never had to open a complex dashboard. They spoke to the computer (or at least used the interface and buttons on the captain’s chair) [*].

There was no need to use a BI app or write a single line of SQL. Kirk or Spock needed only to state their need: ask a question, sometimes add a simple hand gesture. In return, they received an immediate, visual or vocal response. For decades, that fluid, conversational power was pure science fiction.

Today, I ask myself a question:

Are we at the beginning of this particular reality of data analysis?

Data analysis is undergoing a significant transformation. We are moving away from traditional software that requires endless clicking on icons, menus, and windows, learning querying and programming languages or mastering complex interfaces. Instead, we are starting to have simple conversations with our data.

The goal is to replace the steep learning curve of complex tools with the natural simplicity of human language. This opens up data analysis to everyone, not just experts, allowing them to ‘talk with their data.’

At this point, you are probably skeptical about what I have written.

And you have every right to be.

Many of us have tried using ‘the modern era’ AI tools for visualizations or presentations, only to find the results were inferior to what sometimes even a junior analyst could produce. These outputs were often inaccurate. Or even worse: they were hallucinations, far away from the answers we need, or are simply incorrect.

This isn’t just a glitch; there are clear reasons for the gap between promise and reality, which we will address today.

In this article, I delve into a new approach called Natural Language Visualization (NLV). In particular, I will describe how the technology actually works, how we can use it, and what the major challenges are that still need to be solved before we enter our own Star Trek era.

I recommend treating this article as a structured journey through our existing knowledge on this topic. A sidenote: this article also marks a slight return for me to my earlier posts on data visualization, bridging that work with my more recent focus on storytelling.

What I found in the process of writing this particular piece—and what I hope you’ll discover while reading, too—is that this subject seemed perfectly obvious at first glance. However, it quickly revealed a surprising, hidden depth of nuance. Eventually, after reviewing all the cited and non-cited sources, my own reflections, and carefully balancing the facts, I arrived at a fairly unexpected conclusion. Taking this systemic, academic-like approach was a true eye-opener in many ways, and I hope it will be for you as well.

What is Natural Language Visualization?

A critical barrier to understanding this field is the ambiguity of its core terminology. The acronym NLV (Natural Language Visualization) carries two distinct, historical meanings.

• Historical NLV (Text-to-Scene): The older field of generating 2D or 3D graphics from descriptive text [1],[2].
• Modern NLV (Text-to-Viz): The contemporary field of generating data visualizations (like charts) from descriptive text [3].

To maintain precision and allow you to cross-reference ideas and analysis presented in this article, I will use a specific academic methodology used in the HCI and visualization communities:

• Natural Language Interface (NLI): Broad, overarching term for any human-computer interface that accepts natural language as an input.
• Visualization-oriented Natural Language Interface (V-NLI): It is a system that allows users to interact with and analyze visual data (like charts and graphs) using everyday speech or text. Its main purpose is to democratize data by serving as an easy, complementary input method for visual analytics tools, ultimately letting users focus entirely on their data tasks rather than grappling with the technical operation of complex visualization software [4],[5].

V-NLIs are interactive systems that facilitate visual analytics tasks through two primary user interfaces: form-based or chatbot-based. A form-based V-NLI typically uses a text box for natural language queries, sometimes with refinement widgets, but is generally not designed for conversational follow-up questions. In contrast, a chatbot-based V-NLI features a named agent with anthropomorphic traits—such as personality, appearance, and emotional expression—that interacts with the user in a separate chat window, displaying the conversation alongside complementary outputs. While both are interactive, the chatbot-based V-NLI is also anthropomorphic, possessing all the defined chatbot characteristics, whereas the form-based V-NLI lacks the human-like traits [6].

The value proposition of V-NLIs is best understood by contrasting the conversational paradigm with traditional data analysis workflows. These are presented in the infographic below.

This shift represents a move from a static, high-friction, human-gated process to a dynamic, low-friction, automated one. I further illustrate how this new approach could impact how we work with data in Table 1.

Table 1: Comparative Analysis: Traditional BI vs. Conversational Analytics

How does V-NLI work?

To analyze the V-NLI mechanics, I adopted the theoretical framework from the academic survey ‘The Why and The How: A Survey on Natural Language Interaction in Visualization’ [11]. This framework offers a powerful lens for classifying and critiquing V-NLI systems by distinguishing between user intent and dialogue implementation. It dissects two major axes of the V-NLI system: ‘The Why’ and ‘How’. ‘The Why’ axis represents user intent. It examines why users interact with visualizations. The ‘How’ axis represents dialogue structure. It answers the question of how the human-machine dialogue is technically implemented. Each of these axes can be further divided into specific tasks in the case of ‘Why’ and attributes in the case of ‘How’. I list them below.

The four key high-level ‘Why’ tasks are:

1. Present: Using visualization to communicate a narrative, for instance, for visual storytelling or explanation generation.
2. Discover: Using visualization to find new information, for instance, writing natural language queries, performing keyword search, visual question answering (VQA), or analytical conversation.
3. Enjoy: Using visualization for non-professional goals, such as augmentation of images or description generation.
4. Produce: Using visualization to create or record new artifacts, for instance, by making annotations or creating additional visualizations.

The ‘How,’ on the other hand, has three major attributes:

1. Initiative: Answers who drives the conversation. It can be user-initiated, system-initiated, or mixed-initiated.
2. Duration: How long is the interaction? It might be a single turn for a simple query, or a multi-turn conversation for a complex analytical discussion.
3. Communicative Functions: What is the form of the language? The language model supports several interaction forms: users may issue direct commands, pose questions, or engage in a responsive dialogue in which they modify their input based on suggestions from the NLI.

This framework can also help illustrate the most fundamental issue causing our disbelief in NLI. Historically, both commercial and non-commercial Visual Natural Language Interfaces (V-NLIs) operated within a very narrow functional scope. The ‘Why’ was often reduced to Discover task, while the ‘How’ was limited to simple, single-turn queries initiated by the user.

As a result, most ‘talk-to-your-data’ tools functioned as little more than basic ‘ask me a question’ search boxes. This model has proven consistently frustrating for users because it is overly rigid and brittle, often failing unless a query is phrased with perfect precision.

The entire history of this technology is the story of growth in two key ways.

• First, our interactions have been improving, moving from asking just one question at a time to having a full, back-and-forth conversation.
• Second, the reasons for using V-NLIs have been expanding. We have progressed from simply finding information to having the tool automatically create new charts for us, and even explain the data in a written story.

Working using fully all four tasks of ‘Why’ and three attributes of ‘How’ in the future will be the biggest leap of all. The system will stop waiting for us to ask a question and will start the conversation itself, proactively pointing out insights you may have missed. This journey, from a simple search box to a smart, proactive partner, is the main story connecting this technology’s past, present, and future.

Before going further, I would like to make a small course deviation and show you an example of how our interactions with AI could improve. For that purpose I will use a recent post published by my friend Kasia Drogowska, PhD, on LinkedIn.

AI models often become stereotyped, suffering from ‘mode collapse’ because they learn our own biases from their training data. A technique called ‘Verbalized Sampling’ (VS) offers a powerful solution by changing the prompt. Instead of asking for one answer (like ‘Tell me a joke’), you ask for a probability distribution of answers (like ‘Generate five different jokes and their probabilities’). This simple shift not only yields 1.6-2.1x more diverse and creative results but, more importantly, it teaches us to think probabilistically. It shatters the illusion of a single ‘correct answer’ in complex business decisions and puts the power of choice back in our hands, not the model’s.

The image above displays a direct comparison between two AI prompting methods:

• The left side exemplifies direct prompting. On this side I show what happens when you ask the AI the same simple question five times: ‘Tell me a joke about data visualization.’ The result is five very similar jokes, all following the same format.
• The right side exemplifies verbalized sampling. Here I show a different prompting method. The question is changed to ask for a range of answers: ‘Generate five responses with their corresponding probabilities…’ The result is five completely different jokes, each unique in its setup and punchline, and each assigned a probability by the AI (as a matter of fact, it is not true probability, but anyway gives you the idea).

The key benefit of a method like VS is diversity. Instead of just getting the AI’s single ‘default’ answer, it forces the AI to explore a wider spectrum of creative possibilities, letting you choose from the most common to the most unique. This is a perfect example of my point: changing how we interact with these tools can yield very different results.

The V-NLI pipeline

To understand how a V-NLI translates a natural language query, such as ‘show me last quarter’s sales trend,’ into a precise and accurate data visualization, it is necessary to deconstruct its underlying technical architecture. Academics in the V-NLI community have proposed a classic information visualization pipeline as a structured model for these systems [5]. To illustrate the general mechanism of the process, I prepared the following infographic.

For a single ‘text-to-viz’ query, the two most critical and challenging stages are (1) Query Interpretation and (3/4) Visual mapping/encoding. In other words, it is understanding exactly what the user means. The other stages, particularly (6) Dialogue Management, become paramount in more advanced conversational systems.

The older systems consistently failed to grasp this understanding. The reason is that this task is essentially solving two problems instantly:

• First, the system must guess the user’s intent (e.g., is the request to compare sales or to see a trend?).
• Second, it must translate casual words (like ‘best sellers’) into a perfect database query.

If the system misunderstood the user’s intent, it would display a table when the user wanted a chart. If it couldn’t parse user’s words, it would just return an error, or worse, make up something out of the blue.

Once the system understands your question, it must create the visual answer. It should automatically select the best chart for the given intent (e.g., a line chart for a trend) and then map appropriate characteristics to it (e.g., placing ‘Sales’ on the Y-axis and ‘Region’ on the X-axis). Interestingly, this chart-building part evolved in a similar way to the language-understanding part. Both transitioned from old, clunky, hard-coded rules to flexible, new AI models. This parallel evolution set the stage for modern Large Language Models (LLMs), which can now perform both tasks simultaneously.

In fact, the complex, multi-stage V-NLI pipeline described above, with its distinct modules for intent recognition, semantic parsing, and visual encoding, has been significantly disrupted by the advent of LLMs. These models have not just improved one stage of the pipeline; they have collapsed the entire pipeline into a single, generative step.

Why is that, you may ask? Well, the parsers of the previous era were algorithm-centric. They required years of effort by computational linguists and developers to build, and they would break upon encountering a new domain or an unexpected query.

LLMs, in contrast, are data-centric. They offer a pre-trained, simplified solution to the most difficult problem in understanding natural language [13],[14]. This is the great unification: a single, pre-trained LLM can now execute all the core tasks of the V-NLI pipeline simultaneously. This architectural revolution has triggered an equivalent revolution in the V-NLI developer’s workflow. The core engineering challenge has undergone a fundamental shift. Previously, the challenge was to build a perfect, domain-specific semantic parser [11]. Now, the new challenge is to create the ideal prompt and curate the perfect data to guide a pre-trained LLM.

Three key techniques power this new, LLM-centric workflow. The first is Prompt Engineering, a new discipline focused on carefully structuring the text prompt—sometimes using advanced strategies like ‘Tree-of-Thoughts’—to help the LLM reason through a complex data query instead of just making a quick guess. A related method is In-Context Learning (ICL), which primes the LLM by placing a few examples of the desired task (like sample text-to-chart pairs) directly into the prompt itself. Finally, for highly specialized fields, Fine-Tuning is used. This involves re-training the base LLM on a large, domain-specific dataset. These pillars, when in place, enable the creation of a powerful V-NLI that can handle complex tasks and specialized charts that would be impossible for any generic model.

This shift has profound implications for the scalability of V-NLI systems. The old approach (symbolic parsing) required building new, complex algorithms for every new domain. The latest LLM-based approach requires a new dataset for fine-tuning. While creating high-quality datasets remains a significant challenge, it is a data-scaling problem that is far more solvable and economical than the previous algorithmic-scaling problem. This change in fundamental scaling economics is the true and most lasting impact of the LLM revolution.

What is the true meaning of this?

The single biggest promise of ‘talk-to-your-data’ tools is data democratization. They are designed to eliminate the steep learning curve of traditional, complex BI software, which often requires extensive training. ‘Talk-to-your-data’ tools provide a zero-learning-curve entry point for non-technical professionals (like managers, marketers, or sales teams) who can finally get their own insights without having to file a ticket with an IT or data team. This fosters a data-driven culture by enabling self-service for common, high-value questions.

For the business, value is measured in terms of speed and efficiency. The decision lag of waiting for an analyst, lasting days or sometimes weeks, is eliminated. This shift from a multi-day, human-gated process to a real-time, automated one saves an average of 2-3 hours per user per week, allowing the organization to react to market changes instantly.

However, this democratization creates a new and profound socio-technical tension within organizations. The below anecdote illustrates this perfectly: an HR Business Partner (a non-technical user) used one of these tools to present calculations to managers. The managers, however, started discussing… the way we got to the calculation instead of the actual conclusions, because they didn’t trust that HR could ‘actually do the math.’

This reveals the critical conflict: the tool’s primary value is in direct tension with the organization’s fundamental need for governance and trust. When a non-technical user is suddenly empowered to produce complex analytics, it challenges the authority of the traditional data gatekeepers, creating a conflict that is a direct consequence of the technology’s success.

Which current LLM-based AI assistant is the best as a ‘talk-to-your-data’ tool?

You might expect to see a ranking of the best assistants using LLMs for V-NLI here, but I chose not to include one. With numerous tools available, it’s impossible to review them all and rank them objectively and in a competent manner.

My own experience is mainly with Gemini, ChatGPT, and built-in assistants like Microsoft Copilot or Google Workspace. Still, using a few online sources, I’ve put together a brief overview to highlight the key factors you should evaluate when selecting the option that’s most suitable for you. In the end, you’ll need to explore the possibilities yourself and consider aspects such as performance, cost, payment model, and—above all—safety.

The table below outlines several tools with short descriptions. Later, I focus especially on Gemini and ChatGPT, which I know best.

Table 2. Examples of LLMs that could serve as V-NLI

Both Gemini and ChatGPT exemplify the new wave of powerful, visualization-oriented V-NLIs, each with a distinct strategic advantage. Gemini’s primary bonus is its deep integration within the Google ecosystem; it works directly with BigQuery and Google Suite. For example, you can open a PDF attachment directly from Gmail and perform a deep analysis using the Gemini assistant interface, using either a pre-built agent or ad-hoc prompts. Its core strength lies in translating simple, everyday language not just into data points, but directly into interactive visualizations and dashboards.

ChatGPT, in contrast, can serve as a more general-purpose yet equally powerful V-NLI for analytics, capable of handling various data formats, such as CSVs and Excel files. This makes it an ideal tool for users who want to make informed decisions without diving into complex software or coding. Its Natural Language Visualization (NLV) function is explicit, allowing users to ask it to summarize data, identify patterns, or even generate visualizations.

The true, shared strength of both platforms is their ability to handle interactive conversations. They allow users to ask follow-up questions and refine their queries. This iterative, conversational approach makes them highly effective V-NLIs that don’t just answer a single question, but enable a full, exploratory data analysis workflow.

Application example: Gemini as V-NLI

Let’s do a small experiment and see, step by step, how Gemini (version 2.5 Pro) works as a V-NLI. For the purpose of this experiment, I used Gemini to generate a set of artificial daily sales data, split by product, region, and sales representative. Then I asked it to simulate an interaction between a non-technical user (e.g., a sales manager) and a V-NLI. Let’s see what the outcome was.

Generated data sample:

Date,Region,Salesperson,Product,Category,Quantity,UnitPrice,TotalSales
2022-01-01,North,Alice Smith,Alpha-100,Electronics,5,1500,7500
2022-01-01,South,Bob Johnson,Beta-200,Electronics,3,250,750
2022-01-01,East,Carla Gomez,Gamma-300,Apparel,10,50,500
2022-01-01,West,David Lee,Delta-400,Software,1,1000,1000
2022-01-02,North,Alice Smith,Beta-200,Electronics,2,250,500
2022-01-02,West,David Lee,Gamma-300,Apparel,7,50,350
2022-01-03,East,Carla Gomez,Alpha-100,Electronics,3,1500,4500
2022-01-03,South,Bob Johnson,Delta-400,Software,2,1000,2000
2023-05-15,North,Eva Green,Alpha-100,Electronics,4,1600,6400
2023-05-15,East,Frank White,Epsilon-500,Services,1,5000,5000
2023-05-16,South,Bob Johnson,Beta-200,Electronics,5,260,1300
2023-05-16,West,David Lee,Gamma-300,Apparel,12,55,660
2023-05-17,North,Alice Smith,Delta-400,Software,1,1100,1100
2023-05-17,East,Carla Gomez,Epsilon-500,Services,1,5000,5000
2024-11-20,South,Grace Hopper,Alpha-100,Electronics,6,1700,10200
2024-11-20,West,David Lee,Beta-200,Electronics,10,270,2700
2024-11-21,North,Eva Green,Gamma-300,Apparel,15,60,900
2024-11-21,East,Frank White,Delta-400,Software,3,1200,3600
2024-11-22,South,Grace Hopper,Epsilon-500,Services,2,5500,11000
2024-11-22,West,Alice Smith,Alpha-100,Electronics,4,1700,6800

Experiment:

My typical workflow starts with a high-level query for a broad overview. If that initial view looks normal, I might stop. However, if I suspect an underlying issue, I’ll ask the tool to dig deeper for anomalies that aren’t visible on the surface.

Next, I focused on the North region to see if I could spot any anomalies.

For the last query, I shifted my perspective to analyze the daily sales progression. This new view serves as a launchpad for subsequent, more detailed follow-up questions.

As a matter of fact, the above examples were fairly simple and not far away from the ‘Old-era’ NLIs. But let’s see what happens, if the chatbot is empowered to take initiative during the discussion.

This demonstrates a more advanced V-NLI capability: not just answering the question, but also providing context and identifying underlying patterns or outliers that the user might have missed.

This small experiment hopefully demonstrates that AI assistants, such as Gemini, can effectively serve as V-NLIs. The simulation began with the model successfully interpreting a high-level natural-language query about sales data and translating it into an appropriate visualization. The process showcased the model’s ability to handle iterative, conversational follow-ups, such as drilling down into a specific data segment or shifting the analytical perspective to a time series. Most significantly, the final experiment demonstrated proactive capability, in which the model not only answered the user’s query but also independently identified and visualized a critical data anomaly. This indicates that such AI tools can transcend the role of simple executors, acting instead as interactive partners in the data exploration process. But it’s not that they will do that on their own: they must first be empowered through an appropriate prompt.

So is this world really so ideal?

Despite the promise of democratization, V-NLI tools are plagued by fundamental challenges that have led to their past failures. The first and most significant is the Ambiguity Problem, the ‘Achilles’ heel’ of all natural language systems. Human language is inherently imprecise, which manifests in several ways:

• Linguistic ambiguity: Words have multiple meanings. A query for ‘top customers’ could mean top by revenue, volume, or growth, and a wrong guess instantly destroys user trust.
• Under-specification: Users are often vague, asking ‘show me sales’ without specifying the time frame, granularity, or analytical intent (such as a trend versus a total).
• Domain-specific context: A generic LLM might be useless for a specific business because it doesn’t understand internal jargon or company-specific business logic [16], [17].

Second, even if a tool provides a correct answer, it is socially useless if the user cannot trust it. This is the ‘Black Box’ problem, as cited above in the story of the HR business partner. Because the HR user couldn’t explain the ‘why’ behind the ‘what,’ the insight was rejected. This ‘chain of trust’ is critical. When the V-NLI is an opaque black box, the user becomes a ‘data parrot,’ unable to defend the numbers and rendering the tool unusable in any high-stakes business context.

Finally, there is the ‘Last Mile’ problem of technical and economic feasibility. A user’s simple-sounding question (e.g., ‘show me the lifetime value of customers from our last campaign’) may require a hyper-complex, 200-line SQL query that no current AI can reliably generate. LLMs are not a magic fix for this. Even to be remotely useful, they must be trained on a company-specific, prepared, cleaned, and properly described dataset. Unfortunately, this is still an enormous and recurring expense. This leads to the most important conclusion:

The only viable path forward is a hybrid future.

An ungoverned ‘ask anything box’ is a no-go.

The future of V-NLI is not a generic, all-powerful LLM; it is a flexible LLM (for language) operating on top of a rigid, curated semantic model (for governance, accuracy, and domain-specific knowledge) [18], [19]. Instead of ‘killing’ BI and dashboards, LLMs and V-NLI will be the opposite: a powerful catalyst. They won’t replace the dashboard or static report. They will enhance it. We should expect them to be integrated as the next generation of user interface, dramatically improving the quality and utility of data interaction.

What will the future bring?

The future of data interaction points toward a hypothetical paradigm shift, moving well beyond a simple search box to a Multi-Modal Agentic System. Imagine a system that operates more like a collaborator and less like a tool. A user, perhaps wearing an AR/VR headset, might ask, ‘Why did our last campaign fail?’ Then the AI agent would reason over all available data. Not just the sales database, but also unstructured customer feedback emails, the ad creative images themselves, and website logs. Instead of a simple chart, it would proactively present an augmented reality dashboard and offer a predictive conclusion, such as, ‘The creative performed poorly with your target demographic, and the landing page had a 70% bounce rate.’ The crucial evolution is the final ‘agentic’ step: the system wouldn’t stop at the insight but would bridge the gap to action, perhaps concluding:

I have already analyzed Q2’s top-performing creatives, drafted a new A/B test, and alerted DevOps to the page-load issue.

Would you like me to deploy the new test? Y/N_

As scary as it may sound, this vision completes the evolution from simply ‘talking to data’ to actively ‘collaborating with an agent about data’ to achieve an automated, real-world outcome [20].

I realize this last statement opens up even more questions, but this seems like the right place to pause and turn the conversation over to you. I’m eager to hear your opinions on this. Is a future like this realistic? Is it exciting, or frankly, a little scary? And in this advanced agentic system, is that final human ‘yes or no’ truly necessary? Or is it the safety mechanism we will always want / need to keep? I look forward to the discussion.

Concluding remarks

So, will conversational interaction make the data analyst—the one who painstakingly writes queries and manually builds charts—jobless? My conclusion is that the question isn’t about replacement but redefinition.

The pure ‘Star Trek’ vision of an ‘ask anything’ box will not happen. It is plagued by its ‘Achilles’ heel’ of human language ambiguity and the ‘Black Box’ problem that destroys the trust it needs to function. Hence, the future, therefore, is not a generic, all-powerful LLM.

Instead, the only viable path forward is a hybrid system that combines the flexibility of an LLM with the rigidity of a curated semantic model. This new paradigm doesn’t replace the analysts; it elevates them. It frees them from being a ‘data plumber’. It empowers them as a strategic partner, working with a new, multi-modal agentic system that can finally bridge the chasm between data, insight, and automated action.

References

[1] Priyanka Jain, Hemant Darbari, Virendrakumar C. Bhavsar, Vishit: A Visualizer for Hindi Text – ResearchGate

[2] Christian Spika, Katharina Schwarz, Holger Dammertz, Hendrik Lensch, AVDT – Automatic Visualization of Descriptive Texts

[3] Skylar Walters, Arthea Valderrama, Thomas Smits, David Kouřil, Huyen Nguyen, Sehi L’Yi, Devin Lange, Nils Gehlenborg, GQVis: A Dataset of Genomics Data Questions and Visualizations for Generative AI

[4] Rishab Mitra, Arpit Narechania, Alex Endert, John Stasko, Facilitating Conversational Interaction in Natural Language Interfaces for Visualization

[5] Shen Leixian, Shen Enya, Luo Yuyu, Yang Xiaocong, Hu Xuming, Zhang Xiongshuai, Tai Zhiwei, Wang Jianmin, Towards Natural Language Interfaces for Data Visualization: A Survey – PubMed

[6] Ecem Kavaz, Anna Puig, Inmaculada Rodríguez, Chatbot-Based Natural Language Interfaces for Data Visualisation: A Scoping Review

[7] Shah Vaishnavi, What is Conversational Analytics and How Does it Work? – ThoughtSpot

[8] Tyler Dye, How Conversational Analytics Works & How to Implement It – Thematic

[9] Apoorva Verma, Conversational BI for Non-Technical Users: Making Data Accessible and Actionable

[10] Ust Oldfield, Beyond Dashboards: How Conversational AI is Transforming Analytics

[11] Henrik Voigt, Özge Alacam, Monique Meuschke, Kai Lawonn and Sina Zarrieß, The Why and The How: A Survey on Natural Language Interaction in Visualization

[12] Jiayi Zhang, Simon Yu, Derek Chong, Anthony Sicilia, Michael R. Tomz, Christopher D. Manning, Weiyan Shi, Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity

[13] Saadiq Rauf Khan, Vinit Chandak, Sougata Mukherjea, Evaluating LLMs for Visualization Generation and Understanding

[14] Paula Maddigan, Teo Susnjak, Chat2VIS: Generating Data Visualizations via Natural Language Using ChatGPT, Codex and GPT-3 Large Language Models – SciSpace

[15] Best 6 Tools for Conversational AI Analytics

[16] What are the challenges and limitations of natural language processing? – Tencent Cloud

[17] Arjun Srinivasan, John Stasko, Natural Language Interfaces for Data Analysis with Visualization: Considering What Has and Could Be Asked

[18] Will LLMs make BI tools obsolete?

[19] Fabi.ai, Addressing the limitations of traditional BI tools for complex analyses

[20] Sarfraz Nawaz, Why Conversational AI Agents Will Replace BI Dashboards in 2025

[*] Star Trek analogy was generated in ChatGPT, might not accurately reflect the characters’ actions in the series. I haven’t watched it for roughly 30 years .

Disclaimer

This post was written using Microsoft Word, and the spelling and grammar were checked with Grammarly. I reviewed and adjusted any modifications to ensure that my intended message was accurately reflected. All other uses of AI (analogy, concept, image, and sample data generation) were disclosed directly in the text.