💬 NLP

Representing data in words: A context engineering approach

This paper introduces "wordalisations," a context engineering methodology that transforms numerical data into natural, accurate narratives, demonstrating its effectiveness across diverse domains like sports scouting and personality assessment through rigorous human and automated evaluations.

Original authors: Amandine M. Caut, Amy Rouillard, Beimnet Zenebe, Matthias Green, Ágúst Pálmason Morthens, David J. T. Sumpter

Published 2026-03-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amandine M. Caut, Amy Rouillard, Beimnet Zenebe, Matthias Green, Ágúst Pálmason Morthens, David J. T. Sumpter

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive spreadsheet filled with cold, hard numbers about a soccer player, a person's personality, or a whole country. To a data scientist, these numbers tell a story. But to the rest of us, they look like a boring wall of digits that's hard to read and even harder to understand.

This paper introduces a new tool called "Wordalisation." Think of it as a translator that turns the language of "Math" into the language of "Humans."

Here is a simple breakdown of how it works, using some everyday analogies:

1. The Problem: The "Robot" vs. The "Storyteller"

Large Language Models (like the AI you might chat with) are great at writing stories, but they often get confused by numbers. If you just ask an AI, "Tell me about Harry Kane," it might make things up because it's guessing based on what it read on the internet. If you just give it a list of numbers, it might write a robotic, boring report that doesn't make sense.

The authors wanted a way to make the AI tell a true, accurate, yet engaging story based strictly on the data.

2. The Solution: The "Four-Step Recipe"

Instead of trying to retrain the AI (which is like trying to teach a dog to speak French by changing its DNA), the authors use a clever recipe called Context Engineering. They give the AI a specific set of instructions, like a chef following a recipe card.

Here are the four steps of their recipe:

Step 1: "Tell it who it is" (The Costume)
- Analogy: Imagine you are hiring an actor. You don't just say "Act." You say, "You are a grumpy but knowledgeable British football scout."
- What it does: This sets the AI's "persona." It tells the AI exactly what role to play so it uses the right tone and vocabulary.
Step 2: "Tell it what it knows" (The Cheat Sheet)
- Analogy: Before the actor goes on stage, you give them a cheat sheet that explains the rules of the game. "If a player scores a lot, they are a 'goal machine.' If they miss a lot, they are 'struggling'."
- What it does: This gives the AI the definitions and context it needs so it doesn't hallucinate (make things up). It teaches the AI how to interpret the data correctly.
Step 3: "Tell it what data to use" (The Raw Ingredients)
- Analogy: You hand the chef the raw ingredients (the numbers). But instead of just handing over a bag of flour, you first measure it and say, "This is 2 cups of flour, which is a lot for this recipe."
- What it does: The system first converts the raw numbers into a "statistical description" (like saying "This player is 2 standard deviations above average"). This turns cold numbers into a "normative" statement (e.g., "Outstanding" or "Below Average").
Step 4: "Tell it how to answer" (The Final Polish)
- Analogy: You tell the chef, "Now, write a short, exciting paragraph for the menu that highlights the best ingredients but mentions the weak ones too. Don't use a list; write a story."
- What it does: The AI takes the "statistical description" from Step 3 and the "cheat sheet" from Step 2 to write a smooth, natural-sounding paragraph.

3. The Result: From "Z-Scores" to "Zest"

The paper tested this on three very different things:

Football Scouts: Turning Harry Kane's stats into a paragraph saying he's a "clinical finisher" but "weak on assists."
Personality Tests: Turning a test score into a description of someone's personality traits.
Country Surveys: Turning survey data about a country (like Malaysia) into a summary of its culture.

The Magic:

Visuals vs. Words: Just as a pie chart makes data easy to see, a Wordalisation makes data easy to read.
Accuracy: When they tested it, the AI using this method was much less likely to lie or make up facts compared to just asking the AI to guess or just giving it raw numbers.
Engagement: Humans preferred these stories over dry statistical reports. They were more interesting to read and just as accurate.

4. Why This Matters

The authors argue that we shouldn't just rely on "tests" to see if AI is good. Instead, we should be transparent about how we built the "recipe."

Think of it like a food label. Instead of just saying "Delicious," the paper suggests we should say: "This story was made using a specific recipe that turns numbers into words. Here is exactly how we did it, and here are the limits of what it can tell you."

The Bottom Line

Wordalisation is a way to take a spreadsheet and turn it into a compelling, honest story without losing the truth. It's like having a personal data journalist who never lies, never gets bored, and can explain complex numbers to anyone in plain English.

One Catch: Currently, it only works with numbers. It can't yet read a picture or a video file, but the authors hope to expand it in the future. For now, it's a powerful tool for turning "data" into "narrative."

1. Problem Statement

Large Language Models (LLMs) have demonstrated strong reasoning capabilities but struggle to faithfully represent numerical data in text. This limitation arises because LLMs often treat numbers as text tokens rather than quantitative information, leading to hallucinations, factual inaccuracies, or a lack of nuanced interpretation. While traditional "data-to-text" generation often relies on heavy model fine-tuning or structured inputs (like tables), existing methods frequently fail to produce engaging, natural narratives that accurately reflect statistical insights without misleading the user. The authors identify a gap in generating stylistically natural narratives that are both accurate to the data and accessible to non-technical audiences.

2. Methodology: Wordalisation

The authors propose a methodology called Wordalisation, which converts raw numerical data into descriptive, engaging text using Context Engineering rather than model fine-tuning. The process relies on a pre-trained LLM and a four-step prompt construction framework.

The Workflow

The generation pipeline consists of three stages:

Statistical Model: Raw data is processed to calculate informative statistics (e.g., Z-scores) to standardize metrics against a population mean.
Normative Model: The statistical values are converted into a "synthetic text" using a template. This step maps numerical thresholds to qualitative descriptors (e.g., a Z-score > 1.5 becomes "outstanding," -0.5 to 0.5 becomes "average"). This creates a bland but accurate intermediate text.
Wordalisation (LLM Generation): The synthetic text is fed into an LLM via a structured prompt to generate the final, natural-language narrative.

The Four-Step Prompt Engineering Framework

The core innovation lies in the specific construction of the prompt, divided into four distinct instructions:

Tell it who it is: A system prompt defining the LLM's persona (e.g., "You are a UK-based football scout"). This sets the tone and role.
Tell it what it knows: A series of Question-Answer (Q&A) pairs providing background knowledge and domain expertise (e.g., defining what a "forward" is or explaining Big Five personality traits). This ensures the LLM understands the context and limitations of the data.
Tell it what data to use: The conversion of the dataset and specific data points into the "synthetic text" generated by the Normative Model. This step translates numbers into qualitative ranges (e.g., "below average") without using raw numbers, preventing the LLM from hallucinating specific values.
Tell it how to answer: Instructions on the output structure (e.g., "3 concise sentences") combined with few-shot prompting. The prompt includes examples of statistical descriptions paired with ideal narrative responses to guide the LLM on style, tense, and emphasis.

3. Key Contributions

Concept of Wordalisation: Coined a new term for the process of abstracting data insights into descriptive text, analogous to how visualizations abstract data into graphs.
Context Engineering Framework: Demonstrated that a structured, four-step prompting strategy (incorporating persona, background knowledge, normative translation, and few-shot examples) can outperform standard prompting and fine-tuning for data-to-text tasks.
Application Versatility: Validated the method across three distinct domains:
- Football Scouting: Analyzing player performance metrics (e.g., Harry Kane's Z-scores).
- Personality Tests: Interpreting Big Five personality trait results.
- International Surveys: Summarizing sociodemographic data from the World Values Survey.
Open Source Prototype: Released a Streamlit application (wordalisations.streamlit.app) and public code repository to demonstrate the workflow and enable reproducibility.

4. Evaluation and Results

Due to the lack of standardized benchmarks for this specific subjective task, the authors employed a dual-evaluation strategy:

A. LLM-as-a-Judge

Method: Used Google's "gemini-2.0-flash" to evaluate whether generated texts contained unsupported implications compared to the ground truth statistical description.
Results:
- Control (No Data): 91% inaccuracy rate.
- Statistical Text (Raw Z-scores): 50% inaccuracy rate.
- Wordalisation (Full Prompt): 18% inaccuracy rate.
- Statistical tests confirmed that the Wordalisation method significantly increased accuracy compared to both control and raw statistical inputs ( $p < 0.001$ ).
- Reconstruction Test: The ability to reconstruct the original normative categories from the generated text was high for personality and survey data but lower for football data, suggesting that complex, condensed narratives (common in sports scouting) are harder for LLMs to reverse-engineer accurately.

B. Human-as-a-Judge

Method: 121 responses from 26 human raters evaluated 90 texts (Control, Statistical, and Wordalisation) based on Faithfulness, Engagement, and Usefulness.
Results:
- Faithfulness & Usefulness: Wordalisation and Statistical methods performed similarly and significantly better than the Control.
- Engagement: Wordalisation significantly outperformed the Statistical method ( $p = 0.04$ ). Humans preferred the natural, narrative style of Wordalisation over dry statistical descriptions, even though the statistical method was equally accurate.
- Hallucinations: Wordalisation texts contained fewer unsupported claims than the Control but were more varied and less formulaic than the Statistical method.

5. Significance and Future Directions

Shift from Benchmarking to Documentation: The authors argue that traditional NLP benchmarks are insufficient for data-to-text tasks. Instead, they advocate for Model Cards and transparent documentation of the statistical and normative models used. This aligns with "Context Engineering," focusing on best practices for providing LLMs with the right tools and instructions.
Accessibility: Wordalisation makes complex data accessible to non-experts, bridging the gap between data science and general communication (similar to data journalism).
Limitations: The current implementation relies solely on raw numerical data (Z-scores) and does not yet support categorical variables or time series. The authors note the current version is educational and not yet suitable for high-stakes professional use (e.g., recruitment) without further refinement.
Future Work: Plans include integrating more advanced statistical models (regression, PCA) and expanding the framework to handle diverse data types.

In conclusion, the paper demonstrates that Context Engineering—specifically the structured translation of data through statistical and normative models before LLM generation—produces text that is both faithful to the data and engaging to the reader, offering a robust alternative to fine-tuning for data communication tasks.