Introduction to Text as Data

<p class="kicker">GRA 4164 · Lecture 1 · Fall 2026</p>

## Introduction to<br>Text as Data

### Leif Anders Thorsrud & Vegard H. Larsen — BI Norwegian Business School

---

Most of the information firms sit on is **text** — not tables.

This course: turning text into **quantitative business insight**.

---

## Who we are

<div class="cols">
<div class="callout">
<h3>Leif Anders Thorsrud</h3>
<p>Department of Economics · leif.a.thorsrud@bi.no</p>
</div>
<div class="callout">
<h3>Vegard H. Larsen</h3>
<p>Data Science and Analytics · vegard.h.larsen@bi.no</p>
</div>
</div>

We research text as data — news, central bank communication, business cycles.

---

## About you

- Your **background** — what did you study before?
- **Programmed in Python before?** — yes / a little / no
- **Statistics or machine learning?** — courses / some / none
- What do you **expect** from this course?

---

## Why text as data

---

## What is text as data?

- Information captured in **written form** — news, reviews, filings, social media
- Analysed with **computational techniques**, not by reading
- Used across **finance, marketing, political science**, and more

---

## What can you do with it?

- **Predict markets** — news sentiment moves asset prices
- **Listen to customers** — sentiment in reviews at scale
- **Map politics** — policy positions from speeches
- **Detect deception** — fake news from language patterns

---

## From raw text to decisions

![Five-step pipeline from raw text to business decisions](../figures/text-as-data-workflow.svg)

---

How you turn text into **numbers** is a **modelling decision** — not cleanup.

---

## Why it's hard

---

## Four challenges

- **Scale** — billions of documents need efficient processing
- **Noise** — typos, slang, emojis, no fixed structure
- **Ambiguity** — meaning depends on context
- **Ethics** — bias, privacy, black-box decisions

---

## Ambiguity and context

<div class="cols">
<div class="callout">
<h3>"bank"</h3>
<p>Financial institution — or riverbank? Words alone don't carry meaning.</p>
</div>
<div class="callout">
<h3>"Great service…"</h3>
<p>Praise — or sarcasm? Sentiment lives in context, not words.</p>
</div>
</div>

---

## Ethics is part of the job

**Models learn the biases in their training text.** Hiring tools trained on biased CVs reproduce the bias — at scale and with a straight face.

</div>

Privacy, consent, and explainability follow the same pattern.

---

## How the course works

---

## The 12-lecture arc

![Course timeline: counting words, learning from prediction, LLMs, wrap-up](../figures/course-timeline.svg)

---

## Assessment

<div class="cols">
<div class="stat stat--blue">
<span class="stat-value">30%</span>
<span class="stat-label">Submission — 4 assignments, groups of 1–3</span>
</div>
<div class="stat">
<span class="stat-value">70%</span>
<span class="stat-label">School exam — individual, 3 h, written</span>
</div>
</div>

<p class="muted">Both graded ECTS. You present the assignments in lecture 12.</p>

---

## The four assignments

- Solved in **groups of 1–3** — three is the sweet spot
- Four **Jupyter notebooks** through the semester
- Collected into **one PDF** submission
- Presented and discussed in **lecture 12**

</div>

---

**Form your group this week.** Sign up via the group allocation on Itslearning — 1–3 students, 3 preferred.

</div>

---

## Tools

**Core**   `Python 3.11+` `Jupyter` `VS Code`

**Libraries**   `pandas` `NLTK` `scikit-learn` `spaCy` `PyTorch`

</div>

<p class="mono">gra4164.vegardlarsen.com/tools</p>

---

## Compute for the LLM part

**A course server for running language models** — BERT, Llama, Mistral and friends, on your own data — is available during the LLM part of the course.

</div>

---

## Reading

- **Jurafsky & Martin, *Speech and Language Processing*** (3rd ed. draft) — free online, the course backbone
- Selected **papers and cases** per lecture — see the reading map on the course site

---

## By November, you can

- **Choose** the right NLP method for a business problem — and say why
- **Build** text-as-data pipelines in Python, from raw text to results
- **Explain** what large language models do — and where they fail

---

## Questions

## The 12 lectures

| # | Topic | # | Topic |
|---|-------|---|-------|
| 1 | Introduction to text as data | 7 | LLMs and tokenization |
| 2 | Bag-of-words, TF-IDF, cosine | 8 | Attention and transformers |
| 3 | Bag-of-words-based methods | 9 | LLM training and fine-tuning |
| 4 | N-gram language models | 10 | Practical day 2 |
| 5 | Practical day 1 | 11 | Prompting, alignment, summary |
| 6 | Word embeddings | 12 | Presentations and feedback |

## Further free reading

- Eisenstein — *Natural Language Processing* (lecture notes)
- Goldberg — *A Primer on Neural Network Models for NLP*
- Goodfellow, Bengio & Courville — *Deep Learning*
- Tunstall, von Werra & Wolf — *NLP with Transformers*

## Scale and noise, in practice

- **Scale** — archives of news or social media run to billions of documents; storage, indexing, and parallel processing become part of the analysis
- **Noise** — typos, slang, abbreviations, emojis; informal text breaks tools built for clean prose

---

## What's next

- **Lecture 2 (Tue):** bag-of-words, TF-IDF, and cosine similarity — text becomes numbers
- **Before then:** install Python + Jupyter (see the Tools page) and run the *Working with text data* notebook
- Reading: **J&M chapter 2**