Chapter 10 of 17
Summarization, Task Automation and AI for Learning Content
Covers summarisation (extractive and abstractive), task automation with triggers, AI steps and human checks, and AI-generated learning content pitched at different learner levels. Includes ROUGE scoring and the Shiksha copilot pilot in Karnataka.
In this chapter
10.1 Summarisation
Summarisation reduces a long text to its key points while keeping the meaning.
| Type | How it works | Strength | Weakness |
|---|---|---|---|
| Extractive | Selects and joins the most important sentences from the original | Wording is faithful to the source | Can be choppy; may miss connections |
| Abstractive | Writes new sentences that capture the meaning | Fluent and concise | May add or distort facts (hallucination) |
Generative AI tools produce abstractive summaries by default. To reduce risk, ask the model to quote or cite the source section for each key point.
Types of summary to ask for
- Executive summary: the conclusion and recommendation first, for decision-makers.
- Bullet summary: five to seven key points for quick reading.
- Structured summary: fixed headings, for example Problem, Method, Findings, Limitations for a research paper.
- Audience-specific summary: the same report summarised differently for a CEO, an engineer and a customer.
- Comparative summary: what two documents agree and disagree on.
- Progressive summary: a 300-word, then 100-word, then one-line version.
Example prompt for a research paper
Summarise the attached paper for a final-year B.Tech student in 200 words.
Use headings: Problem, Method, Key Result (with the main number), Limitation.
Use only information in the paper. If a heading has no information, write "Not stated".
Summarising long material. If a document is longer than the context window, split it into sections, summarise each, then summarise the summaries. This "map-reduce" approach is how many tools handle books and long reports.
10.2 Task automation with AI
Task automation means letting software carry out repetitive steps with little or no human action. Generative AI adds the ability to handle unstructured text such as emails, forms and documents.
Building blocks of an automated workflow
| Block | Meaning | Example |
|---|---|---|
| Trigger | The event that starts the workflow | A new email arrives; a form is submitted |
| Action | A step the system performs | Create a spreadsheet row; send a message |
| AI step | A generative model reads or writes text | Classify the email; draft a reply; extract fields |
| Condition | A rule that decides the path | If category is "Urgent", notify the manager |
| Human check | A person approves before anything irreversible | Manager approves the reply before sending |
Common no-code tools: Microsoft Power Automate, Zapier, Make, and Google Apps Script for Workspace users. Each offers AI steps that call a language model.
Example workflow: handling admission enquiries
- Trigger: a new enquiry arrives through the college website form.
- AI step: extract the name, course of interest and city, and classify the question as Fees, Eligibility, Hostel or Other.
- Action: add a row to the enquiries sheet.
- AI step: draft a reply using only the approved FAQ.
- Human check: the admission officer reviews and sends.
- Condition: if the enquiry is not answered within 24 hours, send a reminder to the officer.
Personal task automation with AI
- Turning a messy list of notes into a prioritised to-do list.
- Drafting a weekly schedule from deadlines and fixed commitments.
- Creating reusable templates for routine documents.
- Converting a meeting's action items directly into calendar tasks.
Cautions: never let an AI step send money, delete data or send external messages without human approval; log every automated action; and keep personal data inside approved systems.
10.3 Generative AI in education: creating content
| Content type | Example prompt | Teacher's check |
|---|---|---|
| Lesson notes | "Prepare lesson notes on Ohm's law for Class 10 with two real-life examples and a summary box." | Accuracy, syllabus fit |
| Worked examples | "Give three solved numericals on simple interest, increasing in difficulty." | Re-solve every numerical |
| Analogies | "Explain a computer's RAM and storage using a kitchen analogy." | Does the analogy mislead? |
| Case studies | "Write a short case of a small Kanpur leather exporter facing a currency change, for an MBA class." | Realism; mark as fictional |
| Translation | "Translate these notes into simple Hindi for first-generation learners." | Terminology |
| Presentation slides | "Create an outline for 10 slides on the water cycle." | Visual accuracy |
10.4 Summarisation and explanation for different learners
One topic can be explained at many levels. This helps teachers support mixed-ability classes.
Example: explaining inflation at three levels
| Level | Prompt | What a good answer looks like |
|---|---|---|
| Class 6 | "Explain inflation to an 11-year-old using a samosa price example." | Short sentences; a samosa that cost ₹10 now costs ₹15, so the same money buys less |
| Undergraduate | "Explain demand-pull and cost-push inflation for a B.Com student." | Definitions, causes, one example each |
| Postgraduate | "Explain how the RBI's inflation-targeting framework uses the repo rate, for an MA Economics student." | Policy mechanism, CPI target and tolerance band, limitations |
Techniques that help learners
- Feynman technique: "Explain this as if to someone with no background, then list any gaps in the explanation."
- Socratic tutoring: "Do not give me the answer. Ask me guiding questions until I solve it."
- Misconception check: "List three common misconceptions students have about photosynthesis and how to correct each."
- Revision aid: "Turn this chapter into ten flashcards with a question on one side and the answer on the other."
10.5 Benefits and risks in education
| Benefits | Risks |
|---|---|
| Saves teacher preparation time | Factual errors in notes and numericals |
| Personalised explanations at the learner's level | Students may copy instead of learning |
| Content in regional languages | Bias or culturally unsuitable examples |
| Round-the-clock help for students | Over-reliance; weaker independent thinking |
| Quick generation of practice material | Privacy of student data |
UNESCO's guidance. UNESCO's Guidance for Generative AI in Education and Research (2023) calls for a human-centred approach, data-privacy protection, and an age limit of 13 for independent classroom use of these tools.
10.6 Worked numericals
Numerical 10.1: Evaluating a summary with ROUGE.
ROUGE measures overlap between a machine summary and a human reference summary.
Reference: "the bank raised interest rates to control inflation" (8 words) Candidate: "the bank increased interest rates to fight inflation" (8 words)
Shared words (unigrams): the, bank, interest, rates, to, inflation, which is 6.
Word pairs (bigrams): the reference has 7 and the candidate has 7. Shared bigrams are "the bank", "interest rates" and "rates to", which is 3.
The candidate means almost the same thing, yet ROUGE-2 is only 0.43 because it uses different words. ROUGE rewards matching words, not matching meaning, so human judgement is still needed.
Numerical 10.2: Compression ratio.
A 3,000-word report is summarised in 300 words.
Numerical 10.3: Break-even of an automation.
Setting up the admission-enquiry workflow takes 6 hours. It saves 5 minutes on each of 30 enquiries a week.
After about two and a half weeks, the automation has paid back its setup time.
Numerical 10.4: Teacher time saved (from a reported pilot).
In the Shiksha copilot pilot (below), a lesson plan that took a teacher up to 40 minutes by hand took about 10 minutes with AI. A teacher preparing 5 plans a week for 40 weeks saves:
10.7 Industry use cases
- Shiksha copilot, Karnataka. Microsoft Research India and the Sikshana Foundation built an AI assistant that drafts lesson plans, activities and assessments in English and Kannada for government schoolteachers. In a pilot with 30 teachers in 30 schools, teachers reported lesson planning falling from up to 40 minutes, or more, to about 10 minutes (Microsoft Source Asia, February 2024).
- Khanmigo, Khan Academy. In March 2023 Khan Academy began piloting Khanmigo, a GPT-4-based tutor that guides students with questions rather than giving answers, and helps teachers plan lessons (THE Journal).
- Meeting and document summaries in workplaces. Summarising emails, meetings and documents is among the most-used features in Microsoft 365 Copilot and Gemini for Workspace, as the UK government trial in Chapter 9 also found.
10.8 Review questions
- Compare extractive and abstractive summarisation. Which carries more hallucination risk and why?
- Explain the map-reduce approach to summarising a document longer than the context window.
- Design an automated workflow, with trigger, AI steps, conditions and human checks, for handling leave applications in a company.
- Write prompts to explain "compound interest" at Class 8, B.Com and MBA levels.
- Compute ROUGE-1 recall for reference "students must submit the assignment by Friday" and candidate "the assignment must be submitted by Friday". (Answer: shared words must, the, assignment, by, Friday = 5 of 7 reference words, recall ≈ 0.71.)
- What does UNESCO recommend on generative AI in education?
Knowledge sources
- Lin, C.-Y., ROUGE: A Package for Automatic Evaluation of Summaries, ACL Workshop, 2004.
- UNESCO, Guidance for Generative AI in Education and Research, 2023.
- Microsoft, India's schoolteachers are drafting better lesson plans faster, thanks to a copilot, 8 February 2024.
- THE Journal, Khan Academy Pilots GPT-4 Powered Tool Khanmigo for Teachers, 14 March 2023.
- Microsoft, Power Automate documentation; Zapier, AI automation guides.
Independent study notes for learning purposes. Not affiliated with any university or course provider. Facts and case-study figures are linked to their original sources. ChatGPT, Gemini, Copilot, Jasper, Synthesia, Runway, Lumen5 and other product names are trademarks of their respective owners.