Advanced Techniques for ChatGPT in education
Author’s note: I’ve started syndicating my posts from Leonfurze.com over to Medium. You’ll sometimes notice that the dates don’t quite match up — I’m reposting these verbatim so they might be out of sync with the real world. Either way, I hope you enjoy the posts!
In a previous post on Practical Strategies for ChatGPT, I explored some of the basics of prompting the AI language model. I also organised education tasks into six areas:
- Planning
- Refreshing
- Improvising
- Personalising
- Collaborating
- Communicating
In this post I’ll go deeper into prompt techniques, and I’m organising around five key skills that could be used to lift your prompts in any of those task areas:
- Layering
- Outlining
- Transferring
- Simplifying
- Analysing
The skills you learn in this post can be applied to any of the tasks from the previous — lesson and unit planning, personalised learning, producing lesson resources, writing emails — and anything else you can imagine.
Before you go on, if you’re not familiar with the basics of ChatGPT then you should check out this post. To help decide whether you want to use ChatGPT at all, read this post about the complex ethics of AI.
Layering
Sometimes an input into ChatGPT gets you exactly what you want on the first try. That can be true for simple prompts, or where you have provided an example of the output you’d like. More often than not, however, it’s not quite right.
Prompt layering means building up a series of prompts to craft the final output. Remember that, at its core, ChatGPT has been trained as a chatbot. This means it is designed to work in dialogue, including a limited memory of around 4000 “tokens” (bits of a word).
Prompt layering makes use of this memory function and allows you to refine and perfect the model’s final output. Layering is a great technique for brainstorming and refining ideas.
Here’s an example:
The first prompt yielded an unsatisfactory result, so I added in some information from the Project Zero website. With that new information in the system, I can start applying the thinking routine:
Note that I didn’t prompt it how to sort. It has inferred from the original instruction that we want to “ideas according to how central or tangential they are” to the main concept. For this thinking routine, we really want more of a spectrum. Let’s adjust the prompt:
Much better. I’m not saying I agree with the model’s ranking, but it seems to have correctly judged that data privacy is less of an environmental concern than energy consumption. Two more layers to go in our GSCE routine.
Practical applications for prompt layering:
- Gradually refining a material to use as a resource to fit with specific criteria, such as creating a generic story and then layering in prompts to add specific language features or imagery
- Working in dialogue with a student, passing ideas back and forth to refine a piece of work
- Correcting the model when it makes mistakes or fabricates information
- Gradually building out ideas in a brainstorming session, for example using this Generate — Sort — Connect — Elaborate thinking routine
- Taking an idea from start to finish, for example following through a design thinking process
Outlining
Outlining is a powerful ChatGPT prompt technique that makes use of prompt layering and allows you to focus on building ideas in the direction you want. It is a great way to compensate for ChatGPT’s tendency to wander off-track or misinterpret prompts, as it gives you a greater control.
You can request basic outlines (such as dot-points and numerical lists), but I like to trigger another feature: markdown formatting. Markdown is a “lightweight code” that can be used to format text. I stumbled across it when I was trying to figure out why ChatGPT can sometimes produce tables, and sometimes says it can’t. The answer: it’s not technically producing a table, but rather formatting the text in tabular format with markdown (which to be honest seems like a bit of AI pedantry…).
I wrote about markdown in a LinkedIn post, with examples, so I won’t go into it here. For outlining, I like prompts like “use roman numerals and markdown to format headings and subheadings” or “use markdown to bold important key words in the list”.
Outlining becomes more powerful when you start to manipulate the output. For example, you can use shorthand to re-order the original list, combine items, or remove them entirely:
Once you have an outline you’re happy with, you can start to build out those points. Follow the techniques from “layering” above to keep adding depth to the different sections. Remember — ChatGPT can’t and shouldn’t do all the thinking for you here. It’s output will reach a point where it becomes generic, inaccurate, or both. Before that happens, take the outline and start adding in your own content.
Transferring
Text Style Transfer (TST) is a technique in Natural Language Processing which essentially means taking the form, style, and “voice” of one text and applying it to the output.
With ChatGPT, you can leverage its ability for Style Transfer in a few ways. For example, you could use the “role play” technique I’ve discussed elsewhere to aim for a particular style:
Or you could use the style you would like as an example in your prompt:
Or you could prime ChatGPT with questions and prompt layering to get it to “understand” the style you’re looking for before the final output:
Produce lesson suggestions for IIIA which follow the SOLO taxonomy. Display the output as a table using markdown.
Here is a table with lesson suggestions that align with the SOLO taxonomy for Section III.A “Electronic Waste and Data Privacy: Production and Disposal of Electronic Devices”:
SOLO LevelLearning ObjectiveEssential QuestionVocabularyActivityPre-StructuralStudents will become familiar with the concept of electronic waste.What is electronic waste?Electronic wasteIntroduction to electronic waste through images and real-life examples. Group discussion to identify examples of electronic waste.Uni-StructuralStudents will be able to identify the different types of electronic waste.What are the different types of electronic waste?Electronic waste, e-wasteGroup discussion and categorization of different types of electronic waste. Creation of a visual representation of e-waste categories.Multi-StructuralStudents will understand the dangers of improper disposal of electronic waste.Why is it important to properly dispose of electronic waste?Electronic waste, e-waste, disposal, hazardous materialsResearch and discussion on the dangers of improper disposal of electronic waste. Creation of a visual representation of the dangers.RelationalStudents will understand the relationship between the production of electronic devices and the generation of electronic waste.How does the production of electronic devices contribute to electronic waste?Electronic waste, e-waste, productionResearch and discussion on the relationship between production and electronic waste. Creation of a visual representation of the relationship.Extended AbstractStudents will be able to make informed decisions about their own electronic waste practices.How can I reduce my own contribution to electronic waste?Electronic waste, e-waste, disposal, reductionResearch and discussion on ways to reduce electronic waste. Creation of a personal plan for reducing electronic waste.Unedited output copied from ChatGPT. Note how it has “remembered” the earlier UDL prompt and merged the UDL template and the SOLO taxonomy, and has produced the markdown table.
Simplifying
Language models are great at taking complex input and summarising or simplifying the text. In fact, although ChatGPT hit the headlines and raised public awareness, the technology for summarising and simplifying text has been around for a long time. For a couple of years, students have been using tools like Writesonic, Spinbot and Wordtune to use GPT-2 and 3 for paraphrasing.
With ChatGPT, it just became even easier to get consistent responses. It is also possible to use a combination of techniques such as style transfer and outlining to get simplified text in various outputs. For example:
Explain this at a grade 8 level:
Summarise the key points from this text as a dot-point list:
Simplify this article. Rewrite the article so that it is aimed at 13 year olds. Do not summarise or explain the article. Keep the content but reduce the complexity.
Simplified version of To succeed in an AI world, students must learn the human traits of writing — Lucinda McKnight (2021)
There are many applications of this beyond creating abstracts for complex texts. I spent a lot of time as an English teacher finding suitable texts for students at various levels. Often, in a single class, there might be a reading-age difference of many years. ChatGPT and other models can be used to take a single text and adjust it up or down accordingly, making it much easier to differentiate the materials.
Analysing
The final technique is analysis, which covers not only numerical data, but semantic and sentiment analysis. In fact, as I write, a notification popped up to say that ChatGPT has improved its mathematical faculties, so you can imagine that data analysis is going to continue to improve.
Semantic analysis means analysing a text for its meaning. There’s an excellent guide at MonkeyLearn for understanding how it works, but without getting too complicated it is the process by which Natural Language Processing can extract meaning and “make sense” of words and phrases. It’s the reason that ChatGPT can do all of the things above, from writing outlines to understanding your prompts.
Semantic analysis is also the “Google killer” feature — internet search that can accurately interpret meaning is the Holy Grail for search engine development. Google will release its own model as soon as they can guarantee its accuracy, and you can bet that the ability to make sense and meaning will be a key feature.
Sentiment analysis is a sub-feature of semantics. It specifically means analysing the ‘affective’ or emotional content of texts. It’s by no means perfect, but it can be used to help analyse data to look for positive and negative feedback, the mood of a group of people (such as business customers), and so on.
Here are a few examples of semantic and sentiment analysis that might be useful in education:
Provide a sentiment analysis of these comments. Percentage positive, percentage negative, and three dot-points summarising the overall feelings of the class:
What is the tone of this article?
Other uses could include breaking down staff feedback after a review, analysing feedback from parents, getting a “big picture” view of student assessment comments, and anything else where large chunks of text require analysing for general themes and mood.
Conclusion
My primary focus is on exploring the potential of ChatGPT for teachers, rather than students. While there are various benefits for students such as using ChatGPT for purposes other than cheating, I believe that teachers should first understand how to effectively use technology like ChatGPT. This will ensure that when students begin using it in the classroom, teachers are equipped to handle the outcomes and challenges that may arise.
I’ve found it difficult to put together “how to use ChatGPT posts”. Not because it’s hard to find use cases for education: the exact opposite. It’s hard to identify places where language models can’t help in some way. That’s why this post has focused on techniques, rather than specific use cases.
2023 is going to be an exciting year for AI in education. By going in with a creative but critical mindset, teachers will be able to make the most of these new technologies.
Practical Strategies for ChatGPT in Education ran as a live webinar in February. To access a recording of the webinar, click here:
If you’ve got questions about how to use ChatGPT, or would like to have a conversation, then get in touch using the form below:
