By exploring a spell checker program, students discover that AI relies on data-driven algorithms. When we provide more representative data to the computer, data-driven algorithms generally produce a higher quality output.
Lesson Goals |
Students will be able to…
|
Student-facing Lesson Goals |
|
Materials |
|
🔗The First Spell Checker: Intelligent or Not? 15 minutes
Overview
Students consider one specific example of machine learning: evaluating how spell checkers' programmed behaviors compare and contrast with the behaviors of humans.
Launch
The spell checker built into your web browser or word processing system is a classic example of machine learning.
We are so used to receiving spell-checking help that sometimes we don’t even notice that’s it’s there. But copy-editing a document without the help of a spell checker would be laborious and error-prone!
-
Do humans and computers use similar spell checking strategies or approach the task from entirely different angles?
-
How do spell-checking outcomes from humans and computers compare?
As with discussions about other forms of AI, we might wonder whether a spell-checker exhibits any kind of "intelligence" and what intelligence actually is…
-
Complete Human Spell Checking.
-
Be prepared to summarize the method that you used for spell checking:
-
How do you locate misspelled words?
-
How do you correct misspelled words?
-
-
You probably noticed that is was difficult to articulate how exactly you were able to locate and correct spelling mistakes. Why?
When humans spell check, they rely on context clues, experience, and background. There were probably students in your class who said that the spelling mistakes jumped off the page. They couldn’t not see the mistakes!
Of course, the process of spell-checking varies not just by person, but by situation, too. For example:
-
If a complex, technical word had been misspelled you might have relied on a different spell checking strategy.
-
If you were from the U.K. rather than the U.S., you might spell some words differently from others (e.g. "color" v. "colour").
Investigate
Modern spell checkers descend from a program written over fifty years ago. Its creators provided the spell checker program with a 10,000-word dictionary and a precise algorithm to follow. After scanning the text and extracting the words it contained, the spell checker would compare each word with a known list of correctly spelled words. Here is what the algorithm did next:
-
First, it would develop possible alternatives for the misspelled word (input) by making one of the following adjustments: (1) replace a single letter with a different one, or (2) swap the positions of two adjacent letters.
-
Next, it would search its dictionary of correct spellings to see which of the alternatives were real words.
-
Finally, it would produce a list of real, correctly spelled words for the user to choose from.
Exploring this particular algorithm can offer us interesting insights about one of the earliest forms of AI/ML.
-
In Part 1 of The First Spell Checker, you will be the computer, following an early spell checker’s algorithm.
-
Reflect on the activity and how it compares to your own strategies in Part 2.
After students have completed the activity, invite them to share their responses to the three reflection questions (below) in small groups.
-
What do you think are some limitations of the first spell checker’s algorithm?
-
There is an extremely limited number of ways in which a word can be misspelled. For example, if two separate letters in a word are incorrect, the spell checker has no chance of proposing the correct word!
-
How is the first spell checker’s algorithm similar to whatever strategy you use when spell checking your writing?
-
See what your students propose, and discuss. If they come up with any similarities, we’d love to hear what they are!
-
How is the first spell checker’s algorithm different from whatever strategy you use when spell checking your writing?
-
When you check spelling, you rely a lot on context, instinct, and previously acquired knowledge.
-
The first spell checking program did not have, or use, any of these strengths.
-
In contrast, however, a human would not have the time, ability, or concentration to systematically list out thousands of potential words, whereas the computer has no difficulty doing so.
-
Therefore, the strategies used by the human and computer may be extremely different for a similar task.
Synthesize
-
In what ways do spell checkers succeed at mimicking humans?
-
Based on the above, they do not at all! Perhaps the outcomes are similar, depending on a wide variety of variables.
-
Would it be a good idea for a human to try mimic a spell checker’s algorithm? Why or why not?
-
The program we just discussed relies on computational brute force: primitive spell checking programs will repeatedly edit / transform a misspelled word, generating dozens of different words to verify in the dictionary. Were a human to undergo this process, it would be time-consuming, laborious, and highly error prone.
🔗Spell Checking in Pyret 25 minutes
Overview
Students explore both the algorithm and the datasets that power a Pyret-based spell checker, discovering that data-driven algorithms are at the heart of AI.
Launch
By now, we have a decent sense of the extensive work that is happening behind the scenes when we spell check our writing. We have not, however, discussed an essential truth:
Spell checkers, and in fact all modern AI, are "data-driven".
-
Where have you encountered the term "data-driven" before, if at all?
-
Sample responses:
-
data-driven decision making is informed by collecting or analyzing data
-
data-driven health care involves using data to think about the effectiveness of different treatments
-
data-driven teachers will reteach topics that students struggled with
-
-
Have you ever met someone who is "data-driven"? (Teachers? Coaches? Parents?) How so?
-
What do you think it means to be "data-driven"?
-
Responses will vary, but should highlight the general idea that data informs how things are done.
But how exactly is a spell checker data-driven?
Take a look at this screenshot of a text messaging app’s suggestions for the possibly misspelled word "Cose".
-
What alternative words does the text messaging app provide the user to choose from?
-
The original word "Cose" — indicating that no error was made
-
Two alternatives: "Code" and "Close"
-
What sort of algorithm do you think the app used in developing possible alternative words?
-
Responses will vary.
-
Students may refer to the algorithms discussed in the first half of the lesson.
-
Students might also imagine more complicated algorithms — for instance, algorithms that consider the proximity of letters on the keyboard!
-
What sort of data do you think the spell checking app used in developing possible alternative words?
-
Responses will vary.
-
Which words the user is most likely to type
-
The topic of the text conversation and the most probable next word
-
The dictionary of words the app draws from
Students will discuss a similar screenshot of a text messaging app in our lesson on Statistical Language Modeling: Generating Text. During that lesson, however, students explore how generative AI uses data-driven algorithms to determine what word to produce next.
Investigate
-
Let’s take a look at a spell checking program written in Pyret.
-
This program includes a built-in function called
alt-words, which implements a spell-checking algorithm similar to the algorithm you already explored. -
Open the Spell Checker Starter File and click "Run".
-
Complete A Pyret Spell Checker: The Algorithm to discover how the spell checker works.
As you were interacting with the Spell Checker Starter File, you observed that it only proposed five-letter words. This is because the dictionary it draws from is actually a dictionary from the game "Wordle"!
Are you familiar with Wordle? If not, you can quickly learn the rules and play it here. Before moving on with the lesson, be sure to check for students' familiarity with the game via a show of hands. If your students have not played Wordle before, play one round as a class before proceeding.
|
Let’s consider this partially-played Wordle game. The player has attempted three words so far: "WORTH", "MEDIA", and "GAMES". With each turn, we have learned something new. At this point, we know that:
The player has just three turns left! |
-
What word would you try next?
-
Responses will vary; keep a list of student proposals.
-
Each of the words you proposed was probably 2 edits away from "GAMES", the user’s third guess. Why?
-
Three of the letters are correct; we just need to substitute in different letters for g and s.
-
The player of this partially-completed Wordle game wants some Pyret "assistance".
They runalt-words("games", WORDS). Try it. Is Pyret able to produce the winning word? -
Pyret produces two words:
cameoandgamut. -
We know "cameo" is incorrect because it contains the (rejected) letter "o".
-
We know "gamut" is incorrect because "e" must occupy the fourth space.
Disappointingly, Pyret did not provide the correct Wordle solution. But why?
There are basically two "parameters" that our spell-checking program used:
-
the function : what is outside of the parentheses
-
the dictionary : provided inside the parentheses, as an argument
We’ve discussed some ways we could make the function better. (e.g. Maybe try swapping out an additional letter?)
But it’s also possible to improve the quality of the output without changing how the function works… by improving the dictionary argument! Let’s explore how directing Pyret to access differently-sized dictionaries influences the quality of the program’s output.
-
Complete A Pyret Spell Checker: Exploring Different Dataset Sizes using the Spell Checker Starter File.
-
If you finish early, try the two challenges at the bottom of the page.
-
What did you discover about the dictionaries in Spell Checker Starter File?
-
Why did the same
alt-wordsfunction not always return the same results?
When we offered more representative data to our rudimentary Pyret spell checker, we got better results without changing the spell checker’s code.
Synthesize
-
In this lesson, you discovered that providing more data often produces better results. Think about some of the different recommendation systems you have interacted with (e.g., YouTube, Spotify, etc). In your experience, how does the amount of data provided influence the quality of the recommendations made?
-
A brand new YouTube user has not provided any data about what sort of videos they like to watch. YouTube cannot make specific recommendations without this data! As a user watches more videos, the system collects data about the user’s interests, preferences, and more. With more data, YouTube can provide better recommendations.
These materials were developed partly through support of the National Science Foundation, (awards 1042210, 1535276, 1648684, 1738598, 2031479, and 1501927). 
Bootstrap by the Bootstrap Community is licensed under a Creative Commons 4.0 Unported License. This license does not grant permission to run training or professional development. Offering training or professional development with materials substantially derived from Bootstrap must be approved in writing by a Bootstrap Director. Permissions beyond the scope of this license, such as to run training, may be available by contacting contact@BootstrapWorld.org.