AI & Digital Ethics Grade 7 Natural Language Processing How Chatbots Work

Study Guide: Natural Language Processing – How Chatbots Work
Grade 7 | Computer Science & Digital Ethics

1. The Driving Question

"If you text a chatbot like Siri or a customer service AI, how does it actually understand what you’re saying—and why does it sometimes get your question totally wrong, like when you ask for ‘a pizza with no cheese’ and it suggests a vegan pizza instead?" What’s really happening inside the chatbot’s "brain" to turn your words into answers, and where does it go wrong?

2. The Core Idea – Built, Not Listed

Imagine you’re playing a game of Telephone with a robot. You whisper, "I need help resetting my password," but the robot doesn’t hear sounds—it only sees letters. Its job is to break your sentence into tiny pieces (like LEGO bricks), figure out what each brick means in context, and then rebuild them into a response that makes sense. That’s what Natural Language Processing (NLP) does: it’s the set of rules and tricks computers use to "read" human language.

Here’s how it works in a chatbot like the one you’d use to order pizza: 1. Tokenization: The chatbot chops your sentence into words or parts of words ("I" / "need" / "help" / "resetting" / "my" / "password").
2. Parsing: It maps how the words relate ("resetting" is the action, "password" is the thing being reset).
3. Intent Recognition: It guesses why you’re saying this (you want help, not just to chat about passwords).
4. Response Generation: It picks a pre-written answer or builds one from its training data ("I’ll send a reset link to your email").

But here’s the catch: the chatbot doesn’t understand language like you do. It’s more like a super-fast librarian who’s read millions of books but doesn’t feel the stories. If you say "I’m so mad my account’s locked!" it might miss the emotion and just reply, "Here’s how to unlock it." That’s why chatbots sometimes sound robotic—or worse, tone-deaf.

Key Vocabulary:
- Natural Language Processing (NLP)
Definition: The branch of AI that helps computers read, interpret, and respond to human language.
Example: When Google Translate turns "¿Dónde está la biblioteca?" into "Where is the library?" without a human translator.
Grade 7 Note: In high school, you’ll learn how NLP powers things like spam filters and voice assistants—but also how it can inherit biases from the data it’s trained on.

• Token
  Definition: A single unit of text (a word, punctuation mark, or even part of a word) that a computer processes.
  Example: In the sentence "AI’s cool!", the tokens are "AI", "’s", "cool", and "!".
  Grade 7 Note: Tokens can be tricky—"New York" is one token, but "New" and "York" are two. This matters for how chatbots count words!

• Intent
  Definition: The goal or purpose behind what a user says (e.g., asking a question, making a request, expressing an emotion).
  Example: If you text a weather chatbot "Is it gonna rain tomorrow?" the intent is "get weather forecast," not "complain about rain." Grade 7 Note: Misreading intent is why chatbots sometimes give weird answers—like when you say "I’m bored" and it replies "Here’s a Wikipedia article on boredom."

• Training Data
  Definition: The massive collection of text (books, websites, conversations) used to teach an AI how language works.
  Example: A chatbot trained on customer service logs might learn that "My order’s late" usually means "I want a refund," not "I’m curious about shipping times." Grade 7 Note: In high school, you’ll explore how biased training data (e.g., mostly written by one group of people) can make chatbots less accurate for others.

3. Assessment Translation

How this appears in class:
- Formative Assessments (Exit Tickets, Short Responses):
- "Explain in 2–3 sentences how a chatbot might misinterpret the sentence ‘I’m not happy with my order.’ What part of NLP is failing here?"
- Proficient Response: "The chatbot might focus on the word ‘order’ and assume the intent is to track a package, not complain. This is a problem with intent recognition—it’s missing the emotion (‘not happy’) and context."
- Developing Response: "It gets confused because it doesn’t understand feelings." (Lacks connection to NLP steps.) - "Label the tokens in this sentence: ‘Can you reset my password by 3 PM?’"
- Proficient: Lists tokens correctly ("Can", "you", "reset", "my", "password", "by", "3", "PM", "?").
- Developing: Misses punctuation or splits contractions ("Can" and "you" as one token).

• State Standardized Tests (e.g., Computer Science End-of-Course):
• Multiple Choice: "Which NLP step is most likely to fail if a chatbot replies ‘I don’t understand’ to the sentence ‘My flight’s delayed—what are my options?’"
  • Distractors:
  • "Tokenization" (too basic—chatbot can split words).
  • "Response generation" (it can generate a response, just not the right one).
  • "Parsing" (partly true, but not the main issue).
  • Correct Answer: "Intent recognition" (the chatbot isn’t sure if you’re asking for rebooking, compensation, or just venting).
• Short Answer: "A chatbot trained only on formal emails is asked, ‘Yo, why’s my account locked?’ Describe two ways its NLP might struggle and why."
  • Proficient: "1) Tokenization: It might not recognize ‘Yo’ as a greeting. 2) Intent recognition: It’s trained on formal requests like ‘Please assist me,’ so it might miss the urgency in ‘why’s.’"

Model Proficient Response (Short Answer):
Prompt: "Explain how a chatbot’s training data could lead to biased responses. Give one real-world example." Response: "If a chatbot’s training data is mostly from one country or group of people, it might not understand slang, dialects, or cultural references from others. For example, a chatbot trained mostly on American English might not recognize that ‘pants’ in the UK means ‘underwear,’ leading to confusing answers. This is a problem because the chatbot’s responses could be less helpful—or even offensive—to people outside that group."

4. Mistake Taxonomy

Mistake 1: Over-Simplifying NLP Steps
- Prompt: "Describe how a chatbot processes the sentence ‘I hate this app—fix it now!’" - Common Wrong Response: "The chatbot reads the words and knows you’re mad, then fixes the app." - Why It Loses Credit: - Misses all NLP steps (tokenization, parsing, intent).
- Treats the chatbot like a human ("knows you’re mad").
- Correct Approach: 1. Tokenization: Splits into "I", "hate", "this", "app", "—", "fix", "it", "now", "!".
2. Parsing: Identifies "hate" and "fix" as key actions, "app" as the subject.
3. Intent Recognition: Guesses the intent is "complain + demand solution" (not just venting).
4. Response: Might say, "I’m sorry to hear that. Here’s how to report a bug."

Mistake 2: Ignoring Context in Intent
- Prompt: "Why might a chatbot reply ‘Here’s a coupon’ when you say ‘I’m never shopping here again’?" - Common Wrong Response: "The chatbot is stupid and doesn’t listen." - Why It Loses Credit: - Doesn’t connect to NLP concepts.
- Assumes malice instead of explaining the mechanism (training data, intent rules).
- Correct Approach: - The chatbot’s training data likely pairs "shopping here again" with "offer discount" (e.g., "Will you shop here again?" → "Yes, if you give me a coupon").
- It misses the negation ("never") and emotion ("I’m angry"), so it defaults to a sales script.

Mistake 3: Confusing Tokens with Words
- Prompt: "How many tokens are in the sentence ‘Don’t text me after 11 PM’?" - Common Wrong Response: "7 tokens (words)." - Why It Loses Credit: - Counts words, not tokens (misses "Don’t" as two tokens: "Do" + "n’t").
- Doesn’t account for punctuation ("PM" is one token, but "11" and "PM" are two).
- Correct Approach: - Split into: "Do", "n’t", "text", "me", "after", "11", "PM".
- Total: 7 tokens (including the contraction and number-unit pair).

5. Connection Layer

1. Within Computer Science → Machine Learning
   Why it matters: NLP is a type of machine learning where the "patterns" the AI learns are language rules (e.g., "‘not happy’ usually means complain"). Understanding NLP helps you see how all ML works—it’s about finding patterns in data, whether that data is words, images, or numbers.

2. Across Subjects → Linguistics (English/Social Studies)
   Why it matters: NLP forces computers to deal with the same language quirks humans do—ambiguity ("I saw the man with the telescope"), sarcasm, and dialects. Studying NLP makes you notice how weird human language really is, which is what linguists study too.

3. Outside School → Customer Service Chatbots
   Why it matters: Next time you message a brand’s chatbot and it says "I’m sorry for the inconvenience" instead of actually helping, you’ll know it’s not trying to be useless—it’s just bad at intent recognition. You can even "hack" it by using the words its training data expects (e.g., "I want to speak to a manager" instead of "This is ridiculous").

6. The Stretch Question

"If a chatbot is trained only on conversations from 2010, how might it fail to understand a 13-year-old in 2024? Give three specific examples—and explain which NLP step would break down in each case."

Pointer Toward the Answer:
1. Slang: Words like "rizz" or "sigma" didn’t exist in 2010. The chatbot’s tokenization might not recognize them, or its intent recognition would misclassify them (e.g., "That’s so sigma" → "What’s a sigma?" instead of "I agree").
2. Memes/References: If a teen says "Skibidi Toilet moment," the chatbot’s training data has no context for this. Its response generation would default to "I don’t understand" or a generic reply.
3. Emoji as Punctuation: In 2010, emojis were rare; now, "I’m fine ?" is common. The chatbot’s parsing might treat the emoji as decoration, missing the sarcasm (intent: "I’m not fine").

The deeper question: Could a chatbot ever truly "understand" language, or is it just really good at faking it? (Spoiler: Even AI researchers argue about this!)