Machine translation

What is the first step when testing an AI translation model?

When you observe a model behaving unexpectedly, what should you do?

The detective mindset emphasizes that failures are often more informative than successes because they reveal the model's hidden _____ than successes because they reveal the model's hidden

assumptions (or "biases" / "default behaviors")

Name the two things you should do before you start testing any AI translation app.

1) Think of inputs where you know the desired result. 2) Plan to test one change at a time so you can track cause and effect.

Why is noticing when the model fails or does not fit the pattern more helpful than when it succeeds?

Failures reveal the AI's hidden assumptions, limits, and default biases—teaching you how it truly "thinks" under pressure.

The lesson shows that if the surrounding verb describes a physical action (e.g., "I saw"), the AI leans toward a _____ The lesson shows that if the surrounding verb describes a physical action (e.g., "I saw"), the AI leans toward a _____ translation.

• Physical action → direct translation

• Mental state → indirect translation

What is the technical term for the process where the AI identifies and links a pronoun back to the noun it refers to?

Coreference Resolution

Why do homonyms (like "llamas" = flames, animals, OR "to call") pose a specific threat to translation models?