How AI Translation Works

Open the exhibit by posing the everyday miracle: you point your phone at a foreign sentence and seconds later it reads like it was always in your language. Promise visitors they will peek inside the machine that does this.

• Translation feels instant, but inside it is a long chain of clever tricks
• Word-for-word swapping famously fails
• By the end of the exhibit you will see the whole pipeline

Visitors type a short sentence and watch it pass through three different translation machines: an old rule-based dictionary translator, a statistical phrase-matcher, and a modern neural translator. The contrast reveals why neural won.

• Rule-based: dictionary lookups plus grammar rules
• Statistical: choose the most likely phrase match
• Neural: read the whole meaning, then write

Bridge from history to the modern pipeline. Explain that before a machine can translate, it must chop text into tokens and place each token as a dot inside a giant map of meaning called embedding space.

• Text is split into tokens
• Tokens are turned into coordinate vectors
• Words with similar meanings land near each other

Visitors enter a sentence and watch it light up as tokens. The corresponding dots then appear on a 2D map where related words huddle together, making the abstract idea of an embedding suddenly visible.

• Tokens are highlighted on the sentence
• Their embeddings appear as points on a map
• Clusters show that the machine senses meaning

A hands-on station where visitors drag the stages of an encoder-decoder pipeline into the correct order, then press play to watch a sentence travel from source language to target language one step at a time.

• Encoder compresses the source meaning
• Decoder writes the target sentence
• Every output token depends on the encoded meaning

Visitors are shown a naive word-by-word translation of an idiom such as 'kick the bucket' and must tap every word pair that fools the reader. A reveal button explains why the literal swap fails.

• Idioms break word-for-word translation
• Meaning rarely maps one-to-one across languages
• Context beats vocabulary

Set up the idea of attention as a spotlight. While the decoder writes each new word, it briefly glances back at the source sentence and brightens the words it cares most about.

• The decoder writes one token at a time
• Attention weights highlight relevant source words
• The spotlight shifts with every new token

Visitors scrub a slider step by step through a translation. At each step, an animated line thickens between the current output word and the source word the model is paying most attention to, making the hidden focus visible.

• See which source word the model is focusing on
• Watch the focus jump as new words are written
• Thicker lines mean stronger attention

A candid wall of examples showing the famous weaknesses: ambiguous sentences, idioms, rare languages with little training data, and confident-sounding nonsense called hallucination.

• Many translations are equally valid
• Low-resource languages lag behind
• Models can sound sure and still be wrong

A visitor challenge: three machine translations of the same tricky sentence appear side by side. Visitors choose the best one and flag the specific error in the losers, justifying with concepts from the exhibit.

• Read translations critically, not passively
• Spot dropped meaning, awkward phrasing, and invented facts
• Explain the choice using exhibit concepts

A closing mural that strings the entire pipeline into a single flowing diagram: sentence, tokens, embeddings, encoder, attention, decoder, translated tokens. Reinforces that the same blueprint powers chatbots, summaries, and more.

• All stages connect into one pipeline
• The same pattern powers chatbots and summaries
• Translation is prediction, not lookup