AI summary by TwinMind
Why Computers Suck At Translation
Tom Scott
5m 5s
Demonstration of Translation Failures
- Speaker demonstrates translation difficulties by reading introduction after Google Translate round-trip
- Chinese translation introduced subtle errors
- French translation changed "it" to "he" for the introduction
- Simple sentences still contain translation errors despite Google being "one of the best in the world"
Common Misconceptions About Translation
- Monolingual speakers (including speaker himself) assume translation is simple
- Belief that dictionary lookup + grammar adjustment = functional sentence
- Reality: very few real-world sentences are actually simple enough for this approach
Google's Vector Space Approach
- Google released paper on machine translation using vector space mathematics
- Represents concepts in abstract language space
- Maps concepts to words effectively
- Handles homographs (identical words, different meanings) through context
- Example: "hydraulic ram" no longer mistranslated as "water sheep"
- Works well for formal technical documents
- Still struggles with casual communication
Cultural and Linguistic Nuances
British vs American English
- Not just vocabulary differences (hood vs bonnet)
- "That's a brave idea" = insult in British English (means "you're a prat, idea is impossible")
- Not compliment as in American usage
Concept Mismatches Between Languages
- Italian "buona notte" vs "buona sera"
- Both mean "good night/evening" literally
- Buona notte: specifically for bedtime
- Buona sera: for hello/goodbye after dark
- French formal/informal pronouns
- "Vous" for respectful situations
- "Tu" for casual conversation or addressing God
- Computers crush both to "you" in English, lose context for reverse translation
Complex Honorific Systems
- Korean has complicated pronoun systems
- "Opa" example from Gangnam Style
- Literal translation: woman's older brother
- Actual usage: complex, fuzzy rules based on relationship dynamics
- Psy refers to himself in third person (sounds weird when translated)
- No way to translate full meaning into single English sentence
Cultural Time Concepts
Monochronic vs Polychronic Time
- English-speaking cultures: monochronic
- 11:00 AM appointment = expected at 11:00 AM
- Exception: social gatherings ("party starts at 6:00" = people arrive 6:30-9:00)
- Polychronic cultures: flexible time approach
- Two appointments same time = acceptable
- Expanded to all aspects of daily life
- Culture clash occurs when different time concepts meet
- Machine translation won't add cultural context notes (e.g., "they really mean 7 PM")
Requirements for Accurate Translation
- Beyond word-to-concept mapping
- Need understanding of:
- Social structures
- Subtext
- Nuance
- Innuendo
- Theory of mind (speaker/listener beliefs and desires)
- Translators need ability to ask original author questions
- Must verify subtleties reflect author's intention
Core Problem Analysis
- Issue isn't language messiness (computers handle messy data well, solve CAPTCHAs better than humans)
- Real problem: language relies on:
- Intent
- Shared secrets
- Group identity
- Hidden knowledge
- Machine translation useful as tool
- Trying to exceed human translation quality is "a brave idea" (using British meaning)
