How thinking about translation can help you think about bots

Esther Seyffarth

BotSummit 2016


To view my slides on your own device, go to:

About me





  1. Some thoughts about bots and language(s)
  2. Humans navigating language(s)
  3. The Turing Test and Diversity

Bots and Language(s)

Two chatbots walk into a bar...

  • Recent article in a big German paper about two bots talking to each other: Rose and Mitsuku
  • Why were the commenters so unhappy with the bots' performance?

The problem

humans train a bot, other humans interact with the bot as it is trained

The problem

humans train a bot, somebody translates the bot output, other humans read the translated version of the bot output

Idea: Remove idioms & cultural references - just use... like... normal words

Navigating language(s)


“The limits of my language mean the limits of my world.”

Change of perspective

“The limits of my world mean the limits of my language.”

My reality shapes my language...

  • German words for refugee:
    • Flüchtling (grammatically male, slight negative connotation)
    • Asylant (grammatically male, heavily insulting term)
    • Asylsuchende (gender-neutral with neutral connotations)
    • Refugee (neutral term, no compound word)
  • Territorial naming disputes
    • Macedonia or Northern Macedonia?
    • Judea and Samaria or West Bank?
    • Taiwan or Republic of China?

... and my language shapes my bot

How will my bot know which term to use to talk about refugees?

The Turing Test

"Sounding human" is not objective

  • Commonly asked question: Does this bot sound human enough?
  • More interesting question: Why do so many prize-winning bots sound the same? Who's this mysterious default human?

Personalities of Loebner prize winners

  • American (e.g. Rose)
  • Alien/Robot (Izar, Linguo)
  • Child (Eugene Goostman)

Diversity \o/

    Make more diverse bots!