Bad Translator

Proposal

We plan to make use of randomness and indeterminacy to generate comedic situations by playing a telephone-like game with the audience and making use of improv comedy.  The game of telephone is played by having one person whisper a phrase in someone's ear, who then listens and whispers what they think the person said into someone else's ear and so on - and at the end of the sequence of people, the last person says the phrase out loud, often resulting in a garbled version of the original phrase with humorous results.  This kind of interaction is similar to the bad translator, which allows people to input a phrase to be translated into many different languages and eventually back to the original language. 

Machines inherit human biases in translation.  We plan to show and play on these biases to show that computers are more human than we give them credit for due to being programmed by humans. Comedic situations can arise from both human mistranslation and computer mistranslation (when trying to spread a message, things get lost in translation due to differences in ways of “understanding”).

For our project, we can either make use of existing technologies such as Google Image search or generate images based off of a phrase to add a visual component to this game and project the results on a screen.  We can then have the performers react to or riff from what's presented to them and use it as the input for the next round.   In this way, the piece will be interactive and create different results every time.

Intention

Our project intend to explore the role of technology in communication in modern days. As we human understand/interpret a message in very diverse ways, computers/machines inherit such biases. By amplifying comedic randomness of machine translation, we want to show the flaw and absurdity in such human-computer-human interaction.

Performance

The creation of this project was largely in two parts: the creation of the translations, and the creation of the interface and game that displayed them.

The translations, for the most part, were made before any of us had ever enrolled in this class. One of the members of the group had been putting songs through several layers of google translate for fun for the past couple years and displaying them online. Most of the songs were too obscure for this audience, though of those that were recognizable to the average American, Anaconda by Nicki Minaj and Low by Flo Rida were selected for the project since they were deemed the funniest. Closer by The Chainsmokers feat. Halsey was the only one made specifically for this project. Also made for this project (albeit not used since there were funnier options) were Call Me Maybe by Carly Rae Jepsen, Wildest Dreams by Taylor Swift, and Friday by Rebecca Black.

The translations were made by the following process: The songs were each translated in five batches that went through 10, 15, 26, 15, or 22 layers of languages before returning to english. Each layer of translations was made using Yandex Translator. The languages were divided into two groups. Primary languages are languages that Yandex has been translating for a long time, and are therefore translated more accurately, like Spanish, Norwegian, Afrikaans, etc. Secondary languages are the remainder which are either new to the site or in beta, like Hill Mari, Punjabi, and Tajik. These languages are more prone to dramatic errors that can spice up translations, but in excess lead to changes so drastic that it seems more confusing than funny, since nothing ties together the original and translation. In each batch, a list is made of all the primary language and three randomly selected secondary languages. From this list a random sequence of languages are chosen for translation, which are then manually translated through Yandex Translator. When the five batches are complete, then for each line of the original, the funniest translation from each batch is chosen to go into the master translation, which is the final product. The reason why there are five batches of different levels of translation is that often lines are changed minimally in translation or are translated so much that all that is left is one word or just a punctuation mark. Five batches stops most of these cases from going into the translation. Also, different lines need different amounts of translation, since if a line goes through too few translations, it isn't changed enough to be as funny as it could be, but if it goes through too many, it is shortened drastically. The differences in number of translations between batches ensures that different lines get something close to the necessary happy medium. After doing dozens of translations in the past, these numbers were found to be close to what an average good number of translations is for any given song.

Context

We wanted to play with the concept of humor derived from a computer program that everyone knows about but isn't made for humorous purposes. We are very aware that others have played with bad translations, as seen in engrish memes and the work of youtuber Malinda Kathleen Reese (shown below), who puts popular songs through several layers of google translate and then performs them in music videos. However we put our own spin on it by making it into a game to identify the original, which we had not seen previously.

Process

We originally approached the exercise with the idea of a telephone/bad translator game, but we didn't know how to approach it.  We didn't think a game of telephone would be feasible given the large number of people in the class and the quiet nature of the game except at the end.  If we chose a subset of people, it would limit participation options and not be very fun to watch.  We also weren't very sure how to incorporate the technology/media aspect of the performance.

Upon talking to Daragh during the first cut critique sessions, we decided that a better way to go to achieve our comedic situation and incorporate technology would be to build a translation and have a computer voice read it out loud.  Simply translating a piece from English to other languages and back through an automatic translation service created a very similar effect to the telephone game with a garbled outcome that was often humorous just to read, and the extra layer of the computer voice gave it more character.

However, we now needed a way to involve the audience because we wanted participation and play to be part of our performance to emphasize the connection between human and computers.  What better way to achieve that than a game?  Our original intent was to do a live translation through a randomly generated set of languages, but we soon found that funny results were only produced under certain situations that weren't able to be replicated with pure randomness, so we decided to create translations for a few select songs and have the audience try to guess what they were originally.  Through this interaction, we hoped that we could involve people organically through a voting system.  We also needed a way to delve into the game and give the computer voice more of a character, so we decided to make a small script to lead in and introduce the audience to the voice so it didn't seem disconnected from the rest of the performance.

Evaluation

We feel like the translations themselves were effectively humorous, and we captured the essence of the loss of translation we wanted to get, though perhaps the presentation was not the most effective, since the audience didn't respond to it as well as we hoped. One theory about this is that part of the humor came from seeing the connection between an original line and its translation, and not showing them simultaneously made us lose this connection. Another is that the beginning of our performance wasn't clear (looking at a random webcomic, having the computer voice get the attention of the performer to play a game).  The audience was also much more enthusiastic than we had anticipated (we thought nobody would want to participate and we'd have to force them), and the lack of effective rehearsal time showed with the many technical obstacles we encountered with the sound equipment and timing.

If we had to do it again, we would probably ditch the identification game idea and display the original text and the translated text more simultaneously.  We would also practice and refine the script and make the computer voice (Alex) more of an identifiable character, perhaps with a visual representation or with autotuning when he is "singing".

Group Reflection

We learned that:

• There is a man in our computers that can speak to us.
  
• Interfacing with online translators is a very slow process
• Some texts produce more humorous bad translations than others
• Random sentences are much more amusing with context, and we should always make that context clear.
• Practicing extensively (and in person) beforehand is imperative to performance pieces.
• Adding more layers doesn't necessarily make a performance better; in fact, it complicates a performance and muddies the point if not well-integrated (we had a separate script and game that weren't well-integrated)

Sources 

Created HTML/JavaScript Game using resources from:

http://materializecss.com/

http://jquery.com/

Gif from Giphy:

https://media.giphy.com/media/v6hlgfuFftgMU/giphy.gif

Translation by Yandex:

https://translate.yandex.com/

Also momentarily showed Buzzfeed:

https://www.buzzfeed.com/

References

Malinda Kathleen Reese, connoisseur of bad translation, as shown above.

Telephone game and bad translator, for inspiring our interactivity.