Let's Pivot!

For this problem, you will need to use the information on page (1/3) of problem (C). Otherwise, problems (C) and (D) are not related — you don't need to solve either one in order to solve the other.

That data is reproduced here. Each row is a language; each column is an object, and the cell gives the colour word that language uses for it.

Imagine that you know a way to translate reliably from language A to language B, and also a way to translate from B to language C. Then you also know how to translate from A to C: just translate from A to B first, and next from B to C. When a machine translation system uses this strategy, B is called a pivot language.¹

NACLO Labs is about to release its groundbreaking new app, NacloTranslate. NacloTranslate can translate between more language pairs than any of its competitors, and it does so by using pivot languages.

Unfortunately, NacloTranslate doesn't always translate totally reliably. In a study, NACLO Labs scientists tried translating ebi from Ejagham to Seri, using Karajá as a pivot language. They hoped to see the app give only the answers kʔe'el and kwassool, which are the results that could be produced from direct (no pivot) translation. But it was observed producing each of these three results on different test runs:

D1. Below are three more NacloTranslate test runs, from a different study. A pivot language was used for exactly one of the three language pairs—which one? Give a possibility for what the pivot language could be.

D2. NACLO Labs is considering its options for having NacloTranslate translate paban from Tsafiki to English.

a. What are the results of a direct translation? Select all possible answers.
b. What are the results if Karajá is used as a pivot? Select all possible answers.
c. NACLO Labs selects a different pivot language, and the results for paban are green, blue, and black (and no others). Which languages could NACLO Labs be using as pivot? Give all three possible answers.

D3. Here are six test runs of NacloTranslate, from a different study. From the options below, identify the pivot language used in each test run by matching options A-F to test runs 1-6. Each option was used once.

Options: A. Apinaye   B. Bassa   C. Bété   D. English   E. Teribe   F. Tsafiki

¹ This strategy can be useful when training data is hard to find. Modern machine translators may need to be trained on huge amounts of parallel text — that is, text in one language given alongside its translation in another (much like in some NACLO problems!). Sometimes, A-C parallel text will be hard to find, but A-B and B-C parallel text will be much more common.