The limitation of RNN seq2seq model for numeral translation - Part 2 - Deductive reasoning about draft

[fracting]

So why does Google Translate fail on the above example? An easy explanation is Google Translate is a sentence based model, it doesn't capture context across sentences.

This is true, but could we go one step forward? Assume that we have enough computational resources to support a larger seq2seq model which can be trained and evaluated with multi-sentence examples rather than single-sentence examples, would that solve the problem?

In fact translating the word cousin correctly is extremely difficult, considering the following sub-problems involved:

• deducting according to the gender information of the two people, which might be expressed as binary logic
• deducting according to the type of family connection between the two people
• comparing the order of two natural numbers, i.e. the order of birth of Mary and Mike

[fracting]

The pseudo code describing the problem to solve:

a.age = 10
a.gender = female

b.age = 8
b.gender = male

a.parents.male.parents = b.parents.male.parents

typeOfRelationship(a, b) = ?