NLP as a Poisson Regression Problem?

Suppose I encode the sequence to sequence dataset as a vector of integer $$\{0,1\cdots N\}$$ where $$N$$ is the maximun number of unique words in the dataset.

For example, I am planning to build a chatbot with inputs maximum of 8 words:

Q:Where can I buy these apples ? $$[8,19,23,5,76,10,2,0]$$ A:At the market downtown $$[44,7,12,90,0,0,0,0]$$

Can I use multivariate Poisson Regression to fit the data such that when I inputted an unseen sentence

New Input: Where to buy apples ? $$[8,65,32,10,2,0,0,0]$$ Output: At the market downtown $$[44,7,12,90,0,0,0,0]$$

Is there any violations in the conditions for using the MVP regression?

• The words have no inherent ordering (maybe alphabetical, but you don’t even seem to be using that), so I am skeptical that this makes sense. For instance, why is it that $market=12$ and $downtown=90?$
– Dave
Commented Oct 3, 2022 at 14:24
• @Dave I am using the predictors as an identification token through an integer which could be arbitrary (no ordering) but is consistent. Commented Oct 3, 2022 at 14:48