In the seminal paper A Neural Probabilistic Language Model, Yoshua Bengio and his colleagues make the following point:

If one wants to model the joint probability distribution of 10 consecutive words in a natural language with a vocabulary $V$ of size $100,000$, there are potentially $100,000^{10}-1$ free parameters.

I guess it's related to degrees of freedom and joint distributions but I just can't get my hands on the exact formula that was used here to come up with $100,000^{10}-1$.

  • $\begingroup$ $100000$ is probably an estimate of the total number of words in the language. $\endgroup$ – kjetil b halvorsen Dec 1 '15 at 14:16
  • $\begingroup$ @kjetilbhalvorsen 1e5 is the size of the vocabulary, so yes, 1e5 is the total number of unique words in the language, no problem here $\endgroup$ – Antoine Dec 1 '15 at 14:21

The estimate $100 000^{10}-1$ comes from assuming a discrete model for the $10$ consecutive words, without any simplifications or restrictions, thus using all interactions up to and including order $10$.

It is not important that the words are consecutive, we would get the same count for any ten specified word positions. For each position, it can be any of the $100000$ words, so we need that number of probabilities. So you can build up a cube in $10$-space, wity each dimension cut up in $100000$ boxes. Taking all the combinations, that give $100000^{10}$ boxes, each box giving on possible $10$-word sequence, such as " am I writing now holy blue crap green integrated ideas", which would be a sequence of fairly low probability. Then subtract $1$ to account for the fact that the probabilities must sum to one!

In practice, a fairly high number of the probabilities would be zero, because they correspond to non-grammatical utterings.

| cite | improve this answer | |
  • $\begingroup$ please elaborate. Is the fact that words are consecutive important? How does joint distributions and degrees of freedom come into play? What is the formula that is used? $\endgroup$ – Antoine Dec 1 '15 at 14:30
  • $\begingroup$ Many thanks for elaborating. In other words, $100,000^{10}$ gives all the possible 10-element combinations from the vocabulary of size $100,000$. This number is huge but still finite since we're in the discrete case. And estimating the joint probability mass function of a specific 10-word sequence comes down to assigning a probability (probabilities=parameters here I assume) to each portion of that huge but finite discrete sample space. Is that correct? Also, sorry if I'm being dumb but I still don't get why subtracting 1 ensures that the probabilities will sum to one. $\endgroup$ – Antoine Dec 1 '15 at 15:34
  • $\begingroup$ subtracting one does not by itself secure that the probabilities will sum to one, but it takes care of that restriction . After freely choosing the $100000^{10}-1$ probabilities, you can calculate the last one! $\endgroup$ – kjetil b halvorsen Dec 1 '15 at 15:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.