
Let’s talk about KneserNey Smoothing,

one of the most sophisticated forms of smoothing,

but also one with a beautiful and elegant intuition.

Remember that from GoodTuring, we talked about the c*’s,

the discounted counts you end up from GoodTuring,

and we discounted each of the counts —

the count of one was discounted to point four,

and the count of two discounted to 1.26 and so on —

in order to save mass to replace the zero counts with some low number.

And if you look at the actual values of these counts — 8.25 for nine and 7.24 for eight —

you’ll notice that in a very large number of cases,

the discounted count has a very close relationship with the original count.

It’s really the original count minus .75, or somewhat close to that.

So, in practice what GoodTuring often does is produce a fixed small discount from the counts.

And that intuition, that of a fixed small discount, can be applied directly.

When we do this, we call this absolute discounting, and absolute discounting is a popular kind of smoothing.

And here we’re showing you absolute discounting interpolation.

And again, the intuition is just: We’ll save some time and have to compute all those complicated GoodTuring numbers

and we’ll just subtract .75, or maybe it will be a different discount value for different corpora.

Now here’s the equation for absolute discounting.

So we’re doing diagrams again.

So the probability, absolute discounted, of a word, given the previous word,

will be some discounted bigram, interpolated with some interpolation weight, with the unigram probability.

So we have a unigram probability P(w), and then the bigram probability,

and we just subtract a fixed amount, let’s say it’s .75, from the count.

And otherwise compute the bigram probability in the normal way.

So, we have a discounted bigram probability, mixed with some weight,

which I shall talk later about how to set this weight, with a unigram.

And maybe we might keep a couple of extra values of d for counts one and two.

Counts one and two, we saw on a previous slide, weren’t quite subtracting .75,

so we can model this more carefully by having separate counts for those.

But the problem with absolute discounting is the unigram probability itself.

And I want to talk about changing the unigram probability.

And that’s the fundamental intuition of KneserNey.

So in KneserNey smoothing, the idea is keep that same interpolation that we saw in absolute discounting,

but use a better estimate of probabilities of the lower unigrams.

And the intuition for that, we can go back and look at the classic Shannon games.

Remember, in the Shannon game we’re predicting a word from previous words,

so we see a sentence, “I can’t see without my reading .”

What’s the most likely next word?

Well, “glasses” seems pretty likely.

Well, how about instead the word “Francisco”?

Well, that seems very unlikely in this situation.

And yet “Francisco”, as just a unigram, is more common than “glasses”.

But the reason why “Francisco” seems like a bad thing after “reading”,

one intuition we might be able to get is that “Francisco” always follows “San” or very often follows “San”.

So while “Francisco” is very frequent, it’s frequent in the context of the word “San”.

Now, unigrams in an interpolation model, where we’re mixing a unigram and a bigram,

are specifically useful — they’re very helpful — just in case where we haven’t seen a bigram.

So it’s unfortunate that just in the case where we haven’t seen the bigram “reading Francisco”,

we’re trusting “Francisco”’s unigram weight which is just where we shouldn’t trust it.

So instead of using the probability of w, how likely is a word,

our intuition is going to be when we’re backing off to something

we should instead use the continuation probability.

We’re going to call it P continuation of a word,

how likely is the word to appear as a novel continuation.

Well, how do we measure novel continuation?

Well, for each word we’ll just count the number of bigram types it completes.

How many different bigrams does it create by appearing after another word.

In other words, each bigram type is a novel continuation the first time we see this new bigram.

In other words, the continuation probability is going to be proportional to the cardinality of this set,

the number of words of preceding words, i minus one, that occur with our word.

So, how many words occur before this word in a bigram.

How many preceding words are there.

That will be the cardinality of that set, that’s a number we would like our continuation probability to be proportional to.

So how many times does w appear as a novel continuation?

We need to turn that into a probability.

So we just divide by the total number of word bigram types.

So, of all word bigrams that occur more than zero times, what’s the cardinality of that set?

How many different word bigram types are there,

and we’re just going to divide the two to get a probability of continuation,

of all the number of word bigram types how many of those have w as a novel continuation.

Now it turns out that there’s an alternative metaphor for KneserNey with the same equations

so again we can see the numerator as the total number of word types that precede w,

how many word types can w follow and we’re going to normalize it by the number of words that could precede all words.

So, this sum over all words of the number of word types that can precede the word.

And these two are the same:

The number of this denominator and the denominator we saw on the previous slide are the same

because the number of possible bigram types is the same as the number of word type that can precede all words summed over all words.

If you think about that for a second, you’ll realize that’s true.

So in other words, with this kind of KneserNey model,

a frequent word like “Francisco” that occurs only in one context like “San” will have a low continuation probability.

So if we put together the intuition of absolute discounting with the KneserNey probability for the lower order ngram,

we have the KneserNey smoothing algorithm.

So, for the bigram itself we just have absolute discounting —

We take the bigram count, we subtract some d discount,

and I’ve just shown here that we take the max of that and zero

because, obviously, if the discount happens to be higher than the probability we don’t want a negative probability.

And we’re just gonna interpolate that with this same continuation probability that we just saw,

p continuation of w sub i.

And, the lambda, now let’s talk about how to compute that lambda.

The lambda is gonna take all that probability mass from all those normalized discounts

that we took out of these higherorder probabilities,

and use those to weight how much probability we should assign to the unigram.

We’re gonna combine those.

So that lambda is the amount of the discount weight divided by the denominator there.

So, it’s the normalized discount.

And then, we’re gonna multiply that by the total number of word types that can follow this context, w_i minus one.

In other words, how many different word types did we discount

or how many times did we apply this normalized discount.

And we multiply those together

and we know how much probably mass total we can now assign to the continuation of the word.

Now, this is the bigram formulation for KneserNey.

Now in this slide, we’re showing you the general recursive formulation for ngrams in general,

and here we have to make a slight change to deal with all the higherorder ngrams.

So here we’re just showing the KneserNey probability of a word given the prefix of the word.

And, just like KneserNey we saw before,

we’re just interpolating a higherorder ngram which is discounted with a lambda weight and a lowerorder probability.

But now we need to distinguish between the very first toplevel time that we use a count and these lower order counts.

So we’re going to use the actual count for the very highest order bigram,

and we’re going to use the continuation value that we just defined earlier for all the lower order probabilities.

So we’ll define this new thing count KneserNey of dot which will mean actual count.

This will be actual count, let’s say we’re doing trigrams.

For the trigram, and then when we recurse and have the KneserNey probability for the lower order things.

When we get down to the bigrams and unigrams, we’ll be using the continuation count,

that’s again the single word context that we defined earlier.

So KneserNey smoothing, a very excellent algorithm.

It’s very commonly used in speech recognition and machine translation

and yet it has a very beautiful and elegant intuition and I hope you appreciate it.