How can I use "Interpolated Absolute Discounting" for a bigram model in language modeling? - nltk

Question

I want to compare two smoothing methods for a bigram model:
Add-one smoothing
Interpolated Absolute Discounting
For the first method, I found some codes.
def calculate_bigram_probabilty(self, previous_word, word):
bigram_word_probability_numerator = self.bigram_frequencies.get((previous_word, word), 0)
bigram_word_probability_denominator = self.unigram_frequencies.get(previous_word, 0)
if self.smoothing:
bigram_word_probability_numerator += 1
bigram_word_probability_denominator += self.unique__bigram_words
return 0.0 if bigram_word_probability_numerator == 0 or bigram_word_probability_denominator == 0 else float(
bigram_word_probability_numerator) / float(bigram_word_probability_denominator)
However, I found nothing for the second method except for some references for 'KneserNeyProbDist'. However, this is for trigrams!
How can I change my code above to calculate it? The parameters of this method must be estimated from a development-set.

Answer 1

In this answer I just clear up a few things that I just found about your problem, but I can't provide a coded solution.
with KneserNeyProbDist you seem to refer to a python implementation of that problem: https://kite.com/python/docs/nltk.probability.KneserNeyProbDist
There exists an article about Kneser–Ney smoothing on wikipedia: https://en.wikipedia.org/wiki/Kneser%E2%80%93Ney_smoothing
The article above links this tutorial: https://nlp.stanford.edu/~wcmac/papers/20050421-smoothing-tutorial.pdf but this has a small fault on the most important page 29, the clear text is this:
Modified Kneser-Ney
Chen and Goodman introduced modified Kneser-Ney:
Interpolation is used instead of backoff. Uses a separate discount for one- and two-counts instead of a single discount for all counts. Estimates discounts on held-out data instead of using a formula
based on training counts.
Experiments show all three modifications improve performance.
Modified Kneser-Ney consistently had best performance.
Regrettable the modified Version is not explained in that document.
The original documentation by Chen & Goodman luckily is available, the Modified Kneser–Ney smoothing is explained on page 370 of this document: http://u.cs.biu.ac.il/~yogo/courses/mt2013/papers/chen-goodman-99.pdf.
I copy the most important text and formula here as screenshot:
So the Modified Kneser–Ney smoothing now is known and seems being the best solution, just translating the description beside formula in running code is still one step to do.
It might be helpful that below the shown text (above in screenshot) in the original linked document is still some explanation that might help to understand the raw description.