This page is really work in progress!
This page is still work in progress and probably mainly SMT focused, if you want to make a contribution, mail me

B Bayes Rule	C Cept
D Decoding Deficiency	F Fertility
G Gold standard	P Pegging Perplexity

Bayes Rule
Bayes Rule is the "normal" Bayes Rule in the probability theory, which dictates that:
P(A|B)=( P(B|A) * P(A) ) / P(B)
This is interesting because it is used quite often in statistical machine translation, where we try to calculate P(T|S) with T is a sentence in the target language, and S is a sentence in the source language. Because often we try to find the T which maximises P(T|S) we can use Bayes Rule, and because the single P(S) is not used in maximising we try to maximise T over P(T)*P(S|T). The most interesting feature is that we can split our translation problem in a translation model and a language model.

Cept
"When we look at a passage, we cannot see the concepts directly but only the words that hey leave behind. To show that these words are related to a concept but are not quite the whole story, we say that they form a cept".
(Chapter 3 [BRO93])
"A word is a cept if its fertility is greater than zero"
(Chapter 4.6 [BRO93])

Decoding

Deficiency
In statistical machine translation, when a model wastes some of its probability mass on impossible strings it is called deficient. Here we make the distinction between:
Well formed strings:
"This is possible"
Ill-formed, but possible strings:
"This possible is"
and impossible strings:

(Here "The" and "possible" are positioned at the same place in the string)
According to the definition a model is deficient if not all of its probability is concentrated on events of interest.
IBM Model 3 and higher are known to have deficiency, although Model 5 tries to tackle this problem somewhat.
(Chapter 4.5 [BRO93])

Fertility
When in machine translations operations are done on a word level (typical statistical machine translation) the term fertility refers to how many words in a target language are needed for a word in the source language. Words with fertility of zero disappear in the target language, while compound words in a source language with a higher fertilty are expanded to several words in the target language.
In statistical machine translation there are fertility tables which indicate what the probality for a certain word is to disappear, stay one word, or to get expanded in more words in the target language.
(Chapter 5 [BRO93])

Gold standard
Term typically used in evaluation (of machine translation). The sentence that are defined as being a good solution for a problem (translation), in machine translation human translation are regarded to be gold standards.
Other(machine) translations are compare to the gold standard to evaluate how good they are.

Pegging

Perplexity
When ngram models are used, longer sentences tend to get much lower probabilities than shorter sentence, because the probability of the entire sentence is the product of all the individual probabilites.
To remedy this perplexity can be used, if P(e) is the probability of an event, which goes down by the by N (length of the event) then
-₂log(P(e))
     N
is called the perplexity. As P(e) increases, perplexity decreases, usualy a good model tries to maximize P(e) or minimize the perplexity.

---

BRO93 - P. Brown, S.A. Della Pietra, V.J. Della Pietra, and R.L. Mercer, 1993. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263-311 http://acl.ldc.upenn.edu/J/J93/J93-2003.pdf http://www-i6.informatik.rwth-aachen.de/Colleagues/och/ACL00.ps