Lexical Translation Models 1I

Transcription

1 Lexical Translation Models 1I Machine Translation Lecture 5 Instructor: Chris Callison-Burch TAs: Mitchell Stern, Justin Chiu Website: mt-class.org/penn

2 Last Time... X p( Translation)= p(, Translation) Alignment = X Alignment Alignment p( p( Alignment) Translation Alignment) {z } {z } X z } { z } { p(e f,m)= a2[0,n] m p(a f,m) p(e i f ai )

3 X p(e f,m)= a2[0,n] m p(a f,m) p(e i f ai ) Alternate ways of defining p(e i f ai,f ai 1 ) the translation probability p(e i f,f 1) ai ai p(e i f ai,e i 1 ) p(e i,e i+1 f ai ) What is the problem here?

4 X p(e f,m)= a2[0,n] m p(a f,m) p(e i f ai ) = X 1 p(e i f ai ) a2[0,n] m 1+n {z } p(a f,m) = X a2[0,n] m 1 Can we do something better here? 1+n p(e i f ai ) = X a2[0,n] m p(a i ) p(e i f ai )

5 0 NULL das Haus ist klein the house is small NULL das Haus ist klein house is small the

6 p(e f,m) = X p(a i ) p(e i f ai ) a2[0,n] m Model 2 = X p(a i i, m, n) p(e i f ai ) a2[0,n] m

7 Model 2 = X p(a i i, m, n) p(e i f ai ) a2[0,n] m Model alignment with an absolute position distribution Probability of translating a foreign word at position to generate the word at position i (with target length mand source length n) p(a i i, m, n) a i EM training of this model is almost the same as with Model 1 (same conditional independencies hold)

8 Model 2 = X p(a i i, m, n) p(e i f ai ) a2[0,n] m natürlich ist das haus klein natürlich natürlich das haus ist klein of course the house is small

9 Model 2 = X p(a i i, m, n) p(e i f ai ) a2[0,n] m Pros Non-uniform alignment model Fast EM training / marginal inference Cons Absolute position is very naive How many parameters to model p(a i i, m, n)

10 }m = 6 Model 2 = X p(a i i, m, n) p(e i f ai ) a2[0,n] m How much do we know when we only know the source & target lengths and the current position? null j 0 =1 j 0 =2 j 0 =3 How many parameters j 0 =4 j 0 =5 do we actually need to model this? i =3 i =2 i =1 i =4 i =5 i =6 }n = 5

11 }m = 6 Model 2 = X p(a i i, m, n) p(e i f ai ) a2[0,n] m pos in target pos in source h(j, i, m, n) = i m j n null j 0 =1 target len source len j 0 =2 j 0 =3 b(j i, m, n) = exp h(j, i, m, n) P j 0 exp h(j 0, i, m, n) j 0 =4 j 0 =5 i =4 i =3 i =2 i =1 i =6 i =5 }n = 5 p(a i i, m, n) = ( p 0 if a i =0 (1 p 0 )b(a i i, m, n) otherwise

12

13

14 Words reorder in groups. Model this!

15 p(e f,m) = X p(a i ) p(e i f ai ) a2[0,n] m Model 2 = X p(a i i, m, n) p(e i f ai ) a2[0,n] m HMM = X p(a i a i 1 ) p(e i f ai ) a2[0,n] m

16 HMM = X p(a i a i 1 ) p(e i f ai ) a2[0,n] m Insight: words translate in groups Condition on previous alignment position Probability of translating a foreign word at position given that the previous position translated was p(a i a i 1 ) EM training of this model using forward-backward algorithm (dynamic programming) a i 1 a i

17 HMM = X p(a i a i 1 ) p(e i f ai ) a2[0,n] m Improvement: model jumps through the source sentence p(a i a i 1 )=j(a i a i 1 ) Relative position model rather than absolute position model

18 HMM = X p(a i a i 1 ) p(e i f ai ) a2[0,n] m Be careful! NULLs must be handled carefully. Here is one option (due to Och): p(a i a i ni )= ( p 0 if a i =0 (1 p 0 )j(a i a i ni ) otherwise n i is the index of the first non-null aligned word in the alignment to the left of i.

19 HMM = X p(a i a i 1 ) p(e i f ai ) a2[0,n] m Other extensions: certain word-types are more likely to be reordered j( f) j( C(f)) Condition the jump probability on the previous word translated j( f,e) j( A(f), B(e)) Condition the jump probability on the previous word translated, and how it was translated

20 Fertility Models The models we have considered so far have been efficient This efficiency has come at a modeling cost: What is to stop the model from translating a word 0, 1, 2, or 100 times? We introduce fertility models to deal with this

21 IBM Model 3

22 Fertility Fertility: the number of English words generated by a foreign word Modeled by categorical distribution Examples: n( f) Unabhaengigkeitserklaerung zum = (zu + dem) Haus

23 Fertility X p(e f,m)= a2[0,n] m p(a f,m) p(e i f ai ) Fertility models mean that we can no longer exploit conditional independencies to write p(a f,m) as a series of local alignment decisions. How do we compute the statistics required for EM training?

24 EM Recipe reminder If alignment points were visible, training fertility models would be easy We would and n( =3 f = Unabhaenigkeitserklaerung) = count(3, Unabhaenigkeitserklaerung) count(unabhaenigkeitserklaerung) But, alignments are not visible n( =3 f = Unabhaenigkeitserklaerung) = E[count(3, Unabhaenigkeitserklaerung)] E[count(Unabhaenigkeitserklaerung)]

25 Expectation & Fertility We need to compute expected counts under p(a f,e,m) Unfortunately p(a f,e,m) doesn t factorize nicely. :( Can we sum exhaustively? How many different a s are there? What to do?

26 Sampling Alignments Monte-Carlo methods Gibbs sampling Importance sampling Particle filtering For historical reasons Use model 2 alignment to start (easy!) Weighted sum over all alignment configurations that are close to this alignment configuration Is this correct? No! Does it work? Sort of.

27

28 Lexical Translation IBM Models 1-5 [Brown et al., 1993] Model 2: absolute position model Model 3: fertility Model 5: non-deficient model Widely used Giza++ toolkit Model 1: lexical translation, uniform alignment Model 4: relative position model (jumps in target string) HMM translation model [Vogel et al., 1996] Relative position model (jumps in source string) Latent variables are more useful these days than the translations

29 Pitfalls of Conditional Models IBM Model 4 alignment Our model's alignm

30 A few tricks... p(f e) p(e f)

31 A few tricks... p(f e) p(e f)

32 A few tricks... p(f e) p(e f)

33 Suggestions for HW1 Matching the baseline will get you a B Implement IBM Model 2 in addition to IBM Model 1 Try the heuristics for merging the many-toone and one-to-many alignments Try to reduce sparse counts by preprocessing your training data Other ideas?

34 Reading Read Chapter 4 from the textbook (today we covered 4.4 through 4.6)

35 Announcements HW1 leaderboard submissions are due by Tuesday at 11:59pm HW1 write ups and code are due 24 hours later