Title: Building Lexicons
1Building Lexicons
- Jae Dong Kim
- Matthias Eck
2Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
3Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
4Definitions
- Translational equivalence A relation that holds
between two expressions with the same meaning,
where two expressions are in different languages. - Statistical Translation Models statistical
models of translational equivalence - Empirical estimation of statistical translation
models is typically based on parallel texts or
bitexts - Word-to-Word Lexicon
- A list of word pairs
(source word, target word ) - Bidirectional
- Probabilistic word-to-word lexicon (source word,
target word, prob.)
5Additional Universal Property
- Translation models benefit from the best of both
the empiricist and rationalist traditions - Models to be proposed
- Most word tokens translate to only one word
token. Approximated by one-to-one assumption -
Method A - Most text segments are not translated word for
word. Explicit Noise Model - Method B - Different linguistic objects have statistically
different behavior in translation. Translation
models on different word classes. - Method C - Human judgment has shown that each of three
estimation biases improves translation model
accuracy over a baseline knowledge-free model
6Applications of Translation Models
- Where word order is not important
- Cross-language information retrieval
- Multilingual document filtering
- Computer-assisted language learning
- Certain machine-assisted translation tools
- Concordancing for bilingual lexicography
- Corpus linguistics
- crummy machine translation
- Where word order is important
- Speech transcription for translation
- Bootstrapping of OCR systems for new languages
- Interactive translation
- Fully automatic high-quality machine translation
7Advantages of translation models
- Compared to handcrafted models
- The possibility of better coverage
- The possibility of frequent updates
- More accurate information about relative
importance of different translations
IRDB
Q
T
Qi
IR
Uniform Importance?
8Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
9Models of Co-occurrence
- Intuition words that are translations of each
other are more likely to appear in corresponding
bitext regions than other pairs of words. - A boundary-based model assumes that both halves
of the bitext have been segmented into s
segments, so that segment Ui in one half of the
bitext and segment Vi in the other half are
mutual translations, 1ltilts - Co-occurrence count by Brown et al
- Co-occurrence count by Melamed
10Nonprobabilistic Translation Lexicons (1)
- Summary of non-probabilistic translation lexicon
algorithms - Choose a similarity function S between word types
in L1 and word types L2 - Compute association scores S(u,v) for a set of
word type pairs (u,v) ? (L1 x L2) that occur in
training data - Sort the word pairs in descending order of their
association scores - Discard all word pairs for which S(u,v) is less
than a chosen threshold. The remaining word pairs
become the entries in the translation lexicon - Main difference choice of similarity function
- Those functions are based on a model of
co-occurrence with some linguistically motivated
filtering
11Nonprobabilistic Translation Lexicons (2)
- Problem independence assumption in step 2
- Models of translational equivalence that are
ignorant of indirect association have a tendency
to be confused by collocates - If all the entries in a translation lexicon are
sorted by their association scores, the direct
associations will be very dense near the top of
the list, and sparser towards the bottom
He nods his
head Il hoche la
tete
Direct association
Indirect association
12Nonprobabilistic Translation Lexicons (3)
- The very top of the list can be over 98 correct
- Gale and Church (1991) - Gleaned lexicon entries for about 61 of the word
tokens in a sample of 800 English sentences - Selected only entries with high association score
- 61 word tokens represent 4.5word types
- 71.6 precision with top 23.8 of noun-noun
entries - Fung(1995) - Automatic acquisition of 6,517 lexicon entries
with 86 precision from 3.3-million-word corpus -
Wu Xia (1994) - 19 recall
- Weighted precision in (E1,C1,0.533),
(E1,C2,0.277), (E1,C3,0.190), if (E1,C3,0.190)
is wrong, we have precision of 0.810 - Higher than unweighted one
13Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
14Decomposition of Translation Model (1)
- Two stage decomposition of sequence-to-sequence
model - First stage
- Every sequence L is just an ordered bag, and the
bag B can be modeled independently of its order O
15Decomposition of Translation Model (2)
- First Stage
- Let L1 and L2 be two sequences and let A be a
one-to-one mapping between the elements of L1 and
the elements of L2
16Decomposition of Translation Model (2)
- First Stage
- Let L1 and L2 be two sequences and let A be a
one-to-one mapping between the elements of L1 and
the elements of L2
17Decomposition of Translation Model (3)
- First Stage
- Bag-to-bag translation model
18Decomposition of Translation Model (4)
- Second Stage
- From bags of words to the words that they contain
- Bag pair generation process - how word-to-word
model is embeded - Generate a bag size l. l is also the assignment
size - Generate l language-independent concepts C1,,Cl.
- From each concept Ci, 1ltiltl, generate a pair of
word sequences from L1 x L2, according to
the distribution , to lexicalize
the concept in the two languages. Some concepts
are not lexicalized in some languages, so one of
ui and vi may be empty. - Bags
- An assignment (i1,j1),,(il,jl)
19Decomposition of Translation Model (5)
- Second Stage
- The probability of generating a pair of bags
(B1,B2)
20Decomposition of Translation Model (5)
- Second Stage
- The probability of generating a pair of bags
(B1,B2) - is zero for all concepts
except one - is symmetric unlike the models
of Brown et al.
21The One-to-One Assumption
- and may consist of at most one word each
- A pair of bags containing m and n nonempty words
can be generated by a process where the bag size
l is anywhere between max(m,n) and mn - Not as restrictive as it may appear. What if we
extend a word to include spaces?
22Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
23Reestimated Seq.-to-Seq. Trans. Model (1)
- Variations on the theme proposed by Brown et al.
- Conditional probabilities, but can be compared to
symmetric models if the letter are normalized
marginally - Only Co-occurrence Information
- EM
- When information about segment lengths is not
available
24Reestimated Seq.-to-Seq. Trans. Model (2)
- Word Order Correlation Biases
- In any bitext, the positions of words relative to
the true bitext map correlate with the positions
of their translations - The word order correlation bias is most useful
when it has high predictive power - Absolute word positions - Brown et al. 1988
- A much smaller set of relative offset parameters
- Dagan, Church, and Gale. 1993 - Even more efficient parameter estimation using
HMM with some additional assumptions - Vogel,
Ney, and Tillman. 1996
25Reestimated Bag-to-Bag Trans. Models
- Another Bag-to-Bag model by Hiemstra. 1996
- The same one-to-one assumption
- The difference empty words are allowed in only
one of the two bags, the one representing the
shorter sentence - Iterative Proportional Fitting Procedure(IPFP)
for parameter estimation - IPFP is subjective to initial conditions
- With the most advantageous, more accurate than
Model 1
26Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
27Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
28Parameter Estimation
- Methods for estimating the parameters of a
symmetric word-to-word translation model from a
bitext. - Interested in probability trans(u,v) Probability
to jointly generate the pair of words (u,v) - trans(u,v) cannot be directly inferred It is
unknown which words were generated together - Observable in bitext is only cooc(u,v)
(co-occurrence count)
29Definitions
- Link counts links(u,v) hypothesis about the
number of times u and v were generated together - Link token Ordered Pair of word tokens
- Link type Ordered Pair of word types
- links(u,v) ranges over Link types
- trans(u,v) can be calculated using links(u,v)
30Definitions (continued)
- score(u,v) chance u and v can ever be mutual
translationssimilar to trans(u,v), convenient
for estimation - Relationship between trans(u,v) and score(u,v)
can be direct (depending on model)
31General outline for all Methods
- Initialize the score parameter to a first
approximation based only on cooc(u,v) - REPEAT
- Approximate links(u,v) based on score and cooc
- Calculate trans(u,v), Stop if only little change
- Reestimate score(u,v) based on links and cooc
32EM-Algorithm!
- Initialize the score parameter to a first
approximation based only on cooc(u,v) - REPEAT
- Approximate links(u,v) based on score and cooc
- Calculate trans(u,v), Stop if only little change
- Re-estimate score(u,v) based on links and cooc
Initial E-Step
M-Step
E-Step
33EM Maximum Likelihood Approach
- Find the parameters that maximize the probability
of the given bitext - Assignments cannot be decomposed due to the
one-to-one assumption (compare to Brown et al.
1993) - MLE approach is infeasible
- Approximating EM is necessary
34Maximum a Posteriori
- Evaluate Expectations using the single most
probable assignment only (Maximum a posteriori
(MAP) assignment)
35Maximum a Posteriori
- Evaluate Expectations using the single most
probable assignment (Maximum a posteriori (MAP)
assignment) - l number of Concepts, number of produced words
36Maximum a Posteriori
- Evaluate Expectations using the single most
probable assignment (Maximum a posteriori (MAP)
assignment)
37Maximum a Posteriori
- Evaluate Expectations using the single most
probable assignment (Maximum a posteriori (MAP)
assignment) - l, Pr(l) constant
38Maximum a Posteriori
- Evaluate Expectations using the single most
probable assignment (Maximum a posteriori (MAP)
assignment)
39Bipartite Graph
- Represent bitext as bipartite graph
- Find solution for weighted maximum matching
- Still too expensive to solve
- Competitive Linking Algorithm approximates
u
log(trans(u,v))
v
40Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
41Method A Competitive Linking
- Step 1
- Co-occurrence counts
- Use whole table information
- Initialize score(u,v) to G2(u,v) (similar to
Chi-square) - Good-Turing Smoothing gives improvements
42Step 2 Estimation of link counts
- Competitive Linking algorithm is employed
- Greedy approximation of the MAP approximation
- Algorithm
- Sort all score(u,v) from the highest to the
lowest - For each score(u,v) in order
- Link all co-occurring token pairs (u,v) in the
bitext(If u is NULL consider all tokens of v in
the bitext linked to NULL and vice versa) - One-to-One assumption Linked words cannot be
linked againRemove all linked words from the
bitext
43Example Competitive Linking
u
a
b
c
d
v
44Competitive Linking
X
X
X
u
X
X
X
X
X
X
a
X
b
X
X
X
c
d
v
45Competitive Linking
X
X
X
X
X
X
u
X
X
X
X
X
X
a
X
X
X
X
X
X
b
X
X
X
X
X
X
c
d
v
46Competitve Linking per sentence
b
a
links(a,c) links(b,d)
c
d
a
b
links(a,d) links(b,e)
c
d
e
47Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
48Method B
- Most texts are not translated word-for-word
- Why is that a problem with Method A?
a
b
x
c
d
e
f
49Method B
- Most texts are not translated word-for-word
- Why is that a problem with Method A?
a
b
x
Competitive Linking
c
d
e
f
a
b
x
We are forced to connect (b,d)!
c
d
e
f
50Method B
- After one iteration of Method A on 300k sentences
Hansard
- links cooc
- often, probably correct
- links lt cooc
- rare, might be correct
- links ltlt cooc
- often, probably incorrect
51Method B
- Use information links(u,v)/cooc(u,v) to bias
parameter estimation - Introduce p(u,v) as the probability of u and v
being linked when they co-occur. - Leads to binomial process for each co-occurrence
(either linked or not linked) - Too sparse data to model p(u,v)
- Just 2 cases
If u,v are mutual translations (Rate of true
positives)
If u,v are not mutual translations (Rate of false
positives)
52Method B
53Maximum Likelihood Estimation
54Maximum Likelihood Estimation
- on 300k sentences Hansard
55Method B
- Overall score calculation for Method B
- Probability for generating correct links(u,v)
given cooc(u,v) - Probability for generating incorrect links(u,v)
given cooc(u,v) - Score is ratio
56Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
57Method C
- Improved Estimation using Preexisting Word
Classes - Method A, B
- All word pairs that co-occur the same number of
times and are linked the same number of times are
assigned the same score - But Frequent words are translated less
consistently than rare words - Introduce classes to get Statistics per class
58Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
59Method C for Evaluation
- We have to choose classes
- EOS End of sentence punctuation
- EOP End of phrase punctuation (, )
- SCM Subordinate clause markers ( ()
- SYM Symbols ( )
- NU NULL word
- C Content words
- F Function words
60Experiment 1
- Training Data
- 29,614 sentence pairs French, English (Bible)
- Test Data
- 250 hand linked sentences (gold standard)
- Procedure
- Single Best Models guess one translation per
word on each side - Whole Distribution Model outputs all possible
translation with probabilities
61Experiment 1 Results
- Single Best All links (95 confidence
intervals)
62Experiment 1 Results
- Single Best open-class links only (just the
content words)
63Experiment 1 Results
- Whole Distribution All Links
64Experiment 1 Results
- Whole Distribution open-class links only (just
the content words)
65Experiment 2
- Influence of training data size
- Model A is 102 more correct than Model 1 when
trained on only 250 sentence pairs - Overall up to 125 improvements
66Evaluation at the Link Type Level
- Sorted scores for all link types
- 1/1, 2/2 and 3/3 correspond to links/cooc
67Coverage vs. Accuracy
- incomplete Lexicon contains only part of correct
phrase
68Building Lexicons
- Introduction
- Previous Work
- Translation Model Decomposition
- Reestimated Models
- Parameter Estimation
- Method A
- Method B
- Method C
- Evaluation
- Conclusion
69Conclusion - Overview
a
b
x
- IBM Model 1 co-occurrence information only
- Method A one-to-one assumption
- Method B Noise Model
- Method C condition auxiliary parameters on
word classes
c
d
e
f
a
b
x
c
d
e
f
a
b
x
c
d
e
f