An integrated platform for high-accuracy word alignment

1
An integrated platform for high-accuracy word
alignment

• Dan Tufis, Alexandru Ceausu, Radu Ion, Dan
  Stefanescu
• RACAI Research Institute for Artificial
  Intelligence, Bucharest

2
COWAL

• The main task of COWAL is to combine the output
  of two or more comparable word-aligners
• In order to achieve this task, COWAL is also an
  integrated platform with modules for
  tokenization, POS-tagging, lemmatization,
  collocation detection, dependency annotation,
  chunking and word sense disambiguation.

3
Word alignment algorithms (YAWA)

• YAWA starts with all plausible links (those with
  ll-score higher than 11)
• Then, using a competitive linking strategy,
  retains the links that maximizes sentence
  translation equivalence score, and minimizing the
  number of crossing links
• In this way, it generates only 1-1 alignments.
  N-M alignments are possible only with chunking
  and/or dependency linking available.

4
Word alignment algorithms (MEBA)

• MEBA iterates several times over each pair of
  aligned sentences, at each iteration adding only
  the highest score links.
• The links already established in previous
  iterations give support or create restrictions
  for the links to be added in a subsequent
  iteration.
• MEBA uses different weights and different
  significance thresholds on each feature and
  iteration step.

5
Features characterizing a link

• A link ltToken1 Token2gt is characterized by a set
  of features, the values of which are real numbers
  in the 0,1 interval.
• context independent features CIF, they refer
  only to the tokens of the current link
• context dependent features CDF, they refer to
  the properties of the current link with respect
  to the rest of links in a bi-text

6
Context independent features

• Translation equivalents (lemma and/or wordform )
• Translation equivalents entropy (lemma)
• Part-of-Speech affinity
• Cognates

7
Translation equivalents (TE)

• YAWA, TREQ-AL use competitive linking based on
  ll-scores, plus the Ro-En aligned wordnets
• MEBA uses GIZA generated candidates filtered
  with a log-likelihood threshold (11).
• The TE candidates search space is limited by
  lemmatization and POS meta-classes (e.g.
  meta-class 1 includes only N, V, Aj and Adv
  meta-class 8 includes only proper names)
• For a pair of languages translation equivalents
  are computed in both directions. The value of the
  TE feature of a candidate link ltTOKEN1 TOKEN2gt is
  1/2 (PTR(TOKEN1, TOKEN2) PTR(TOKEN2, TOKEN1).

8
Entropy Score (ES)

• The entropy of a word's translation equivalents
  distribution proved to be an important hint on
  identifying highly reliable links (anchoring
  links)
• Skewed distributions favored against uniform ones
• For a link ltA Bgt, the link feature value is
  0.5(ES(A)ES(B))

9
Part-of-speech affinity (PA)

• An important clue in word alignment is the fact
  that the translated words tend to keep their
  part-of-speech and when they have different
  POSes, this is not arbitrary.
• Tried to use GIZA (replacing tokens with their
  respective POSes) but there was too much noise!
• The information was computed based on a gold
  standard (the revised NAACL2003), in both
  directions (source-target and target-source).
• For a link ltA,Bgt PA0.5(P(cat(A)cat(B))P(cat(B)
  cat(A))

10
Cognates (COG)

• The cognates feature assigns a string similarity
  (using Levenstein distance) to the tokens of a
  candidate link
• We estimated the probability of a pair of
  orthographically similar words, appearing in
  aligned sentences, to be cognates, with different
  string similarity thresholds. For the threshold
  0.6 we didnt find any exception. Therefore, the
  value of this feature is either 1 (if the
  similarity score is above the threshold or 0
  otherwise).
• Before computing the string similarity score, the
  words are normalized (duplicate letters are
  removed, diacritics are removed, some suffixes
  are discarded).

11
Context dependent features

• Locality
• Links crossed
• Relative position/Distortion
• Collocation/Fertility
• Coherence

12
Collocation

• Bi-gram lists (only content words) were built
  from each monolingual part of the training
  corpus, using the log-likelihood score (threshold
  of 10) and minimal occurrence frequency (3) for
  candidates filtering. Collocation probabilities
  are estimated for each surviving bi-gram.
• If neither token of a candidate link has a
  relevant collocation score with the tokens in its
  neighborhood, the link value of this feature is
  0. Otherwise the value is the maximum of the
  collocation probabilities of the links tokens.
  Competing links (starting or finishing in the
  same token) are licensed only and only if at
  least one of them have a non-null collocation
  score

13
Distorsion/Relative position

• Each token in both sides of a bi-text is
  characterized by a position index, computed as
  the ratio between the relative position in the
  sentence and the length of the sentence. The
  absolute value of the difference between tokens
  position indexes, gives the links obliqueness
• The distorsion feature of a link is its
  obliqueness D(link)OBL(SWi, TWj)

14
Localization

• This feature is relevant with or without chunking
  or dependency parsing modules. It accounts for
  the degree of the cohesion of links.
• With the chunking module is available, and the
  chunks are aligned via the linking of their
  respective heads, the links starting in one chunk
  should finish in the aligned chunk.
• When chunking information is not available, the
  link localization is judged against a window, the
  span of which is dependent on the aligned
  sentences length.
• Maximum localization (1) is the one with all the
  tokens in the source window are linked to all
  tokens in the target window

15
Crossed links

• The crossed links feature computes (for a window
  size depending on the categories of the
  candidates and the sentences lengths) the links
  that were crossed.
• The normalization factor (maximum number of
  crossable links) is empirically set, based on
  categories of the links tokens

16
EVALUATIONOfficial ranking
U.RACAI.Combined L.ISI.Run5.vocab.grow
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
Word alignment combiners

• The COWAL(ACL2005) combiner is fine-tuned for the
  concerned language pair (rule-based)
• The SMV filter is a language independent combiner
  (trainable on positive and negative examples)
• Trade-off between human introspection and
  performance

21
SVM filter

• Combining word alignments requires the ability to
  distinguish among correct links and incorrect
  links of the two ore more merged alignments. SVM
  technology is specifically adequate for this
  task
• The SVM combiner is a classifier trained on both
  positive and negative examples.

22
SVM filter evaluation
MEBA COWAL MEBA filtered YAWA MEBA filtered
Precision 0.9122 0.8795 0.9315 0.8830
Recall 0.6976 0.7775 0.6712 0.7713
F-measure 0.7924 0.8254 0.7802 0.8234
SVM filtering results.The SVM model was trained
on NAACL 2003 gold standard.
23
Romanian Acquis

• The available Romanian documents were downloaded
  from CCVISTA (over 12000 Microsoft word
  documents)
• We kept only 11228 files (some of them were
  different versions of the same document)
• The remaining documents were converted into the
  same XML format of the ACQUIS corpus
• From the 11228 Romanian files only 6256 are
  available for English in the JRC distribution

24
Romanian Acquis

• Tokenization
• Sentence splitting
• POS-tagging
• Lemmatization
• Chunking
• Sentence aligning

25
(No Transcript)