Cognates and Word Alignment in Bitexts

1
Cognates and Word Alignment in Bitexts

• Greg Kondrak
• University of Alberta

2
Outline

• Background
• Improving LCSR
• Cognates vs. word alignment links
• Experiments results

3
Motivation

• Claim words that are orthographically similar
  are more likely to be mutual translations than
  words that are not similar.
• Reason existence of cognates, which are usually
  orthographically and semantically similar.
• Use Considering cognates can improve word
  alignment and translation models.

4
Objective

• Evaluation of orthographic similarity measures in
  the context of word alignment in bitexts.

5
MT applications

• sentence alignment
• word alignment
• improving translation models
• inducing translation lexicons
• aid in manual alignment

6
Cognates

• Similar in orthography or pronunciation.
• Often mutual translations.
• May include
• genetic cognates
• lexical loans
• names
• numbers
• punctuation

7
The task of cognate identification

• Input two words
• Output the likelihood that they are cognate
• One method compute their orthographic/phonetic/se
  mantic similarity

8
Scope

• The measures that we consider are
• language-independent
• orthography-based
• operate on the level of individual letters
• binary identity function

9
Similarity measures

• Prefix method
• Dice coefficient
• Longest Common Subsequence Ratio (LCSR)
• Edit distance
• Phonetic alignment
• Many other methods

10
IDENT

• 1 if two words are identical, 0 otherwise
• The simplest similarity measure
• e.g. IDENT(colour, couleur) 0

11
PREFIX

• The ratio of the longest common prefix of two
  words to the length of the longer word
• e.g. PREFIX(colour, couleur) 2/7 0.28

12
DICE coefficient

• The ratio of the number of common letter bigrams
  to the total number of letter bigrams
• e.g. DICE(colour, couleur) 6/11 0.55

13
Longest Common Sub-sequence Ratio (LCSR)

• The ratio of the longest common subsequence of
  two words to the length of the longer word.
• e.g. LCSR(colour, couleur) 5/7 0.71

14
LCSR

• Method of choice in several papers
• Weak point insensitive to word length
• Example
• LCSR(walls, allés) 0.8
• LCSR(sanctuary, sanctuaire) 0.8
• Sometimes a minimal word length imposed
• A principled solution?

15
The random model

• Assumption strings are generated randomly from a
  given distribution of letters.
• Problem what is the probability of seeing k
  matches between two strings of length m and n?

16
A special case

• Assumption k0 (no matches)
• t alphabet size
• S(n,i) - Stirling number of the second kind

17
The problem

• What is the probability of seeing k matches
  between two strings of length m and n?
• An exact analytical formula is unlikely to exist.
• A very similar problem has been studied in
  bioinformatics as statistical significance of
  alignment scores.
• Approximations developed in bioinformatics are
  not applicable to words because of length
  differences.

18
Solutions for the general case

• Sampling
• Not reliable for small probability values
• Works well for low k/n ratios (uninteresting)
• Depends on a given alphabet size and letter
  frequencies
• No insight
• Inexact approximation
• Works well for high k/n ratios (interesting)
• Easy to use

19
Formula 1

• - probability of a match

20
Formula 1

• Exact for kmn
• Inexact in general
• Reason implicit independence assumption
• Lower bound for the actual probability
• Good approximation for high k/n ratios.
• Runs into numerical problems for larger n

21
Formula 2

• Expected number of pairs of k-letter substrings.
• Approximates the required probability for high
  k/n ratios.

22
Formula 2

• Does not work for low k/n ratios.
• Not monotonic.
• Simpler than Formula 1.
• More robust against numerical underflow for very
  long words.

23
Comparison of both formulas

• Both are exact for kmn
• For k close to max(m,n)
• both formulas are good approximations
• their values are very close
• Both can be quickly computed using dynamic
  programming.

24
LCSF

• A new similarity measure based on Formula 2.
• LCSR(X,Y) k/n
• LCSF(X,Y)
• LCSF is as fast as LCSR because its values that
  depend only on k and n can be pre-computed and
  stored

25
Evaluation - motivation

• Intrinsic evaluation of orthographic similarity
  is difficult and subjective.
• My idea extrinsic evaluation on cognates and
  word aligned bitexts.
• Most cross-language cognates are orthographically
  similar and vice-versa.
• Cognation is binary and not subjective

26
Cognates vs alignment links

• Manual identification of cognates is tedious.
• Manually word-aligned bitexts are available, but
  only some of the links are between cognates.
• Question 1 can we use manually-constructed word
  alignment links instead?

27
Manual vs automatic alignment links

• Automatically word-aligned bitext are easily
  obtainable, but a good fraction of the links are
  wrong.
• Question 2 can we use machine-generated word
  alignment links instead?

28
Evaluation methodology

• Assumption a word aligned bitext
• Treat aligned sentences as bags of words
• Compute similarity for all word pairs
• Order word pairs by their similarity value
• Compute precision against a gold standard
• either a cognate list or alignment links

29
Test data

• Blinker bitext (French-English)
• 250 Bible verse pairs
• manual word alignment
• all cognates manually identified
• Hansards (French-English)
• 500 sentences
• manual and automatic word-alignment
• Romanian-English
• 248 sentences
• manually aligned

30
Blinker results
31
Hansards results
32
Romanian-English results
33
Contributions

• We showed that word alignment links can be used
  instead of cognates for evaluating word
  similarity measures.
• We proposed a new similarity measure which
  outperforms LCSR.

34
Future work

• Extend our approach to length normalization to
  edit distance and other similarity measures.
• Incorporate cognate information into statistical
  MT models as an additional feature function.

35
Thank you
36
Applications

• Recognition of Cognates
• Historical linguistics
• Machine translation
• Sentence and word alignment
• Confusable Drug Names
• Edit Distance Tasks
• Spelling error correction

37
Improved word alignment quality

• GIZA trained on 50,000 sentences from Hansards.
• Tested on 500 manually aligned sentences
• 10 reduction of the error rate when cognates are
  added.

38
Blinker results
39
Problems with links (1)

• The lion(1) killed enough for his cubs and
  strangled the prey for his mate(2)
• Le lion(1) déchirait pour ses petits, etranglait
  pour ses lionnes(2)

40
Problems with links (2)

• But let justice(1) roll on like a river ,
  righteousness(2) like a never -failing stream.
• Mais que la droiture(1) soit comme un courant de
  eau , et la justice(2) comme un torrent qui
  jamais ne tarit.