Named Entity Transliteration with Comparable Corpora

1
Named Entity Transliteration with Comparable
Corpora

• Richard Sproat, Tao Tao, ChengXiang Zhai
• University of Illinois at Urbana-Champaign
• Presented by Ye Shani
• 2006-12-8

2
Named Entity Transliteration with Comparable
Corpora

• Given comparable corpora (C1, , Cn), where Cis
  are corpora in different languages
• Given a name X in Ci , the task is to find
  transliterations of X in C1,Ci-1, Ci, Cn
• Possible applications
• Cross-lingual entity retrieval
• Cross-lingual information integration
• Cross-lingual trend analysis

3
Previous Work

• Transliteration Knight Graehl 1998 Meng et
  al. 2001 Gao et al. 2004 inter alia.
• Comparable corpora Fung, 1995 Rapp 1995 Tanaka
  and Iwasaki, 1996 Franz et al.,1998 Ballesteros
  and Croft, 1998 Masuichi et al., 2000 Sadat et
  al., 2003 Tao and Zhai, 2005.
• Mining transliterations from multilingual web
  pages Zhang Vines, 2004

4
Method

• We assume that we have comparable corpora,
  consisting of newspaper articles in English and
  Chinese from the same day, or almost the same
  day.
• In our experiments we use data from the English
  and Chinese stories from the Xinhua News agency
  for about 6 months of 2001.

5
Method

• Identify NEs english names using method of Li et
  al. 2001 (based on SNoW Carlson et al. 1999)
• Identify NEs chinese names using a list of
  characters that frequently used for
  transliterating foreign names.

6
Method

• The general three-step approach
• Given an English name, identify candidate Chinese
  character n-grams as possible transliterations.
• Score each candidate based on how likely the
  candidate is to be a transliteration of the
  English name.
• Two initial scoring methods.
• Phonetic scoring,
• Frequency profile of the candidate pair over
  time.
• Propagate scores of all the candidate
  transliteration pairs globally based on their
  cooccurences in document pairs in the comparable
  corpora.

7
Overview
8
Method 1 Phonetic Transliteration

• Much work using the source-channel approach
• Cast as a problem where you have a clean
  source e.g. a Chinese name and a noisy
  channel that corrupts the source into the
  observed form e.g. an English name
• P(EC)P(C)
• Seek to estimate P(EC)
• E.g. P(fi,E fi1,E fi2,E fin,E sC)
• Chinese characters represent syllables (s) we
  match these to sequences of English phonemes (f)

9
Phonetic Transliteration Estimation
10
Phonetic Transliteration General Approach

• Train a transliteration model from a dictionary
  of known transliterations (720 entries)
• Identify names in English news text for a given
  day using an existing named entity recognizer
• Process same day of Chinese text looking for
  sequences of characters used in foreign names
• Do an all-pairs match using the transliteration
  model to find possible transliteration pairs

11
Phonetic Transliteration Some Automatically
Found Pairs

• Pairs found in same day of newswire text

12
Method 2 Frequency Correlation

• We pool all documents in a single day to form a
  large pseudo-document.
• compute each transliterations frequency in each
  of those pseudo-documents and obtain a raw
  frequency vector
• normalize the raw frequency vector so that it
  becomes a frequency distribution over all the
  time points (days).
• Using The Pearson correlation coefficient

13
Frequency Correlation

• Example
• lte1,c1gt lt1,12gt lt7,8gt..
• Normalizelt1/sum-e,12/sum-cgt..
• The Pearson correlation coefficient

14
Method 2 Frequency Correlation
15
(No Transcript)
16
Method 3 Combining Phonetic and Time Correlation
Methods

• Two methods exploit complementary resources, thus
  combining them might further help.
• Phonetic filter use the phonetic model to filter
  out (clearly impossible) candidates and then use
  the frequency correlation method to rank the
  candidates.
• Score combination mean of normalized scores

17
Method 4 Score Propagation

• The methods so far score each transliteration
  pair independently
• But knowing that two transliteration pairs
  co-occur in the same cross-lingual document pair
  should increase our confidence on both
  transliteration pairs
• Similarly, document pairs that contain lots of
  plausible transliteration pairs are likely
  comparable content-wise
• Thus, document/transliteration pairs reinforce
  each other

18
Score Propagation
19
Score Propagation
20
Score Propagation
21
Estimate of P(ji)

• Two ways
• Cooccurrences (CO) in whole collection
• Mutual Information (MI)

22
Evaluation

• We take one days worth of comparable news
  articles (234 Chinese stories and 322 English
  stories) from Chinese English Gigaword Corpus
  (LDC)
• Generate about 600 English names with the entity
  recognizer (Li et al., 2004)
• Find potential Chinese names by looking for
  strings of characters that are commonly used in
  transliteration
• This generates 627 Chinese candidates
• In principle any of the 600 x 627 pairings could
  be correct
• Use phonetic and time-correlation methods to rank
  the candidate pairings.
• Evaluate using Mean Reciprocal Rank (MRR)

23
Evaluation Further Details

• Small number of English names do not seem to have
  any standard transliteration according to the
  resources that we consulted.
• Removing these, we have a list of 490 out of
  original 600 English names
• Furthermore, about 20 of answers are not in
  Chinese candidate list
• Either they are really not there
• Or our candidate selection process missed them.
• This motivates two scores
• AllMRR using the original list of 600 English
  names
• CoreMRR for just those where the English names
  are also in our Chinese candidate list

24
Evaluation

• Phonetic correspondence yielded an AllMRR score
  of 0.3 and a CoreMRR score of 0.89
• Time-correlation scores yielded results as
  follows, with different correlation measures

25
Upper Bound Analysis

• Results when we manually added the correct
  transliterations to the Chinese candidate list

26
Score Propagation (Core)
27
Score Propagation
28
Score Combination
29
Score Propagation (All)
30
Summary

• We propose several complementary methods for
  transliteration that rely on relatively light
  linguistic resources
• One of these is a novel score propagation method
• We show that different methods can be combined to
  further improve performance
• It is feasible to perform transliteration over
  comparable corpora without much manual effort