---
title: "Transliteration of Proper Names in Cross-Lingual Information"
description: "The authors build a fully data-driven English-to-Chinese name transliterator. 10 references, cited by 63 papers here."
authors: ["Paola Virga", "Sanjeev Khudanpur"]
year: 2003
venue: "Proceedings of the ACL 2003 Workshop on Multilingual and Mixed-language Named Entity Recognition"
source: "https://aclanthology.org/W03-1508/"
doi: "10.3115/1119384.1119392"
licence: "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)"
licence_url: "https://creativecommons.org/licenses/by-nc-sa/3.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/virga-2003-transliteration-proper-names-cross-lingual/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/virga-2003-transliteration-proper-names-cross-lingual/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2003](https://md.research.elamur.ai/free/en/papers/2003/) / Transliteration of Proper Names in Cross-Lingual Information Retrieval

# Transliteration of Proper Names in Cross-Lingual Information Retrieval

**Authors:** Paola Virga and [Sanjeev Khudanpur](https://md.research.elamur.ai/free/en/authors/sanjeev-khudanpur/) · **Year:** 2003 · **Venue:** Proceedings of the ACL 2003 Workshop on Multilingual and Mixed-language Named Entity Recognition · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY-NC-SA 3.0 (ACL Anthology, before 2016)](https://creativecommons.org/licenses/by-nc-sa/3.0/)

**Links:** [Publisher's page](https://aclanthology.org/W03-1508/) · [PDF](https://aclanthology.org/W03-1508.pdf) · [DOI 10.3115/1119384.1119392](https://doi.org/10.3115/1119384.1119392) · [OpenAlex](https://openalex.org/W2055543848)

**Citations:** 169 citations (OpenAlex, 2026-10-14) · [63](https://md.research.elamur.ai/free/en/papers/virga-2003-transliteration-proper-names-cross-lingual/cited-by/) in our library, of the 141,840 papers whose bibliographies we read

## Summary
The paper tackles a gap in cross-lingual retrieval: proper names are often missing from bilingual lexicons, yet they are among the most informative query terms. The authors build a fully data-driven English-to-Chinese name transliterator. They use no hand-written phonological rules. An English name is first converted to phonemes with the Festival TTS system. A first IBM-style source-channel model maps the phonemes to Mandarin generalized initials and finals (GIFs). The GIFs are then deterministically turned into toneless pin-yin, and a second statistical model maps the pin-yin to Chinese characters. Training starts from about 3875 name pairs from Meng et al. and is later enlarged with filtered pairs from the noisy LDC Chinese-English named entity list. On pin-yin error rate, the approach matches or slightly beats the rule-seeded method of Meng et al. Selecting LDC data with a model trained on clean data cuts pin-yin errors from 51.1% to 42.5%. In English-query retrieval of Mandarin TDT-2 audio, mAP rises modestly from 0.501 to 0.517.

## Key points
- **Pipeline and tools:**
  - Four steps: Festival phonemes, then GIFs, then pin-yin, then characters.
  - The statistical steps are trained with GIZA++ (Model 1, Model 2, HMM, Model 4 iterations) and decoded with the ISI ReWrite decoder.
  - Both stages use trigram language models built with the CMU toolkit (Good-Turing smoothing, Katz back-off).
- **Vocabularies:** 43 English phonemes versus 58 GIF symbols (21 initials and 37 finals). Only 409 initial-final combinations are legal pin-yin. The pin-yin/character corpus has about 282 and about 680 distinct symbols.
- **Comparison with Meng et al. (2233 training / 1541 test pairs):**
  - Pin-yin error rate is 50.8% versus 52.5% for Meng et al.
  - Training on 3625 pairs gives 49.1% on a 250-pair test set.
  - End-to-end character error rate is 57.4%.
  - The pin-yin-to-character stage alone has a 12.6% character error rate.
- **LDC data filtering:**
  - From roughly 1M person/place name pairs, filtering with a self-trained model kept 343,738 training pairs and gave 51.5% pin-yin error, which is no real gain.
  - Filtering with the clean-data Big MT model (about 309K pairs) gave 42.5%.
- **CLIR on TDT-2:**
  - Baseline without name transliteration: mAP 0.501.
  - Big MT transliterator: 0.515.
  - Huge MT (Big MT-filtered) transliterator: 0.517, with p = 0.080.
  - Substituting LDC-list translations for most names reached only 0.506, which was not significant.

## Interesting ideas
- **Use a phonetic pivot.** Transliteration is treated as two "translation" problems over sound units rather than over spellings. This removes the need for target-language phonology rules. The same design could suit Latin-to-Cyrillic name mapping, for example English to Ukrainian.
- **Learn insertion symbols from the model.** The symbols that the decoder may insert freely (zero-fertility items) are read directly off the trained model. They turn out to be the vowel nuclei Mandarin needs to break up English consonant clusters. Linguistic knowledge thus emerges from alignment statistics instead of being hand-coded.
- **Filter noisy data with a clean-data model.** Alignment scores from a model trained on a small, trusted set are used to rank a large, noisy corpus. Scoring the noisy data with a model trained on that same noisy data did not help. This is a reusable recipe for cleaning web-scale or beta-quality parallel lists.
- **Repair illegal outputs cheaply.** When a smoothed language model allows invalid outputs, such as two initials in a row, a simple deterministic fix is applied. The system tries inserting a default vowel until the sequence becomes legal. This is a pragmatic guard for any constrained output space.
- **Check intrinsic gains extrinsically.** A large drop in pin-yin error translated into only small retrieval gains. The authors suspect speech recognition errors on out-of-vocabulary names are the cause. They also conjecture that using 10-best transliterations rather than a single output may add robustness against these errors. The broader lesson is to verify intrinsic improvements in the downstream task, especially over ASR output.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [Johns Hopkins University](https://md.research.elamur.ai/free/en/universities/johns-hopkins-university/) — University, Baltimore, United States · 187 works · 92,194 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,086 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,889 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works

🧭 Topics (2): [🌍 Multilingual & cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/) · 467 works · [🔍 Information retrieval & search](https://md.research.elamur.ai/free/en/topics/information-retrieval/) · 145 works

## Cite

Cite as: Paola Virga and Sanjeev Khudanpur. 2003. Transliteration of Proper Names in Cross-Lingual Information Retrieval. Proceedings of the ACL 2003 Workshop on Multilingual and Mixed-language Named Entity Recognition.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{virga2003transliteration,
  title = {{Transliteration of Proper Names in Cross-Lingual Information Retrieval}},
  author = {Paola Virga and Sanjeev Khudanpur},
  booktitle = {Proceedings of the ACL 2003 Workshop on Multilingual and Mixed-language Named Entity Recognition},
  year = {2003},
  doi = {10.3115/1119384.1119392},
  url = {https://doi.org/10.3115/1119384.1119392}
}
```

- [← previous Fast Methods for Kernel-Based Text Analysis](https://md.research.elamur.ai/free/en/papers/kudo-2003-fast-methods-kernel-based-text/)
- [next → Loosely Tree-Based Alignment for Machine Translation](https://md.research.elamur.ai/free/en/papers/gildea-2003-loosely-tree-based-alignment-machine/)
- [↑ back to papers of 2003](https://md.research.elamur.ai/free/en/papers/2003/)

## References

10 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Peter F. Brown, Stephen A. Della Pietra, Vincent J. Della Pietra, and Robert L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263-311. [find](https://scholar.google.com/scholar?q=Peter+F%2E+Brown%2C+Stephen+A%2E+Della+Pietra%2C+Vincent+J%2E+Della+Pietra%2C+and+Robert+L%2E+Mercer%2E+1993%2E+The+mathematics+of+statistical+machine+translation%3A+Parameter+estimation%2E+Computational+Linguistics%2C+19%282%29)
2. Sung Young Jung, SungLim Hong, and Eunok Paek. 2000. An English to Korean Transliteration Model of Extended Markov Window. Proceedings of COLING. [find](https://scholar.google.com/scholar?q=Sung+Young+Jung%2C+SungLim+Hong%2C+and+Eunok+Paek%2E+2000%2E+An+English+to+Korean+Transliteration+Model+of+Extended+Markov+Window%2E+Proceedings+of+COLING%2E)
3. K. Knight and J. Graehl. 1997. [Machine Transliteration](https://md.research.elamur.ai/free/en/papers/knight-1998-machine-transliteration/). Proceedings of ACL.
4. Paul McNamee and Jim Mayfield. 2001. JHU/APL Experiments at CLEF-2001: Translation Resources and Score Normalization. Proceedings of CLEF. [find](https://scholar.google.com/scholar?q=Paul+McNamee+and+Jim+Mayfield%2E+2001%2E+JHU%2FAPL+Experiments+at+CLEF%2D2001%3A+Translation+Resources+and+Score+Normalization%2E+Proceedings+of+CLEF%2E)
5. Paul McNamee and Jim Mayfield. 2002. Comparing Cross-Language Query Expansion Techniques by Degrading Translation Resources. Proceedings of SIGIR. [find](https://scholar.google.com/scholar?q=Paul+McNamee+and+Jim+Mayfield%2E+2002%2E+Comparing+Cross%2DLanguage+Query+Expansion+Techniques+by+Degrading+Translation+Resources%2E+Proceedings+of+SIGIR%2E)
6. Helen M. Meng et al.̇ 2000. Mandarin-English Information (MEI): Investigating Translingual Speech Retrieval. Technical Report for the Johns Hopkins Univ. Summer Workshop. [find](https://scholar.google.com/scholar?q=Helen+M%2E+Meng+et+al%2E%CC%87+2000%2E+Mandarin%2DEnglish+Information+%28MEI%29%3A+Investigating+Translingual+Speech+Retrieval%2E+Technical+Report+for+the+Johns+Hopkins+Univ%2E+Summer+Workshop%2E)
7. Helen M. Meng, Wai-Kit Lo, Berlin Chen, and Karen Tang. 2001. Generating Phonetic Cognates to Handle Named Entities in English-Chinese Cross-Language Spoken Document Retrieval. Proceedings of ASRU. [find](https://scholar.google.com/scholar?q=Helen+M%2E+Meng%2C+Wai%2DKit+Lo%2C+Berlin+Chen%2C+and+Karen+Tang%2E+2001%2E+Generating+Phonetic+Cognates+to+Handle+Named+Entities+in+English%2DChinese+Cross%2DLanguage+Spoken+Document+Retrieval%2E+Proceedings+of+ASRU%2E)
8. Practical Chinese Reader, Book I. The Commercial Press LTD. 1981. [find](https://scholar.google.com/scholar?q=Practical+Chinese+Reader%2C+Book+I%2E+The+Commercial+Press+LTD%2E+1981%2E)
9. Bonnie Glover Stalls and Kevin Knight. 1998. Translating Names and Technical Terms in Arabic Text. Proceedings of the COLING/ACL Workshop on Computational Approaches to Semitic Languages. [find](https://scholar.google.com/scholar?q=Bonnie+Glover+Stalls+and+Kevin+Knight%2E+1998%2E+Translating+Names+and+Technical+Terms+in+Arabic+Text%2E+Proceedings+of+the+COLING%2FACL+Workshop+on+Computational+Approaches+to+Semitic+Languages%2E)
10. Djoerd Hiemstra. 2001. Using Language Models for Information Retrieval. Ph.D. thesis,University of Twente, Netherlands. [find](https://scholar.google.com/scholar?q=Djoerd+Hiemstra%2E+2001%2E+Using+Language+Models+for+Information+Retrieval%2E+Ph%2ED%2E+thesis%2CUniversity+of+Twente%2C+Netherlands%2E)

*Updated 24 October 2026*