Abstract
Foreign name translations typically include multiple spelling variants. These variants cause data sparseness problems, increase Out-of-Vocabulary (OOV) rate, and present challenges for machine translation, information extraction and other NLP tasks. This paper aims to identify name spelling variants in the target language using the source name as an anchor. Based on word-to-word translation and transliteration probabilities, as well as the string edit distance metric, target name translations with similar spellings are clustered. With this approach tens of thousands of high precision name translation spelling variants are extracted from sentence-aligned bilingual corpora. When these name spelling variants are applied to Machine Translation and Information Extraction tasks, improvements over strong baseline systems are observed in both cases.
Introduction
In natural language processing, foreign names present a significant challenge due to transliteration inconsistencies. When names are moved from one script to another—such as Arabic or Chinese to English—multiple valid but inconsistent spellings often emerge. For instance, the Arabic name for a specific region might be rendered as al-Kharroub, Al-Kharrub, Al-Kharub, or al-Khurub.
These inconsistencies lead to three primary issues:
- Data Sparseness: Frequency counts for a single entity are split across multiple variants, hindering pattern recognition.
- Out-of-Vocabulary (OOV) Errors: Systems may fail to recognize a known entity if it appears with a slightly different spelling.
- Document Incoherence: Machine Translation (MT) systems may inconsistently translate the same name within a single document, confusing the reader.
This research proposes a method to normalize these variants by mapping them to a single canonical form using the source-language name as a stable "anchor."
Methodology
The authors developed a cross-lingual pipeline that operates without pre-existing named entity taggers, relying instead on three main stages:
1. Candidate Generation
Using sentence-aligned bilingual corpora, the system collects word co-occurrence statistics to identify potential target-language translations for a given source name.
2. Clustering
Target-language words with high string similarity are grouped using an agglomerative clustering algorithm based on edit distance. This initial grouping identifies words that look similar in the target script.
3. Transliteration Validation
To distinguish actual name variants from semantically unrelated but orthographically similar words, the system calculates a transliteration cost. A character-based Hidden Markov Model (HMM) is trained to align phonetic components between languages. If the phonetic "cost" of transforming the source name into the target cluster is low, the cluster is validated as a set of name translations.
For Chinese names, the researchers introduced a phonetic intermediate step: converting characters to Pinyin. Given the vast number of Chinese characters, this Romanized representation significantly improved the accuracy of the transliteration cost calculation.
Data and Evaluation
The study leveraged large-scale datasets from the Linguistic Data Consortium (LDC):
- Arabic-English: 5.6 million sentence pairs.
- Chinese-English: 8.7 million sentence pairs.
Evaluation was conducted on two fronts:
- Machine Translation: Tested on the GALE 2007 Development set.
- Information Extraction (IE): Evaluated using ACE 2007 (NIST) benchmarks for mention detection.
Results and Discussion
The proposed method achieved high precision in extracting name variants across both language pairs. The top 22,000 Arabic-English name pairs were extracted with 96.9% precision. For Chinese-English, the Pinyin-based model reached 93.6% precision, outperforming the direct character-based approach.
Impact on Downstream Tasks
- Information Extraction: The system provided a statistically significant improvement in mention detection. In English, the F-measure rose from 82.4 to 82.7.
- Machine Translation: While the BLEU score increase was modest (0.1), the Translation Edit Rate (TER) improved by 0.26. An "oracle" experiment indicated the system corrected approximately 73% of the potential name errors in the dataset.
A key insight is the benefit of "probability merging." In standard MT, translation probabilities are diluted across variants. By merging these into a canonical form, the correct translation receives a higher probability score, aiding the decoder's selection process.
Conclusion
While the improvements in MT were constrained by the fact that the baseline was trained on the same data used for variant discovery, the research demonstrates a robust method for name normalization. The system effectively distinguishes phonetic similarity (transliteration) from semantic meaning (translation), preventing incorrect clustering of related but non-name words. Future work will explore applying these techniques to cross-document coreference and more complex information extraction tasks.
Summary updated