TL;DR
This paper introduces TermMine, a novel system that automatically learns to find English-to-Chinese transliterations of proper nouns on the web. By intelligently expanding search queries with relevant morphemes, TermMine significantly improves the precision and recall of transliteration extraction compared to traditional machine translation systems.
What Problem Does This Paper Address?
Machine translation (MT) systems often struggle with proper nouns (like names of people, places, and organizations) because they are not always present in standard bilingual dictionaries. This "out-of-vocabulary" (OOV) problem can lead to poor translation quality. While some online services and specialized tools exist for transliteration, they often rely on manually compiled dictionaries or parallel corpora, which can be insufficient for the vast and ever-growing number of proper nouns. The paper aims to address this gap by developing a method to effectively and automatically discover transliterations of proper nouns directly from the web.
Methods
The core of TermMine's approach is a morphological query expansion strategy. Instead of just searching for the original proper noun, the system learns relationships between English name prefixes/suffixes and their corresponding Chinese transliteration morphemes.
Here's a breakdown of the key techniques:
- Learning Morphological Relationships: The system is trained on a bilingual list of name-transliteration pairs. It analyzes prefixes and suffixes of both the English names and their Chinese transliterations to build probabilistic models. For example, it learns that an English prefix like "par-" is likely to correspond to Chinese prefixes like "巴" (Ba) or "帕" (Pa).
- Homophone Classes: To handle data sparseness and variations in transliteration (e.g., different Chinese characters that sound similar), the system groups homophonic characters into "homophone classes." This allows for more robust probability estimation.
- Probabilistic Transliteration Model: A probabilistic function is used to estimate the likelihood of a Chinese string being a transliteration of an English name, considering both prefix and postfix relationships.
- Query Expansion: At runtime, given an English proper noun, TermMine generates a set of expanded queries by combining the original name with its most probable transliteration morphemes. For instance, to find transliterations for "Acton," it might generate queries like "Acton 阿," "Acton 亞," etc.
- Search and Extraction: These expanded queries are sent to a search engine. The system then filters the returned snippets to ensure they contain a sufficient proportion of target-language text and extracts potential transliterations.
- Ranking: Extracted transliterations are ranked using the learned probabilistic model.
Datasets
- Training Data: The system was trained on a dataset of 23,615 name-transliteration pairs, sourced from NICT, Taiwan. This dataset includes Romanized names from various languages and their Chinese counterparts, which are sometimes a mix of translation and transliteration.
- Testing Data: A subset of 500 randomly selected place names was used for testing.
Key Insights and Ideas
- Web as a Corpus for Transliteration: The paper demonstrates that the vast amount of information on the web can be effectively leveraged to discover transliterations, even for proper nouns not found in traditional dictionaries.
- Morphological Clues are Powerful: Focusing on the sub-lexical (prefix/suffix) relationships between names and their transliterations is a highly effective strategy for query expansion.
- Handling Ambiguity with Homophones: The use of homophone classes is a clever way to mitigate data sparsity and improve the robustness of transliteration probability estimation.
- Query Expansion for Targeted Search: By intelligently expanding queries, the system biases search engines towards retrieving relevant transliterated content, significantly improving efficiency and accuracy.
Key Results
- Improved Applicability: With query expansion (QE+), TermMine achieved an applicability rate of 60%, meaning it found transliterations for 60% of the test cases, a significant improvement over the 48% without QE.
- High Precision: TermMine achieved a precision rate of 0.89 (with QE), meaning that when it returned a transliteration, it was correct 89% of the time. This is substantially higher than Google Translate (0.46) and Yahoo! Babelfish (0.52).
- Higher Recall: TermMine returned 300 transliterations for the 500 test names, far exceeding Google Translate (146) and Yahoo! Babelfish (44).
- Outperforms Commercial Systems: The paper concludes that TermMine outperforms both Google Translate and Yahoo! Babelfish in the area of machine transliteration of proper names.
Limitations and Future Work
- Limitations:
- The system's performance is dependent on the quality and coverage of the training data.
- Snippets with a very low proportion of target-language text are discarded, which might lead to missing some transliterations.
- The current approach focuses on proper nouns and might need adaptation for general text translation.
- Future Work:
- Extending the method to support back transliteration and noun phrase translation.
- Integrating this IR-based approach with statistical machine translation for sentence-level translation, potentially by pre-processing sentences to replace proper nouns with their transliterations.
- Applying the method to other proper name types (e.g., products, organizations).
Summary updated