TL;DR
This paper introduces a novel approach to automatically discover transliterations of proper nouns on the web. By leveraging morphological query expansion and information retrieval, the system effectively expands search queries to retrieve relevant transliterations, significantly improving recall and addressing the limitations of traditional machine translation dictionaries.
What Problem Does This Paper Address?
Machine translation systems often struggle with proper nouns and technical terms due to their limited vocabulary coverage. This "out-of-vocabulary" (OOV) problem leads to poor translation quality. While dedicated transliteration services exist, they may also face challenges with comprehensive coverage. The authors aim to address this by developing a method to automatically learn and find transliterations of proper nouns directly from the vast resources available on the World Wide Web, thereby enhancing the robustness of machine translation and cross-lingual information retrieval.
Methods
The core of the proposed method is morphological query expansion for machine transliteration. This approach involves two main stages:
- Training Stage: The system learns morphological relationships between name prefixes/suffixes and their transliterated counterparts. This is achieved by pairing morphemes from a training dataset of names and transliterations, calculating statistical associations, and identifying pairs with high association. This stage effectively builds a model of how names are transliterated.
- Runtime Stage: When a given name needs transliteration, the system automatically expands the query by incorporating relevant prefixing morphemes learned during training. This expanded query is then submitted to a search engine. After retrieving search snippets, the system extracts potential transliterations based on redundancy, proximity to the original name, and cross-lingual morphological relationships.
This method is presented as a general approach applicable to both translation and transliteration, with a specific focus on transliteration in this paper.
Datasets
- Training Data: The authors mention using a collection of 23,615 place names and transliterations pairs for training.
- Evaluation Data: The prototype system was evaluated using a list of 500 proper names.
The paper does not introduce a new benchmark dataset but rather utilizes existing web data for its retrieval-based approach.
Key Insights and Ideas
- Leveraging the Web as a Transliteration Dictionary: The paper demonstrates that the web, with its vast amount of mixed-language content, can serve as a rich source for discovering transliterations, overcoming the limitations of manually compiled dictionaries.
- Morphological Query Expansion is Key: By intelligently expanding queries with learned morphological components (e.g., "Parthian" becomes "Parthian 巴" or "Parthian 帕"), the system biases search engines towards retrieving relevant transliterated results.
- Cross-Lingual Morphological Relationships: The method effectively captures how phonetic or structural elements of a name in one language correspond to elements in its transliteration in another, even without explicit bilingual lexicon content.
- Improving Recall Significantly: The proposed method shows a substantial improvement in recall, indicating its effectiveness in finding transliterations that might otherwise be missed.
Key Results
- The system achieved a recall rate of 0.60, a significant improvement from the baseline of 0.48 without query expansion.
- The top 1, 2, and 3 results returned by the system had precision scores of 0.88, 0.93, and 0.94, respectively, indicating high accuracy in the retrieved transliterations.
- The system demonstrated that 60% of the time, there was sufficient relevant data on the web to perform effective machine transliteration.
- The prototype system supports functionalities like searching and extracting transliterations, listing alternative transliterations, browsing snippets, and saving results.
Limitations and Future Work
- Limitations:
- The effectiveness of the system is dependent on the availability of sufficient relevant data on the web for a given proper noun.
- While the recall improved, it is not 100%, suggesting that some transliterations may still be missed.
- Future Work:
- The authors suggest extending this approach to support other machine translation subtasks, such as back-transliteration and noun phrase translation.
- They propose combining this IR-based approach with statistical machine translation for potentially improved translation quality, for example, by using a preprocessing unit to replace proper names with transliterations before final translation.
- The method's independence from bilingual lexicon content makes it adaptable to other types of proper names (e.g., person, product, organization names).
Summary updated