TL;DR
This paper tackles the challenging problem of translating rarely occurring named entities (NEs) by combining phonetic and semantic similarity measures. Their novel approach leverages a context vector model and cross-lingual retrieval to achieve significantly improved translation accuracy, outperforming state-of-the-art statistical machine translation systems.
What Problem Does This Paper Address?
Translating named entities (NEs) like people, locations, and organizations is crucial for many natural language processing tasks, including cross-lingual information retrieval and question answering. However, NE translation is particularly difficult because NEs are often phonetically transliterated or semantically translated, and there isn't always a direct one-to-one mapping between languages. Existing methods, like pre-compiled dictionaries, fail to handle the vast number of rarely occurring or new NEs. For instance, a significant percentage of automatically tagged Chinese NEs are not found in standard translation lexicons. This leads to incorrect translations that not only lose information but also distort the overall context of a translated text.
Methods
The core of this paper's contribution lies in its hybrid approach to NE translation. They propose a two-pronged strategy:
- Phonetic Similarity: This is estimated using a surface string transliteration model. This model learns how to convert characters or sounds from one language to another, focusing on how names sound.
- Semantic Similarity: This is calculated using a context vector semantic model. For a given NE, they construct a "context vector" representing the surrounding words and their importance (based on Part-of-Speech tags and proximity to the NE). The semantic similarity between two NEs is then defined as the "mutual translation probability" of their respective context vectors, essentially measuring how well the surrounding context of one NE translates to the context of another.
The overall process involves:
- Query Generation: Translating the source document containing the NE into the target language using a statistical machine translation system.
- Cross-lingual Retrieval: Using these translated documents as queries to search a large target language corpus (e.g., English news archives) for relevant documents.
- Candidate NE Identification: Automatically tagging NEs within the retrieved target language documents.
- Similarity Matching: Comparing the source NE with each retrieved target NE using both the phonetic and semantic similarity measures.
- Selection: The NE with the highest combined similarity score is selected as the correct translation.
Datasets
- Training Data: The paper mentions using a manually compiled Chinese-English dictionary from LDC containing 54,131 unique Chinese words for training the transliteration model. They also utilize a large bilingual corpus (6M words) for training phrase transducers.
- Corpus for Retrieval: A large corpus of 963,478 English documents from the Xinhua News Agency (over 7.3 million sentences, 200 million words) was indexed for cross-lingual retrieval.
- Evaluation Data: The NIST 2002 Machine Translation Evaluation test data, consisting of 100 Chinese documents, 878 sentences, and 25,430 words, was used for evaluation. This dataset contained 2469 automatically tagged NEs.
Key Insights and Ideas
- The Power of Hybridity: Combining phonetic and semantic similarity is crucial for robust NE translation, especially for transliterated names where sound is paramount but context is key to disambiguation.
- Context is King (for Semantics): The "context vector" approach effectively captures the topical relevance of surrounding words to an NE, providing a strong semantic signal. The distribution of word weights in the context vector shows a Gaussian-like pattern, emphasizing words closer to the NE.
- Cross-lingual Retrieval as a Bridge: Leveraging large target-language corpora through cross-lingual retrieval is an effective way to find potential translations for rare NEs that might not exist in bilingual dictionaries or parallel corpora.
- Iterative Improvement: The architecture allows for an iterative process where newly translated NEs can be fed back into the system to improve subsequent translation quality, creating a virtuous cycle.
Key Results
The proposed approach achieves a significant improvement in NE translation accuracy:
- 67% Accuracy: The method achieves 67% accuracy on translating rarely occurring NEs.
- Consistent Improvement: It consistently improves translation quality on different tasks compared to a state-of-the-art statistical machine translation system.
- Breakdown of Improvement:
- Using only the transliteration model increased precision by about 30%.
- Adding the context vector semantic model further improved accuracy by about 10%.
- Coverage: Document-based query/indexing for retrieval covered about 70% of correct NE translations, outperforming sentence-based methods (60%).
Limitations and Future Work
- Retrieved Document Coverage: A significant limitation (contributing to 50% of errors) is the limited coverage of retrieved documents, meaning correct NE translations might not be included or retrieved from the indexed corpus.
- Iterative Model Refinement: The paper demonstrates an iterative process for building NE translation dictionaries, suggesting that continued iteration could further refine the models.
- Integration into MT Systems: The authors highlight the potential for integrating translated NE pairs back into machine translation engines to improve overall translation quality, suggesting this as a direction for future work.
Summary updated