TL;DR
This paper introduces a novel method for automatically discovering name translation pairs (like "国家情报局" and "National Intelligence Service") from comparable corpora. By constructing and aligning "bilingual information networks," the authors achieve high accuracy in identifying translations for people, geopolitical entities, and organizations, offering a more robust approach than traditional phonetic transliteration.
What Problem Does This Paper Address?
Accurate translation of named entities (people, places, organizations) is fundamental for many cross-lingual information processing tasks, such as information retrieval and machine translation. While previous methods often relied on phonetic similarity to find name translations, this approach is limited to names that are transliterated phonetically. Many names, especially organizations, are translated semantically (e.g., "解放之虎" to "Liberation Tiger"), and some translations are context-dependent. Furthermore, comparable corpora—texts about similar topics but not direct translations—are abundant (e.g., news articles released simultaneously in different languages). However, these corpora lack direct document or sentence-level alignments, making it challenging to extract reliable name translations. This paper aims to bridge this gap by leveraging the rich, albeit unaligned, information within comparable corpora.
Methods
The core of the paper's contribution lies in its novel approach of creating and aligning "bilingual information networks." Here's a breakdown:
- Bilingual Information Extraction: The authors employ state-of-the-art information extraction (IE) systems to process comparable corpora (Chinese and English in this case). These systems identify named entities (persons, GPEs, organizations), relations between them (e.g., "leader," "located"), and events, along with their confidence scores.
- Information Network Creation: For each language, the extracted entities and relations are used to construct directed graphs, forming "information networks." Nodes represent named entities, and edges represent attributes derived from relations (e.g., "leader," "located") and events (e.g., "Arrest/2001-06-25"). These attributes are designed to be language-independent.
- Information Network Alignment (Bootstrapping): The key innovation is aligning these bilingual information networks. Starting with a small set of known name translation pairs (seeds), the algorithm traverses the networks. If two entities in different language networks are linked by the same language-independent attribute and satisfy certain conditions (e.g., similar entity types, and for persons, a low Pinyin edit distance), their corresponding linked entities are considered potential new translation pairs. This process is repeated iteratively, adding newly discovered high-confidence pairs to the seed set, effectively bootstrapping the discovery of more translations.
Datasets
The authors used documents from the following corpora for their experiments:
- ACE (2004, 2005) training corpora: Used for manual evaluation.
- TDT-5 corpora: Used for manual evaluation.
- CHSet1, CHSet3: Chinese corpora from different news agencies and time frames.
- ENSet1, ENSet2, ENSet3, ENSet4: English corpora from different news agencies and time frames.
The paper does not introduce new benchmark datasets but rather utilizes existing comparable corpora to demonstrate its approach.
Key Insights and Ideas
- Leveraging Structure Over Phonetics: The paper demonstrates that by building and aligning structured information networks, one can overcome the limitations of purely phonetic transliteration, enabling the discovery of semantically translated names.
- Language-Independent Attributes are Key: The use of attributes derived from relations and events (e.g., "located," "Arrest/2001-06-25") is crucial. These attributes act as bridges between the language-specific information networks, allowing for cross-lingual alignment.
- Bootstrapping for Scalability: The iterative bootstrapping approach allows the system to gradually expand its knowledge of name translations, starting from a small seed set and progressively discovering more pairs.
- Importance of Corpus Quality: The results highlight that temporal and source similarity of comparable corpora significantly impact the performance of name translation mining, suggesting that more aligned corpora yield better results.
Key Results
- High Accuracy: After four iterations, using 100 seed name pairs from CH-Set3 and EN-Set3, the approach discovered 968 new name translation pairs with an accuracy of 82.9%.
- Entity Type Breakdown: The discovered pairs included 361 persons (76.4% accuracy), 384 geo-political names (87.5% accuracy), and 223 organization names (85.2% accuracy).
- Impact of Corpus Alignment:
- Using comparable corpora from the same time frame (CHSet1-ENSet1) yielded 42 name pairs with 81.0% accuracy.
- Using corpora from different time frames (CHSet1-ENSet2) resulted in fewer pairs (17) and lower accuracy (76.5%).
- Using corpora from different news sources also showed a performance drop.
- Impact of IE Errors: The paper notes that errors in the initial information extraction step can propagate. For instance, Chinese name boundary detection errors caused about 30% of incorrect name pairs. When using perfect IE annotations, accuracy increased significantly to 90.5%.
Limitations and Future Work
- Sensitivity to IE Errors: The performance is directly impacted by the quality of the initial information extraction. Errors in name tagging, relation extraction, or event extraction can lead to incorrect alignments and name translations.
- Dependence on Seed Set: While the bootstrapping approach is effective, the initial seed set size and quality can influence the overall discovery process.
- Future Work:
- Importing more efficient graph mining and alignment algorithms.
- Investigating unsupervised relation extraction to reduce reliance on supervised IE pipelines.
- Exploring the trade-off between coverage and accuracy by applying the discovered name pairs to downstream tasks like cross-lingual name search and machine translation.
- Using mis-aligned sub-graphs to detect novel or local information specific to one language.
Summary updated