Abstract
This paper proposes a novel method for lexicon extraction that extracts translation pairs from comparable corpora by using graph-based label propagation. In previous work, it was established that performance drastically decreases when the coverage of a seed lexicon is small. We resolve this problem by utilizing indirect relations with the bilingual seeds together with direct relations, in which each word is represented by a distribution of translated seeds. The seed distributions are propagated over a graph representing relations among words, and translation pairs are extracted by identifying word pairs with a high similarity in the seed distributions. We propose two types of graphs: a co-occurrence graph, representing co-occurrence relations between words, and a similarity graph, representing context similarities between words. Evaluations using English and Japanese patent comparable corpora show that our proposed graph propagation method outperforms conventional methods. Further, the similarity graph achieved improved performance by clustering synonyms into the same translation.
Introduction
Building bilingual lexicons is a fundamental task for machine translation and cross-lingual information retrieval. While parallel corpora (direct translations) are the gold standard for this task, they are often unavailable for low-resource language pairs or specialized technical domains. Researchers instead rely on comparable corpora—texts that discuss similar topics but are not direct translations.
Traditional extraction methods assume that a word and its translation share similar contexts. However, these methods typically require a "seed lexicon" to map context words between languages. When this seed lexicon is small, the mapping becomes sparse, and extraction accuracy drops significantly. This paper addresses this bottleneck by leveraging indirect relations through graph-based label propagation.
The Proposed Graph-Based Method
The core innovation of this research is the use of Label Propagation (LP) to represent words as distributions of seed translations. Instead of looking only at direct context neighbors, the method allows information to flow through a network of word relationships.
1. Graph Construction
The authors construct graphs for both the source and target languages. Two types of edges were evaluated:
- Co-occurrence Graph: Edges connect words that frequently appear together in a fixed window.
- Similarity Graph: Edges connect words that share similar contexts, even if they never co-occur. This approach is particularly effective at grouping synonyms and spelling variants.
2. Seed Propagation and Extraction
Known translation pairs from a seed lexicon are treated as labels. These labels are iteratively propagated through the graph. After propagation, every word is represented by a distribution of seed words it is related to, either directly or indirectly. To find a translation, the system compares the seed distribution of a Japanese word against all English words using cosine similarity.
Experimental Evaluation
The method was tested using English and Japanese patent documents from the NTCIR-8 task. The researchers compared their method against established baselines (Rapp and Andrade) using two seed lexicon sizes: LexL (large, ~28k pairs) and LexS (small, ~2.7k pairs).
Key Results
The proposed method, specifically using the Similarity Graph (Sim), demonstrated superior performance and robustness:
- Small Seed Performance: With the limited LexS dataset, the
Simmethod achieved 4.1% Top-1 accuracy, more than doubling Rapp's 1.5%. At Top-20,Simreached 11.5%, nearly tripling the baseline. - Large Seed Performance: Even with a comprehensive seed lexicon, the
Simmethod maintained a significant lead, achieving 30.6% Top-20 accuracy compared to Rapp's 17.6%.
Discussion and Conclusion
The results confirm that utilizing indirect relations through label propagation effectively mitigates the sparsity issues inherent in comparable corpora. The Similarity Graph proved particularly powerful because it naturally clusters synonyms, ensuring that related terms receive similar seed distributions.
While challenges remain—such as handling word sense ambiguity (polysemy) and the computational cost of graph propagation—this method provides a robust framework for expanding bilingual dictionaries in specialized domains where parallel data is scarce. Future work will focus on integrating word sense disambiguation and testing the extracted lexicons in end-to-end machine translation systems.
Summary updated