Abstract
This paper proposes a completely unsupervised, corpus-based method for automatically aligning ontologies across different languages—specifically English WordNet and Chinese HowNet. By using "seedword groups" and context vectors derived from non-parallel corpora, the authors successfully map concepts between these structurally diverse ontologies with 66.3% accuracy, outperforming a standard translation-based baseline.
Introduction
Building multilingual ontologies (structured frameworks of concepts and their relationships) is a massive undertaking. For context, the EuroWordNet project for eight European languages took 11 institutions three years to complete. While many languages already have high-quality monolingual ontologies, they are often built using entirely different philosophies and structures, making them difficult to "stitch together."
The core challenge is translation ambiguity. A single word in one language often has multiple meanings that translate to different words in another. Without a way to disambiguate these senses, a simple dictionary-based alignment fails; for example, the authors found that a single definition in the Chinese HowNet could map to an average of 8.1 different synonym sets (synsets) in the English WordNet.
Methodology
The authors introduce a method that measures Cross-lingual Semantic Similarity without requiring parallel (translated) text.
Seedword Groups
Instead of using 1-to-1 word translations as anchors, they use "m-to-n" translation groups. This increases robustness because even if one specific word is rare or translated inconsistently, the group as a whole provides a stable semantic signal.
Context Vectors and Alignment
For every word of interest, they build a vector based on how often it co-occurs with these seedword groups in a large corpus. The weight of each element in the vector is calculated using a variation of TF-IDF. To align a concept:
- For a given concept (synset) in Ontology A, they identify all potential candidate synsets in Ontology B based on dictionary translations.
- They calculate the average similarity between the words in the source synset and the words in the candidate synsets using the cosine similarity of their context vectors.
- The candidate with the highest similarity score is selected as the match.
Evaluation
The method was evaluated using WordNet 1.7 (English) and HowNet (Mandarin Chinese). The corpora used were the Wall Street Journal (15 million English words) and People’s Daily (11.6 million Chinese words). Notably, these corpora are non-parallel and non-comparable, covering different topics and styles.
Quantitative Results
On a test set of 160 randomly chosen HowNet definitions, the algorithm achieved 66.3% accuracy. This significantly outperformed the baseline (58.8%), which simply chose the synset with the highest number of matching translations.
Qualitative Analysis
The system excelled at disambiguating senses. For example, it correctly mapped the Chinese concept for "population" (as in "number of people") to the correct WordNet sense (inhabitants of a territory) rather than the biological sense (a group of organisms) by analyzing the words it co-occurs with in the news.
Discussion and Limitations
- Corpus Mismatch: The authors noted that the Wall Street Journal and People’s Daily are very different in style. This "fan-shaped" divergence in word usage made alignment harder. Using more comparable corpora would likely improve results.
- Ontology Gaps: Some concepts are culturally specific. Chinese idiomatic expressions often have no single-word equivalent in English WordNet, leading to "unalignable" definitions.
- Seedword Sensitivity: Performance is highly dependent on the quality and frequency of the seedwords. Future work could involve more sophisticated automatic selection of these anchors.
Conclusion
This research demonstrates that complex semantic structures can be aligned across languages using unrelated text from two different cultures. By shifting from 1-to-1 word anchors to "seedword groups," the method provides a robust, unsupervised path toward building comprehensive multilingual ontologies.
Summary updated