TL;DR
This paper tackles the challenge of building a Chinese WordNet by automatically linking English WordNet synsets to Chinese translations found in a bilingual dictionary (E-C LDOCE). They introduce novel weighting strategies that leverage keyword matching and hierarchical relationships to improve the accuracy and coverage of this linking process.
What Problem Does This Paper Address?
Building comprehensive lexical databases like WordNet for languages other than English is a significant undertaking. While the EuroWordNet project has made strides in creating multilingual WordNets, the process of linking existing English WordNet to other languages is still costly and time-consuming. This paper focuses on the specific challenge of constructing a Chinese WordNet by leveraging a Machine-Readable Dictionary (MRD) – the Longman English-Chinese Dictionary of Contemporary English (E-C LDOCE). The core problem is how to accurately map English WordNet synsets (groups of synonymous words) to their corresponding Chinese translations within the dictionary, especially given the fine-grained nature of WordNet senses and the potential for ambiguity in both English and Chinese.
Methods
The authors propose a method for linking English WordNet synsets to Chinese translations from the E-C LDOCE. Their approach involves several key steps:
- Keyword Extraction: They extract keywords (nouns, verbs, adjectives, adverbs) from the glosses (definitions) of WordNet synsets and from the definitions in the E-C LDOCE.
- Initial Linking: They use these keywords to find potential matches between WordNet synsets and MRD definitions. The initial linking strategy assigns a "weighting score" based on the number of matching keywords, treating WordNet synsets as queries and MRD senses as documents.
- Competition Weighting: When a WordNet synset has multiple headwords (words associated with it), or when an MRD definition has multiple senses, a competition mechanism is employed. This involves calculating scores for each potential mapping and selecting the one with the highest score.
- Enhanced Weighting: To address limitations in dictionary definitions and leverage hierarchical information, they introduce an "enhanced weighting" function. This function incorporates information from hypernyms (more general terms) of a synset and also considers genus terms (classifying terms) found in ancestor synsets. This helps to disambiguate senses and improve linking accuracy.
Datasets
- WordNet 1.7: Used as the source for English synsets, containing 74,487 noun synsets and 12,753 verb synsets.
- Longman English-Chinese Dictionary of Contemporary English (E-C LDOCE): Used as the bilingual resource for Chinese translations. This dictionary features a controlled vocabulary and provides Chinese translations for English word senses.
Key Insights and Ideas
- Keywords are Crucial: The paper highlights the importance of keywords within dictionary definitions and WordNet glosses as primary clues for establishing links between the two resources.
- Leveraging Hierarchy is Key: Simply matching keywords isn't enough. The authors demonstrate that incorporating information from WordNet's hypernymy hierarchy significantly improves linking accuracy, especially for senses that are less clearly defined.
- Genus Terms Matter: Identifying and weighting "genus terms" (terms that classify a word) from ancestor synsets helps to refine the linking process, as these terms often provide crucial semantic context.
- Chinese Translations Need Post-Processing: Even after successful linking, the paper acknowledges that the direct Chinese translations might not always be perfectly aligned with WordNet's fine-grained senses. For instance, a single Chinese phrase might cover multiple distinct WordNet senses, requiring further processing for precise disambiguation.
Key Results
The paper reports on the performance of their linking strategies for noun and verb synsets:
- Nouns:
- Competition Weighting: Achieved 73.0% coverage with 92% precision.
- Competition Weighting + Hypernymy: Increased coverage to 81.7% with 94% precision.
- Full Enhanced Weighting: Reached 86.1% coverage with 92% precision.
- Verbs:
- Competition Weighting: Achieved 68.5% coverage with 94% precision.
- Competition Weighting + Hypernymy: Increased coverage to 78.8% with 92% precision.
- Full Enhanced Weighting: Reached 83.2% coverage with 92% precision.
They also noted that a significant portion of noun synsets (46,099) share no vocabulary with the E-C LDOCE, indicating a limitation in the dictionary's coverage for certain semantic categories.
Limitations and Future Work
- Dictionary Coverage: A notable limitation is that the E-C LDOCE does not cover all WordNet synsets, particularly in categories like "group, grouping," "plant, flora," and "region." This leaves a portion of WordNet synsets unlinked.
- Ambiguity in Chinese Translations: The paper points out that direct Chinese translations might be too broad and encompass multiple WordNet senses, necessitating further processing to achieve precise disambiguation.
- Future Work: The authors suggest that their work can serve as a post-processing step for other methods that parse Chinese translations and tag semantic components (referencing Ker and Song, 2002). They also envision the potential for creating a new Chinese word hierarchy by extracting semantic elements from these translations.
Summary updated