TL;DR
This paper presents a novel method for automatically building large, accurate semantic taxonomies from conventional, monolingual machine-readable dictionaries (MRDs). By combining unsupervised extraction with a ranking mechanism and leveraging existing structured resources like WordNet, the authors demonstrate that rich lexical knowledge bases can be constructed without relying on highly curated, semantically coded dictionaries.
What Problem Does This Paper Address?
Developing robust Natural Language Processing (NLP) applications hinges on having comprehensive lexical and semantic resources, such as Lexical Knowledge Bases (LKBs). However, a significant bottleneck is the creation of these resources, especially for languages other than English and when using readily available, but less structured, dictionaries. Existing methods often rely on rich, explicitly coded dictionaries (like LDOCE) and struggle to achieve comparable accuracy with more common, monolingual MRDs that lack explicit semantic coding. This paper aims to bridge this gap by proposing a method to build accurate LKBs from such conventional dictionaries, addressing the "lexical bottleneck" for a wider range of languages and resources.
Methods
The core of the proposed method involves a two-stage process:
- Automatic Extraction and Ranking of Taxonomic Links: The system automatically extracts potential "genus" (hypernym) links from dictionary definitions. Crucially, it then ranks these extracted relations. This ranking is key, as it allows for selective manual refinement, making the process more efficient and manageable.
- Leveraging External Semantic Primitives: To overcome the limitations of purely descriptive approaches from MRDs, the method incorporates a minimal, prescribed ontology. Specifically, it uses the 24 semantic files (lexicographer's categories) from WordNet 1.5 as semantic primitives. The Spanish dictionary Diccionario General Ilustrado de la Lengua Española (DGILE) is then mapped to these WordNet primitives.
The process involves:
- Conceptual Distance: A method based on conceptual distance (using WordNet's semantic net) is employed to initially attach DGILE dictionary senses to WordNet semantic files.
- Salient Word Collection: For each semantic file, "salient words" (words that are highly representative of that category) are identified using a formula similar to Mutual Information.
- Enrichment and Labeling: DGILE definitions are then re-labeled based on the presence of these salient words, improving the accuracy of the semantic tagging.
- Filtering: Several filtering criteria (F1, F2, F3) are applied to refine the extracted genus terms, removing those that are less likely to be correct hypernyms or are more relevant to other semantic categories.
Datasets
- DGILE (Diccionario General Ilustrado de la Lengua Española): The primary monolingual Spanish dictionary used for building the taxonomies.
- WordNet 1.5: Used as the source of semantic primitives (24 lexicographer's files) and for calculating conceptual distances.
- Bilingual Spanish/English Dictionary: Used to bridge the gap between the monolingual Spanish dictionary and the English-based WordNet.
- Castellón (1993) results: Used as a baseline for comparison, representing a manually acquired taxonomy for the FOOD domain from DGILE.
Key Insights and Ideas
- MRDs are Valuable: The paper strongly argues that conventional, monolingual MRDs, even without explicit semantic coding, are valuable resources for building large-scale semantic taxonomies.
- Ranking is Crucial: Simply extracting all possible links is insufficient. Ranking these links allows for efficient manual correction, making the process scalable.
- Hybrid Approach Works: Combining a descriptive approach (extracting from MRDs) with a prescriptive one (using external semantic primitives like WordNet categories) yields better results than either alone.
- Salient Words as Indicators: Identifying and using "salient words" within dictionary definitions is an effective way to improve the accuracy of semantic classification.
- Filtering for Precision: Applying carefully designed filters to genus terms significantly improves the precision and quality of the resulting taxonomies.
Key Results
- The proposed method achieves an accuracy of around 100% depending on the degree of coverage selected.
- The initial automatic labeling of DGILE definitions achieved 64% accuracy at the file level. After enrichment with salient words, this improved to 80% coverage with 80% accuracy.
- The constructed FOOD taxonomy using the proposed method (F2+F3>9) resulted in 952 senses, more than twice the size of the manually acquired baseline (392 senses).
- Using a less strict criterion (F2+F3>4), the taxonomy grew to 1,242 senses, over three times larger than the baseline.
- The method successfully builds taxonomies that are significantly larger and more detailed than manually constructed ones, though they tend to be "flatter" due to the filtering of inner taxonomic genus terms.
Limitations and Future Work
- Accuracy of Initial Labeling: The initial automatic labeling process, while improved, still has errors (around 20% at the file level). This can lead to misclassifications, particularly for terms that can belong to multiple semantic categories (e.g., "fish" being both an ANIMAL and FOOD).
- "Flatter" Taxonomies: The filtering process, while improving precision, can lead to taxonomies that are less deep than ideal, as it restricts the selection of inner taxonomic genus terms.
- Future Work: The authors suggest iterating the process by collecting new sets of salient words from the second labeled dictionary to re-estimate semantic tags. They also propose exploring finer-grained semantic primitives beyond the 24 WordNet files, potentially by considering direct hyponyms of synsets or even selecting specific semantic categories based on application needs. Further experiments could also investigate the impact of different filtering criteria and the potential for cross-lingual applications.
Summary updated