TL;DR
This paper presents a novel methodology for automatically building large semantic taxonomies from conventional, unstructured dictionaries (MRDs). By combining unsupervised algorithms for extracting and ranking taxonomic links, the system achieves high accuracy, demonstrating that rich lexical resources can be built without relying on pre-coded semantic information.
What Problem Does This Paper Address?
Developing robust Natural Language Processing (NLP) systems heavily relies on comprehensive lexical and semantic resources like Lexical Knowledge Bases (LKBs) and ontologies. These resources often encode hierarchical "class-subclass" relationships, crucial for inferencing and understanding. However, building these resources is a significant bottleneck, especially for languages other than English, and often requires manually curated, highly structured dictionaries (like LDOCE). This paper tackles the challenge of creating accurate, large-scale taxonomies from readily available, but less structured, monolingual dictionaries that lack explicit semantic coding.
Methods
The proposed system employs a mixed methodology that combines descriptive and prescriptive approaches.
- Automatic Extraction of Taxonomic Links: The core of the method involves automatically extracting "genus" (hypernym) relationships from dictionary definitions. This is achieved through a multi-step process:
- Genus Identification: Parsing definitions to identify the genus term.
- Genus Disambiguation: A crucial step to resolve ambiguity when a genus term can refer to multiple concepts.
- Conceptual Distance: Utilizing a measure of conceptual distance based on shortest paths in a semantic network (WordNet) to link dictionary senses to WordNet synsets.
- Attaching Dictionary Senses to Semantic Primitives: To overcome the limitations of purely descriptive approaches, the system uses a set of pre-defined semantic primitives. In this work, the 24 lexicographer's files (semantic categories) from WordNet 1.5 were used as these primitives. Dictionary senses are then attached to these primitives using:
- Conceptual Distance: As mentioned above, to map dictionary senses to WordNet synsets.
- Salient Words: Identifying words that are highly indicative of a particular semantic category using a mutual information-like formula. Dictionary definitions are then enriched with these salient words to improve classification.
- Ranking and Filtering: Extracted relations are ranked to allow for selective manual refinement. This involves several filtering criteria (F1, F2, F3) applied to genus terms to select the most relevant ones for a given semantic category, ensuring accuracy and coverage.
Datasets
- Diccionario General Ilustrado de la Lengua Española (DGILE): The primary monolingual Spanish dictionary used for building the taxonomies.
- WordNet 1.5: Used as a semantic network for computing conceptual distances and as a source of semantic primitives (lexicographer's files).
- Bilingual Spanish-English Dictionary: Utilized to bridge the gap between the Spanish DGILE and the English WordNet.
Key Insights and Ideas
- Leveraging Unstructured Dictionaries: The paper demonstrates that even conventional, non-semantically coded dictionaries can be a rich source for building high-quality semantic taxonomies, significantly expanding the availability of such resources.
- The Power of Hybrid Approaches: Combining unsupervised extraction with a minimal prescribed structure (WordNet semantic files) proves more effective than purely descriptive methods.
- Conceptual Distance as a Bridge: Using conceptual distance in WordNet effectively links senses from a monolingual dictionary to a structured semantic resource, even when direct translations are not perfect.
- Salient Words for Improved Classification: Identifying and using "salient words" associated with semantic categories significantly improves the accuracy of attaching dictionary senses to these categories.
- Filtering is Key for Accuracy: The proposed filtering criteria (F1, F2, F3) are crucial for refining the extracted taxonomic links and achieving high accuracy, allowing for a trade-off between coverage and precision.
Key Results
- The system achieved an accuracy rate of 80% for semantically labelling DGILE dictionary senses with WordNet lexicographer's files after the second labelling pass.
- The final taxonomy building process, using filtering criteria F2+F3>9, resulted in a FOOD taxonomy with 952 senses, more than double the size of a manually acquired taxonomy.
- Using a less strict criterion (F2+F3>4), the taxonomy size grew to 1,242 senses.
- The method demonstrated that taxonomy building is not limited to highly structured dictionaries like LDOCE.
Limitations and Future Work
- Accuracy Trade-offs: While the system achieves high accuracy, there's an inherent trade-off between coverage and precision, influenced by the chosen filtering thresholds.
- Language Gap: The reliance on a bilingual dictionary introduces potential noise and limitations, especially for less common word pairings.
- Manual Refinement: The paper acknowledges that while the system allows for selective manual refinement, a fully automatic process might still require further improvements.
- Future Work: The authors suggest iterating the process by collecting more salient words from a larger corpus (the second labelled lexicon) and re-estimating semantic tags. They also propose exploring finer-grained semantic primitives for more detailed taxonomies. The potential for building taxonomies for other languages using this methodology is also implied.
Summary updated