TL;DR
This paper presents a novel, fully automatic method for building large semantic taxonomies from conventional, monolingual machine-readable dictionaries (MRDs) without explicit semantic coding. The approach leverages unsupervised algorithms to extract and rank taxonomic links, enabling the creation of valuable lexical knowledge bases (LKBs) from readily available resources, thus addressing the "lexical bottleneck" in NLP.
What Problem Does This Paper Address?
Building comprehensive lexical knowledge bases (LKBs) is crucial for advancing Natural Language Processing (NLP) applications. However, a significant challenge, often termed the "lexical bottleneck," is the scarcity of accurate and broadly covering lexical resources. Existing approaches often rely on highly structured, semantically coded dictionaries (like LDOCE) and are primarily focused on English. This paper tackles the problem of acquiring accurate semantic taxonomies from conventional, monolingual MRDs that lack explicit semantic coding, aiming to make this process applicable to languages other than English and to a wider range of dictionary types. The authors highlight that directly extracting taxonomies from MRDs can lead to issues like circularity, errors, and a mismatch between the dictionary's structure and the knowledge needed for NLP systems.
Methods
The proposed system employs a two-pronged approach:
- Fully Automatic Extraction: The core of the method involves automatically extracting taxonomic links (genus-hyponym relationships) from dictionary definitions. This is achieved through a process that includes parsing definitions to identify genus terms and then using a conceptual distance measure based on WordNet's semantic net to attach dictionary senses to WordNet's semantic files (lexicographer's files).
- Ranking and Refinement: To make the extracted relations useful, the system ranks them. This involves:
- Salient Word Collection: Identifying "salient words" that are highly representative of specific semantic categories within the dictionary. This is done using a mutual information-like formula.
- Enrichment: Re-labeling dictionary definitions based on the presence of these salient words, increasing the confidence in the semantic classification.
- Filtering: Applying a series of filters (F1, F2, F3) to discard incorrect or less relevant genus terms. Filter 1 removes genus terms not mapped to the target semantic file, Filter 2 discards genus terms that are too general (frequent across many semantic categories), and Filter 3 removes infrequent genus terms within the target category.
This mixed methodology combines a descriptive approach (extracting from the MRD) with a minimal prescribed structure (WordNet's semantic primitives) to build the LKB.
Datasets
- Diccionario General Ilustrado de la Lengua Española (DGILE): The primary monolingual Spanish dictionary used for extracting taxonomic information.
- WordNet 1.5: Used as a semantic reference and for calculating conceptual distances between words. The 24 lexicographer's files (semantic files) of WordNet's noun synsets were used as the main semantic primitives.
- Bilingual Spanish/English Dictionary: Utilized to bridge the language gap and enable the attachment of Spanish dictionary senses to English WordNet synsets.
- Castellón (1993) FOOD Taxonomies: Used as a baseline for comparison of the automatically acquired FOOD taxonomies.
Key Insights and Ideas
- Leveraging Conventional MRDs: The paper demonstrates that even dictionaries without explicit semantic coding can be valuable sources for building large-scale semantic taxonomies, significantly expanding the potential for LKB construction.
- The Power of Unsupervised Methods: The success of the fully automatic extraction and ranking process highlights the effectiveness of unsupervised algorithms in tackling complex linguistic tasks.
- Bridging the Gap: The use of a bilingual dictionary and WordNet as an intermediary is a crucial step in connecting monolingual dictionary content to a structured semantic framework.
- Importance of Filtering: The effectiveness of the filtering criteria (F1, F2, F3) in improving accuracy and reducing noise is a key takeaway, showing that post-processing is vital for high-quality results.
- Taxonomy Granularity: The choice of filtering thresholds directly impacts the size and flatness of the resulting taxonomies, offering flexibility for different NLP application needs.
Key Results
- The system achieved an accuracy rate of 64% in the initial semantic labelling of DGILE definitions.
- After applying the salient word enrichment and filtering process, the accuracy improved to 80% at a sense (synset) level and 82% at a file level.
- The method successfully processed 86,759 DGILE definitions, covering over 93% of all noun definitions.
- For the FOOD category, the automatic approach generated taxonomies that were 2 to 3 times larger than those obtained manually by Castellón (1993).
- The filtering criteria (specifically F2+F3>9) were shown to be more accurate than F1, with 97% coverage of correct genus terms.
Limitations and Future Work
- Accuracy of Initial Labelling: The initial labelling accuracy of 64% indicates that there is still room for improvement in the core extraction process.
- Language Gap: While a bilingual dictionary was used, the inherent limitations of such resources can still introduce noise.
- Taxonomy Flatness: The resulting taxonomies tend to be "flatter" than manually constructed ones, as the method restricts inner taxonomic genus terms to those selected by the criteria.
- Future Work: The authors suggest that iterating the salient word collection process (collecting new salient words from the second labelled dictionary and re-estimating tags) could further improve results. They also note that a finer-grained set of semantic primitives than WordNet's 24 lexicographer's files could be beneficial for specific applications. The paper also hints at the possibility of applying this methodology to other languages and dictionary types.
Summary updated