Enriching very large ontologies using the WWW

• The introduction presents shallow corpus-based enrichment and deep processing of specialized texts as complementary approaches to knowledge acquisition.Deep processing can provide specific relations, while shallow techniques can provide generic knowledge.Source: §1 · p. 1
• The paper links WordNet concepts to relevant web document collections, which are processed to address shortcomings in the ontology.Source: §1 · p. 1
• Topic signatures associate each concept with topically related words, but unlike relevancy signatures they are not sentence-based and do not require parsing.They are not intended for information extraction.Source: §1 · p. 1
• For each word, the method hierarchically clusters the concepts that lexicalize it, addressing WordNet’s proliferation of word senses.Source: §1 · p. 1
• The acquired signatures and clusters are assessed through word-sense disambiguation on SemCor, rather than compared with other disambiguation algorithms.Source: §1 · p. 1
• The authors describe the experiments as preliminary, choosing to run the full process first to judge whether the approach is feasible and interesting.They note that several aspects could be improved and optimized.Source: §1 · p. 1

• WordNet 1.6 contains 121,962 words and 99,642 concepts, represented as synonym sets organized around lexical-semantic relations.Source: §2 · p. 1
• WordNet lacks explicit links between semantic variants such as paint–paint and song–sing, and between topically related concepts such as bat–baseball.Source: §2 · p. 2
• WordNet’s fine sense distinctions can be unclear and difficult to justify or use; line, for example, has 32 senses, complicating automatic disambiguation.Source: §2 · p. 2
• The paper builds lists of words topically related to each concept and uses them to cluster word senses, addressing WordNet’s sense-proliferation problem.Source: §2 · p. 2
• The evaluation uses SemCor and a random set of 20 nouns occurring at least 100 times there; these nouns average 6.3 senses.Source: §2 · p. 2

3 BUILDING TOPIC SIGNATURES FOR THE CONCEPTS IN WORDNET

• Each topic signature is a list of words with distinctive frequencies in the document collection for a WordNet sense.The process extracts and counts words in each collection, then compares frequencies across collections.Source: §3 · p. 2, §3.3 · p. 3

3.1 Building the queries

• Queries combine a target word with ontology-derived cuewords for one sense and exclude cuewords for other senses; exhaustive web classification was infeasible.WordNet cuewords include synonyms, gloss terms, and semantic relations; instead of classifying millions of hits, the method submits one query per concept.Source: §3.1 · p. 2

3.2 Search the internet

• Retrieving documents for 20 words yielded tens of thousands of documents, used more than one gigabyte compressed, and took nine days of continuous internet access.Retrieving 1,217 documents for the four senses of boy took three and a half hours.Source: §3.2 · p. 3

3.3 Build topic signatures

• The method forms a word-frequency vector for each sense collection and uses χ2 to weight words that occur more often than expected relative to the other senses.The other word-sense collections form the contrast set.Source: §3.3 · p. 3

4 APPLY SIGNATURES FOR WORD SENSE DISAMBIGUATION

• The evaluation tests whether signatures contain useful information, not whether they outperform complete disambiguation systems; signatures are only one evidence source.The paper notes that existing systems also use collocations, selectional restrictions, topic and domain information, and co-occurrence relations.Source: §4 · p. 3
• For each target occurrence, the method sums signature χ2 values for context words by sense and selects the sense with the highest total.It uses a 100-word window around the target.Source: §4 · pp. 3–4
• Comparisons include random selection and WordNet lists built from synonyms, definitions, or additional semantic relations, scored with the same algorithm.The lists add definition content words and, in the fullest variant, hyponyms, hypernyms, and meronyms.Source: §4 · p. 4
• Signature-based disambiguation generally outperforms random selection and, overall, the other WordNet-based word lists, though some words fall below random.The paper interprets this as evidence that signatures learn topic information absent from WordNet, while sometimes introducing noise.Source: §4 · p. 4

5 CLUSTERING WORD SENSES

• Because signature overlap did not yield interesting results, the authors clustered retrieved documents directly; for boy, young and offspring senses were closest at 0.65 similarity.The colored sense contexts were least similar to the others at 0.46.Source: §5 · p. 4

5.1 Evaluation of word sense clusters on a word sense disambiguation task

• The hierarchy can be applied by combining existing sense signatures along branches or by merging branch document collections and computing new cluster signatures.The methods compare evidence at successive levels of the hierarchy.Source: §5.1 · pp. 4–5
• Disambiguation can proceed from coarse to fine distinctions, with the hierarchy successively separating branches until it reaches the original fine-grained senses.The described levels distinguish two broad groups, intermediate clusters, and the finest senses.Source: §5.1 · p. 5
• Above 0.90 precision for coarser sense distinctions; hierarchy information improves precision even at the fine level, but results do not settle whether recomputing cluster signatures helps.The evaluation compares random, original-signature-plus-hierarchy, and recomputed cluster-signature approaches.Source: Table 4 · p. 5, §5.1 · p. 5

6.1 Building topic signatures

• Prior weighting work reported gains from χ2/LSA and likelihood-ratio methods on clean data, while web-extracted text remained disappointing [9].The paper attributes poor hand-evaluated signature quality to the retrieved documents rather than the signature-building procedure.Source: §6.1 · p. 5

6.2 Concept Clustering

• Word-sense clusters improved WSD results, enhancing precision for fine-level distinctions and exceeding 90% precision at the coarse level.Source: §6.2 · p. 5

6.3 Searching the internet for concepts

• Unlike prior example-harvesting methods [23], this method retrieves concept-related documents, using all close relatives and excluding other senses’ cuewords.Source: §6.3 · pp. 5–6
• Web retrieval is the weakest component: query strategies trade missing senses against excessive, less relevant documents, and web content is unbalanced or sparse.Source: §6.3 · p. 6
• Retrieval delays affected by time, workload, or server location could also affect results.Source: §6.3 · p. 6

• Web documents linked to ontology concepts provide a basis for topic signatures and clustering same-word senses, addressing WordNet’s topical-link and sense-proliferation problems.Source: §7 · p. 6
• Web-derived topic signatures are dirtier than those built from topically organized documents, although both signature construction and clustering seem satisfactory.Source: §7 · p. 6
• The main obstacle to cleaner signatures is linking concepts to relevant web documents; queries need tuning and retrieved texts need filtering for noise and bias.Source: §7 · p. 6
• Cleaner signatures could help identify topically related concepts and extend clustering, while richer ontologies could benefit word sense disambiguation.Source: §7 · p. 6