A search for “driver” might mean a person operating a vehicle or software used by a computer system. Query-sense disambiguation identifies these competing meanings so the search experience can make an intentional choice instead of mixing unrelated results. section 6.2
One approach uses a semantic knowledge graph to traverse from a query term to document categories, then from each category to related terms. Category-relatedness scores suggest likely contexts; the terms reveal what the query means in each one. Known user context can guide the choice. When context is unknown, semantic relatedness or popularity among users can help, but neither makes the choice certain. section 6.2
Inspect the senses of “driver”
A three-step traversal can start at driver in the body field, find matching categories, then retrieve body terms within each category: nodes = [ {"field": "body", "values": ["driver"]}, {"field": "category", "min_occurrences": 5, "limit": 2}, {"field": "body", "min_occurrences": 5, "limit": 10}, ] traversal = skg.traverse(*nodes) The travel category scores 0.38996 and yields terms such as taxi, car, and license. DevOps scores 0.08917 and yields terms including ipam, aufs, and overlayfs. The category limit, term limit, and minimum occurrence threshold control how much evidence is returned. section 6.2
Table 6.3 Contextualized related terms lists by category for the query chef
|
Query: chef
|
|
|---|---|
Context: cooking 0.37731
Keywords:
chef 0.93239
chefs 0.5151
www.pamperedchef.com 0.41292
kitchen 0.39127
restaurant 0.38975
cooking 0.38332
chef's 0.37392
professional 0.36688
nakiri 0.36599
pampered 0.34736
|
Context: devops 0.34959
Keywords:
chef 0.87653
puppet 0.79142
docs.chef.io 0.7865
ansible 0.73888
www.chef.io 0.72073
learn.chef.io 0.71902
default.rb 0.70194
configuration 0.68296
inspec 0.65237
cookbooks 0.61503
|
The “chef” table shows why the highest score need not be treated as the only possible interpretation: cooking scores 0.37731, while DevOps scores 0.34959, with distinct terms under each. section 6.2
Possible responses include grouping results by meaning, selecting the likeliest meaning, mixing senses deliberately for diversity, or suggesting alternative queries. The graph-based approach depends on documents that represent the domain well; user signals can help reveal language and intentions that the document collection underrepresents. section 6.2 section 6.0
