High-Precision Content Classification Using Hierarchy
Content understanding represents each piece of content in the index. Relevance of content is a function of query and content understanding. Query understanding represents each search query as a search intent.
Classification maps a document to one or more predefined categories. We can do so using hand-tuned rules or machine learning.1 The categories can be a flat list, or they can be arranged in a hierarchical (single-hierarchy or faceted) taxonomy2.
If the categories are hierarchical and broadly applicable (I1), then a classifier might take advantage of the hierarchy and more confidently map content to a non-leaf category (e.g., mapping a material to “Semiconductor” rather than “High-Gap Semiconductor” or “III-V Semiconductor”). In general, it’s best to map value objects and entities to leaf categories.
Reducing the number of labels substantially improves the precision of a classifier. But filtering out infrequent labels decreases coverage, and it’s not clear that out-of-scope examples will be recognized in production.(F4) A more robust approach is to leverage the hierarchical nature of a taxonomy and roll up infrequently used labels to their parent or other ancestor categories.
G. Ingersoll and D. Tunkelang, “Course Notes for ‘Search with Machine Learning.’” Corise Education, Jun. 20, 2022. [Online]. Available: https://corise.com/course/search-with-machine-learning/ ↩︎
D. Tunkelang, “Taxonomies and Ontologies,” Medium, Aug. 30, 2017. https://queryunderstanding.com/taxonomies-and-ontologies-8e4812a79cb2 (accessed Jul. 15, 2022). ↩︎