This paper points out that product reviews contain domain-specific knowledge. To capture the hierarchical relationships between product attributes, they introduce a new approach: "hierarchical learning with sentiment ontology tree" (HL-SOT) in order to:

1. identify attributes
2. identify which attributes have sentiment attached to them

This would enable searching for particular attributes in reviews.

Their algorithm is based on H-RLS from Incremental algorithms for hierarchical classification. Evaluations are conducted against a human-labeled data set.