Asymmetric search retrieves a long, specific passage for a short, interrogative, and often vague query. The query and document may answer one another despite sharing little surface form, unlike symmetric search, which compares texts of similar length and structure. section 3.1
The central design contrast is:
- Sparse representations support efficient keyword matching but miss semantic meaning; dense representations encode semantic context across continuous dimensions. section 3.1
- SBERT targets symmetric comparisons such as duplicate questions or paraphrases, whereas DPR uses separate BERT encoders for questions and passages. section 3.1
- DPR learns the asymmetric space from question, correct-passage, and incorrect-passage triplets for open-domain question answering. section 3.1
Table 3.1 Symmetric search vs asymmetric search
|
Feature
|
Symmetric
Search
|
Asymmetric
Search
|
|
Goal
|
Find similar text
(twins)
|
Find answers to
questions
|
|
Input
|
Similar
length/structure
|
Short query vs.
Long doc
|
|
Model
|
SBERT
|
DPR
|
|
Use Case
|
Paraphrase
detection
|
Open Domain QA
|
Compare symmetric and asymmetric search by goal, input shape, model, and use case. section 3.1
