mm.md

Doc2Query−−: When Less is More

You might notice that not all the generated queries look related to the source text. This is due to a defect that can appear in sequence-to-sequence models known as "hallucination".

Doc2Query−− can filter out these low-quality queries by measuring the relevance between them and the text that generated them using a scoring transformer S. It is applied as two transformers that follow the Doc2Query generator:

D

Doc2Query

D

QueryScorer

S

D

QueryFilter

D