#retrieval

Question 1: Single choice

You build a news search engine with BERT. For each training query the single most relevant article is labeled, and the pool of articles is huge. How can negative sampling be used in training?

Classify (query, article) pairs from BERT's <cls> output: labeled article positive, random articles negative1
Feed BERT the query alone and train a softmax over all articles, one output class per article2
Sample noise words inside each article and train BERT's masked language model on them3
Train on the labeled (query, article) pairs only, because a binary relevance classifier does not need any negatives4

Timed test