Post

We find that simply scaling up the transformer architecture still leads to significant effectiveness drops in face of typos, keywords, ordering, and paraphrasing. We further highlight the need for more elaborate query variation datasets, which should retain the queries' semantics.