core Estimated learning time: 5 h

7.9 Seq2seq tasks

You can build translation or summarization pipelines.

Before:06. Deep LearningUnlocks:08. Large Language Models

Seq2seq covers text-to-text tasks — translation, summarisation — and T5's framing shows how one paradigm swallows many tasks whole. BLEU and its relatives are useful, blunt instruments. It sits in the applied run of the module. The failure mode to respect is fluent confabulation in abstractive summarisation: a summary can read beautifully while asserting things the source never said, and only checking against the source catches it.

Work through these

  • Machine translation and BLEU

    Translating between languages and the long-standing automatic measure for it, which compares against reference translations. The measure's weaknesses are well documented and worth knowing.

  • Abstractive vs. extractive summarization

    Writing a new summary against selecting existing sentences, which are different tasks with different failure modes. The first can be wrong in ways the second cannot.

  • T5 and text-to-text framing

    Framing every task as text in and text out, which unified a fragmented field. It is also the framing that current instruction-following models inherit.

  • Hallucination in summarization

    A summary can contain confident statements that appear nowhere in the source, which is the central risk of the abstractive approach. Detecting this is an active problem.

Sign in to keep your progress.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.