Translation and summarization are the classic sequence-to-sequence tasks: the model reads one piece of text and writes a different one. Translation turns a sentence in one language into another, while summarization condenses a long document into a short one that keeps the important points.
The lessons start with an introduction to neural machine translation and a university lecture on how attention improved it. You then look at the two main approaches to summarization. Because there is rarely a single correct output, the module also covers how these systems are scored: BLEU for translation and ROUGE for summaries, both of which compare a model's output with human-written references.
Watch the lessons in order, then continue on to Question Answering & RAG.
Machine translation researcher Philipp Koehn introduces how neural networks are used to translate text from one language into another.
Video 1h 6minA university lecture on how neural networks translate text and how attention fixes the bottleneck of the basic encoder-decoder setup.
Video 1h 21minExplains the two main approaches to summarization, picking key sentences versus writing new ones, and how transformer models are used for it.
Video 16minExplains BLEU, the standard score for machine translation, which compares a model's translation with one or more reference translations.
Video 4minExplains ROUGE, the standard score for summarization, which measures how much a generated summary overlaps with a reference summary.
Video 4min