Summarization with T5
Every NLP task as text-to-text.
Motivation
T5 unifies NLP under a single text-to-text interface: classification, translation, and summarization all become 'read a string, write a string.' This project fine-tunes the encoder–decoder for abstractive summarization.
Method
A pre-trained transformer encoder–decoder is fine-tuned on document/summary pairs, decoding summaries autoregressively. The shared format lets a single model transfer knowledge across tasks.
\[ p(y \mid x) = \prod_{t} p\!\left(y_t \mid y_{<t},\, x\right) \]
Mathematical core
The text-to-text framing exposes the common seq2seq backbone behind modern language models and clarifies how transfer learning reuses a single distributional model across many objectives.