Learn Before
Concept icon
Concept

Controllable Text Generation Using Pre-Trained Language Models

Many controllable text generation methods use autoregressive or sequence-to-sequence pre-trained language models to generate text under a task-specific control condition. Given a control condition CC and output sequence Y=(y1,,yN)Y=(y_1,\dots,y_N), generation is modeled as P(YC)P(Y\mid C). In an autoregressive model, this distribution is factorized as P(Ymid C)=prod_{n=1}^{N}p(y_nmid y_{1:n-1},C), so each token depends on both the preceding tokens and the control condition.

0

1

Concept icon
Updated 2026-08-29

Tags

Deep Learning (in Machine learning)

Data Science

Computing Sciences