Controllable Neural Text Generation

A language model can be adapted to a generation task by learning a small set of steering values while keeping the base model fixed. Lilian Weng's controllable-generation survey describes prefix-tuning, which learns continuous internal values that subsequent text can attend to.

These values need not correspond to ordinary words. The original paper proposes task-specific prefixes with frozen language-model parameters, allowing the main model to be shared across tasks. The adaptation still requires training, but the task-specific result is a prefix rather than another fully modified copy of the model. This makes the location of the learned change central to understanding the method.

“Note that model steerability is still an open research question.”

“The uniform information density hypothesis (UID; Levy and Jaeger, 2007) states that—subject to the constraints of the grammar—humans prefer sentences that distribute information (in the sense of information theory) equally across the linguistic signal, e.g., a sentence.”

“GeDi guided generation in their experiments showed strong controllability and ran 30x faster than PPLM.”

“One obvious drawback of PPLM is that due to multiple passes at every decoding step, the test time computation becomes much more expensive.”