Penalty Function in Controllable Decoding

A developer is using a language model for text summarization. The model's outputs are generally fluent but suffer from excessive repetition of certain phrases. To address this, the developer employs a decoding objective that penalizes repetition, formulated as: argmax [Pr(y|x) - λ * Penalty(x, y)], where Penalty(x, y) increases with the amount of repetition in the output y. How should the developer adjust the hyperparameter λ to make the summaries less repetitive?

Analyzing the Trade-off in Penalized Decoding

Consider the decoding objective for controllable text generation: ŷ = argmax [Pr(y|x) - λ * Penalty(x, y)]. If the hyperparameter λ is set to 0, the objective simplifies to finding the output with the highest conditional probability, effectively ignoring any penalty.

Greedy Search with Penalty Objective

Sampling-based Search with Penalty Objective