Evaluating a Penalty Term for Creative Writing
A developer is using a language model to generate creative story continuations. To encourage novelty, they've added a penalty to the decoding process that strongly penalizes any word that has already appeared in the generated output. However, the resulting text is often ungrammatical and nonsensical, using strange synonyms for common words like 'the' and 'is'. Evaluate the developer's approach. Why is this penalty strategy failing, and what specific modification would you recommend to better achieve the goal of generating creative yet coherent text?
0
1
Tags
Ch.5 Inference - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Evaluation in Bloom's Taxonomy
Cognitive Psychology
Psychology
Social Science
Empirical Science
Science
Related
Decoding Objective with Penalty Term
A language model is being used to generate one-sentence summaries of news articles. The initial outputs are often too long and contain repetitive phrases (e.g., 'The study showed the research indicated that...'). To improve the quality of the summaries, a penalty term is added to the decoding process. Which of the following penalty strategies would be most effective at addressing both of the identified issues?
Evaluating a Penalty Term for Creative Writing
A language model is exhibiting several undesirable behaviors during text generation. Match each problem with the penalty term specifically designed to mitigate it.