Trading Off Diversity and Quality in Natural Language Generation

By Hugh Zhang, Hugh Zhang, Daniel Duckworth, Daphne Ippolito, Arvind Neelakantan. 2020

133 citations recorded by the source corpus.

For open-ended language generation tasks such as storytelling or dialogue, choosing the right decoding algorithm is vital for controlling the tradeoff between generation quality and diversity. However, there presently exists no consensus on which decoding procedure is best or even the criteria by which to compare them. In this paper, we cast decoding as a tradeoff between response quality and diversity, and we perform the first large-scale evaluation of decoding methods along the entire quality-diversity spectrum. Our experiments confirm the existence of the likelihood trap: the counter-intuitive observation that high likelihood sequences are often surprisingly low quality. We also find that when diversity is a priority, all methods perform similarly, but when quality is viewed as more important, nucleus sampling (Holtzman et al., 2019) outperforms all other evaluated decoding algorithms.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

In Evaluation & Benchmarks

Citations in the corpus

Learning to summarize from human feedback (2020, background): “Quality can often be improved significantly by non-uniform sampling strategies such as beam search [51], but these can lead to repetition and other undesirable artifacts [69, 23].”
Evaluation of Text Generation: A Survey (2020, background): “Nevertheless, diversity-based metrics are rarely used on their own, as text diversity can come at the cost of text quality (Montahaei et al., 2019a; Hashimoto et al., 2019; Zhang et al., 2021), and some NLG tasks do not require highly diverse generations.”
Do Language Models Plagiarize? (2022, methodology): “pling is reported to be the most effective decoding method in various generation settings (Ippolito et al., 2019a; Zhang et al., 2020).”
Residual Energy-Based Models for Text (2021, background): “Better perplexity results do not necessarily imply better generations (Zhang et al., 2020; Hashimoto et al., 2019): for example, approaches exhibiting mode-dropping problems such as language GANs might score terribly in terms of perplexity but do well in terms of generation (Yu et al., 2017;…”
Plug-and-Play Conversational Models (2020, methodology): “In all the baselines, we sample 10 different hypotheses using multinomialsampling after a top-k filtering (with k = 10), to ensure response diversity (Zhang et al., 2020a), and we select the hypotheses with the lowest attribute model loss as the response.”
A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions (2023, unknown): “The rationale for incorporating randomness into decoding strategies stems from the realization that high likelihood sequences often result in surprisingly low-quality text, which is called likelihood trap (Stahlberg and Byrne, 2019; Holtzman et al., 2020; Meister et al., 2020; Zhang et al., 2021).”
MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers (2021, background): “However, quantifying these aspects in a principled yet computationally tractable manner is challenging, as the text distributions are high-dimensional and discrete, accessed only through samples or expensive model evaluations [26, 58, 62].”
Is MAP Decoding All You Need? The Inadequacy of the Mode in Neural Machine Translation (2020, background): “In open-ended generation, Zhang et al. (2020) correlate model likelihood with human judgements for a fixed sequence length, thus eliminating any possible length bias issues.”
Learning Natural Language Generation with Truncated Reinforcement Learning (2022, unknown): “When generating text from a trained language model, the quality and diversity of samples depend on the decoding algorithm (Zhang et al., 2020).”
RankGen: Improving Text Generation with Large Ranking Models (2022, background): “Modern LMs suffer from the “likelihood trap” (See et al., 2019; Zhang et al., 2021), in which high likelihood (low perplexity) sequences produced by greedy decoding or beam search tend to be dull and repetitive.”
DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models (2022, background): “Generating high-quality texts with high diversity is an important requirement for many text generation applications and the trade-off between quality and diversity is always a critical concern in open-ended NLG tasks (Zhang et al., 2021).”
Typical Decoding for Natural Language Generation (2022, background): “A number of recent works have tried to offer explanations for this phenomenon: some have attributed it to a diversity-quality tradeoff (Zhang et al., 2021) while others blame shortcomings of model architectures or training strategies (Welleck et al.”
Exposure Bias versus Self-Recovery: Are Distortions Really Incremental for Autoregressive Text Generation? (2019, background): “For generation, we mainly focus on classical ancestral sampling, due to the following reasons: (1) The sampling algorithms (e.g., top-k sampling) are known to trade quality out of diversity (Caccia et al., 2018; Zhang et al., 2020; Nadeem et al., 2020).”