Improving Uncertainty Estimation through Semantically Diverse Language Generation

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Aichberger, Lukas, Schweighofer, Kajetan, Ielanskyi, Mykyta, Hochreiter, Sepp
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866915595765678080
author Aichberger, Lukas
Schweighofer, Kajetan
Ielanskyi, Mykyta
Hochreiter, Sepp
author_facet Aichberger, Lukas
Schweighofer, Kajetan
Ielanskyi, Mykyta
Hochreiter, Sepp
contents Large language models (LLMs) can suffer from hallucinations when generating text. These hallucinations impede various applications in society and industry by making LLMs untrustworthy. Current LLMs generate text in an autoregressive fashion by predicting and appending text tokens. When an LLM is uncertain about the semantic meaning of the next tokens to generate, it is likely to start hallucinating. Thus, it has been suggested that predictive uncertainty is one of the main causes of hallucinations. We introduce Semantically Diverse Language Generation (SDLG) to quantify predictive uncertainty in LLMs. SDLG steers the LLM to generate semantically diverse yet likely alternatives for an initially generated text. This approach provides a precise measure of aleatoric semantic uncertainty, detecting whether the initial text is likely to be hallucinated. Experiments on question-answering tasks demonstrate that SDLG consistently outperforms existing methods while being the most computationally efficient, setting a new standard for uncertainty estimation in LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2406_04306
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Improving Uncertainty Estimation through Semantically Diverse Language Generation
Aichberger, Lukas
Schweighofer, Kajetan
Ielanskyi, Mykyta
Hochreiter, Sepp
Machine Learning
Artificial Intelligence
Large language models (LLMs) can suffer from hallucinations when generating text. These hallucinations impede various applications in society and industry by making LLMs untrustworthy. Current LLMs generate text in an autoregressive fashion by predicting and appending text tokens. When an LLM is uncertain about the semantic meaning of the next tokens to generate, it is likely to start hallucinating. Thus, it has been suggested that predictive uncertainty is one of the main causes of hallucinations. We introduce Semantically Diverse Language Generation (SDLG) to quantify predictive uncertainty in LLMs. SDLG steers the LLM to generate semantically diverse yet likely alternatives for an initially generated text. This approach provides a precise measure of aleatoric semantic uncertainty, detecting whether the initial text is likely to be hallucinated. Experiments on question-answering tasks demonstrate that SDLG consistently outperforms existing methods while being the most computationally efficient, setting a new standard for uncertainty estimation in LLMs.
title Improving Uncertainty Estimation through Semantically Diverse Language Generation
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2406.04306