Skip to main navigation Skip to search Skip to main content

Value Extraction from Policy Texts with Large Language Models

  • Umeå University

Research output: Contribution to conference › Paper › peer-review

Abstract

Socio-ethical values can be understood as common factors that guide human decision-making, preferences between outcomes, and interactions. The value-alignment challenge in artificial intelligence involves designing components that emulate reasoning based on stakeholder values, guiding the system's actions towards outcomes aligned with these preferences. However, directly engaging stakeholders to elicit their values and preferences is difficult and often impractical, especially when the stakeholder group is large. We present an indirect method that leverages the output of collective activities, in the form of policy documents as aggregated contributions and desired outcomes of stakeholders, to infer the underlying values of these groups. An automated value-extraction tool for text facilitates the analysis and comparison of such documents while providing a mechanism for expressing values to autonomous agents. This paper introduces and evaluates a process for automated value extraction using a large language model. Our primary research question is whether values from policy documents can be consistently extracted while accounting for the inherent unpredictability and opacity of large language models. Our evaluation shows how our approach can analyse policy documents and consistently produce descriptive values, demonstrating that this approach can be a useful tool for comparing policy documents, identifying the values they advocate, and serving as a proxy for stakeholders.
Original languageEnglish
Number of pages22
Publication statusE-pub ahead of print - 2025
EventThird International Workshop on Value Engineering in AI - University of Bologna, The Engineering School, Bologna, Italy
Duration: 25 Oct 2025 → 25 Oct 2025
Conference number: W29
https://vale2025.iiia.csic.es/

Workshop

WorkshopThird International Workshop on Value Engineering in AI
Abbreviated titleVALE 2025
Country/TerritoryItaly
CityBologna
Period25/10/25 → 25/10/25
Internet address

Funding

FundersFunder number
UKRI CDT in Accountable, Responsible and Transparent AIEP/S023437/1.

    Keywords

    • Value extraction
    • Software pipeline
    • Socia-technical system

    Fingerprint

    Dive into the research topics of 'Value Extraction from Policy Texts with Large Language Models'. Together they form a unique fingerprint.

    Cite this