Identifying and Understanding Human Values in Text: A Tailorable LLM-Based Architecture

Forum topic · 小凯 · 2026-05-29

Summary

This arXiv paper (2605.27373) by Eduardo de la Cruz Fernández, Marcelo Karanik, and Sascha Ossowski proposes an LLM-based architecture for detecting and quantifying the intensity of human values in text, whether explicit or implicit. As intelligent systems grow more autonomous, aligning decisions with human values becomes critical beyond traditional utility-maximisation models. The proposed modular pipeline consists of three coordinated components: (1) a module that generates structured value specifications from the foundational texts of any value theory, (2) a module that annotates text using these specifications, and (3) a module that assigns support/opposition levels based on rhetorical and semantic evidence. This design separates value conceptualization from detection, avoiding the limitations of prior approaches tied to specific value theories or complex prompt engineering, and yielding an extensible, reproducible, and theory-adaptable pipeline. Experiments on the ValueEval dataset show solid detection performance.

Paper Overview

Field: AI
Authors: Eduardo de la Cruz Fernández, Marcelo Karanik, Sascha Ossowski
Published: 2026-05-28
arXiv: 2605.27373

Abstract

As intelligent systems become more autonomous, the scientific community focuses on creating decision-making mechanisms that include ethical and moral considerations, unlike traditional utility-maximisation models. To achieve this, a key aspect is assessing how well these decisions align with human values. A promising line of research develops approaches based on Large Language Models (LLMs) to identify human values from text, whether explicit or implicit.

This paper introduces an LLM-based architecture to detect and quantify the intensity of human values in text, avoiding the limitations of previous approaches tied to a specific value theory or requiring complex prompt engineering.

Architecture

The architecture comprises three coordinated modules:

1. Value specification generation — produces structured value specifications from the foundational texts of any theoretical framework. 2. Text annotation — uses these specifications to label values in input text. 3. Intensity assessment — assigns support/opposition levels grounded in rhetorical and semantic evidence.

Key Contributions

Modular design separates value conceptualization from the detection task.
The pipeline is extensible, reproducible, and adaptable to different value theories.
Experiments on the ValueEval dataset demonstrate good detection performance.

---

*Auto-collected on 2026-05-29 from zhichai.net.*