Decoding Values in Text with Artificial Intelligence

by

Author: Denis Avetisyan

Researchers have developed a new method leveraging the power of large language models to automatically identify and analyze the human values expressed within written content.

Value theory is conceptualized as the foundational stage in understanding a system's inherent worth and guiding principles.
Value theory is conceptualized as the foundational stage in understanding a system’s inherent worth and guiding principles.

This paper introduces Value Lens, an LLM-based model for detecting and quantifying human values in text, offering competitive performance and insights into value promotion or demotion.

As autonomous systems become increasingly prevalent, ensuring their decisions align with human values presents a significant challenge. This paper introduces ‘Value Lens: Using Large Language Models to Understand Human Values’, a novel text-based model leveraging generative AI to detect and analyze these values within given text. Value Lens operates by first formalizing value theory with an LLM, then employing a dual-LLM process-one to detect values and another to critically review the findings-achieving performance comparable to, and exceeding, existing value detection methods. Could this approach unlock more nuanced understandings of value promotion and demotion, ultimately guiding the development of more ethically aligned AI systems?

Defining the Ethical Landscape: Quantifying Human Values

The capacity to understand and quantify human values extends far beyond abstract philosophical inquiry, proving increasingly vital across a spectrum of practical applications. In the social sciences, a precise understanding of values informs models of human behavior, predicts societal trends, and allows for more effective policy-making. Simultaneously, this capability is paramount in the burgeoning field of artificial intelligence, particularly in the quest for AI alignment – ensuring that increasingly autonomous systems operate in accordance with human ethical frameworks. Without a robust means of defining and measuring values, there remains a significant risk of creating AI that pursues objectives misaligned with human well-being, highlighting the urgent need for interdisciplinary research focused on this fundamental challenge. From tailoring personalized experiences to fostering cross-cultural understanding, the ability to discern and prioritize values is becoming an indispensable tool for navigating an increasingly complex world.

Current methods for identifying human values within natural language often fall short due to the inherent complexity of how these values are communicated. Analyses frequently rely on predefined lexicons or simplistic keyword matching, failing to account for subtle variations in phrasing, cultural context, and the interplay of multiple values within a single statement. A phrase like “responsible innovation” carries significantly more weight than the individual words alone, and its meaning can shift depending on the domain – healthcare versus environmental policy, for example. Furthermore, values are rarely stated directly; they are frequently implied through narratives, metaphors, and emotional language, requiring a deeper understanding of semantic nuance than most current algorithms possess. This limitation hinders accurate value detection and makes it difficult to apply these methods reliably across diverse texts and contexts.

Establishing a reliable system for identifying human values necessitates more than just technical skill; it demands a solid grounding in established ethical and psychological theories. Current approaches often falter because they treat values as static labels, overlooking the inherent complexity and contextual sensitivity of moral judgements. A truly robust framework, therefore, must integrate insights from value theories – such as Schwartz’s theory of basic values or Rokeach’s value survey – and translate these into computational models. Crucially, these models shouldn’t be rigid; they require adaptable detection methods capable of discerning subtle variations in language and recognizing that the same statement can express different values depending on the situation. Such flexibility is paramount for accurately mapping the landscape of human values and applying this understanding to fields ranging from behavioral science to the development of ethically aligned artificial intelligence.

Stage two focuses on identifying valuable states within the environment.
Stage two focuses on identifying valuable states within the environment.

Value Lens: A Modular Architecture for Value Detection

Value Lens utilizes a sequential, three-stage Large Language Model (LLM) architecture to analyze text for the presence and strength of values. The first stage, conceptualization, establishes a framework for value identification. This is followed by detection, where the model identifies instances of these values within a given text. Finally, the intensity assessment stage quantifies both the strength and emotional valence (positive or negative) associated with each detected value, providing a nuanced understanding beyond simple presence or absence. Each of these stages is independently handled by a dedicated LLM, allowing for specialized processing and improved accuracy.

The initial stage of Value Lens utilizes a Large Language Model (LLM1) to generate ‘Value Specifications’ which function as a formalized representation of a selected value theory. These specifications are not simply lists of values; they are detailed constructions defining each value’s facets, indicators, and associated linguistic patterns. This process moves beyond simple keyword matching by providing LLM2 with nuanced criteria for value identification. The resulting specifications enrich the model’s understanding by establishing a structured framework for interpreting text and disambiguating value-laden language, allowing for a more precise and contextually aware analysis of underlying values.

Following value specification construction, Large Language Model 2 (LLM2) performs value detection within input text. This process utilizes the ‘Value Specifications’ generated by LLM1 as guiding parameters, enabling identification of instances where specified values are expressed or implied. Subsequently, Large Language Model 3 (LLM3) quantifies the identified values, assigning both a strength score – indicating the prominence of the value within the text – and a valence score, representing the positive or negative connotation associated with that value expression. These quantified scores provide a numerical representation of value presence and orientation, allowing for comparative analysis and aggregation across texts.

Empirical Validation: Assessing Performance with the Touché24-ValueEval Dataset

Value Lens was subjected to empirical testing using the Touché24-ValueEval Dataset, a publicly available resource specifically designed for evaluating the performance of models in value detection tasks. This dataset is constructed based on Schwartz’s Value Theory, a widely recognized psychological framework that identifies ten motivational values – Universalism, Benevolence, Conformity, Tradition, Security, Power, Achievement, Hedonism, Stimulation, and Self-Direction – which represent basic human motivations. The Touché24-ValueEval Dataset provides annotated text segments categorized according to these ten values, enabling quantitative assessment of a model’s ability to accurately identify and categorize expressions of these values within natural language.

Model performance was quantitatively evaluated using four standard metrics commonly employed in information retrieval and classification tasks. Micro F1-score calculates the harmonic mean of precision and recall, giving equal weight to each instance. Macro F1-score computes the F1-score for each class and then averages those scores, providing a class-balanced assessment. Precision measures the proportion of predicted positive instances that are actually positive, while Recall measures the proportion of actual positive instances that are correctly predicted. These metrics provide a comprehensive evaluation of both the model’s ability to avoid false positives (precision) and false negatives (recall), across the entire dataset and individual value categories.

Evaluation of Value Lens on the Touché24-ValueEval dataset indicates performance comparable to the Hierocles of Alexandria model, as measured by the Micro F1-score. Specifically, Value Lens achieved a Macro F1-score of 0.301, representing an improvement over existing BERT-based models which achieved a score of 0.232. Further analysis reveals that Value Lens surpasses the performance of the top-performing model in 6 of the 19 assessed value categories, indicating nuanced strengths in specific value detection tasks.

Broadening the Ethical Horizon: Implications and Future Directions

Value Lens offers a significant advancement in computational analysis of human values, presenting a versatile tool with broad implications for several rapidly evolving fields. The system’s ability to identify and categorize value expressions within text opens possibilities for automating content moderation, allowing platforms to more effectively flag potentially harmful or biased material. Furthermore, Value Lens can provide valuable insights for social media analysis, enabling researchers to understand public sentiment and identify prevailing value systems within online communities. Perhaps most critically, the technology has direct application in the development of ethical artificial intelligence, offering a means to align AI systems with human values and mitigate the risk of unintended consequences. By providing a quantifiable measure of value alignment, Value Lens helps ensure that AI operates in a manner consistent with societal norms and ethical principles.

Value Lens distinguishes itself through a deliberately modular architecture, enabling seamless adaptation to diverse frameworks of value assessment and linguistic contexts. This design prioritizes flexibility; the core algorithms remain consistent, while modules responsible for identifying and interpreting value-laden terms can be readily swapped or retrained. Consequently, the system isn’t limited to a single, pre-defined value system – it can accommodate utilitarian, deontological, virtue-based, or any other established ethical theory. Furthermore, the modularity extends to language support, allowing for the integration of new lexicons and linguistic rules without requiring substantial code revisions, thus broadening its potential application across varied cultural and communicative landscapes. This adaptability positions Value Lens as a highly versatile tool for navigating the complexities of value expression in a globalized world.

Continued development of Value Lens prioritizes enhancing its nuanced understanding of language, specifically tackling the challenges presented by complex sentence structures and ambiguous phrasing. Researchers aim to refine the model’s capacity to discern subtle value expressions often masked by figurative language or contextual dependence. This includes exploring advanced natural language processing techniques and expanding the training data to encompass a wider range of linguistic styles and subject matter. Crucially, future efforts will shift towards validating Value Lens’s performance with large-scale, real-world datasets – such as social media feeds, news articles, and policy documents – to assess its practical utility and identify areas for further improvement before broader implementation in applications like content moderation or ethical AI systems.

The pursuit of value detection, as demonstrated by Value Lens, echoes a fundamental principle of systemic design: understanding the interconnectedness of elements. This model doesn’t merely identify values; it analyzes their intensity, revealing nuanced gradients that shape human preference. This aligns with the idea that structure dictates behavior – the way values are expressed, and their relative strength, profoundly influences the resulting actions and interpretations. As Bertrand Russell observed, “The good life is one inspired by love and guided by knowledge.” Value Lens, in its attempt to map and understand these guiding forces, offers a pathway towards more aligned and, potentially, ‘good’ systems. Good architecture is invisible until it breaks, and only then is the true cost of decisions visible.

Looking Ahead

The introduction of Value Lens, while a notable step, merely illuminates the complexity of the undertaking. To believe a model can ‘detect’ values is akin to charting the ocean by sampling a single drop – useful, perhaps, for salinity, but insufficient to grasp the currents. The true challenge isn’t identifying that a value is present, but understanding its interplay within a larger system of beliefs, its weighting relative to competing values, and-crucially-the mechanisms by which those weights shift. A value, isolated, is a static concept; it is the network of values, the constant negotiation between them, that dictates behavior.

Future work must move beyond simple detection. The field needs to address the inherent ambiguity of language, the cultural specificity of value expression, and the dynamic nature of value systems. One cannot simply replace a single ‘value detector’ without understanding the entire ‘cognitive bloodstream’. The current focus on intensity analysis is promising, but only a starting point. A more holistic approach requires modeling not just the what of values, but the why and the how-the underlying motivations and the contextual factors that shape them.

Ultimately, the goal shouldn’t be to build a perfect ‘value detector’, but to create a framework for understanding the architecture of human belief. The elegance of a system lies not in its complexity, but in the simplicity of its governing principles. To truly align with human values, any model must first understand the fundamental patterns that bind them, not merely catalog their surface manifestations.

Original article: https://arxiv.org/pdf/2512.15722.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-21 00:04