Uncovering Values: Detecting Latent Moral Content from Natural Language with Explainable and Non-Trained Methods

Luigi Asprino, Luana Bulla, Stefano De Giorgis, Aldo Gangemi, Ludovica Marinucci, Misael MongioviACL 2022

TL;DR

This paper introduces two novel, non-training-based methods for detecting latent moral content in natural language, inspired by Moral Foundation Theory (MFT). By leveraging zero-shot learning and knowledge graph reasoning, the authors demonstrate that it's possible to uncover moral values from text without extensive prior training, offering a more transparent and adaptable approach to understanding human morality in digital communication.

What Problem Does This Paper Address?

Understanding the moral underpinnings of human communication is crucial for building AI systems that can effectively integrate into our social and cultural environments. However, identifying moral content in text is challenging due to its often latent nature, cultural dependence, and the inherent subjectivity involved in annotation. Existing methods often rely on extensive labeled datasets or domain-specific knowledge, limiting their versatility and explainability. This paper aims to address this gap by proposing methods that can detect moral values from natural language without requiring task-specific training data, making the process more transparent and adaptable.

Methods

The paper presents two primary approaches:

  1. Zero-Shot Learning with Emotion Enhancement: This method adapts pre-trained Natural Language Inference (NLI) models (specifically BART-large fine-tuned on MNLI) to classify text based on its semantic similarity to MFT moral values. The core idea is to treat the input text as a premise and generate hypotheses for each MFT moral value, assessing the entailment. To improve performance, a second step incorporates detected emotional components and their valence, augmenting the original text with this information. A third step combines the results from both the basic zero-shot and emotion-enhanced versions. This approach is "zero-shot" because it classifies labels it hasn't been explicitly trained on.
  2. Frame-Based Value Reasoner: This approach utilizes a knowledge graph (KG) generation pipeline. First, the input text is processed by the FRED tool to create a KG that includes FrameNet frames, VerbNet roles, and links to external knowledge bases like DBpedia. Second, this KG is enriched by querying Framester's SPARQL endpoint to link extracted entities to MFT moral values. The resulting KG provides a semantically grounded representation of moral content. This method is designed to be explainable, as the generated knowledge graph explicitly shows the "value triggers" and their connections.

Datasets

The primary dataset used for evaluation is the Moral Foundation Twitter Corpus (MFTC). This corpus consists of 35,000 tweets organized into seven distinct thematic topics, with each tweet annotated by multiple annotators for moral values and violations according to Moral Foundation Theory. The authors evaluated their models on a subset of 6,075 tweets from the MFTC test set.

Key Insights and Ideas

  • Moral Values are Detectable Without Explicit Training: The paper strongly suggests that even without task-specific fine-tuning, models can learn to identify moral concepts, especially when leveraging pre-trained NLI capabilities and structured knowledge.
  • Emotion Enhances Moral Detection: Incorporating emotional signals significantly improves the performance of the zero-shot model, highlighting the intertwined nature of emotion and moral judgment.
  • Knowledge Graphs Offer Explainability: The frame-based approach provides a transparent way to understand why a particular moral value is detected, by explicitly linking textual triggers to moral concepts through a structured knowledge graph.
  • Subjectivity and Nuance are Key Challenges: The authors acknowledge that moral labeling is inherently subjective, leading to variations in annotator agreement and impacting model performance, particularly for more nuanced moral concepts like "purity."
  • Zero-Shot NLI Models are Versatile: The success of adapting NLI models for moral detection suggests their potential for broader applications beyond their original intent, demonstrating an emergent ability to grasp abstract social concepts.

Key Results

The paper reports that both the Emotion-Zero-shot+ and Frame-based methods achieve comparable overall F1 scores of approximately 45% on the MFTC dataset.

  • Emotion-Zero-shot+ and Frame-based models outperform other evaluated methods (Random, Zero-shot, Emotion-Zero-shot) for four out of eleven moral value labels.
  • For specific labels, F1 scores ranged from 0.11 to 0.53, indicating varying degrees of success across different moral dimensions.
  • The Emotion-Zero-shot+ architecture showed a notable improvement of around 10% over the basic Emotion-Zero-shot model and 20% over the simple Zero-shot model, underscoring the benefit of emotion integration.
  • The frame-based detector demonstrated better performance on more generic moral values (e.g., "harm") compared to more nuanced ones (e.g., "purity").

Limitations and Future Work

Limitations:

  • Dataset Size and Efficiency: The evaluation was performed on a subset of the MFTC due to the computationally intensive nature of the FRED-based method. A full evaluation would require significant processing time.
  • Subjectivity of Annotation: The inherent subjectivity of moral judgment leads to annotator disagreement, which is a challenge for any moral detection task.
  • "Non-moral" Label Handling: The frame-based detector struggled with sentences labeled as "non-moral" when they co-occurred with other moral values, requiring a modification of the dataset for consistent evaluation.
  • FRED Tool Limitations: The effectiveness of the frame-based approach can be limited by the FRED tool's ability to fully capture complex sentence structures and implicit moral triggers, especially in fragmented text like tweets.

Future Work:

  • Improving Zero-Shot Models: Developing techniques to better comprehend the intricacies of more contentious moral values.
  • Sentence Structure and Weighting: Giving greater weight to significant aspects of a sentence to more simply detect prevailing moral values.
  • Frame-Based Refinements: Implementing syntax-based heuristics and considering frame structures for more complex inferences, such as detecting values through negation or negative verb polarity.
  • Quantitative/Qualitative Scoring: Introducing scoring systems based on trigger occurrences or weighting different trigger types.
  • Hybrid Approaches: Conjugating the zero-shot and frame-based approaches to leverage their respective strengths, potentially by adding a layer to guide prediction or using extended knowledge bases.
  • Broader Datasets: Testing on different types of datasets and expanding the current dataset to include supervised methods for comparison.

Summary updated

BibTeX
@inproceedings{asprino2022uncovering,
  author = {Luigi Asprino and Luana Bulla and Stefano De Giorgis and Aldo Gangemi and Ludovica Marinucci and Misael Mongiovi},
  title = {{Uncovering Values: Detecting Latent Moral Content from Natural Language with Explainable and Non-Trained Methods}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure