README.md

---
license: apache-2.0
base_model: microsoft/deberta-base
tags:
- deberta
- human value detection
- schwartz values
- moral values
- text classification
- multi-label classification
model-index:
- name: human-value-detection-deberta-baseline
  results: []
---

# Human Value Detection – DeBERTa Baseline

This model is the **baseline 19-way value detector** from the paper:

> *Human Values in a Single Sentence: Moral Presence, Hierarchies, and Transformer Ensembles on the Schwartz Continuum*  
> Víctor Yeste, Paolo Rosso (2026)

It is a **multi-label classifier** over the **19 refined Schwartz basic values**, trained on the **English, machine-translated** portion of the ValueEval'24 / ValuesML corpus.  

- Inputs: a single sentence (news / political text, in English).  
- Outputs: a probability for each of the 19 Schwartz values.  
- Labels: we **collapse “attained” and “constrained”** into a single binary label per value (value is expressed vs. not expressed).

This is the *text-only DeBERTa-base baseline* used in the paper. In the experiments, it is also one of the members of the best-performing ensemble (Baseline + LIWC-22 + Previous-2-Sentences).

---

## Intended use

- Research on **human value detection** and **moral language**.
- Baseline / starting point for work on:
  - Schwartz value theory in NLP
  - Moral/value-aware text analysis in news and political discourse
  - Multi-label classification under class imbalance

The model was not trained or audited for safety-critical or high-stakes decision-making.

---

## Labels

The 19 labels follow the refined Schwartz value continuum:

1. `Self-direction: thought`  
2. `Self-direction: action`  
3. `Stimulation`  
4. `Hedonism`  
5. `Achievement`  
6. `Power: dominance`  
7. `Power: resources`  
8. `Face`  
9. `Security: personal`  
10. `Security: societal`  
11. `Tradition`  
12. `Conformity: rules`  
13. `Conformity: interpersonal`  
14. `Humility`  
15. `Benevolence: caring`  
16. `Benevolence: dependability`  
17. `Universalism: concern`  
18. `Universalism: nature`  
19. `Universalism: tolerance`

---

## How to use

### 1. Quick start: direct use with Transformers

Because this model uses a custom architecture (`EnhancedDebertaForSequenceClassification` with extra feature inputs), it is loaded via `AutoModelForSequenceClassification(..., trust_remote_code=True)` rather than the generic `pipeline("text-classification")`, which only supports a fixed list of built-in model classes.

```python
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "VictorYeste/human-value-detection-deberta-baseline"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(
    model_id,
    trust_remote_code=True,  # important for custom model code
)

values = [
    "Self-direction: thought",
    "Self-direction: action",
    "Stimulation",
    "Hedonism",
    "Achievement",
    "Power: dominance",
    "Power: resources",
    "Face",
    "Security: personal",
    "Security: societal",
    "Tradition",
    "Conformity: rules",
    "Conformity: interpersonal",
    "Humility",
    "Benevolence: caring",
    "Benevolence: dependability",
    "Universalism: concern",
    "Universalism: nature",
    "Universalism: tolerance",
]

id2label = {i: label for i, label in enumerate(values)}

def predict_values(text, threshold=0.50):
    enc = tokenizer(text, return_tensors="pt", truncation=True)
    with torch.no_grad():
        outputs = model(**enc)

    logits = outputs.logits.squeeze(0)          # (19,)
    probs = torch.sigmoid(logits)               # tensor of shape (19,)
    probs = probs.cpu().numpy()

    active = probs >= threshold
    active_labels = [id2label[i] for i, is_on in enumerate(active) if is_on]

    return {
        "probs": {id2label[i]: float(p) for i, p in enumerate(probs)},
        "labels": active_labels,
    }

example = "We must do more to protect the environment and future generations."
print(predict_values(example))
```

This will return something like:
```
{
    'probs': {
        'Self-direction: thought': 0.004236925393342972,
        'Self-direction: action': 0.007529713679105043,
        'Stimulation': 0.014666699804365635,
        'Hedonism': 0.004158752970397472,
        'Achievement': 0.017073791474103928,
        'Power: dominance': 0.006939167156815529,
        'Power: resources': 0.0076741743832826614,
        'Face': 0.0034943644423037767,
        'Security: personal': 0.00695117749273777,
        'Security: societal': 0.012955584563314915,
        'Tradition': 0.00661467807367444,
        'Conformity: rules': 0.0017643438186496496,
        'Conformity: interpersonal': 0.004064192529767752,
        'Humility': 0.0032048451248556376,
        'Benevolence: caring': 0.011124887503683567,
        'Benevolence: dependability': 0.017767170444130898,
        'Universalism: concern': 0.01814778335392475,
        'Universalism: nature': 0.9813610911369324,
        'Universalism: tolerance': 0.0025894937571138144
    },
    'labels': ['Universalism: nature']
}
```

Note: this is a multi-label model that returns all labels with scores in [0,1]. You still need to choose a threshold to decide what counts as “present”.

### 2. Multi-label usage with a custom threshold

For research use, you will often want to:
- Apply a sigmoid over logits
- Use a label-wise or global threshold (e.g., 0.3 instead of 0.5)

The predict_values function above already takes a threshold argument, so you can simply do:

```
predict_values(example, threshold=0.30)
```

and tune the threshold on your own validation set depending on your precision/recall preferences.

---

## Training data

The model was trained on the English, machine-translated portion of the ValueEval’24 / ValuesML dataset:
	•	Domain: news articles and political manifestos
	•	Unit of analysis: individual sentences
	•	Labels: 19 refined Schwartz values
	•	Each value has attained and constrained annotations in the original data
	•	For this model, these are collapsed into a single binary label per value
	•	Presence variable in the paper ((z_s)) is defined as “any of the 19 labels is positive”, but this model directly predicts the 19 values.

Important: the original dataset is distributed under a restricted Data Usage Agreement.
You must obtain the data separately from the ValueEval/ValuesML organisers (e.g. via Zenodo) and respect their license.

---

## Training setup

	•	Base model: microsoft/deberta-base
	•	Task: 19-way multi-label classification
	•	Objective: binary cross-entropy (BCEWithLogits) over the 19 labels
	•	Max sequence length: 512 tokens
	•	Optimizer: AdamW
	•	Effective batch size: 16 (batch 4 × gradient accumulation 4)
	•	Learning rate: 2e-5
	•	Weight decay: 0.15
	•	Epochs: up to 10 with early stopping on validation macro–F1
	•	Hardware: single GPU with ≤ 8 GB VRAM

This is the text-only baseline configuration in the paper (no LIWC or context features).

---

## Performance (paper reference)

On the English ValueEval’24 sentence-level splits used in the paper, this baseline DeBERTa model achieves:
	•	Macro–F₁ (19 values, test): ~0.31 (with a tuned global threshold around 0.30)

The paper reports:
	•	Comparisons with:
	•	Presence-gated hierarchies
	•	Feature-augmented DeBERTa models (LIWC-22, prior-sentence context, topics)
	•	Instruction-tuned LLM baselines (7–9B)
	•	A small soft-voting ensemble of three DeBERTa-based models (including this baseline) obtains the best overall performance (macro–F₁ ≈ 0.33).

For full details, please refer to the paper:

Human Values in a Single Sentence: Moral Presence, Hierarchies, and Transformer Ensembles on the Schwartz Continuum
Víctor Yeste, Paolo Rosso (2026)

---

## Limitations and bias

	•	The model is trained on news and political texts; it may not generalise to:
	•	Social media
	•	Everyday conversations
	•	Other genres or languages
	•	Values are annotated at the sentence level; many real-world value cues are only clear in broader context.
	•	Rare values (e.g., Humility, Hedonism, Universalism: tolerance) have few positive examples and are harder to predict.
	•	No systematic bias or fairness analysis has been conducted; the model should not be used for profiling individuals or making high-stakes decisions.

If you use this model, please:
	•	Keep humans in the loop.
	•	Treat outputs as noisy indicators, especially for rare labels.

---


## License

The model weights and code in this repository are released under the Apache License 2.0.

See the LICENSE file (or the license field in this model card) for details.

Note: This does not grant you any rights over the underlying training data (ValueEval/ValuesML).
Please obtain and use that data under its own license and Data Usage Agreement.

---

## Citation

If you use this model or the associated code in your research, please cite:

```
@misc{yeste2026humanvaluessinglesentence,
      title={Human Values in a Single Sentence: Moral Presence, Hierarchies, and Transformer Ensembles on the Schwartz Continuum}, 
      author={Víctor Yeste and Paolo Rosso},
      year={2026},
      eprint={2601.14172},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.14172}, 
}
```

You may also want to cite the ValueEval / ValuesML dataset:

```
@misc{ValueEval24Zenodo,
  author  = {{The ValuesML Team}},
  title   = {Touch{\'e}24{-}ValueEval},
  year    = {2024},
  month   = {8},
  version = {2024-08-09},
  publisher = {Zenodo},
  doi     = {10.5281/zenodo.13283288},
  url     = {https://doi.org/10.5281/zenodo.13283288}
}
```