|
--- |
|
license: mit |
|
datasets: |
|
- webis/Touche23-ValueEval |
|
language: |
|
- en |
|
metrics: |
|
- f1 |
|
tags: |
|
- social-values |
|
--- |
|
# Schwartz Value Classifier |
|
This classifier is intended to predict the existence of social values from text snippets. |
|
|
|
*Disclaimer: this is not the official repo published by the authors of the paper, and may not truly replicate the performance described in the original study* |
|
|
|
## Value dimensions |
|
In this model we follow the 10-dimensional categorization of the Schwartz values. [link](https://en.wikipedia.org/wiki/Theory_of_basic_human_values) |
|
|
|
1. security – safety, harmony, and stability of society, of relationships, and of self |
|
2. power – social status and prestige, control or dominance over people and resources |
|
3. achievement – personal success through demonstrating competence according to social standards |
|
4. hedonism – pleasure or sensuous gratification for oneself |
|
5. stimulation – excitement, novelty and challenge in life |
|
6. self-direction – independent thought and action—choosing, creating, exploring |
|
7. universalism – understanding, appreciation, tolerance, and protection for the welfare of all people and for nature |
|
8. benevolence – preserving and enhancing the welfare of those with whom one is in frequent personal contact (the 'in-group') |
|
9. conformity – restraint of actions, inclinations, and impulses likely to upset or harm others and violate social expectations or norms |
|
10. tradition – respect, commitment, and acceptance of the customs and ideas that one's culture or religion provides |
|
|
|
## Datasets |
|
This model is finetuned on two datasets: ValueNet (A New Dataset for Human Value Driven Dialogue System, Qiu et al. 2021) and Touche23-ValueEval (The Touché23-ValueEval Dataset for Identifying Human Values behind Arguments, Mirzakhmedova et al., 2023). |
|
|
|
We follow the original paper to convert both datasets into a binary classification task for each dimension. |
|
- ValueNet |
|
- A sentence has a positive label if the original label contains 1 (positive) or -1 (negative), and 0 if the original label is 0. |
|
- ValueEval |
|
- A sentence is assigned a positive label if the original label vector is marked 1 for that dimension. |
|
- Since the original paper follows a 20-dimension refined categorization, we map them back to 10 dimensions. Therefore, the same sentence appears ten times, once for each dimension. |
|
|
|
## How to use |
|
Start your sentence with a label that indicates which dimension to measure. An example would be: |
|
|
|
- \<power> [SEP] staying out late after telling my girlfriend I could be home early |
|
|
|
Please make sure to follow the exact format "<value\_name>" at the beginning of the sentence as this is a special token in the tokenizer: any spaces or different formats will not be encoded correctly. |
|
|
|
## Performances |
|
- macro F1 score |
|
- on ValueNet: 0.648 |
|
- on ValueEval: 0.744 |
|
- Combined: 0.759 |
|
|
|
- ROC-AUC |
|
- on ValueNet: 0.736 |
|
- on ValueEval:0.847 |
|
- Combined: 0.855 |
|
|
|
## Training details |
|
- Base model: bert-base-uncased |
|
- Epochs: 10 w/ early stopping after no F1 increase in 3 epochs |
|
- Learning rate: 5e-5 w/ warmup for 0.03 steps and subsequent linear decay |
|
- Batch size: 32 |
|
- Upsampled training set to maintain 1:1 balance for pos:neg labels. |
|
|
|
## References |
|
- Do Differences in Values Influence Disagreements in Online Discussions? (EMNLP'23) [link](https://aclanthology.org/2023.emnlp-main.992/) |
