@inproceedings{cui-etal-2025-conditional,
 title = "Conditional Dichotomy Quantification via Geometric Embedding",
 author = "Cui, Shaobo and
 Liu, Wenqing and
 Feng, Yiyang and
 Zhou, Jiawei and
 Faltings, Boi",
 editor = "Che, Wanxiang and
 Nabende, Joyce and
 Shutova, Ekaterina and
 Pilehvar, Mohammad Taher",
 booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
 month = jul,
 year = "2025",
 address = "Vienna, Austria",
 publisher = "Association for Computational Linguistics",
 url = "https://aclanthology.org/2025.acl-long.383/",
 doi = "10.18653/v1/2025.acl-long.383",
 pages = "7765--7791",
 ISBN = "979-8-89176-251-0"
}

A lightweight toolkit for measuring how "opposite" two texts are when they share the same context.

Software: PyPI version

Pretrained Models: debate-bert defeasibleNLI-bert causal-reasoning-bert

• ✨ Why should I care?
• 📦 Installation
• 🗂️ Supportive Datasets (Quick Overview)
• Opposite-Score
  • Features
  • Installation
  • 💡 Usage Examples
• 🛡 Responsible Usage
• 📊 Leaderboard: Dichotomy Quantification

Powered by our Opposite-Score embeddings and three rigorously curated datasets (Debate ▪︎ Defeasible NLI ▪︎ Causal Reasoning).

conda create -n dichotomy python=3.10
conda activate dichotomy
## opposite-score is our implemented package: https://pypi.org/project/opposite-score/
pip install opposite-score

Sentence-length distributions

Figure 1. Sentence-length distributions for contexts, positive, negative, and neutral arguments across datasets.

Why it matters
Balanced lengths & human-verified neutrals stop models from "cheating" on superficial cues and keep the focus on genuine oppositional content.

Efficient embeddings and scoring mechanism for detecting contrasting or opposite relationships in text, based on a given context.

Opposite-Score is designed to generate embeddings and compute the opposite-score, which quantifies the degree of contrast or opposition between two textual outputs within the same context. This package is particularly useful in scenarios like debates, legal reasoning, and causal analysis where contrasting perspectives need to be evaluated based on shared input.

• Opposite-Score Calculation: Computes a numerical score representing how opposite two texts are, conditioned on a shared context.
• Opposite-Aware Embeddings: Generates embeddings optimized for contrasting textual relationships.
• Easy to Use: Only a few lines of code to get sentence/token-level embeddings and calculate opposite scores.
• Automatic Model Download: The first initialization automatically downloads and installs the pre-trained Opposite-Score model.

Install Opposite-Score via pip:

pip install opposite-score==0.0.1

from oppositescore.model.dichotomye import DichotomyE
# Example inputs
context = ["A company launches a revolutionary product."]
sentence1 = ["Competitors quickly release similar products, reducing the company's advantage."]
sentence2 = ["The company gains a significant advantage due to its unique product."]
# Initialize the model
opposite_scorer = DichotomyE.from_pretrained('shaobocui/opposite-score-debate-bert', pooling_strategy='cls').cuda()
# Calculate opposite-score (using cosine similarity as an example)
opposite_score = opposite_scorer.calculate_opposite_score(ctx=context, sent1=sentence1, sent2=sentence2)
print('Opposite Score:', opposite_score)
# Output: Opposite Score: 0.11086939

This software is released for research and educational purposes only. It is intended to support studies on argument contrast, causal reasoning, and sentence embeddings.

Please ensure proper attribution when using the code, models, or datasets in publications or derivative work. Commercial use are expected to contact authors for explicit permission.

For questions or collaborations, feel free to contact the authors.

📌 Both metrics benefit from higher values: better classification and stronger geometric contrast.

📄 Full Table with Ablations →