Sbert Synonymy

🚀 Phrase Synonymy Detector

This is a sentence-transformers model designed to determine the synonymy of two short texts, mainly single sentences with a length of up to 10 - 15 words.

🚀 Quick Start

This is a sentence-transformers model aimed at determining the synonymy of two short texts, primarily single sentences of up to 10 - 15 words in length. Part of the training dataset is available in the repository inkoziev/paraphrases.

The model computes 312 - dimensional vectors for the text and the question. The cosine of the angle between these vectors provides an assessment of their semantic similarity. In the dialogue system project, it is used to identify paraphrases of statements and facts.

The model is based on cointegrated/rubert-tiny2. It has a very small size and performs inference quickly even on a CPU.

✨ Features

• Designed for detecting synonymy of short texts.
• Computes 312 - dimensional vectors for semantic assessment.
• Based on a lightweight model, enabling fast inference on CPU.

📦 Installation

For convenience, install sentence-transformers:

pip install -U sentence-transformers

💻 Usage Examples

Basic Usage

To determine the synonymy of sentence pairs, you can use the following code:

import sentence_transformers

sentences = ["Одна голова - хорошо, а две - лучше",
"пара голов имеет преимущество перед единственной",
"Потерявши голову, по волосам не плачут",]

model = sentence_transformers.SentenceTransformer('inkoziev/sbert_synonymy')
embeddings = model.encode(sentences)

s1 = sentences[0]
v1 = embeddings[0]
for i2 in range(1, 3):
    s = sentence_transformers.util.cos_sim(a=v1, b=embeddings[i2]).item()
    print('text1={} text2={} cossim={}'.format(s1, sentences[i2], s))

The result will be approximately as follows:

text1=Одна голова - хорошо, а две - лучше  text2=пара голов имеет преимущество перед единственной  cossim=0.8603419065475464
text1=Одна голова - хорошо, а две - лучше  text2=Потерявши голову, по волосам не плачут  cossim=0.013120125979185104

📚 Documentation

Related Projects

The generative poetic paraphraser is available on the Hugging Face hub: inkoziev/paraphraser. The code for training this model is available in the GitHub repository.

📄 License

This project is under the Unlicense.

📚 Citations

@MISC{rugpt_chitchat,
    author  = {Ilya Koziev},
    title   = {Paraphrase Detection Model},
    url     = {https://huggingface.co/inkoziev/sbert_synonymy},
    year    = 2022
}