The CoSENT (Cosine Sentence) model is designed to address the 'collapse' phenomenon in BERT-based sentence representations, where sentence embeddings are often clustered in a small region of the vector space, leading to high similarity scores even for unrelated sentences.

Unlike Sentence-BERT, which uses a classification-based training approach (concatenating $u$, $v$, and $|u-v|$), CoSENT optimizes the cosine similarity directly. It uses a ranking loss function to ensure that for any positive pair $(x^i, x^{i+})$ and negative pair $(x^i, x^{i-})$, the relationship $\cos(h^i, h^{i+}) > \cos(h^i, h^{i-})$ is maintained. This makes the training process more consistent with the actual prediction task (calculating cosine similarity).

The text2vec project provides several pre-trained CoSENT models optimized for different Chinese semantic matching tasks. Choose the model that best fits your use case:

• General Semantic Matching: Use shibing624/text2vec-base-chinese for general-purpose Chinese semantic matching tasks. It is based on hfl/chinese-macbert-base.
• Sentence-to-Sentence (s2s) Matching: Use shibing624/text2vec-base-chinese-sentence for s2s semantic matching. It is based on nghuyong/ernie-3.0-base-zh and trained on manually selected Chinese STS datasets.
• Sentence-to-Paraphrase (s2p) Matching: Use shibing624/text2vec-base-chinese-paraphrase for s2p tasks or when you need enhanced representation for long texts. It is based on nghuyong/ernie-3.0-base-zh and includes s2p data to achieve SOTA performance on NLI datasets.

To reproduce the Spearman correlation evaluation results reported in the documentation:

1. Download the Chinese matching datasets to the examples/data directory.
2. Run the evaluation script tests/model_spearman.py.

Jina provides a high-performance C/S mode deployment supporting Docker, Kubernetes, and multi-GPU processing.

1. Install Jina:

pip install jina

2. Start the server: Use the following pattern in examples/jina_server_demo.py to start a service using the Text2vecEncoder from JinaHub:

from jina import Flow

port = 50001
f = Flow(port=port).add(
    uses='jinahub://Text2vecEncoder',
    uses_with={'model_name': 'shibing624/text2vec-base-chinese'}
)

with f:
    f.block()

3. Call the service:

from jina import Client
from docarray import Document, DocumentArray

port = 50001
c = Client(port=port)

data = ['如何更换花呗绑定银行卡', '花呗更改绑定银行卡']
r = c.post('/', inputs=DocumentArray([Document(text='如何更换花呗绑定银行卡'), Document(text='花呗更改绑定银行卡')]))
print(r.embeddings)

You can install text2vec using pip. It requires torch to be installed first.

Option 1: Direct installation via PyPI

pip install torch
pip install -U text2vec

Option 2: Installation from source

pip install torch
git clone https://github.com/shibing624/text2vec.git
cd text2vec
pip install --no-deps .

You can train your own semantic matching models using the provided training scripts. The script to use depends on the training method and data format:

• SBERT training: Use examples/training_sup_text_matching_model.py.
• CoSENT training (JSONL data): Use examples/training_sup_text_matching_model_jsonl_data.py for models like text2vec-base-chinese-sentence or text2vec-base-chinese-paraphrase.

Build high-performance services using Jina. This method supports C/S mode, docker cloud native deployment, gRPC/HTTP/WebSocket, simultaneous prediction of multiple models, and multi-GPU processing. The model executor is available on JinaHub.

pip install jina

Start the service

from jina import Flow

port = 50001
f = Flow(port=port).add(
    uses='jinahub://Text2vecEncoder',
    uses_with={'model_name': 'shibing624/text2vec-base-chinese'}
)

with f:
    f.block()

Call the service

from jina import Client
from docarray import Document, DocumentArray

port = 50001
c = Client(port=port)

data = ['如何更换花呗绑定银行卡', '花呗更改绑定银行卡']
r = c.post('/', inputs=DocumentArray([Document(text='如何更换花呗绑定银行卡'), Document(text='花呗更改绑定银行卡')]))
print(r.embeddings)

For a native HTTP service, use FastAPI.

1. Install dependencies:

pip install fastapi uvicorn

2. Start the server:

cd examples
python fastapi_server_demo.py

3. Call the service via curl:

curl -X 'GET' \
  'http://0.0.0.0:8001/emb?q=hello' \
  -H 'accept: application/json'