daveminer/bert-serv

FinBERT pre-trained models as a containerized HTTP service.

 _______  _______  ______    _______         _______  _______  ______    __   __
|  _    ||       ||    _ |  |       |       |       ||       ||    _ |  |  | |  |
| |_|   ||    ___||   | ||  |_     _| ____  |  _____||    ___||   | ||  |  |_|  |
|       ||   |___ |   |_||_   |   |  |____| | |_____ |   |___ |   |_||_ |       |
|  _   | |    ___||    __  |  |   |         |_____  ||    ___||    __  ||       |
| |_|   ||   |___ |   |  | |  |   |          _____| ||   |___ |   |  | | |     |
|_______||_______||___|  |_|  |___|         |_______||_______||___|  |_|  |___|

BERT-serv provides FinBERT sentiment as a service. Send financial text via HTTP and get sentiment analysis back from these pretrained models.

❗ Only this sentiment model is implemented at this time.

Why

• FinBERT analysis made available on-demand to HTTP clients (no Python required!)
• Avoid adding PyTorch dependencies to other parts of your system
• Results are saved and can be queried
• Analysis requests are handled asynchronously

Run the Local Demo

Install Docker Desktop if needed.

git clone https://github.com/daveminer/BERT-serv.git
cd BERT-serv
cp .env.dist .env
docker-compose up

BERT-serv is now running on local port 8000!

How to Use

Request sentiment analysis on text

Send a POST request to the /sentiment/new/ path. A callback_url may be specified in the query parameters for asynchrous use cases and long-running sentiment analyses. This callback will have a JSON object in the body with an array of the new sentiment record ids: '{"ids": [95, 96]}'

The body of the POST request must be a list; the strings inside will be processed synchronously.

curl --request POST \
  --url 'http://localhost:8000/sentiment/new/?callback_url=http%3A%2F%2Fweb%3A8000%2Fcallback%2Fsentiment%2Fnew%2F' \
  --header 'Content-Type: application/json' \
  --data '[
	"there is a shortage of capital, and we need extra financing",
	"year over year growth is increasing"
]'

Look up sentiment results

Response defaults to HTML unless application/json is specified in the Accept header.

Get last 100 sentiments

Make a GET request to the /sentiment/ path.

curl --request GET \
  --url http://localhost:8000/sentiment/ \
  --header 'Accept: application/json'

Get a specific sentiment by index

Add the index of the sentiment resource to the sentiment path:

curl --request GET \
  --url http://localhost:8000/sentiment/1/ \
  --header 'Accept: application/json'

Filter sentiments

By Tag

A comma-separated list of tags may be specified in the tags query parameter.

By Age

The period query parameter accepts an integer that specifies how many days back to look for sentiments.

Pagination

The sentiment index controller uses Django's built in pagination. The /sentiment/ path accepts page_size and a page query parameters.

Development Environment Setup

The make services command will start all of the services besides the app. This allows for the app to be started and stopped (with make app) in the terminal for convenience during development.

Setting up the environment

Local development requires that the local environment is set up alongside the containerized services.

Create the virtualenv

python3 -m virtualenv env

Load

python3 -m venv .venv

Activate

source .venv/bin/activate

Notes:

• make services requires Docker Desktop
• make deps will install dependencies via pip3 and must be run before make app. This can take a few minutes as the PyTorch dependencies are sizable.