Building a Sentiment Analysis Python Microservice with Flair and Flask

Flair delivers state-of-the-art performance in solving NLP problems such as named entity recognition (NER), part-of-speech tagging (PoS), sense disambiguation and text classification. It’s an NLP framework built on top of PyTorch.

In this post, I will cover how to build sentiment analysis Microservice with flair and flask framework.

Step 1: Create Python 3.6 virtualenv

To use Flair you need Python 3.6. We will start by creating a Python 3.6 virtualenv

$ python3.6 -m venv pyeth

Next, we activate the virtualenv

$ source pyeth/bin/activate

Next, you can check Python version

(pyeth) $ python --version

Python 3.6.1

Step 2: Install flair and flask package

To install Flair and Flask we will use pip as shown below

$ pip install flair flask

The above command will install all the required packages needed to build our Microservice. It will also install PyTorch which flair uses to do the heavy lifting.

Step 3: Create a REST API to analyse sentiments

Create a new file called app.py under the application directory.

$ touch app.py

Copy the following source code and paste it in app.py source file

from flask import abort, Flask, jsonify, request
from flair.models import TextClassifier
from flair.data import Sentence

app = Flask(__name__)

classifier = TextClassifier.load('en-sentiment')

@app.route('/api/v1/analyzeSentiment', methods=['POST'])
def analyzeSentiment():
    if not request.json or not 'message' in request.json:
        abort(400)
    message = request.json['message']
    sentence = Sentence(message)
    classifier.predict(sentence)
    print('Sentence sentiment: ', sentence.labels)
    label = sentence.labels[0]
    response = {'result': label.value, 'polarity':label.score}
    return jsonify(response), 200

if __name__ == "__main__":
    app.run()

The code shown above does the following:

1. It imports Flask classes and functions
2. Next, we import TextClassifier and Sentence classes from flair package
3. Next, we load the model related to sentiment analysis en-sentiment. The sentiment analysis model is based on IMDB dataset. When line 7 runs, it will download the sentiment analysis model and store it into the .flair subfolder of the home directory. This will take few minutes depending on your internet speed.
4. Next, we defined a POST route mapping to /api/v1/analyzeSentiment url. This API endpoint will receive the message in a JSON body. We created an instance of Sentence and and passed it to classifier predict method. The result is in the form of label object that has two fields – value and score.

You can now start the app using python app.py

Once application is started, you can test the REST API using on your favourite REST client. I will show how to make REST API using cURL.

We will first check a positive review I could watch The Marriage over and over again. At 90 minutes, it's just so delightfully heartbreaking

curl --request POST \
  --url http://localhost:5000/api/v1/analyzeSentiment \
  --header 'content-type: application/json' \
  --data '{
    "message":"I could watch The Marriage over and over again. At 90 minutes, it'\''s just so delightfully heartbreaking."
}'

The response returned by API

{
"polarity": 1.0,
"result": "POSITIVE"
}

Let’s now look at an example of negative sentence Inoffensive and unremarkable

curl --request POST \
  --url http://localhost:5000/api/v1/analyzeSentiment \
  --header 'content-type: application/json' \
  --data '{
        "message":"Inoffensive and unremarkable."
}'

The response returned by API

{
"polarity": 1.0,
"result": "NEGATIVE"
}

Finally, let’s look at mixed review I don't think Destroyer is a good movie, but it is never boring and often hilarious.

curl --request POST \
  --url http://localhost:5000/api/v1/analyzeSentiment \
  --header 'content-type: application/json' \
  --data '{
        "message":"I don'\''t think Destroyer is a good movie, but it is never boring and often hilarious."
}'

The response returned by API

{
  "polarity": 0.11292144656181335,
  "result": "NEGATIVE"
}

That’s it for today.