Deploy BERT for Sentiment Analsysi with FastAPI

Deploy a pre-trained BERT model for Sentiment Analysis as a REST API using FastAPI

Demo

The API currently support NER and POS tagging as well as a generalized tokenizer model that splits words that Swedish languages usually concatenates together. The models are loaded from the KB/Bert repository at huggingface website: .

Tokenizer

This model combines opus, kb-bert and af models to tokenize the input text and outputs a list of lists of tokens where each item in the outer list corresponds to the text string separated by a space character. The inner list contains all words the tokenizer model has devided the word into. This model should be combined with POS and NER models to reduce vocabulary.

Here's a sample request to the API (note: installation of http-test library can be done on any linux computer by apt-get install httpie):

http POST http://127.0.0.1:8081/tokenize/predict text="Obama bor i vitahuset, han kommer flytta därifrån om tre dagar"

The response you'll get looks something like this:

{
    "tokens": [
        ["Obama"],
        ["bor"],
        ["i"],
        ["vita", "huset", ","],
        ["han"],
        ["kommer"],
        ["flytta"],
        ["där", "ifrån"],
        ["om"],
        ["tre"],
        ["dagar"],
    ]
}

Part-of-Speech

This model downloads and uses KB/bert-base-swedish-cased-pos model to tag each word in the input text with a part-of-speech tag.

Here's a sample request to the API (note: installation of http-test library can be done on any linux computer by apt-get install httpie):

http POST http://127.0.0.1:8081/part-of-speech/predict text="Obama bor i vitahuset, han kommer flytta därifrån om tre dagar"

The response you'll get looks something like this:

{
    "tokens": [
        { "word": "Obama", "start": 0, "end": 4, "entity": "NN", "score": 0.9987 },
        ....
        { "word": "dagar"...}
    ]
}

Named-Entity-Recognition

This model downloads and uses KB/bert-base-swedish-cased-ner model to tag all words in the input text that has a potential named entity. The entiteis the model considers are: Name (PER), Location (LOC), Organisation (ORG), Event (EVT) and Time (TIME). It also happens that the model identifies entities where the entity type cannot be determined and thus these will have an unkown entity attached.

Here's a sample request to the API (note: installation of http-test library can be done on any linux computer by apt-get install httpie):

http POST http://127.0.0.1:8081/named-entity-recognition/predict text="Obama bor i vitahuset, han kommer flytta därifrån om tre dagar"

The response you'll get looks something like this:

{
    "tokens": [
        { "word": "Obama", "start": 0, "end": 4, "entity": "PER", "score": 0.9987 },
        { "word": "vitahuset", "entity": "LOC", ...},
        {"word": "tre", "entity": "TIME", ...},
        {"word": "dagar", "entity": "TIME", ...}
    ]
}

Installation

Install the dependencies:

pipenv install --dev

Test the setup

Start the HTTP server:

bin/start_server

Send a test request:

bin/test_request

Run helm dockerfile

To be able to run application in docker

./bin/build_and_tag
./bin/start_docker_server

Run helm kubernetes

To run in kubernetes first follow instructions at: Then build the docker image and install the helm chart in this repo:

./bin/build_and_tag
helm install bms helm -n bert-models