app.py

import streamlit as st
from annotated_text import annotated_text
import transformers

ENTITY_TO_COLOR = {
    'PER': '#8ef',
    'LOC': '#faa',
    'ORG': '#afa',
    'MISC': '#fea',
}

@st.cache(allow_output_mutation=True, show_spinner=False)
def get_pipe():
    model_name = "dslim/bert-base-NER"
    model = transformers.AutoModelForTokenClassification.from_pretrained(model_name)
    tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)
    pipe = transformers.pipeline("token-classification", model=model, tokenizer=tokenizer, aggregation_strategy="simple")
    return pipe

def parse_text(text, prediction):
    start = 0
    parsed_text = []
    for p in prediction:
        parsed_text.append(text[start:p["start"]])
        parsed_text.append((p["word"], p["entity_group"], ENTITY_TO_COLOR[p["entity_group"]]))
        start = p["end"]
    parsed_text.append(text[start:])
    return parsed_text

st.set_page_config(page_title="Named Entity Recognition")
st.title("Named Entity Recognition")
st.write("Type text into the text box and then press 'Predict' to get the named entities.")

default_text = "My name is John Smith. I work at Microsoft. I live in Paris. My favorite painting is the Mona Lisa."

text = st.text_area('Enter text here:', value=default_text)
submit = st.button('Predict')

with st.spinner("Loading model..."):
    pipe = get_pipe()

if (submit and len(text.strip()) > 0) or len(text.strip()) > 0:

    prediction = pipe(text)

    parsed_text = parse_text(text, prediction)

    st.header("Prediction:")
    annotated_text(*parsed_text)

    st.header('Raw values:')
    st.json(prediction)