app.py

import streamlit as st
import transformers
from annotated_text import annotated_text

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):
    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.")

option = st.selectbox('Select model',
                      ("dslim/bert-base-NER",
                       "dslim/bert-large-NER",
                       "Davlan/bert-base-multilingual-cased-ner-hrl"))
default_text = "Xbox v PlayStation: Giants clash over Call of Duty: Xbox owner Microsoft has hit back at claims its plan to buy the maker of Call of Duty may unfairly affect its rivals, including Sony, which owns PlayStation."
text = st.text_area('Enter text here:', value=default_text)
st.write('Model used for prediction:', option)
submit = st.button('Predict')

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

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)