app.py

"""Demo for NER4OPT."""
import re
import warnings

warnings.filterwarnings("ignore", category=DeprecationWarning)

import spacy
from spacy import displacy
from spacy.training import iob_to_biluo, biluo_tags_to_offsets
from spacy.tokenizer import Tokenizer

import streamlit as st

from simpletransformers.ner import NERModel, NERArgs

HTML_WRAPPER = """<div style="overflow-x: auto; border: 1px solid #e6e9ef; border-radius: 0.25rem; padding: 1rem; margin-bottom: 2.5rem">{}</div>"""


@st.cache_resource
def load_models():
    """Load custom built NER4OPT model."""
    custom_labels = [
        'O',
        'B-CONST_DIR',
        'I-CONST_DIR',
        'B-LIMIT',
        'I-LIMIT',
        'B-VAR',
        'I-VAR',
        'B-OBJ_DIR',
        'B-OBJ_NAME',
        'I-OBJ_NAME',
        'B-PARAM',
        'I-PARAM',
    ]
    # # create model
    model_args = NERArgs()
    model_args.use_early_stopping = True
    model_args.early_stopping_delta = 0.01
    model_args.early_stopping_metric = "eval_loss"
    model_args.early_stopping_metric_minimize = False
    model_args.early_stopping_patience = 5
    model_args.evaluate_during_training_steps = 2000
    model_args.overwrite_output_dir = True
    model_args.reprocess_input_data = True
    model_args.num_train_epochs = 11
    model_args.adafactor_beta1 = 0.9
    model_args.weight_decay = 0.01
    model_args.max_seq_length = 512
    model_args.learning_rate = 4e-5
    model_args.train_batch_size = 1
    model_args.eval_batch_size = 1
    model_args.manual_seed = 123456789
    model_args.output_dir = "trained_transformer_model"
    model_args.use_cuda = True
    model_args.use_multiprocessing = False
    model = NERModel("roberta",
                     "skadio/ner4opt-roberta",
                     labels=custom_labels,
                     use_cuda=True,
                     args=model_args)
    spacy_model = spacy.load("en_core_web_sm")
    spacy_model.tokenizer = Tokenizer(spacy_model.vocab,
                                      token_match=re.compile(r'\S+').match)
    spacy_blank_model = spacy.blank('en')
    spacy_blank_model.tokenizer = Tokenizer(
        spacy_blank_model.vocab, token_match=re.compile(r'\S+').match)
    return model, spacy_model, spacy_blank_model


def main():

    st.sidebar.title("""
        NER4OPT Demo: \nFull code will be available at https://github.com/skadio/Ner4Opt
        """)

    text = st.text_area(
        "Text",
        "Cautious Asset Investment has a total of $ 150,000 to manage and decides to invest it in money market fund , which yields a 2 % return as well as in foreign bonds , which gives and average rate of return of 10.2 % . Internal policies require PAI to diversify the asset allocation so that the minimum investment in money market fund is 40 % of the total investment . Due to the risk of default of foreign countries , no more than 40 % of the total investment should be allocated to foreign bonds . How much should the Cautious Asset Investment allocate in each asset so as to maximize its average return ?"
    )
    if text == "":
        st.write("Please write a valid sentence.")
    model, spacy_model, spacy_blank_model = load_models()

    # Augmented Text
    spacy_doc = spacy_model(text)
    if len(list(spacy_doc.sents)) >= 2:
        last_two_sentences = ' '.join(
            [item.text for item in list(spacy_doc.sents)[-2::]])
    else:
        last_two_sentences = ' '.join(
            [item.text for item in list(spacy_doc.sents)[-1::]])
    to_skip_count = len(last_two_sentences.split())
    augmented_sent = last_two_sentences + " " + text

    if st.button("Get Named Entities"):
        predictions, raw_outputs = model.predict([augmented_sent],
                                                 split_on_space=True)
        transformer_predictions = [
            list(val.values())[0] for val in predictions[0]
        ]
        transformer_predictions = transformer_predictions[to_skip_count::]
        biluo_tags = iob_to_biluo(transformer_predictions)
        doc = spacy_blank_model.make_doc(text)
        entities = biluo_tags_to_offsets(doc, biluo_tags)
        entities_formatted = []
        for tag in entities:
            entities_formatted.append({
                "start": tag[0],
                "end": tag[1],
                "label": tag[2],
                "score": 1.0
            })
        ner_for_display = [{
            "text": doc.text,
            "ents": entities_formatted,
            "title": None
        }]
        st.title("Named Entity Results")
        html_ner = displacy.render(ner_for_display, style="ent", manual=True)
        html_ner = html_ner.replace("\n", " ")
        st.write(HTML_WRAPPER.format(html_ner), unsafe_allow_html=True)


if __name__ == '__main__':
    main()