import os import re from huggingface_hub import snapshot_download import streamlit as st import streamlit.components.v1 as components import spacy from spacy import displacy from spacy.tokens import Span import pandas as pd import numpy as np from utils import add_logo_to_sidebar, add_footer HF_TOKEN = os.environ.get("HF_TOKEN") REPO_ID = "simplexico/cuad-spacy-custom-ner" EXAMPLE_TEXT = """Exhibit 10.16 CONSULTING AGREEMENT This Consulting Agreement (the "Agreement") is made and entered into as of this 2nd day of January 2020, by and between Global Technologies, Ltd (hereinafter the "Company"), a Delaware corporation whose address is 501 1st Ave N., Suite 901, St. Petersburg, FL 33701 and Timothy Cabrera (hereinafter the "Consultant"), an individual whose address is 11718 SE Federal Hwy., Suite 372, Hobe Sound, FL 33455 (individually, a "Party"; collectively, the "Parties").""" ## Layout stuff st.set_page_config( page_title="Extract Demo", page_icon="⛏", layout="wide", initial_sidebar_state="expanded", menu_items={ 'Get Help': 'mailto:hello@simplexico.ai', 'Report a bug': None, 'About': "## This a demo showcasing different Legal AI Actions" } ) add_logo_to_sidebar() st.title('⛏ Extract Demo') st.write(""" This demo shows how AI can be used to extract information from text. We've trained an AI model to extract key pieces of information from a contract recital. """) st.write("**👈 Enter a contract recital on the left** and hit the button **Extract Data** to see the demo in action") @st.cache(allow_output_mutation=True) def load_model(): snapshot_download(repo_id=REPO_ID, token=HF_TOKEN, local_dir='./') nlp = spacy.load('model-best') return nlp text = st.sidebar.text_area('Enter Clause Text', value=EXAMPLE_TEXT, height=250) button = st.sidebar.button('Extract Data', type='primary', use_container_width=True) with st.spinner('⚙️ Loading model...'): nlp = load_model() def check_span_pair_for_overlap(span1, span2): """ Checks a pair of spans for any overlapping ranges Args: span1: (start, end) tuple span2: (start, end) tuple Return: True if overlap, False otherwise """ # remove offset minimum = min(span1[0], span2[0]) span1 = (span1[0] - minimum, span1[1] - minimum) span2 = (span2[0] - minimum, span2[1] - minimum) maximum = max(span1[1], span2[1]) vec1 = np.zeros(maximum) vec1[span1[0]:span1[1]] = 1 vec2 = np.zeros(maximum) vec2[span2[0]:span2[1]] = 1 if np.dot(vec1, vec2): return True return False def add_detected_persons_as_parties(doc): nlp = spacy.load('en_core_web_md') doc_ = nlp(doc.text) original_ents = list(doc.ents) for ent in doc_.ents: if ent.label_ == 'PERSON': if not any([check_span_pair_for_overlap((ent.start, ent.end), (ent_.start, ent_.end)) for ent_ in original_ents]): print(ent) # check for overlapping ents original_ents.append(Span(doc, ent.start, ent.end, label='parties')) doc.ents = original_ents return doc def add_rule_based_entites(doc): """Adds rule based entity spans to document Args: doc (spacy.tokens.doc.Doc) """ patterns = [ ('[0-9]+[\s]+[a-zA-Z0-9.\-\,\#]+[\s]*[a-zA-Z0-9.\-\,\#]+[a-zA-Z0-9\s.\-\,\#]*\s[0-9]+', 'address'), ('Consultant|Company|Party|Parties', 'role'), ] for pattern, label in patterns: ents = [] for match in re.finditer(pattern, doc.text): start, end = match.span() span = doc.char_span(start, end) if span is not None: ents.append((span.start, span.end, span.text)) original_ents = list(doc.ents) for ent in ents: start, end, address = ent per_ent = Span(doc, start, end, label=label) original_ents.append(per_ent) doc.ents = original_ents return doc if button: col1, col2 = st.columns(2) with st.spinner('⚙️ Extracting Data...'): doc = nlp(text) doc = add_rule_based_entites(doc) doc = add_detected_persons_as_parties(doc) with col1: st.subheader('🎨 Highlighted Text') colors = {'party': "#85C1E9", "address": "#ff6961", "agreement_date": "#5de36f", "role": "#b05de3"} options = {"ents": ['party', 'address', 'agreement_date', 'role'], "colors": colors} label_aliases = { 'parties': 'Party', 'address': 'Address', 'agreement_date': 'Agreement Date', 'role': 'Role' } doc.spans["sc"] = [ Span(doc, ent.start, ent.end, label_aliases[ent.label_]) for ent in doc.ents ] html = displacy.render(doc, style="span", options=options) components.html(html, height=400) with col2: # display table data = { 'Text': [], 'Label': [] } st.subheader('📊 Extracted Data') for span in doc.spans['sc']: data['Label'].append(span.label_) data['Text'].append(span.text) df = pd.DataFrame(data) hide_dataframe_row_index = """ """ # Inject CSS with Markdown st.markdown(hide_dataframe_row_index, unsafe_allow_html=True) st.table(df) add_footer()