app.py

import numpy as np
import onnxruntime
import onnx
import gradio as gr
import requests
import json
from extractnet import Extractor
import math
from transformers import AutoTokenizer
import spacy
import os
from transformers import pipeline
import itertools

MODEL_TRANSFORMER_BASED = "distilbert-base-uncased"
MODEL_ONNX_FNAME = "ESG_classifier_batch.onnx"
MODEL_SENTIMENT_ANALYSIS = "ProsusAI/finbert"
#MODEL_SUMMARY_PEGASUS = "oMateos2020/pegasus-newsroom-cnn_full-adafactor-bs6"



#API_HF_SENTIMENT_URL = "https://api-inference.huggingface.co/models/cardiffnlp/twitter-roberta-base-sentiment" 

def _inference_ner_spancat(text, summary, penalty=0.5, normalise=True, limit_outputs=10):
    nlp = spacy.load("en_pipeline")
    doc = nlp(text)
    spans = doc.spans["sc"]
    comp_raw_text = dict( sorted( dict(zip([str(x) for x in spans],[float(x)*penalty for x in spans.attrs['scores']])).items(), key=lambda x: x[1], reverse=True) )
    doc = nlp(summary)
    spans = doc.spans["sc"]
    exceeds_one = 0.0
    for comp_s in spans:
        if str(comp_s) in comp_raw_text.keys():
            comp_raw_text[str(comp_s)] = comp_raw_text[str(comp_s)] / penalty
            temp_max = comp_raw_text[str(comp_s)]if comp_raw_text[str(comp_s)] > 1.0 else 0.0
            exceeds_one = comp_raw_text[str(comp_s)] if temp_max > exceeds_one else exceeds_one 
    #This "exceeds_one" is a bit confusing. So the thing is that the penalty is reverted for each time the company appears in the summary and hence the value can exceed one when the company appears more than once. The normalisation means that all the other scores are divided by the maximum when any value exceeds one
    if normalise and (exceeds_one > 1):
        comp_raw_text = {k: v/exceeds_one for k, v in comp_raw_text.items()}
   
    return dict(itertools.islice(sorted(comp_raw_text.items(), key=lambda x: x[1], reverse=True), limit_outputs))

#def _inference_summary_model_pipeline(text):
#    pipe = pipeline("text2text-generation", model=MODEL_SUMMARY_PEGASUS)
#    return pipe(text,truncation='longest_first')

def _inference_sentiment_model_pipeline(text):
    tokenizer_kwargs = {'padding':True,'truncation':True,'max_length':512}#,'return_tensors':'pt'}
    pipe = pipeline("sentiment-analysis", model=MODEL_SENTIMENT_ANALYSIS )
    return pipe(text,**tokenizer_kwargs)

#def _inference_sentiment_model_via_api_query(payload):
#    response = requests.post(API_HF_SENTIMENT_URL , headers={"Authorization": os.environ['hf_api_token']}, json=payload)
#    return response.json()

def _lematise_text(text):
   nlp = spacy.load("en_core_web_sm", disable=['ner'])
   text_out = []
   for doc in nlp.pipe(text): #see https://spacy.io/models#design
       new_text = ""
       for token in doc:
           if (not token.is_punct
               and not token.is_stop
               and not token.like_url
               and not token.is_space
               and not token.like_email
               #and not token.like_num
               and not token.pos_ == "CONJ"):
                    
                new_text = new_text + " " + token.lemma_

       text_out.append( new_text )
   return text_out

def sigmoid(x):
  return 1 / (1 + np.exp(-x))

def to_numpy(tensor):
    return tensor.detach().cpu().numpy() if tensor.requires_grad else tensor.cpu().numpy()

def is_in_archive(url):
    try:
        r = requests.get('http://archive.org/wayback/available?url='+url)
        archive = json.loads(r.text)
    
        if archive['archived_snapshots'] :
            archive['archived_snapshots']['closest']
            return {'archived':archive['archived_snapshots']['closest']['available'], 'url':archive['archived_snapshots']['closest']['url'],'error':0}
        else:
            return {'archived':False, 'url':"", 'error':0}
    except:
        print(f"[E] Quering URL ({url}) from archive.org")
        return {'archived':False, 'url':"", 'error':-1}

#def _inference_ner(text):
#    return labels

def _inference_classifier(text):
    tokenizer = AutoTokenizer.from_pretrained(MODEL_TRANSFORMER_BASED)
    inputs = tokenizer(_lematise_text(text), return_tensors="np", padding="max_length", truncation=True) #this assumes head-only!
    ort_session = onnxruntime.InferenceSession(MODEL_ONNX_FNAME)
    onnx_model = onnx.load(MODEL_ONNX_FNAME)
    onnx.checker.check_model(onnx_model)

    # compute ONNX Runtime output prediction
    ort_outs = ort_session.run(None, input_feed=dict(inputs))

    return sigmoid(ort_outs[0])

def inference(input_batch,isurl,use_archive,limit_companies=10):
    input_batch_content = []
    if isurl:
        for url in input_batch:
            if use_archive:
                archive = is_in_archive(url)
                if archive['archived']:
                    url = archive['url']
            #Extract the data from url
            extracted = Extractor().extract(requests.get(url).text)
            input_batch_content.append(extracted['content'])
    else:
        input_batch_content = input_batch
    
    prob_outs = _inference_classifier(input_batch_content)
    #sentiment = _inference_sentiment_model_via_api_query({"inputs": extracted['content']})
    #sentiment = _inference_sentiment_model_pipeline(input_batch_content )[0]
    #summary = _inference_summary_model_pipeline(input_batch_content )[0]['generated_text']
    #ner_labels = _inference_ner_spancat(input_batch_content ,summary, penalty = 0.8, limit_outputs=limit_companies)

    return prob_outs #ner_labels, {'E':float(prob_outs[0]),"S":float(prob_outs[1]),"G":float(prob_outs[2])},{sentiment['label']:float(sentiment['score'])},"**Summary:**\n\n" + summary

title = "ESG API Demo"
description = """This is a demonstration of the full ESG pipeline backend where given a URL (english, news) the news contents are extracted, using extractnet, and fed to three models:

- An off-the-shelf sentiment classification model (ProsusAI/finbert)
- A custom NER for the company extraction
- A custom ESG classifier for the ESG labeling of the news (the extracted text is also lemmatised prior to be fed to this classifier) 

API input parameters:
- URL: text. Url of the news (english)
- `use_archive`: boolean. The model will extract the archived version in archive.org of the url indicated. This is useful with old news and to bypass news behind paywall
- `limit_companies`: integer. Number of found relevant companies to report.

"""
examples = [[['https://www.bbc.com/news/uk-62732447',
            'https://www.bbc.com/news/business-62747401',
            'https://www.bbc.com/news/technology-62744858',
            'https://www.bbc.com/news/science-environment-62758811',
            'https://www.theguardian.com/business/2022/sep/02/nord-stream-1-gazprom-announces-indefinite-shutdown-of-pipeline',
            'https://www.bbc.com/news/world-europe-62766867',
            'https://www.bbc.com/news/business-62524031',
            'https://www.bbc.com/news/business-62728621',
            'https://www.bbc.com/news/science-environment-62680423'],'url',False,5]]
demo = gr.Interface(fn=inference, 
                    inputs=[gr.Dataframe(label='input batch', col_count=1, datatype='str', type='array', wrap=True),
                            gr.Dropdown(label='data type', choices=['text','url'], type='index'),
                            gr.Checkbox(label='if url parse cached in archive.org'),
                            gr.Slider(minimum=1, maximum=10, step=1, label='Limit NER output')],
                    outputs=[gr.Dataframe(label='output raw', col_count=1, datatype='number', type='array', wrap=True)],
                             #gr.Label(label='Company'),
                             #gr.Label(label='ESG'),
                             #gr.Label(label='Sentiment'),
                             #gr.Markdown()],
                    title=title,
                    description=description,
                    examples=examples)
demo.launch()