Create app.py

app.py

@@ -0,0 +1,67 @@
+import streamlit as st
+from sentence_transformers import SentenceTransformer, util
+from transformers import (AutoModelForQuestionAnswering,
+                          AutoTokenizer, pipeline)
+
+import pandas as pd
+import regex as re
+from urllib import request
+
+# Select model for question answering
+model_name = "deepset/roberta-base-squad2"
+
+# Load model & tokenizer
+model = AutoModelForQuestionAnswering.from_pretrained(model_name)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+# Create pipeline
+pipe = pipeline('question-answering', model=model_name, tokenizer=model_name)
+
+# Load Harry Potter book corpus from link
+url = ("https://raw.githubusercontent.com/formcept/whiteboard/master/nbviewer/notebooks/data/harrypotter/Book%201%20-%20The%20Philosopher's%20Stone.txt")
+response = request.urlopen(url)
+book1_raw_0 = response.read().decode('utf8')
+
+# Text pre-processing
+# Remove page statements
+book1_raw_1 = re.sub(r'Page \| [0-9]+ Harry Potter [a-zA-Z \-]+J.K. Rowling',
+                     '', book1_raw_0)
+
+# Remove newlines
+book1_raw_1 = re.sub(r'\n', '', book1_raw_1)
+
+# Remove periods; this will relevant in the regrouping later
+book1_raw_1 = re.sub(r'Mr. ', 'Mr ', book1_raw_1)
+book1_raw_1 = re.sub(r'Ms. ', 'Ms ', book1_raw_1)
+book1_raw_1 = re.sub(r'Mrs. ', 'Mrs ', book1_raw_1)
+
+# Group into 3 sentences-long parts
+paragraphs = re.findall("[^.?!]+[.?!][^.?!]+[.?!][^.?!]+[.?!]", book1_raw_1)
+
+# Type in HP-related query here
+query = st.text_area("Hello muggle! What is your question?")
+
+# Perform sentence embedding on query and sentence groups
+model_embed_name = 'sentence-transformers/multi-qa-MiniLM-L6-cos-v1'
+
+model_embed = SentenceTransformer(model_embed_name)
+doc_emb = model_embed.encode(paragraphs)
+query_emb = model_embed.encode(query)
+
+#Compute dot score between query and all document embeddings
+scores = util.cos_sim(query_emb, doc_emb)[0].cpu().tolist()
+
+#Combine docs & scores
+doc_score_pairs = list(zip(paragraphs, scores))
+
+#Sort by decreasing score and get only 3 most similar groups
+doc_score_pairs = sorted(doc_score_pairs, key=lambda x: x[1],
+                         reverse=True)[:3]
+
+# Join these similar groups to form the context 
+context = "".join(x[0] for x in doc_score_pairs)
+
+# Perform the querying
+QA_input = {'question': query, 'context': context}
+out = pipe(QA_input)
+st.json(out)