I-Comprehend / app.py
miiiciiii's picture
Create app.py
54e92ac verified
import os
import sys
import warnings
# Suppress specific warnings
warnings.filterwarnings("ignore", message="This sequence already has </s>.")
# Append path for module imports
scripts_path = os.path.abspath(os.path.join('..', 'scripts'))
sys.path.append(scripts_path)
# Standard library imports
import random
import string
# Third-party imports
import json
import numpy as np
import pandas as pd
import torch
import nltk
from dateutil.parser import parse
from nltk.stem import PorterStemmer
from nltk.corpus import stopwords, wordnet as wn
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
# from textdistance import levenshtein
from rapidfuzz import fuzz
from rapidfuzz.distance import Levenshtein as levenshtein
from sense2vec import Sense2Vec
from transformers import T5ForConditionalGeneration, T5Tokenizer
from sentence_transformers import SentenceTransformer
# Download necessary NLTK data
nltk.download('omw-1.4')
nltk.download('stopwords')
nltk.download('punkt')
nltk.download('brown')
nltk.download('wordnet')
from typing import List, Dict
import re
# Initialize models
t5ag_model = T5ForConditionalGeneration.from_pretrained("miiiciiii/I-Comprehend_ag")
t5ag_tokenizer = T5Tokenizer.from_pretrained("miiiciiii/I-Comprehend_ag", legacy=False)
t5qg_model = T5ForConditionalGeneration.from_pretrained("miiiciiii/I-Comprehend_qg")
t5qg_tokenizer = T5Tokenizer.from_pretrained("miiiciiii/I-Comprehend_qg", legacy=False)
s2v = Sense2Vec().from_disk(S2V_MODEL_PATH)
sentence_transformer_model = SentenceTransformer("sentence-transformers/LaBSE")
def answer_question(question, context):
"""Generate an answer for a given question and context."""
input_text = f"question: {question} context: {context}"
input_ids = t5ag_tokenizer.encode(input_text, return_tensors="pt", max_length=512, truncation=True)
with torch.no_grad():
output = t5ag_model.generate(input_ids, max_length=512, num_return_sequences=1, max_new_tokens=200)
return t5ag_tokenizer.decode(output[0], skip_special_tokens=True).capitalize()
def get_passage(passage):
"""Generate a random context from the dataset."""
return passage.sample(n=1)['context'].values[0]
def get_question(context, answer, model, tokenizer):
"""Generate a question for the given answer and context."""
answer_span = context.replace(answer, f"<hl>{answer}<hl>", 1) + "</s>"
inputs = tokenizer(answer_span, return_tensors="pt")
question = model.generate(input_ids=inputs.input_ids, max_length=50)[0]
return tokenizer.decode(question, skip_special_tokens=True)
def get_keywords(passage):
"""Extract keywords using TF-IDF."""
try:
vectorizer = TfidfVectorizer(stop_words='english')
tfidf_matrix = vectorizer.fit_transform([passage])
feature_names = vectorizer.get_feature_names_out()
tfidf_scores = tfidf_matrix.toarray().flatten() # type: ignore
word_scores = dict(zip(feature_names, tfidf_scores))
sorted_words = sorted(word_scores.items(), key=lambda x: x[1], reverse=True)
keywords = [word for word, score in sorted_words]
return keywords
except Exception as e:
print(f"Error extracting keywords: {e}")
return []
def classify_question_type(question: str) -> str:
"""
Classify the type of question as literal, evaluative, or inferential.
Parameters:
question (str): The question to classify.
Returns:
str: The type of the question ('literal', 'evaluative', or 'inferential').
"""
# Define keywords or patterns for each question type
literal_keywords = [
'what', 'when', 'where', 'who', 'how many', 'how much',
'which', 'name', 'list', 'identify', 'define', 'describe',
'state', 'mention'
]
evaluative_keywords = [
'evaluate', 'justify', 'explain why', 'assess', 'critique',
'discuss', 'judge', 'opinion', 'argue', 'agree or disagree',
'defend', 'support your answer', 'weigh the pros and cons',
'compare', 'contrast'
]
inferential_keywords = [
'why', 'how', 'what if', 'predict', 'suggest', 'imply',
'conclude', 'infer', 'reason', 'what might', 'what could',
'what would happen if', 'speculate', 'deduce', 'interpret',
'hypothesize', 'assume'
]
question_lower = question.lower()
# Check for literal question keywords
if any(keyword in question_lower for keyword in literal_keywords):
return 'literal'
# Check for evaluative question keywords
if any(keyword in question_lower for keyword in evaluative_keywords):
return 'evaluative'
# Check for inferential question keywords
if any(keyword in question_lower for keyword in inferential_keywords):
return 'inferential'
# Default to 'unknown' if no pattern matches
return 'unknown'
def filter_same_sense_words(original, wordlist):
"""Filter words that have the same sense as the original word."""
try:
base_sense = original.split('|')[1] # Ensure there is a sense part
except IndexError:
print(f"Warning: The original phrase '{original}' does not have a sense part.")
return wordlist # Return all words if the sense part is missing
return [word[0].split('|')[0].replace("_", " ").title().strip() for word in wordlist if word[0].split('|')[1] == base_sense]
def extract_similar_keywords(input_phrases, topn=5):
"""Call get_distractors and extract only the similar_keywords values."""
distractors_result = get_distractors(input_phrases, topn)
similar_keywords_list = [result["similar_keywords"] for result in distractors_result]
return similar_keywords_list
def get_max_similarity_score(wordlist, word):
"""Get the maximum similarity score between the word and a list of words."""
return max(levenshtein.normalized_similarity(word.lower(), each.lower()) for each in wordlist)
def mmr(doc_embedding, word_embeddings, words, top_n, lambda_param):
"""Maximal Marginal Relevance (MMR) for keyword extraction."""
try:
word_doc_similarity = cosine_similarity(word_embeddings, doc_embedding)
word_similarity = cosine_similarity(word_embeddings)
keywords_idx = [np.argmax(word_doc_similarity)]
candidates_idx = [i for i in range(len(words)) if i != keywords_idx[0]]
for _ in range(top_n - 1):
candidate_similarities = word_doc_similarity[candidates_idx, :]
target_similarities = np.max(word_similarity[candidates_idx][:, keywords_idx], axis=1)
mmr = (lambda_param * candidate_similarities) - ((1 - lambda_param) * target_similarities.reshape(-1, 1))
mmr_idx = candidates_idx[np.argmax(mmr)]
keywords_idx.append(mmr_idx)
candidates_idx.remove(mmr_idx)
return [words[idx] for idx in keywords_idx]
except Exception as e:
print(f"Error in MMR: {e}")
return []
def format_phrase(phrase):
"""Format phrases by replacing spaces with underscores and adding default |n."""
return phrase.replace(" ", "_") + "|n"
def is_valid_distractor(distractor, input_phrase):
"""Check if the distractor is valid by ensuring it's alphabetic and relevant."""
if not re.match(r'^[a-zA-Z\s]+$', distractor):
return False
word_count = len(distractor.split())
if word_count < 1 or word_count > 4:
return False
return True
def filter_distractors(input_phrase, similar_keywords, topn):
"""Filter distractors to ensure they match word count, aren't identical to the input,
and aren't too similar to each other or the input (e.g., stem similarity)."""
word_count = len(input_phrase.split())
filtered_keywords = []
stemmer = PorterStemmer()
input_stem = stemmer.stem(input_phrase.lower())
for keyword in similar_keywords:
keyword_stem = stemmer.stem(keyword.lower())
if (len(keyword.split()) == word_count and
keyword.lower() != input_phrase.lower() and
keyword_stem != input_stem and
is_valid_distractor(keyword, input_phrase)):
if all(stemmer.stem(kw.lower()) != keyword_stem for kw in filtered_keywords):
filtered_keywords.append(keyword)
if len(filtered_keywords) == topn:
break
return filtered_keywords
def get_distractors(input_phrases, topn=5):
"""Find similar keywords for a list of input phrases using Sense2Vec and WordNet."""
result_list = []
for phrase in input_phrases:
formatted_phrase = format_phrase(phrase)
# Check if the phrase exists in the Sense2Vec model
if formatted_phrase in s2v:
# Get similar phrases from Sense2Vec
similar_phrases = s2v.most_similar(formatted_phrase, n=topn * 2) # Get more to filter later
similar_keywords = [item[0].split("|")[0].replace("_", " ") for item in similar_phrases]
else:
# List similar keys that might exist in the model for exploration
print(f"'{formatted_phrase}' not found in the model. Exploring similar available keys...")
available_keys = [key for key in s2v.keys() if phrase.split()[0] in key or phrase.split()[-1] in key]
print(f"Available keys related to '{phrase}': {available_keys}")
# Use WordNet to find synonyms if available keys are empty
if not available_keys:
print(f"No close match in the model for '{phrase}'. Trying WordNet for synonyms...")
synonyms = set()
for syn in wn.synsets(phrase.replace(" ", "_")):
for lemma in syn.lemmas():
synonyms.add(lemma.name().replace("_", " "))
similar_keywords = list(synonyms)[:topn * 2] if synonyms else ["No match found"]
else:
# Provide available keys as similar suggestions
similar_keywords = [key.split("|")[0].replace("_", " ") for key in available_keys[:topn * 2]]
# Filter distractors to match word count, avoid identical or stem-similar words, and check format
final_distractors = filter_distractors(phrase, similar_keywords, topn)
# Further filter out words with the same sense
final_distractors = filter_same_sense_words(phrase, final_distractors)
result_list.append({
"phrase": phrase,
"similar_keywords": final_distractors
})
return result_list
def get_mca_questions(context, qg_model, qg_tokenizer, sentence_transformer_model, num_questions=5, max_attempts=2) -> List[Dict]:
"""
Generate multiple-choice questions for a given context.
Parameters:
context (str): The context from which questions are generated.
qg_model (T5ForConditionalGeneration): The question generation model.
qg_tokenizer (T5Tokenizer): The tokenizer for the question generation model.
s2v (Sense2Vec): The Sense2Vec model for finding similar words.
sentence_transformer_model (SentenceTransformer): The sentence transformer model for embeddings.
num_questions (int): The number of questions to generate.
max_attempts (int): The maximum number of attempts to generate questions.
Returns:
list: A list of dictionaries with questions and their corresponding distractors.
"""
output_list = []
imp_keywords = get_keywords(context)
print(f"[DEBUG] Length: {len(imp_keywords)}, Extracted keywords: {imp_keywords}")
generated_questions = set()
generated_answers = set()
attempts = 0
while len(output_list) < num_questions and attempts < max_attempts:
attempts += 1
for keyword in imp_keywords:
if len(output_list) >= num_questions:
break
question = get_question(context, keyword, qg_model, qg_tokenizer)
print(f"[DEBUG] Generated question: '{question}' for keyword: '{keyword}'")
# Encode the new question
new_question_embedding = sentence_transformer_model.encode(question, convert_to_tensor=True)
is_similar = False
# Check similarity with existing questions
for generated_q in generated_questions:
existing_question_embedding = sentence_transformer_model.encode(generated_q, convert_to_tensor=True)
similarity = cosine_similarity(new_question_embedding.unsqueeze(0), existing_question_embedding.unsqueeze(0))[0][0]
if similarity > 0.8:
is_similar = True
print(f"[DEBUG] Question '{question}' is too similar to an existing question, skipping.")
break
if is_similar:
continue
# Generate and check answer
t5_answer = answer_question(question, context)
print(f"[DEBUG] Generated answer: '{t5_answer}' for question: '{question}'")
# Skip answers longer than 3 words
if len(t5_answer.split()) > 3:
print(f"[DEBUG] Answer '{t5_answer}' is too long, skipping.")
continue
if t5_answer in generated_answers:
print(f"[DEBUG] Answer '{t5_answer}' has already been generated, skipping question.")
continue
generated_questions.add(question)
generated_answers.add(t5_answer)
# Generate distractors
distractors = extract_similar_keywords([t5_answer], topn=5)[0]
print(f"list of distractors : {distractors}")
print(f"length of distractors {len(distractors)}")
print(f"type : {type(distractors)}")
# Remove any distractor that is the same as the correct answer
distractors = [d for d in distractors if d.lower() != t5_answer.lower()]
print(f"Filtered distractors (without answer): {distractors}")
# Ensure there are exactly 3 distractors
if len(distractors) < 3:
# Fill with random keywords from the imp_keywords list until we have 3 distractors
while len(distractors) < 3:
random_keyword = random.choice(imp_keywords)
# Ensure the random keyword isn't the same as the answer or already a distractor
if random_keyword.lower() != t5_answer.lower() and random_keyword not in distractors:
distractors.append(random_keyword)
# Limit to 3 distractors
distractors = distractors[:3]
print(f"[DEBUG] Final distractors: {distractors} for question: '{question}'")
choices = distractors + [t5_answer]
choices = [item.title() for item in choices]
random.shuffle(choices)
print(f"[DEBUG] Options: {choices} for answer: '{t5_answer}'")
# Classify question type
question_type = classify_question_type(question)
output_list.append({
'answer': t5_answer,
'answer_length': len(t5_answer),
'choices': choices,
'passage': context,
'passage_length': len(context),
'question': question,
'question_length': len(question),
'question_type': question_type
})
print(f"[DEBUG] Generated {len(output_list)} questions so far after {attempts} attempts")
return output_list[:num_questions]