|
|
|
import gradio as gr |
|
import torch |
|
import torch.nn as nn |
|
import numpy as np |
|
import pandas as pd |
|
import spacy |
|
import textstat |
|
from nltk.tokenize import word_tokenize |
|
import nltk |
|
import re |
|
import joblib |
|
from transformers import BertTokenizerFast, BertForSequenceClassification |
|
from sentence_transformers import SentenceTransformer |
|
|
|
|
|
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") |
|
FINETUNE_MODEL_NAME = 'bert-base-uncased' |
|
MAX_LEN_BERT = 128 |
|
print(f"Using device: {DEVICE}") |
|
NLP = spacy.load('en_core_web_sm', disable=['ner']) |
|
SCALER = joblib.load('scaler_mlp_discrete.joblib') |
|
|
|
|
|
class AdvancedMLP(nn.Module): |
|
|
|
def __init__(self, input_dim, num_classes=2): |
|
super(AdvancedMLP, self).__init__() |
|
self.layer_1 = nn.Linear(input_dim, 512) |
|
self.relu1 = nn.ReLU() |
|
self.batchnorm1 = nn.BatchNorm1d(512) |
|
self.dropout1 = nn.Dropout(0.3) |
|
self.layer_2 = nn.Linear(512, 128) |
|
self.relu2 = nn.ReLU() |
|
self.batchnorm2 = nn.BatchNorm1d(128) |
|
self.dropout2 = nn.Dropout(0.3) |
|
self.output_layer = nn.Linear(128, num_classes) |
|
|
|
def forward(self, x): |
|
x = self.layer_1(x); x = self.relu1(x); x = self.batchnorm1(x); x = self.dropout1(x) |
|
x = self.layer_2(x); x = self.relu2(x); x = self.batchnorm2(x); x = self.dropout2(x) |
|
x = self.output_layer(x) |
|
return x |
|
|
|
|
|
print("Loading models and artifacts...") |
|
try: |
|
nltk.download('punkt', quiet=True) |
|
nltk.download('punkt_tab', quiet=True) |
|
|
|
TOKENIZER = BertTokenizerFast.from_pretrained(FINETUNE_MODEL_NAME) |
|
|
|
bert_for_seq_clf = BertForSequenceClassification.from_pretrained(FINETUNE_MODEL_NAME, num_labels=2) |
|
|
|
bert_for_seq_clf.load_state_dict(torch.load("best_bert_finetuned_fold_4.bin", map_location=DEVICE)) |
|
BERT_EMBEDDING_MODEL = bert_for_seq_clf.bert.to(DEVICE).eval() |
|
|
|
INPUT_DIM_MLP = 768 + 19 |
|
MLP_MODEL = AdvancedMLP(input_dim=INPUT_DIM_MLP).to(DEVICE) |
|
MLP_MODEL.load_state_dict(torch.load("best_mlp_combined_features_ZuCo.bin", map_location=DEVICE)) |
|
MLP_MODEL.eval() |
|
|
|
NLP = spacy.load('en_core_web_sm', disable=['ner']) |
|
|
|
|
|
SCALER = joblib.load('scaler_mlp_discrete.joblib') |
|
|
|
print("All models and artifacts loaded successfully.") |
|
|
|
except FileNotFoundError as e: |
|
print(f"ERROR: A required file was not found: {e.name}") |
|
print("Please ensure 'best_bert_finetuned_fold_4.bin', 'best_mlp_combined_features_ZuCo.bin', and 'scaler_mlp_discrete.joblib' are in the same directory.") |
|
exit() |
|
|
|
|
|
def clean_text(text): |
|
text = str(text).lower() |
|
return re.sub(r'\\s+', ' ', text).strip() |
|
|
|
|
|
def get_discrete_features(sentence, nlp_model): |
|
"""Calculates all 19 discrete linguistic features for a single sentence.""" |
|
features = {} |
|
|
|
|
|
features['char_count'] = len(sentence) |
|
words = sentence.split() |
|
features['word_count'] = len(words) |
|
features['avg_word_length'] = features['char_count'] / features['word_count'] if features['word_count'] > 0 else 0 |
|
features['flesch_ease'] = textstat.flesch_reading_ease(sentence) |
|
features['flesch_grade'] = textstat.flesch_kincaid_grade(sentence) |
|
features['gunning_fog'] = textstat.gunning_fog(sentence) |
|
tokens = word_tokenize(sentence) |
|
features['ttr'] = len(set(tokens)) / len(tokens) if tokens else 0 |
|
features['lex_density_proxy'] = sum(1 for w in tokens if len(w) > 6) / len(tokens) if tokens else 0 |
|
|
|
|
|
doc = nlp_model(sentence) |
|
dep_distances = [abs(token.i - token.head.i) for token in doc if token.head is not token] |
|
pos_counts = doc.count_by(spacy.attrs.POS) |
|
|
|
features['num_subord_clauses'] = sum(1 for token in doc if token.dep_ == 'mark') |
|
features['num_conj_clauses'] = sum(1 for token in doc if token.dep_ == 'cc' and token.head.pos_ == 'VERB') |
|
features['avg_dep_dist'] = np.mean(dep_distances) if dep_distances else 0 |
|
features['max_dep_dist'] = np.max(dep_distances) if dep_distances else 0 |
|
features['num_verbs'] = pos_counts.get(spacy.parts_of_speech.VERB, 0) |
|
features['num_nouns'] = pos_counts.get(spacy.parts_of_speech.NOUN, 0) + pos_counts.get(spacy.parts_of_speech.PROPN, 0) |
|
features['num_adjectives'] = pos_counts.get(spacy.parts_of_speech.ADJ, 0) |
|
features['num_adverbs'] = pos_counts.get(spacy.parts_of_speech.ADV, 0) |
|
features['num_prepositions'] = pos_counts.get(spacy.parts_of_speech.ADP, 0) |
|
features['num_conjunctions'] = pos_counts.get(spacy.parts_of_speech.CCONJ, 0) + pos_counts.get(spacy.parts_of_speech.SCONJ, 0) |
|
|
|
feature_order = [ |
|
'char_count', 'word_count', 'avg_word_length', 'ttr', 'lex_density_proxy', |
|
'flesch_ease', 'flesch_grade', 'gunning_fog', 'num_subord_clauses', |
|
'num_conj_clauses', 'avg_dep_dist', 'max_dep_dist', 'num_verbs', |
|
'num_nouns', 'num_adjectives', 'num_adverbs', 'num_prepositions', 'num_conjunctions', |
|
'ollama_llm_rating' |
|
] |
|
features['ollama_llm_rating'] = 3.0 |
|
return np.array([features[k] for k in feature_order]).reshape(1, -1) |
|
|
|
def get_bert_embedding(sentence): |
|
|
|
encoded = TOKENIZER.encode_plus(sentence, add_special_tokens=True, max_length=MAX_LEN_BERT, return_token_type_ids=False, padding='max_length', truncation=True, return_attention_mask=True, return_tensors='pt') |
|
input_ids, attention_mask = encoded['input_ids'].to(DEVICE), encoded['attention_mask'].to(DEVICE) |
|
with torch.no_grad(): |
|
outputs = BERT_EMBEDDING_MODEL(input_ids, attention_mask=attention_mask) |
|
embedding = outputs.last_hidden_state[:, 0, :].cpu().numpy() |
|
return embedding |
|
|
|
|
|
def predict_cognitive_state(sentence): |
|
if not sentence.strip(): |
|
return {"Normal Reading (NR)": 0, "Task-Specific Reading (TSR)": 0} |
|
|
|
cleaned = clean_text(sentence) |
|
|
|
|
|
discrete_features = get_discrete_features(cleaned, NLP) |
|
|
|
scaled_discrete_features = SCALER.transform(discrete_features) |
|
bert_embedding = get_bert_embedding(cleaned) |
|
combined_features = np.concatenate((bert_embedding, scaled_discrete_features), axis=1) |
|
|
|
features_tensor = torch.tensor(combined_features, dtype=torch.float32).to(DEVICE) |
|
with torch.no_grad(): |
|
logits = MLP_MODEL(features_tensor) |
|
probabilities = torch.softmax(logits, dim=1).cpu().numpy()[0] |
|
|
|
labels = ["Normal Reading (NR)", "Task-Specific Reading (TSR)"] |
|
confidences = {label: float(prob) for label, prob in zip(labels, probabilities)} |
|
|
|
return confidences |
|
|
|
|
|
title = "🧠 Cognitive State Analysis from Text" |
|
description = ( |
|
"Enter a sentence to predict its cognitive state. This demo uses a fine-tuned BERT model for semantic " |
|
"embeddings combined with 19 discrete linguistic features. These features are fed into a Multi-Layer Perceptron (MLP) " |
|
"to classify the input as either:\n\n" |
|
"- **Normal Reading (NR):** Casual reading without a specific goal—like reading a story or browsing news.\n" |
|
"- **Task-Specific Reading (TSR):** Purpose-driven reading—such as searching for an answer or following instructions.\n\n" |
|
"The model is trained on text data from the ZuCo dataset, using only linguistic features—no EEG or eye-tracking signals are used." |
|
) |
|
example_list = [ |
|
["Through his son Timothy Bush, Jr., who was also a blacksmith, descended two American Presidents -George H. W. Bush and George W. Bush."], |
|
["He received his bachelor's degree in 1965 and master's degree in political science in 1966 both from the University of Wyoming."], |
|
["What does the abbreviation Ph.D. stand for?"], |
|
["What is the name of the director of the 2003 American film 'The Haunted Mansion'?"], |
|
] |
|
|
|
demo = gr.Interface( |
|
fn=predict_cognitive_state, |
|
inputs=gr.Textbox(lines=3, label="Input Sentence", placeholder="Type a sentence here..."), |
|
outputs=gr.Label(num_top_classes=2, label="Prediction"), |
|
title=title, |
|
description=description, |
|
examples=example_list, |
|
allow_flagging="never" |
|
) |
|
|
|
if __name__ == "__main__": |
|
|
|
demo.launch(debug=True) |