feat: working with negatiive emotion

app.py

@@ -1,18 +1,17 @@
 import streamlit as st
-from transformers import BertForSequenceClassification, BertTokenizerFast, pipeline
+from transformers import BertForSequenceClassification, BertTokenizerFast
+from emotion_utils import predict  # Custom module for prediction
 
 # Load the BERT model and tokenizer
 model_path = "./model/"
 model = BertForSequenceClassification.from_pretrained(model_path)
 tokenizer = BertTokenizerFast.from_pretrained(model_path)
-nlp = pipeline("sentiment-analysis", model=model, tokenizer=tokenizer)
 
 # Function to update sentiment analysis
 def analyze_sentiment(text):
     if text.strip():
-        result = nlp(text)
-        label = result[0]['label']
-        score = result[0]['score']
+        probs, _, label = predict(text, model, tokenizer)
+        score = probs.max().item()  # Get the highest probability score
         return label, score
     else:
         return None, None
@@ -69,7 +68,7 @@ st.markdown("""
 # Title and description
 st.title("🌟 G-Bert: Emotion Analysis")
 st.markdown("""
-    G-Bert is a bangla sentiment analysis tool that uses a pre-trained BERT model to analyze the emotion of any bengali or religious (gita) text.
+    G-Bert is a Bangla sentiment analysis tool that uses a pre-trained BERT model to analyze the emotion of any Bengali or religious (Gita) text.
     It can detect emotions like Anger, Astonished, Optimistic, and Sadness with a confidence score.
 """)
 # Text input

emotion_utils.py

@@ -0,0 +1,80 @@
+import torch
+
+def load_words_from_file(file_path):
+    """
+    Load words from a text file and return them as a list.
+    Each word should be on a separate line in the text file.
+    """
+    with open(file_path, 'r', encoding='utf-8') as file:
+        words = file.read().splitlines()
+    return words
+
+def preprocess_with_negation_v2(text):
+    from emotion_utils import load_words_from_file
+    negation_words = load_words_from_file('./model/stopwords/negation_words.txt')
+    emotion_words = load_words_from_file('./model/stopwords/emotion_words.txt')
+
+    # Tokenize the sentence into words
+    words = text.split()
+
+    modified_words = words[:]  # Create a copy to modify
+
+    # Iterate through all words to detect negation-emotion pairs
+    for i, word in enumerate(words):
+        if word in negation_words:
+            # Check the previous 3 words for an emotion word
+            for j in range(1, 4):
+                if i - j >= 0 and words[i - j] in emotion_words:
+                    # Mark the detected emotion with a negation label
+                    modified_words[i - j] = f"{words[i - j]} (Negative context)"
+                    break
+
+    # Reconstruct the text
+    return " ".join(modified_words)
+
+
+def predict(text, model, tokenizer):
+    from emotion_utils import preprocess_with_negation_v2
+    """
+    Predict the sentiment for a given text with advanced negation handling.
+    """
+    # Ensure the model is on the correct device
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model = model.to(device)
+
+    # Preprocess the text for advanced negation handling
+    processed_text = preprocess_with_negation_v2(text)
+    # print(processed_text)
+
+    # Tokenize the text
+    inputs = tokenizer(
+        processed_text,
+        padding=True,
+        truncation=True,
+        max_length=512,
+        return_tensors="pt"
+    ).to(device)
+
+    # Perform inference
+    with torch.no_grad():
+        outputs = model(**inputs)
+
+    # Compute probabilities
+    probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
+
+    # Get the class with the highest probability
+    pred_label_idx = probs.argmax(dim=-1).item()
+
+    # Map the index to the label
+    pred_label = model.config.id2label[pred_label_idx]
+
+    # Adjust prediction for negation context
+    negation_map = {
+        "Sadness": "Optimistic",
+        "Optimistic": "Sadness",
+        "Anger": "Optimistic",
+    }
+    if "(Negative context)" in processed_text:
+        pred_label = negation_map.get(pred_label, pred_label)
+
+    return probs, pred_label_idx, pred_label

model/README.md

@@ -1,13 +0,0 @@
----
-title: Gita Bengali Sentiment Predicition
-emoji: 🦀
-colorFrom: green
-colorTo: blue
-sdk: streamlit
-sdk_version: 1.40.1
-app_file: app.py
-pinned: false
-license: apache-2.0
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

model/stopwords/emotion_words.txt

@@ -0,0 +1,74 @@
+ভালবাসা
+ভালোবাসি
+ভালোবাসো
+ভালোবাসে
+ভালো
+খারাপ
+দুঃখ
+দুঃখী
+আনন্দ
+রাগ
+হতাশ
+হতাশা
+অপমান
+অপমানিত
+হেলা
+অবহেলা
+অবহেলিত
+হিংসা
+শান্ত
+শান্তি
+প্রেম
+ভয়
+ঘৃণা
+কষ্ট
+বিজয়
+সুখ
+সুখী
+হাসি
+যত্ন
+মমতা
+স্বপ্ন
+খুশী
+খুশি
+সন্তুষ্ট
+অসন্তুষ্ট
+আশা
+আশাবাদ
+প্রেরণা
+উন্নতি
+সাফল্য
+সম্ভাবনা
+বিজয়
+নির্ভরতা
+সাহস
+আত্মবিশ্বাস
+আনন্দদায়ক
+পরিতোষ
+অর্জন
+সৌভাগ্য
+খুশিময়
+উদ্যম
+উদ্ভাবনী
+আনন্দময়
+ভরসা
+নিরাশা
+হতাশাজনক
+কষ্টকর
+দুঃখজনক
+অভিমান
+ব্যথা
+বিচ্ছেদ
+নীরবতা
+বেদনা
+শূন্যতা
+হাহাকার
+বিষাদ
+অশ্রু
+পীড়া
+ক্ষতি
+বিপর্যয়
+অনুশোচনা
+দুঃসহ
+অস্বস্তি
+ভারাক্রান্ত

model/stopwords/negation_words.txt

@@ -0,0 +1,4 @@
+না
+নেই
+নয়
+নাই