Update app.py

app.py

@@ -8,11 +8,14 @@ tokenizer = AutoTokenizer.from_pretrained(model_name)
 model = AutoModelForSequenceClassification.from_pretrained(model_name)
 model.eval()  # Set the model to evaluation mode
 
-# Function to get relevance score and relevant excerpt while preserving order
+# Threshold for attention relevance
+THRESHOLD = 0.02  # Adjust as needed based on observations
+
+# Function to get relevance score and relevant excerpt with bolded tokens
 def get_relevance_score_and_excerpt(query, paragraph):
     if not query.strip() or not paragraph.strip():
         return "Please provide both a query and a document paragraph.", ""
-    
+
     # Tokenize the input
     inputs = tokenizer(query, paragraph, return_tensors="pt", truncation=True, padding=True)
     
@@ -23,17 +26,17 @@ def get_relevance_score_and_excerpt(query, paragraph):
     logit = output.logits.squeeze().item()
     relevance_score = torch.sigmoid(torch.tensor(logit)).item()
 
-    # Extract attention scores (last attention layer)
+    # Extract attention scores (last layer)
     attention = output.attentions[-1]  # Shape: (batch_size, num_heads, seq_len, seq_len)
 
-    # Average attention across heads and batch dimension
+    # Average across heads and batch dimension
     attention_scores = attention.mean(dim=1).mean(dim=0)  # Shape: (seq_len, seq_len)
 
-    # Get tokenized query and paragraph separately
+    # Tokenize query and paragraph separately
     query_tokens = tokenizer.tokenize(query)
     paragraph_tokens = tokenizer.tokenize(paragraph)
 
-    query_len = len(query_tokens) + 2  # +2 for special tokens [CLS] and [SEP]
+    query_len = len(query_tokens) + 2  # +2 for [CLS] and first [SEP]
     para_start_idx = query_len
     para_end_idx = len(inputs["input_ids"][0]) - 1  # Ignore final [SEP] token
 
@@ -41,22 +44,27 @@ def get_relevance_score_and_excerpt(query, paragraph):
     if para_end_idx <= para_start_idx:
         return round(relevance_score, 4), "No relevant tokens extracted."
 
+    # Extract paragraph attention scores
     para_attention_scores = attention_scores[para_start_idx:para_end_idx, para_start_idx:para_end_idx].mean(dim=0)
 
     if para_attention_scores.numel() == 0:
         return round(relevance_score, 4), "No relevant tokens extracted."
 
-    # Get indices of top-k attended tokens while preserving order
-    top_k = min(5, para_attention_scores.size(0))  # Ensure top-k does not exceed available tokens
-    top_indices = para_attention_scores.topk(top_k).indices.sort().values  # Sort indices to preserve order
+    # Filter tokens based on threshold and preserve order
+    relevant_indices = (para_attention_scores > THRESHOLD).nonzero(as_tuple=True)[0].tolist()
 
-    # Extract highlighted tokens from the paragraph
-    highlighted_tokens = [paragraph_tokens[i] for i in top_indices.tolist()]
+    # Reconstruct paragraph with bolded relevant tokens
+    highlighted_text = ""
+    for idx, token in enumerate(paragraph_tokens):
+        if idx in relevant_indices:
+            highlighted_text += f"**{token}** "
+        else:
+            highlighted_text += f"{token} "
 
-    # Convert tokens back to a readable string
-    excerpt = tokenizer.convert_tokens_to_string(highlighted_tokens)
+    # Convert tokens to readable format (handling special characters)
+    highlighted_text = tokenizer.convert_tokens_to_string(highlighted_text.split())
 
-    return round(relevance_score, 4), excerpt
+    return round(relevance_score, 4), highlighted_text
 
 # Define Gradio interface
 interface = gr.Interface(
@@ -67,10 +75,10 @@ interface = gr.Interface(
     ],
     outputs=[
         gr.Textbox(label="Relevance Score"),
-        gr.Textbox(label="Most Relevant Excerpt")
+        gr.HTML(label="Highlighted Document Paragraph")
     ],
-    title="Cross-Encoder Relevance Scoring with Ordered Excerpt Extraction",
-    description="Enter a query and a document paragraph to get a relevance score and a relevant excerpt in original order.",
+    title="Cross-Encoder Relevance Scoring with Highlighted Excerpt",
+    description="Enter a query and a document paragraph to get a relevance score and see relevant tokens in bold.",
     allow_flagging="never",
     live=True
 )