Update app.py
Browse files
app.py
CHANGED
@@ -9,71 +9,89 @@ model = AutoModelForSequenceClassification.from_pretrained(model_name)
|
|
9 |
model.eval()
|
10 |
|
11 |
# Function to compute relevance score (in logits) and dynamically adjust threshold
|
12 |
-
def get_relevance_score_and_excerpt(query,
|
13 |
-
if not query.strip() or not
|
14 |
-
return "Please provide both a query and
|
15 |
|
16 |
-
|
17 |
-
inputs = tokenizer(query, paragraph, return_tensors="pt", truncation=True, padding=True)
|
18 |
|
19 |
-
|
20 |
-
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
|
26 |
-
|
27 |
-
|
28 |
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
|
33 |
-
|
34 |
-
|
35 |
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
|
40 |
-
|
41 |
-
|
42 |
|
43 |
-
|
44 |
|
45 |
-
|
46 |
-
|
47 |
|
48 |
-
|
49 |
-
|
50 |
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
|
59 |
-
|
60 |
|
61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
|
63 |
-
# Define Gradio interface with a slider for threshold adjustment
|
64 |
interface = gr.Interface(
|
65 |
fn=get_relevance_score_and_excerpt,
|
66 |
inputs=[
|
67 |
gr.Textbox(label="Query", placeholder="Enter your search query..."),
|
68 |
-
gr.Textbox(label="Document Paragraph", placeholder="Enter a paragraph to match..."),
|
|
|
|
|
69 |
gr.Slider(minimum=0.02, maximum=0.5, value=0.1, step=0.01, label="Attention Threshold")
|
70 |
],
|
71 |
outputs=[
|
72 |
-
gr.Textbox(label="Relevance
|
73 |
-
gr.HTML(label="Highlighted Document
|
74 |
],
|
75 |
-
title="Cross-Encoder Attention Highlighting",
|
76 |
-
description="Adjust the attention threshold to control token highlighting sensitivity.",
|
77 |
allow_flagging="never",
|
78 |
live=True
|
79 |
)
|
|
|
9 |
model.eval()
|
10 |
|
11 |
# Function to compute relevance score (in logits) and dynamically adjust threshold
|
12 |
+
def get_relevance_score_and_excerpt(query, *paragraphs, threshold_weight):
|
13 |
+
if not query.strip() or not any(p.strip() for p in paragraphs):
|
14 |
+
return "Please provide both a query and at least one document paragraph.", ""
|
15 |
|
16 |
+
ranked_paragraphs = []
|
|
|
17 |
|
18 |
+
# Process each paragraph and calculate its logits and highlighted text
|
19 |
+
for paragraph in paragraphs:
|
20 |
+
# Tokenize the input
|
21 |
+
inputs = tokenizer(query, paragraph, return_tensors="pt", truncation=True, padding=True)
|
22 |
+
|
23 |
+
with torch.no_grad():
|
24 |
+
output = model(**inputs, output_attentions=True)
|
25 |
|
26 |
+
# Extract logits (no sigmoid applied)
|
27 |
+
logit = output.logits.squeeze().item()
|
28 |
+
base_relevance_score = logit # Relevance score in logits
|
29 |
|
30 |
+
# Dynamically adjust the attention threshold based on user weight
|
31 |
+
dynamic_threshold = max(0.02, threshold_weight)
|
32 |
|
33 |
+
# Extract attention scores (last layer)
|
34 |
+
attention = output.attentions[-1]
|
35 |
+
attention_scores = attention.mean(dim=1).mean(dim=0)
|
36 |
|
37 |
+
query_tokens = tokenizer.tokenize(query)
|
38 |
+
paragraph_tokens = tokenizer.tokenize(paragraph)
|
39 |
|
40 |
+
query_len = len(query_tokens) + 2 # +2 for special tokens [CLS] and first [SEP]
|
41 |
+
para_start_idx = query_len
|
42 |
+
para_end_idx = len(inputs["input_ids"][0]) - 1
|
43 |
|
44 |
+
if para_end_idx <= para_start_idx:
|
45 |
+
continue
|
46 |
|
47 |
+
para_attention_scores = attention_scores[para_start_idx:para_end_idx, para_start_idx:para_end_idx].mean(dim=0)
|
48 |
|
49 |
+
if para_attention_scores.numel() == 0:
|
50 |
+
continue
|
51 |
|
52 |
+
# Get indices of relevant tokens above dynamic threshold
|
53 |
+
relevant_indices = (para_attention_scores > dynamic_threshold).nonzero(as_tuple=True)[0].tolist()
|
54 |
|
55 |
+
# Reconstruct paragraph with bolded relevant tokens using HTML tags
|
56 |
+
highlighted_text = ""
|
57 |
+
for idx, token in enumerate(paragraph_tokens):
|
58 |
+
if idx in relevant_indices:
|
59 |
+
highlighted_text += f"<b>{token}</b> "
|
60 |
+
else:
|
61 |
+
highlighted_text += f"{token} "
|
62 |
|
63 |
+
highlighted_text = tokenizer.convert_tokens_to_string(highlighted_text.split())
|
64 |
|
65 |
+
ranked_paragraphs.append({
|
66 |
+
"logit": logit,
|
67 |
+
"highlighted_text": highlighted_text
|
68 |
+
})
|
69 |
+
|
70 |
+
# Sort paragraphs by logit (descending)
|
71 |
+
ranked_paragraphs.sort(key=lambda x: x["logit"], reverse=True)
|
72 |
+
|
73 |
+
# Prepare output
|
74 |
+
relevance_scores = [round(p["logit"], 4) for p in ranked_paragraphs]
|
75 |
+
highlighted_texts = [p["highlighted_text"] for p in ranked_paragraphs]
|
76 |
+
|
77 |
+
return "\n".join([f"Relevance Score: {score}" for score in relevance_scores]), "\n\n".join(highlighted_texts)
|
78 |
|
79 |
+
# Define Gradio interface with a slider for threshold adjustment and ability to add multiple paragraphs
|
80 |
interface = gr.Interface(
|
81 |
fn=get_relevance_score_and_excerpt,
|
82 |
inputs=[
|
83 |
gr.Textbox(label="Query", placeholder="Enter your search query..."),
|
84 |
+
gr.Textbox(label="Document Paragraph 1", placeholder="Enter a paragraph to match...", lines=4),
|
85 |
+
gr.Textbox(label="Document Paragraph 2 (optional)", placeholder="Enter another paragraph...", lines=4),
|
86 |
+
gr.Textbox(label="Document Paragraph 3 (optional)", placeholder="Enter another paragraph...", lines=4),
|
87 |
gr.Slider(minimum=0.02, maximum=0.5, value=0.1, step=0.01, label="Attention Threshold")
|
88 |
],
|
89 |
outputs=[
|
90 |
+
gr.Textbox(label="Relevance Scores (Logits)"),
|
91 |
+
gr.HTML(label="Highlighted Document Paragraphs")
|
92 |
],
|
93 |
+
title="Cross-Encoder Attention Highlighting with Reranking",
|
94 |
+
description="Adjust the attention threshold to control token highlighting sensitivity. Multiple paragraphs can be added and reranked based on their logits.",
|
95 |
allow_flagging="never",
|
96 |
live=True
|
97 |
)
|