Update app.py
Browse files
app.py
CHANGED
@@ -2,23 +2,23 @@ import os
|
|
2 |
import numpy as np
|
3 |
import pandas as pd
|
4 |
import streamlit as st
|
5 |
-
from
|
6 |
-
|
7 |
-
AutoTokenizer,
|
8 |
-
AutoModelForSequenceClassification,
|
9 |
-
AutoModelForSeq2SeqLM
|
10 |
-
)
|
11 |
from keybert import KeyBERT
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
-
# ─── Sentiment & Keyword Models ─────────────────────────────────────────────
|
14 |
@st.cache_resource
|
15 |
def load_sentiment_pipeline():
|
16 |
model_name = "mayf/amazon_reviews_bert_ft"
|
17 |
tok = AutoTokenizer.from_pretrained(model_name, use_auth_token=True)
|
18 |
-
mdl = AutoModelForSequenceClassification.from_pretrained(
|
19 |
-
model_name,
|
20 |
-
use_auth_token=True
|
21 |
-
)
|
22 |
return pipeline(
|
23 |
"sentiment-analysis",
|
24 |
model=mdl,
|
@@ -30,26 +30,6 @@ def load_sentiment_pipeline():
|
|
30 |
def load_keybert_model():
|
31 |
return KeyBERT(model="all-MiniLM-L6-v2")
|
32 |
|
33 |
-
# ─── BlenderBot Response Components ─────────────────────────────────────────
|
34 |
-
@st.cache_resource
|
35 |
-
def load_response_components():
|
36 |
-
# Load tokenizer and model directly for manual generation with truncation
|
37 |
-
tok = AutoTokenizer.from_pretrained(
|
38 |
-
"facebook/blenderbot-400M-distill",
|
39 |
-
use_fast=True
|
40 |
-
)
|
41 |
-
mdl = AutoModelForSeq2SeqLM.from_pretrained("facebook/blenderbot-400M-distill")
|
42 |
-
return tok, mdl
|
43 |
-
:
|
44 |
-
# Use BlenderBot 400M Distill for text generation
|
45 |
-
return pipeline(
|
46 |
-
"text2text-generation",
|
47 |
-
model="facebook/blenderbot-400M-distill",
|
48 |
-
tokenizer="facebook/blenderbot-400M-distill",
|
49 |
-
max_new_tokens=150,
|
50 |
-
do_sample=False
|
51 |
-
)
|
52 |
-
|
53 |
LABEL_MAP = {
|
54 |
"LABEL_0": "Very Negative",
|
55 |
"LABEL_1": "Negative",
|
@@ -69,22 +49,33 @@ def main():
|
|
69 |
st.warning("Please enter a review to analyze.")
|
70 |
return
|
71 |
|
72 |
-
#
|
73 |
-
|
|
|
|
|
|
|
74 |
sentiment_pipeline = load_sentiment_pipeline()
|
|
|
|
|
|
|
|
|
|
|
75 |
raw_scores = sentiment_pipeline(review)[0]
|
|
|
76 |
sentiment_results = {LABEL_MAP[item['label']]: float(item['score']) for item in raw_scores}
|
|
|
77 |
|
78 |
-
#
|
79 |
-
|
80 |
keywords = kw_model.extract_keywords(
|
81 |
review,
|
82 |
keyphrase_ngram_range=(1, 2),
|
83 |
stop_words="english",
|
84 |
top_n=3
|
85 |
)
|
|
|
86 |
|
87 |
-
# Display
|
88 |
col1, col2 = st.columns(2)
|
89 |
with col1:
|
90 |
st.subheader("Sentiment Scores")
|
@@ -94,51 +85,44 @@ def main():
|
|
94 |
for kw, score in keywords:
|
95 |
st.write(f"• {kw} ({score:.4f})")
|
96 |
|
97 |
-
# Bar
|
98 |
-
|
99 |
-
|
100 |
-
orient='index',
|
101 |
-
columns=['score']
|
102 |
-
)
|
103 |
df_scores.index.name = 'Sentiment'
|
104 |
st.bar_chart(df_scores)
|
|
|
105 |
|
106 |
-
# Highlight
|
107 |
max_label, max_score = max(sentiment_results.items(), key=lambda x: x[1])
|
108 |
st.markdown(f"**Highest Sentiment:** **{max_label}** ({max_score:.4f})")
|
109 |
-
# ────────────────────────────────────────────────────────────────────
|
110 |
-
|
111 |
-
# Generate appropriate reply using manual tokenization & generation
|
112 |
-
tok, mdl = load_response_components()
|
113 |
-
if max_label in ["Positive", "Very Positive"]:
|
114 |
-
prompt_text = (
|
115 |
-
f"You are a friendly customer success representative. The customer said: \"{review}\". "
|
116 |
-
"Write two sentences to express gratitude and highlight their positive experience."
|
117 |
-
)
|
118 |
-
else:
|
119 |
-
prompt_text = (
|
120 |
-
f"You are a helpful customer support specialist. The customer said: \"{review}\". "
|
121 |
-
f"Identified issues: {', '.join([kw for kw, _ in keywords])}. "
|
122 |
-
"First, ask 1-2 clarifying questions to understand their situation. "
|
123 |
-
"Then provide two concrete suggestions or next steps to address these issues."
|
124 |
-
)
|
125 |
-
# Tokenize with truncation to avoid out-of-range embeddings
|
126 |
-
inputs = tok(
|
127 |
-
prompt_text,
|
128 |
-
return_tensors="pt",
|
129 |
-
truncation=True,
|
130 |
-
max_length=tok.model_max_length
|
131 |
-
)
|
132 |
-
outputs = mdl.generate(
|
133 |
-
**inputs,
|
134 |
-
max_new_tokens=150,
|
135 |
-
do_sample=False
|
136 |
-
)
|
137 |
-
reply = tok.decode(outputs[0], skip_special_tokens=True).strip()
|
138 |
|
139 |
-
|
140 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
141 |
|
|
|
|
|
|
|
142 |
|
143 |
-
if __name__ ==
|
144 |
main()
|
|
|
2 |
import numpy as np
|
3 |
import pandas as pd
|
4 |
import streamlit as st
|
5 |
+
from huggingface_hub import login
|
6 |
+
from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification
|
|
|
|
|
|
|
|
|
7 |
from keybert import KeyBERT
|
8 |
+
from openai import AzureOpenAI # new
|
9 |
+
|
10 |
+
# ─── Azure OpenAI Client ─────────────────────────────────────────────────────
|
11 |
+
openai_client = AzureOpenAI(
|
12 |
+
api_key = "fbca46bfd8814334be46a2e5c323904c", # use your key here
|
13 |
+
api_version = "2023-05-15", # apparently HKUST uses a deprecated version
|
14 |
+
azure_endpoint = "https://hkust.azure-api.net" # per HKUST instructions
|
15 |
+
)
|
16 |
|
|
|
17 |
@st.cache_resource
|
18 |
def load_sentiment_pipeline():
|
19 |
model_name = "mayf/amazon_reviews_bert_ft"
|
20 |
tok = AutoTokenizer.from_pretrained(model_name, use_auth_token=True)
|
21 |
+
mdl = AutoModelForSequenceClassification.from_pretrained(model_name, use_auth_token=True)
|
|
|
|
|
|
|
22 |
return pipeline(
|
23 |
"sentiment-analysis",
|
24 |
model=mdl,
|
|
|
30 |
def load_keybert_model():
|
31 |
return KeyBERT(model="all-MiniLM-L6-v2")
|
32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
LABEL_MAP = {
|
34 |
"LABEL_0": "Very Negative",
|
35 |
"LABEL_1": "Negative",
|
|
|
49 |
st.warning("Please enter a review to analyze.")
|
50 |
return
|
51 |
|
52 |
+
# Initialize progress bar
|
53 |
+
progress = st.progress(0)
|
54 |
+
|
55 |
+
# Load models
|
56 |
+
progress.text("Loading models...")
|
57 |
sentiment_pipeline = load_sentiment_pipeline()
|
58 |
+
kw_model = load_keybert_model()
|
59 |
+
progress.progress(20)
|
60 |
+
|
61 |
+
# Run sentiment analysis
|
62 |
+
progress.text("Analyzing sentiment...")
|
63 |
raw_scores = sentiment_pipeline(review)[0]
|
64 |
+
# Map labels
|
65 |
sentiment_results = {LABEL_MAP[item['label']]: float(item['score']) for item in raw_scores}
|
66 |
+
progress.progress(40)
|
67 |
|
68 |
+
# Extract keywords
|
69 |
+
progress.text("Extracting keywords...")
|
70 |
keywords = kw_model.extract_keywords(
|
71 |
review,
|
72 |
keyphrase_ngram_range=(1, 2),
|
73 |
stop_words="english",
|
74 |
top_n=3
|
75 |
)
|
76 |
+
progress.progress(60)
|
77 |
|
78 |
+
# Display scores and keywords side by side
|
79 |
col1, col2 = st.columns(2)
|
80 |
with col1:
|
81 |
st.subheader("Sentiment Scores")
|
|
|
85 |
for kw, score in keywords:
|
86 |
st.write(f"• {kw} ({score:.4f})")
|
87 |
|
88 |
+
# Bar chart
|
89 |
+
progress.text("Rendering chart...")
|
90 |
+
df_scores = pd.DataFrame.from_dict(sentiment_results, orient='index', columns=['score'])
|
|
|
|
|
|
|
91 |
df_scores.index.name = 'Sentiment'
|
92 |
st.bar_chart(df_scores)
|
93 |
+
progress.progress(80)
|
94 |
|
95 |
+
# Highlight highest sentiment
|
96 |
max_label, max_score = max(sentiment_results.items(), key=lambda x: x[1])
|
97 |
st.markdown(f"**Highest Sentiment:** **{max_label}** ({max_score:.4f})")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
|
99 |
+
# GPT-Driven Analysis & Suggestions
|
100 |
+
progress.text("Generating insights...")
|
101 |
+
prompt = f"""
|
102 |
+
You are an analytical amazon feedback expert.
|
103 |
+
Review: \"{review}\"
|
104 |
+
Sentiment Scores: {sentiment_results}
|
105 |
+
Top Keywords: {[kw for kw, _ in keywords]}
|
106 |
+
Tasks:
|
107 |
+
1. Analysis: Write a concise paragraph (3 sentences) interpreting customer sentiment by combining the scores and keywords.
|
108 |
+
2. Recommendations: Three separate paragraphs with actionable suggestions (max 30 words each).
|
109 |
+
"""
|
110 |
+
|
111 |
+
response = openai_client.chat.completions.create(
|
112 |
+
model="gpt-35-turbo",
|
113 |
+
messages=[
|
114 |
+
{"role": "system", "content": "You are a product-feedback analyst."},
|
115 |
+
{"role": "user", "content": prompt}
|
116 |
+
],
|
117 |
+
temperature=0.7,
|
118 |
+
max_tokens=200
|
119 |
+
)
|
120 |
+
gpt_reply = response.choices[0].message.content.strip()
|
121 |
+
st.markdown(gpt_reply)
|
122 |
|
123 |
+
# Complete
|
124 |
+
progress.progress(100)
|
125 |
+
progress.text("Done!")
|
126 |
|
127 |
+
if __name__ == "__main__":
|
128 |
main()
|