Update app.py

app.py

@@ -1,26 +1,22 @@
-import os
-import zipfile
-import gradio as gr
 import torch
 from transformers import AutoTokenizer, AutoModelForSequenceClassification
+from transformers import BertTokenizer, BertForSequenceClassification
+import gradio as gr
+import os
 
-# 情感模型（京东）
+# --------- Sentiment Model (Binary, expanded to 3 classes) ---------
 sentiment_model_name = "uer/roberta-base-finetuned-jd-binary-chinese"
 sentiment_tokenizer = AutoTokenizer.from_pretrained(sentiment_model_name)
 sentiment_model = AutoModelForSequenceClassification.from_pretrained(sentiment_model_name)
 sentiment_model.eval()
 
-# 解压你自己的多标签模型
-if not os.path.exists("result"):
-    with zipfile.ZipFile("model_output.zip", "r") as zip_ref:
-        zip_ref.extractall(".")
+# --------- Multi-label Classification Model (Your model) ---------
+label_dir = "./result"
+label_tokenizer = BertTokenizer.from_pretrained(label_dir)
+label_model = BertForSequenceClassification.from_pretrained(label_dir)
+label_model.eval()
 
-# 加载你的多标签分类模型
-custom_tokenizer = AutoTokenizer.from_pretrained("result")
-custom_model = AutoModelForSequenceClassification.from_pretrained("result", use_safetensors=True)
-custom_model.eval()
-
-# 多标签类别
+# Label categories
 label_map = {
     0: "Landscape & Culture",
     1: "Service & Facilities",
@@ -29,45 +25,71 @@ label_map = {
     4: "Interactive Activities",
     5: "Price & Consumption"
 }
+threshold = 0.5
+
+# --------- Inference Function ---------
+def analyze(text):
+    if not text.strip():
+        return "Please enter a valid comment.", "Please enter a valid comment."
 
-# 推理函数
-def analyze(text, threshold=0.5):
-    # 情感分析
-    inputs = sentiment_tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=128)
+    # --- Sentiment Analysis ---
+    sent_inputs = sentiment_tokenizer(text, return_tensors='pt', truncation=True, padding=True, max_length=128)
     with torch.no_grad():
-        outputs = sentiment_model(**inputs)
-        probs = torch.softmax(outputs.logits, dim=1).squeeze().tolist()
-    sentiment = "积极 (Positive)" if torch.argmax(outputs.logits) == 1 else "消极 (Negative)"
-    sentiment_result = f"{sentiment}\nPositive: {probs[1]:.2f}, Negative: {probs[0]:.2f}"
+        sent_outputs = sentiment_model(**sent_inputs)
+        probs = torch.softmax(sent_outputs.logits, dim=1).squeeze().tolist()
 
-    # 多标签分类
-    inputs = custom_tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=128)
+    pos_prob, neg_prob = probs[1], probs[0]
+    if abs(pos_prob - neg_prob) < 0.2:
+        sentiment_label = "Neutral"
+    elif pos_prob > neg_prob:
+        sentiment_label = "Positive"
+    else:
+        sentiment_label = "Negative"
+
+    sentiment_result = (
+        f"Prediction: {sentiment_label}\n\n"
+        f"Sentiment Scores:\n"
+        f"Positive: {pos_prob:.2f}\n"
+        f"Neutral: {1 - abs(pos_prob - neg_prob):.2f} (estimated)\n"
+        f"Negative: {neg_prob:.2f}"
+    )
+
+    # --- Label Prediction ---
+    label_inputs = label_tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=128)
     with torch.no_grad():
-        logits = custom_model(**inputs).logits
+        label_outputs = label_model(**label_inputs)
+        logits = label_outputs.logits
         probs = torch.sigmoid(logits).squeeze().tolist()
-    if isinstance(probs, float):  # 单个标签时
+
+    if isinstance(probs, float):
         probs = [probs]
-    results = [
-        f"{label_map[i]} ({probs[i]:.2f})"
-        for i in range(len(probs)) if probs[i] >= threshold
-    ]
-    if results:
-        label_result = "\n".join(results)
+
+    selected_labels = [label_map[i] for i, p in enumerate(probs) if p >= threshold]
+    if selected_labels:
+        label_result = "Predicted Tags:\n" + "\n".join([f"{label_map[i]} ({probs[i]:.2f})" for i in range(len(probs)) if probs[i] >= threshold])
     else:
-        label_result = "The model was unable to identify the correct labels."
+        label_result = "No confident labels identified by the model."
+
+    return sentiment_result, label_result
+
+# --------- Gradio Web UI ---------
+with gr.Blocks(title="Sentiment + Tag Analysis System") as demo:
+    gr.Markdown("## 🌟  Comment Analyzer")
+    gr.Markdown(
+        "This tool analyzes **Chinese product reviews** using deep learning models. "
+        "It predicts both **sentiment polarity** (Positive / Neutral / Negative) and **semantic category tags** (6 themes)."
+    )
 
-    return f"【Sentiment analysis】\n{sentiment_result}\n\n【Category of topic】\n{label_result}"
+    with gr.Row():
+        with gr.Column():
+            input_box = gr.Textbox(label="Enter a JD.com review", placeholder="e.g., The park is peaceful and the staff are friendly...", lines=4)
+            submit_btn = gr.Button("🔍 Analyze")
+        with gr.Column():
+            sentiment_output = gr.Textbox(label="Sentiment Result", lines=6)
+            label_output = gr.Textbox(label="Tag Classification Result", lines=6)
 
-# 创建 Gradio 页面
-demo = gr.Interface(
-    fn=analyze,
-    inputs=[
-        gr.Textbox(lines=3, label="请输入评论内容"),
-        gr.Slider(minimum=0.1, maximum=0.9, step=0.05, value=0.5, label="分类标签阈值")
-    ],
-    outputs="text",
-    title="中文评论分析器",
-    description="使用京东情感模型 + 自定义多标签模型，对评论内容进行双重分析"
-)
+    submit_btn.click(fn=analyze, inputs=input_box, outputs=[sentiment_output, label_output])
 
-demo.launch()
+# --------- Run App ---------
+if __name__ == "__main__":
+    demo.launch()