Update app.py

app.py

@@ -5,8 +5,6 @@ from PIL import Image
 import pytesseract
 import pandas as pd
 import plotly.express as px
-import re
-from collections import defaultdict
 
 # ✅ Step 1: Emoji 翻译模型（你自己训练的模型）
 emoji_model_id = "JenniferHJF/qwen1.5-emoji-finetuned"
@@ -39,74 +37,50 @@ with st.sidebar:
 if "history" not in st.session_state:
     st.session_state.history = []
 
-# 映射 label 到雷達圖分類（可依模型微調）
-label_to_category = {
-    "toxic": "Abuse",
-    "offensive": "Insult",
-    "insult": "Insult",
-    "threat": "Hate Speech",
-    "obscene": "Vulgarity",
-    "hate": "Hate Speech",
-    "discrimination": "Discrimination"
-}
-
-# ✅ 冒犯性分析函數（逐元素）
-def classify_text_elements(text: str):
-    elements = re.split(r"[，。,、！!？?\s\n]", text)
-    elements = [e for e in elements if e.strip()]
-
-    results = []
-    radar_scores = defaultdict(float)
-
-    for element in elements:
-        prompt = f"输入：{element}\n输出："
-        input_ids = emoji_tokenizer(prompt, return_tensors="pt").to(emoji_model.device)
-        with torch.no_grad():
-            output_ids = emoji_model.generate(**input_ids, max_new_tokens=64, do_sample=False)
-        decoded = emoji_tokenizer.decode(output_ids[0], skip_special_tokens=True)
-        translated = decoded.split("输出：")[-1].strip() if "输出：" in decoded else decoded.strip()
-
-        classification = classifier(translated)[0]
-        label = classification["label"].lower()
-        score = classification["score"]
-
-        category = label_to_category.get(label, "Others")
-        radar_scores[category] += score
-
-        reasoning = f"'{element}' was flagged as '{label}' → '{category}' due to potential offensiveness."
-        results.append({
-            "text": element,
-            "translated": translated,
-            "label": label,
-            "category": category,
-            "score": score,
-            "reason": reasoning
-        })
-
-    st.session_state.history.extend(results)
-    return results, radar_scores
-
-# 主页面
+# 分类函数
+def classify_emoji_text(text: str):
+    prompt = f"输入：{text}\n输出："
+    input_ids = emoji_tokenizer(prompt, return_tensors="pt").to(emoji_model.device)
+    with torch.no_grad():
+        output_ids = emoji_model.generate(**input_ids, max_new_tokens=64, do_sample=False)
+    decoded = emoji_tokenizer.decode(output_ids[0], skip_special_tokens=True)
+    translated_text = decoded.split("输出：")[-1].strip() if "输出：" in decoded else decoded.strip()
+
+    result = classifier(translated_text)[0]
+    label = result["label"]
+    score = result["score"]
+    reasoning = (
+        f"The sentence was flagged as '{label}' due to potentially offensive phrases. "
+        "Consider replacing emotionally charged, ambiguous, or abusive terms."
+    )
+
+    st.session_state.history.append({
+        "text": text,
+        "translated": translated_text,
+        "label": label,
+        "score": score,
+        "reason": reasoning
+    })
+    return translated_text, label, score, reasoning
+
+# 主页面：输入与分析共存
 st.title("🚨 Emoji Offensive Text Detector & Analysis Dashboard")
 
 # 文本输入
 st.subheader("1. 输入与分类")
-default_text = "你是🐷，太垃圾了，滚开！"
+default_text = "你是🐷"
 text = st.text_area("Enter sentence with emojis:", value=default_text, height=150)
 
 if st.button("🚦 Analyze Text"):
     with st.spinner("🔍 Processing..."):
         try:
-            analysis_results, radar_scores = classify_text_elements(text)
-
-            st.markdown("### ✨ Element-wise Classification")
-            for item in analysis_results:
-                st.markdown(f"- 🔹 **Input:** {item['text']}")
-                st.markdown(f"   - ✨ **Translated:** {item['translated']}")
-                st.markdown(f"   - ❗ **Label:** {item['label']} → **{item['category']}** ({item['score']:.2%})")
-                st.markdown(f"   - 🔧 **Reasoning:** {item['reason']}")
-
-            st.success("✅ Analysis complete!")
+            translated, label, score, reason = classify_emoji_text(text)
+            st.markdown("**Translated sentence:**")
+            st.code(translated, language="text")
+            st.markdown(f"**Prediction:** {label}")
+            st.markdown(f"**Confidence Score:** {score:.2%}")
+            st.markdown("**Model Explanation:**")
+            st.info(reason)
         except Exception as e:
             st.error(f"❌ An error occurred:\n{e}")
 
@@ -122,12 +96,13 @@ if uploaded_file:
         if ocr_text:
             st.markdown("**Extracted Text:**")
             st.code(ocr_text)
-            analysis_results, radar_scores = classify_text_elements(ocr_text)
-            for item in analysis_results:
-                st.markdown(f"- 🔹 **Input:** {item['text']}")
-                st.markdown(f"   - ✨ **Translated:** {item['translated']}")
-                st.markdown(f"   - ❗ **Label:** {item['label']} → **{item['category']}** ({item['score']:.2%})")
-                st.markdown(f"   - 🔧 **Reasoning:** {item['reason']}")
+            translated, label, score, reason = classify_emoji_text(ocr_text)
+            st.markdown("**Translated sentence:**")
+            st.code(translated, language="text")
+            st.markdown(f"**Prediction:** {label}")
+            st.markdown(f"**Confidence Score:** {score:.2%}")
+            st.markdown("**Model Explanation:**")
+            st.info(reason)
         else:
             st.info("⚠️ No text detected in the image.")
 
@@ -135,24 +110,59 @@ if uploaded_file:
 st.markdown("---")
 st.subheader("3. Violation Analysis Dashboard")
 if st.session_state.history:
+    # 展示历史记录
     df = pd.DataFrame(st.session_state.history)
-
-    st.markdown("### 🧾 Offense History Summary")
+    st.markdown("### 🧾 Offensive Terms & Suggestions")
     for item in st.session_state.history:
-        st.markdown(f"- **Input:** {item['text']}")
-        st.markdown(f"   - 🔠 Translated: {item['translated']}")
-        st.markdown(f"   - 🏷️ Label: {item['label']} → {item['category']}, Score: {item['score']:.2%}")
-
-    # 累积雷达分数
-    category_list = ["Insult", "Abuse", "Discrimination", "Hate Speech", "Vulgarity"]
-    radar_data = {
-        "Category": category_list,
-        "Score": [min(st.session_state.history.count(c)/len(st.session_state.history), 1.0)
-                  if c in radar_scores else 0.0 for c in category_list]
-    }
-    radar_df = pd.DataFrame(radar_data)
+        st.markdown(f"- 🔹 **Input:** {item['text']}")
+        st.markdown(f"   - ✨ **Translated:** {item['translated']}")
+        st.markdown(f"   - ❗ **Label:** {item['label']} with **{item['score']:.2%}** confidence")
+        st.markdown(f"   - 🔧 **Suggestion:** {item['reason']}")
+
+    # 雷达图
+    radar_df = pd.DataFrame({
+        "Category": ["Insult","Abuse","Discrimination","Hate Speech","Vulgarity"],
+        "Score": [0.7,0.4,0.3,0.5,0.6]
+    })
     radar_fig = px.line_polar(radar_df, r='Score', theta='Category', line_close=True, title="⚠️ Risk Radar by Category")
     radar_fig.update_traces(line_color='black')
     st.plotly_chart(radar_fig)
+
+    # —— 新增：单词级冒犯性相关性分析 —— #
+    st.markdown("### 🧬 Word-level Offensive Correlation")
+
+    # 取最近一次翻译文本，按空格拆分单词
+    last_translated_text = st.session_state.history[-1]["translated"]
+    words = last_translated_text.split()
+
+    # 对每个单词进行分类并收集分数
+    word_scores = []
+    for word in words:
+        try:
+            res = classifier(word)[0]
+            word_scores.append({
+                "Word": word,
+                "Label": res["label"],
+                "Score": res["score"]
+            })
+        except Exception:
+            continue
+
+    if word_scores:
+        word_df = pd.DataFrame(word_scores)
+        word_df = word_df.sort_values(by="Score", ascending=False).reset_index(drop=True)
+
+        max_display = 5
+        # Streamlit 1.22+ 支持 st.toggle，若版本不支持可改用 checkbox
+        show_more = st.toggle("Show more words", value=False)
+
+        display_df = word_df if show_more else word_df.head(max_display)
+        # 隐藏边框并渲染 HTML 表格
+        st.markdown(
+            display_df.to_html(index=False, border=0),
+            unsafe_allow_html=True
+        )
+    else:
+        st.info("❕ No word-level analysis available.")
 else:
     st.info("⚠️ No classification data available yet.")