Aspect-Based Sentiment Analysis with TextCNN

research/arxiv_papers/PL_FGSA/README.md

@@ -0,0 +1,107 @@
+# 🧠 Aspect-Based Sentiment Analysis with TextCNN (MindSpore)
+
+本项目基于 **MindSpore 框架**，实现了一个轻量级的 **情感极性分类系统**，支持多个典型情感分析数据集的加载、处理、训练与评估，并采用 **TextCNN** 架构进行建模。
+
+## ✅ 特性 Highlights
+
+- 📦 **支持数据集**：
+  - SemEval-2014 Task 4 (Laptops)
+  - MAMS-ATSA
+  - SST-2（自动兼容 `.parquet` 转 JSON）
+
+- 🔨 **统一数据预处理模块**：
+  - 多格式支持（CSV / XML / JSONL / Parquet）
+  - 处理结果统一转为 JSON 格式
+
+- ⚙️ **模型架构**：
+  - 轻量化 TextCNN
+  - 支持 Dropout / Weight Decay 正则化
+  - 可快速在 CPU 上训练
+
+- 📊 **完整评估指标输出**：
+  - 准确率 / 精确率 / 召回率 / F1
+  - 每个数据集独立日志 & 最终汇总为 CSV
+
+## 📁 项目结构
+
+```
+├── data/                       # 数据预处理（多格式 → JSON）
+│   ├── MAMS                    # MAMS数据集
+│   ├── SemEval_2014_Task_4     # SemEval数据集
+│   └── sst-2                   # sst-2数据集
+├── processed/                  # 存储处理好的 JSON 格式数据集
+│   ├── mams                    # MAMS数据集
+│   ├── semeval                 # SemEval数据集
+│   └── sst2                    # sst-2数据集
+├── train.py                    # TextCNN 模型训练入口
+├── dataset.py                  # 数据加载器 + 词表构建
+└── README.md
+```
+
+## 🚀 使用说明
+
+### ✅ 1. 数据预处理
+
+请将原始数据放入指定目录，然后运行：
+
+```bash
+python dataset.py
+```
+
+将自动处理以下数据集并转换为 JSON 格式：
+
+- `data/SemEval_2014_Task_4/`
+- `data/MAMS/MAMS-ATSA/raw/`
+- `data/SST2/`
+
+### ✅ 2. 模型训练
+
+运行主训练脚本：
+
+```bash
+python train.py
+```
+
+将依次训练以下数据集：
+
+- `semeval`
+- `mams`
+- `sst2`
+
+每个数据集：
+
+- 自动构建词表；
+- 划分训练/验证集（比例 8:1:1）；
+- 输出日志至 `log_*.txt`；
+- 将评估指标汇总至 `results_summary.csv`。
+
+## 📊 示例评估结果（`results_summary.csv`）
+
+| 数据集  | Acc  | Prec | Recall | F1   |
+| ------- | ---- | ---- | ------ | ---- |
+| semeval | 0.71 | 0.69 | 0.68   | 0.69 |
+| mams    | 0.61 | 0.59 | 0.60   | 0.59 |
+| sst2    | 0.86 | 0.85 | 0.86   | 0.85 |
+
+## 🛠️ 环境依赖
+
+- Python ≥ 3.9
+- MindSpore ≥ 2.5.0 (CPU)
+- numpy, pandas, scikit-learn, tqdm
+- pyarrow（用于读取 `.parquet`）
+
+安装依赖：
+
+```bash
+pip install -r requirements.txt
+```
+
+## 📌 注意事项
+
+- 如果不使用 `sst2` 数据集，可在 `train.py` 中移除对应行；
+- SST-2 默认使用 `train.json` 划分训练/验证集
+
+## 📚 参考文献
+
+- Kim, Y. (2014). *Convolutional Neural Networks for Sentence Classification*.
+- Socher, R. et al. (2013). *Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank*.

research/arxiv_papers/PL_FGSA/dataset.py

@@ -0,0 +1,179 @@
+import csv
+import json
+import xml.etree.ElementTree as ET
+from pathlib import Path
+from tqdm import tqdm
+
+def unify_label(label: str) -> int:
+    label = str(label).lower()
+    if label in ["positive", "pos", "2"]:
+        return 2
+    elif label in ["neutral", "neu", "1"]:
+        return 1
+    elif label in ["negative", "neg", "0"]:
+        return 0
+    else:
+        raise ValueError(f"Unknown label: {label}")
+
+def load_semeval_dir(folder_path: str, output_folder: str):
+    output_folder = Path(output_folder) / "laptops"
+    folder = Path(folder_path)
+    output = Path(output_folder)
+    output.mkdir(parents=True, exist_ok=True)
+    for file in tqdm(list(folder.glob("*.csv")), desc="Processing SemEval"):
+        if file.name not in {
+            "Laptop_Train_v2.csv",
+            "Laptops_Test_Data_PhaseA.csv",
+            "Laptops_Test_Data_PhaseB.csv"
+        }:
+            continue
+
+        data = []
+        with open(file, newline='', encoding='utf-8') as f:
+            reader = csv.DictReader(f)
+            headers = reader.fieldnames
+            if headers is None:
+                continue
+
+            is_labeled = "Aspect Term" in headers and "polarity" in headers
+
+            for row in reader:
+                text = row.get("Sentence")
+                if is_labeled:
+                    aspect = row.get("Aspect Term")
+                    polarity = row.get("polarity")
+                    if text and aspect and polarity:
+                        try:
+                            label = unify_label(polarity)
+                            data.append({"text": text, "aspect": aspect, "label": label})
+                        except ValueError:
+                            continue
+                else:
+                    if text:
+                        data.append({
+                            "text": text,
+                            "aspect": None,
+                            "label": None,
+                            "inference_only": True
+                        })
+        save_to_json(data, output / file.with_suffix(".json").name)
+
+def load_ecare_dir(folder_path: str, output_folder: str):
+    folder = Path(folder_path)
+    output = Path(output_folder)
+    output.mkdir(parents=True, exist_ok=True)
+
+    for file in tqdm(list(folder.glob("*.jsonl")), desc="Processing e-CARE"):
+        data = []
+        with open(file, encoding='utf-8') as f:
+            for line in f:
+                try:
+                    entry = json.loads(line)
+                    premise = entry.get("premise")
+                    h1 = entry.get("hypothesis1")
+                    h2 = entry.get("hypothesis2")
+                    label = entry.get("label")
+                    ask_for = entry.get("ask-for", "")
+
+                    if premise and h1 and h2 and label is not None:
+                        # 正确的 hypothesis → label=1，错误的 → label=0
+                        data.append({
+                            "text": f"{premise} [SEP] {h1}",
+                            "aspect": ask_for,
+                            "label": 1 if label == 0 else 0
+                        })
+                        data.append({
+                            "text": f"{premise} [SEP] {h2}",
+                            "aspect": ask_for,
+                            "label": 0 if label == 0 else 1
+                        })
+                except Exception:
+                    continue
+        save_to_json(data, output / file.with_suffix(".json").name)
+
+
+def load_mams_dir(folder_path: str, output_folder: str):
+    folder = Path(folder_path)
+    output = Path(output_folder)
+    output.mkdir(parents=True, exist_ok=True)
+    for file in tqdm(list(folder.glob("*.xml")), desc="Processing MAMS"):
+        data = []
+        tree = ET.parse(file)
+        root = tree.getroot()
+        for sentence in root.findall("sentence"):
+            text = sentence.find("text").text
+            aspect_terms = sentence.find("aspectTerms")
+            if aspect_terms is not None and list(aspect_terms):
+                for aspect in aspect_terms:
+                    try:
+                        term = aspect.attrib["term"]
+                        polarity = unify_label(aspect.attrib["polarity"])
+                        data.append({"text": text, "aspect": term, "label": polarity})
+                    except (KeyError, ValueError):
+                        continue
+            else:
+                data.append({
+                    "text": text,
+                    "aspect": None,
+                    "label": None,
+                    "inference_only": True
+                })
+        save_to_json(data, output / file.with_suffix(".json").name)
+
+
+def load_sst2_dir(folder_path: str, output_folder: str):
+    import pandas as pd
+
+    folder = Path(folder_path)
+    output = Path(output_folder)
+    output.mkdir(parents=True, exist_ok=True)
+
+    for file in tqdm(list(folder.glob("*.parquet")), desc="Processing SST-2"):
+        data = []
+        df = pd.read_parquet(file)
+        for _, row in df.iterrows():
+            text = row.get("sentence")
+            label = row.get("label")
+            if text is not None and label is not None:
+                try:
+                    data.append({
+                        "text": text,
+                        "aspect": None,
+                        "label": int(label)
+                    })
+                except ValueError:
+                    continue
+        save_to_json(data, output / file.with_suffix(".json").name)
+
+
+
+def save_to_json(data, output_path):
+    with open(output_path, 'w', encoding='utf-8') as f:
+        json.dump(data, f, ensure_ascii=False, indent=2)
+    print(f"Saved {len(data)} samples to {output_path}")
+
+def preprocess_all_datasets(dataset_dirs: dict, output_root: str):
+    """
+    dataset_dirs: dict with keys 'semeval', 'ecare', 'mams'
+    output_root: directory to save all processed files, grouped by dataset
+    """
+    output_root = Path(output_root)
+    if "semeval" in dataset_dirs:
+        load_semeval_dir(dataset_dirs["semeval"], output_root / "semeval")
+    if "ecare" in dataset_dirs:
+        load_ecare_dir(dataset_dirs["ecare"], output_root / "ecare")
+    if "mams" in dataset_dirs:
+        load_mams_dir(dataset_dirs["mams"], output_root / "mams")
+    if "sst2" in dataset_dirs:
+        load_sst2_dir(dataset_dirs["sst2"], output_root / "sst2")
+
+if __name__ == '__main__':
+    preprocess_all_datasets(
+        dataset_dirs={
+            "semeval": "data/SemEval_2014_Task_4/",
+            "ecare": "data/e-CARE/data/",
+            "mams": "data/MAMS/MAMS-ATSA/raw",
+            "sst2": "data/sst-2/"
+        },
+        output_root="processed/"
+    )

research/arxiv_papers/PL_FGSA/requirements.txt

@@ -0,0 +1,4 @@
+tqdm
+scikit-learn
+numpy
+pyarrow