SentimentAnalysis

返回首页

---

项目概述

• 项目参考了Yoon Kim的论文Convolutional Neural Networks for Sentence Classification的实现方法，利用CNN卷积神经网络完成语句情感分析。
• 项目结构
  • data/*：json数据
  • parse_data：生成数据的json文件
  • cnn.py：CNN网络模型
  • train.py：训练脚本
  • test.py：测试脚本
• 参数

    ALLOW_SOFT_PLACEMENT=True
    BATCH_SIZE=50
    CHECKPOINT_EVERY=100
    DEV_DATA_FILE=./data/dev.json
    DROPOUT_KEEP_PROB=0.5
    EMBEDDING_DIM=300
    EVALUATE_EVERY=100
    FILTER_SIZES=3,4,5
    L2_REG_LAMBDA=3.0
    LOG_DEVICE_PLACEMENT=False
    NUM_CHECKPOINTS=5
    NUM_EPOCHS=200
    NUM_FILTERS=100
    TEST_DATA_FILE=./data/test.json
    TRAIN_DATA_FILE=./data/train.json
  

步骤

• 模型数据：https://cloud.tsinghua.edu.cn/d/e3da1c00a9e84a5d9132/
• Train with dropout.

    $ python test.py  --checkpoint_dir="./runs/1577169494/checkpoints/"
    Total number of test examples: 2210
    Accuracy: 0.400905
  

• Train with 256 or 512 hidden size.
  • 修改参数NUM_FILTERS = 256

      $ python test.py  --checkpoint_dir="./runs/1577171119/checkpoints/"
      Total number of test examples: 2210
      Accuracy: 0.40905
    

  • 修改参数NUM_FILTERS = 512

      $ python test.py  --checkpoint_dir="./runs/1577172661/checkpoints/"
      Total number of test examples: 2210
      Accuracy: 0.39819
    

• Train with a different number of the hidden layer. (The number of hiddenlayer should be set to 1 and 3)
  • 修改参数FILTER_SIZES = 3

      $ python test.py  --checkpoint_dir="./runs/1577174708/checkpoints/"
      Total number of test examples: 2210
      Accuracy: 0.384163
    

  • 修改参数FILTER_SIZES = 3,4

      $ python test.py  --checkpoint_dir="./runs/1577175634/checkpoints/"
      Total number of test examples: 2210
      Accuracy: 0.39819
    

  • 修改参数FILTER_SIZES = 3,4,5

      $ python test.py  --checkpoint_dir="./runs/1577169494/checkpoints/"
      Total number of test examples: 2210
      Accuracy: 0.400905
    

• Train with pre-trained word embedding. (We supply GloVe pretrainedword embedding with 300-dimension for your experiments and you canexplore the model performance with the same dimension without pre-trained word embeddings.)

    $ python test.py  --checkpoint_dir="./runs/1577022248/checkpoints/"
    Total number of test examples: 2210
    Accuracy: 0.414027