The code is mainly used for finetuning pre-train models on ABSA datasets, where the pipe.py is the data preprocessing code; the finetune.py is used to fine-tune and get the fine-tuned model.

Usage

1. Install the fastNLp and the fitlog. Our code heavily rely on these two packages.

pip install fastNLP fitlog

2. Run the code with command:

  python finetune.py --data_dir /user/project/dataset/ --dataset Restaurant

• The --data_dir and --dataset arguments are explained in the following Notes.

Notes

0. The code are based on fastNLP and Fitlog. More can be found in fastNLP and fitlog.
1. The dataset path in our original code is:

(--data_dir)/user/project/dataset/
(--dataset) |---Restaurant
            |------Train.json
            |------Test.json
(--dataset) |---Laptop
            |------Train.json
            |------Test.json
(--dataset) |---fr
            |------Train.json
            |------Test.json
...

2. The Train/Test.json data files are in the following format：

{
  "sentence": "BEST spicy tuna roll , great asian salad .",
  "token": [
    "BEST",
    "spicy",
    "tuna",
    "roll",
    ",",
    "great",
    "asian",
    "salad",
    "."
  ],
  "pos": [
    "PROPN",
    "ADJ",
    "NOUN",
    "NOUN",
    "PUNCT",
    "ADJ",
    "ADJ",
    "NOUN",
    "PUNCT"
  ],
  "deprel": ["dep", "root", "dep", "dep", "dep", "dep", "dep", "dep", "punct"],
  "head": [2, 0, 4, 2, 4, 5, 8, 6, 7],
  "dependencies": [
    ["dep", 2, 1],
    ["root", 0, 2],
    ["dep", 4, 3],
    ["dep", 2, 4],
    ["dep", 4, 5],
    ["dep", 5, 6],
    ["dep", 8, 7],
    ["dep", 6, 8],
    ["punct", 7, 9]
  ],
  "aspects": [
    { "term": ["asian", "salad"], "polarity": "positive", "from": 6, "to": 8 },
    {
      "term": ["spicy", "tuna", "roll"],
      "polarity": "positive",
      "from": 1,
      "to": 4
    }
  ]
}