This folder mainly contains codes for fine-tuning models on ABSA datasets.
pipe.pypre-processes data.finetune.pyfine-tunes models, which also save the fine-tuned models after fine-tuning.
-
Install fastNLP and fitlog. Our codes rely on this two packages.
-
Run the codes with command:
python finetune.py --data_dir {/your/dataset_filepath/} --dataset {dataset_name}-
The codes are based on fastNLP and Fitlog. More can be found in fastNLP and fitlog.
-
The codes will use data files named as
Train/Test.jsonunder the{data_dir}/{dataset}folder, e.g.{RoBERTaABSA/Dataset}/{Restaurant}, so please make sure this two files are available. -
The
Train/Test.jsondata 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
}
]
}