Dataset Card for "tomaarsen/setfit-absa-semeval-restaurants"

Dataset Summary

This dataset contains the manually annotated restaurant reviews from SemEval-2014 Task 4, in the format as understood by SetFit ABSA.

For more details, see https://aclanthology.org/S14-2004/

Data Instances

An example of "train" looks as follows.

{"text": "But the staff was so horrible to us.", "span": "staff", "label": "negative", "ordinal": 0}
{"text": "To be completely fair, the only redeeming factor was the food, which was above average, but couldn't make up for all the other deficiencies of Teodora.", "span": "food", "label": "positive", "ordinal": 0}
{"text": "The food is uniformly exceptional, with a very capable kitchen which will proudly whip up whatever you feel like eating, whether it's on the menu or not.", "span": "food", "label": "positive", "ordinal": 0}
{"text": "The food is uniformly exceptional, with a very capable kitchen which will proudly whip up whatever you feel like eating, whether it's on the menu or not.", "span": "kitchen", "label": "positive", "ordinal": 0}
{"text": "The food is uniformly exceptional, with a very capable kitchen which will proudly whip up whatever you feel like eating, whether it's on the menu or not.", "span": "menu", "label": "neutral", "ordinal": 0}

Data Fields

The data fields are the same among all splits.

• text: a string feature.
• span: a string feature showing the aspect span from the text.
• label: a string feature showing the polarity of the aspect span.
• ordinal: an int64 feature showing the n-th occurrence of the span in the text. This is useful for if the span occurs within the same text multiple times.

Data Splits

Training ABSA models using SetFit ABSA

To train using this dataset, first install the SetFit library:

pip install setfit

And then you can use the following script as a guideline of how to train an ABSA model on this dataset:

from setfit import AbsaModel, AbsaTrainer, TrainingArguments
from datasets import load_dataset
from transformers import EarlyStoppingCallback

# You can initialize a AbsaModel using one or two SentenceTransformer models, or two ABSA models
model = AbsaModel.from_pretrained("sentence-transformers/all-MiniLM-L6-v2")

# The training/eval dataset must have `text`, `span`, `polarity`, and `ordinal` columns
dataset = load_dataset("tomaarsen/setfit-absa-semeval-restaurants")
train_dataset = dataset["train"]
eval_dataset = dataset["test"]

args = TrainingArguments(
    output_dir="models",
    use_amp=True,
    batch_size=256,
    eval_steps=50,
    save_steps=50,
    load_best_model_at_end=True,
)

trainer = AbsaTrainer(
    model,
    args=args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=5)],
)
trainer.train()

metrics = trainer.evaluate(eval_dataset)
print(metrics)

trainer.push_to_hub("tomaarsen/setfit-absa-restaurants")

You can then run inference like so:

from setfit import AbsaModel

# Download from Hub and run inference
model = AbsaModel.from_pretrained(
    "tomaarsen/setfit-absa-restaurants-aspect",
    "tomaarsen/setfit-absa-restaurants-polarity",
)

# Run inference
preds = model([
    "The best pizza outside of Italy and really tasty.",
    "The food here is great but the service is terrible",
])

Citation Information

@inproceedings{pontiki-etal-2014-semeval,
    title = "{S}em{E}val-2014 Task 4: Aspect Based Sentiment Analysis",
    author = "Pontiki, Maria  and
      Galanis, Dimitris  and
      Pavlopoulos, John  and
      Papageorgiou, Harris  and
      Androutsopoulos, Ion  and
      Manandhar, Suresh",
    editor = "Nakov, Preslav  and
      Zesch, Torsten",
    booktitle = "Proceedings of the 8th International Workshop on Semantic Evaluation ({S}em{E}val 2014)",
    month = aug,
    year = "2014",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/S14-2004",
    doi = "10.3115/v1/S14-2004",
    pages = "27--35",
}