tomaarsen/setfit-absa-bge-small-en-v1.5-restaurants-aspect

SetFit Aspect Model with BAAI/bge-small-en-v1.5 on SemEval 2014 Task 4 (Restaurants)

This is a SetFit model trained on the SemEval 2014 Task 4 (Restaurants) dataset that can be used for Aspect Based Sentiment Analysis (ABSA). This SetFit model uses BAAI/bge-small-en-v1.5 as the Sentence Transformer embedding model. A LogisticRegression instance is used for classification. In particular, this model is in charge of filtering aspect span candidates.

The model has been trained using an efficient few-shot learning technique that involves:

1. Fine-tuning a Sentence Transformer with contrastive learning.
2. Training a classification head with features from the fine-tuned Sentence Transformer.

This model was trained within the context of a larger system for ABSA, which looks like so:

1. Use a spaCy model to select possible aspect span candidates.
2. Use this SetFit model to filter these possible aspect span candidates.
3. Use a SetFit model to classify the filtered aspect span candidates.

Model Details

Model Description

• Model Type: SetFit
• Sentence Transformer body: BAAI/bge-small-en-v1.5
• Classification head: a LogisticRegression instance
• spaCy Model: en_core_web_lg
• SetFitABSA Aspect Model: tomaarsen/setfit-absa-bge-small-en-v1.5-restaurants-aspect
• SetFitABSA Polarity Model: tomaarsen/setfit-absa-bge-small-en-v1.5-restaurants-polarity
• Maximum Sequence Length: 512 tokens
• Number of Classes: 2 classes
• Training Dataset: SemEval 2014 Task 4 (Restaurants)
• Language: en
• License: apache-2.0

Model Sources

• Repository: SetFit on GitHub
• Paper: Efficient Few-Shot Learning Without Prompts
• Blogpost: SetFit: Efficient Few-Shot Learning Without Prompts

Model Labels

Label	Examples
aspect	'staff:But the staff was so horrible to us.' "food: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." "food: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."
no aspect	"factor: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." "deficiencies: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." "Teodora: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."

Evaluation

Metrics

Label	Accuracy
all	0.8623

Uses

Direct Use for Inference

First install the SetFit library:

pip install setfit

Then you can load this model and run inference.

from setfit import AbsaModel

# Download from the 🤗 Hub
model = AbsaModel.from_pretrained(
    "tomaarsen/setfit-absa-bge-small-en-v1.5-restaurants-aspect",
    "tomaarsen/setfit-absa-bge-small-en-v1.5-restaurants-polarity",
)
# Run inference
preds = model("The food was great, but the venue is just way too busy.")

Training Details

Training Set Metrics

Training set	Min	Median	Max
Word count	4	19.3576	45

Label	Training Sample Count
no aspect	170
aspect	255

Training Hyperparameters

• batch_size: (256, 256)
• num_epochs: (5, 5)
• max_steps: 5000
• sampling_strategy: oversampling
• body_learning_rate: (2e-05, 1e-05)
• head_learning_rate: 0.01
• loss: CosineSimilarityLoss
• distance_metric: cosine_distance
• margin: 0.25
• end_to_end: False
• use_amp: True
• warmup_proportion: 0.1
• seed: 42
• eval_max_steps: -1
• load_best_model_at_end: True

Training Results

Epoch	Step	Training Loss	Validation Loss
0.0027	1	0.2498	-
0.1355	50	0.2442	-
0.2710	100	0.2462	0.2496
0.4065	150	0.2282	-
0.5420	200	0.0752	0.1686
0.6775	250	0.0124	-
0.8130	300	0.0128	0.1884
0.9485	350	0.0062	-
1.0840	400	0.0012	0.183
1.2195	450	0.0009	-
1.3550	500	0.0008	0.2072
1.4905	550	0.0031	-
1.6260	600	0.0006	0.1716
1.7615	650	0.0005	-
1.8970	700	0.0005	0.1666
2.0325	750	0.0005	-
2.1680	800	0.0004	0.2086
2.3035	850	0.0005	-
2.4390	900	0.0004	0.183
2.5745	950	0.0004	-
2.7100	1000	0.0036	0.1725
2.8455	1050	0.0004	-
2.9810	1100	0.0003	0.1816
3.1165	1150	0.0004	-
3.2520	1200	0.0003	0.1802

• The bold row denotes the saved checkpoint.

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Carbon Emitted: 0.018 kg of CO2
• Hours Used: 0.303 hours

Training Hardware

• On Cloud: No
• GPU Model: 1 x NVIDIA GeForce RTX 3090
• CPU Model: 13th Gen Intel(R) Core(TM) i7-13700K
• RAM Size: 31.78 GB

Framework Versions

• Python: 3.9.16
• SetFit: 1.0.0.dev0
• Sentence Transformers: 2.2.2
• spaCy: 3.7.2
• Transformers: 4.29.0
• PyTorch: 1.13.1+cu117
• Datasets: 2.15.0
• Tokenizers: 0.13.3

Citation

BibTeX

@article{https://doi.org/10.48550/arxiv.2209.11055,
    doi = {10.48550/ARXIV.2209.11055},
    url = {https://arxiv.org/abs/2209.11055},
    author = {Tunstall, Lewis and Reimers, Nils and Jo, Unso Eun Seo and Bates, Luke and Korat, Daniel and Wasserblat, Moshe and Pereg, Oren},
    keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
    title = {Efficient Few-Shot Learning Without Prompts},
    publisher = {arXiv},
    year = {2022},
    copyright = {Creative Commons Attribution 4.0 International}
}