m-e5-small-vlsp2018-restaurant

Overview

Vietnamese aspect-category sentiment classification model for restaurant reviews from the VLSP 2018 sentiment analysis benchmark.

Model Details

• Base model: intfloat/multilingual-e5-small
• Architecture: absa
• Checkpoint source: vlsp-2018-restaurant-e5-small-best.pt
• Sequence length used during training/inference pipeline: 256
• Number of aspect categories: 12

Label Schema

• 0: aspect not mentioned
• 1: positive
• 2: negative
• 3: neutral

Aspect Categories

• AMBIENCE#GENERAL
• DRINKS#PRICES
• DRINKS#QUALITY
• DRINKS#STYLE&OPTIONS
• FOOD#PRICES
• FOOD#QUALITY
• FOOD#STYLE&OPTIONS
• LOCATION#GENERAL
• RESTAURANT#GENERAL
• RESTAURANT#MISCELLANEOUS
• RESTAURANT#PRICES
• SERVICE#GENERAL

Dataset

• Dataset: VLSP 2018 Restaurant Reviews This model is trained on the restaurant subset of the VLSP 2018 aspect-based sentiment analysis benchmark.

Data Format

• Review is the input text column.
• Each aspect-category column is encoded as 0/1/2/3 for none, positive, negative, or neutral.

Splits

• Train: 2961 samples
• Validation: 1290 samples
• Test: 500 samples

Checkpoint Metrics

• loss: 0.3680
• accuracy: 0.8747

Usage

Load the model with trust_remote_code=True because this repository contains custom modeling code.

from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo_id = "NeoCyber/m-e5-small-vlsp2018-restaurant"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(
    repo_id,
    trust_remote_code=True,
)

texts = ["Đồ ăn ngon nhưng phục vụ hơi chậm."]
inputs = tokenizer(texts, return_tensors="pt", truncation=True, padding=True)
outputs = model(**inputs)
predictions = model.decode_predictions(outputs.logits)
print(predictions)

Notes

• The repository includes custom configuration_*.py and modeling_*.py files required by transformers AutoClasses.
• outputs.logits has shape [batch_size, num_aspects, 4] and model.decode_predictions(...) maps logits back to aspect-level labels.