models/onnx/README.md

metadata

language:
  - en
  - hi
tags:
  - sentiment-analysis
  - aspect-based-sentiment-analysis
  - onnx
  - int8
  - xlm-roberta

Multilingual ABSA (Aspect-Based Sentiment Analysis)

Model Description

This repository contains INT8-quantized ONNX models for Multilingual Aspect-Based Sentiment Analysis (ABSA). It uses a two-stage pipeline:

1. Aspect Extraction: Token classification model to identify aspects in text.
2. Sentiment Classification: Sequence classification model to determine sentiment (Positive, Negative, Neutral, Conflict) for extracted aspects.

Both models are based on xlm-roberta-base, fine-tuned using QLoRA, and exported to ONNX for CPU-optimized inference.

Languages Supported

• English (en)
• Hindi (hi)
• Hinglish (code-mixed)

Performance Metrics (Phase 4)

• English: Macro-F1 > 78%
• Hindi: Macro-F1 > 65%
• Latency (INT8 CPU): P95 < 300ms

Usage

from optimum.onnxruntime import ORTModelForTokenClassification, ORTModelForSequenceClassification
from transformers import AutoTokenizer

model_id = "YOUR_HF_USERNAME/multilingual-absa"
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Load Aspect Extraction Model
aspect_model = ORTModelForTokenClassification.from_pretrained(
    model_id, 
    subfolder="aspect_extraction_int8"
)

# Load Sentiment Model
sentiment_model = ORTModelForSequenceClassification.from_pretrained(
    model_id,
    subfolder="sentiment_int8"
)

Training Data

Fine-tuned on combined SemEval (English) and translated/native Hindi product review datasets.