Repository for Our Paper: DS$^2$-ABSA: Dual-Stream Data Synthesis with Label Refinement for Few-Shot Aspect-Based Sentiment Analysis

This repository contains all the synthetic data and prompts used in our paper for reference. The code is currently being organized and will be released in the future.

File Structure

• brainstorm_output: Contains candidate values for each attribute generated by LLMs in the key-point-driven data synthesis.

  • categories.py: Attribute categories.
  • aspects.py: Aspects of the data.
  • opinions.py: Opinions corresponding to aspects.
  • sentiments.py: Sentiment labels for aspects.
  • objects.py: Objects related to aspects.

• data: Original ABSA datasets and few-shot datasets.

  • Subdirectories represent dataset names: lap, res, res15, res16.
  • Each dataset has its train, val, and test splits.
  • Files like sample2_all.json and sample5_all.json correspond to 2%-shot and 5%-shot randomly sampled subsets used in the paper.

• data_synthetic_norm: GPT-3.5 Turbo synthetic data w/ label normalization.

  • Subdirectories for dataset names: lap, res, res15, res16.
  • Files with _key.json correspond to key-point-driven synthesis, and _ins.json corresponds to instance-driven synthesis.

• data_synthetic_refined: Refined synthetic data, improved through the label refinement module (label normalization + noisy self-training).

  • Subdirectories for ABSA models: instructabsa, paraphrase.
  • Each model has subdirectories for datasets: lap, res, res15, res16.

• prompts: All prompt templates used in the paper.

  • keypoint_driven.py and instance_driven.py contain the full set of prompts for the respective strategies.
  • templates.py includes prompts for baseline methods and utility functions.

• utils_prompt: Functions to populate prompt templates.

  • get_keypoint_prompts.py: Detailed implementation of attribute prompting in key-point-driven synthesis.
  • get_instance_prompts.py: Implements sample combination and masking strategies in selective reconstruction.