BvSP: Broad-view Soft Prompting for Few-Shot Aspect Sentiment Quad Prediction

Introduction • Results • Quick Start • Core Code

✨ Introduction

We construct a few-shot ASQP dataset ($\mathtt{FSQP}$) and propose Broad-view Soft Prompting (BvSP).

BvSP, a various templates-based soft prompt learning method that improves quad prediction by taking into account the correlation between the different templates.

A new few-shot ASQP dataset $\mathtt{FSQP}$ contains richer categories and is more balanced for the few-shot study. To the best of our knowledge, this is the first work to explore the few-shot ASQP problem.

## 🎯 Results

Experimental results under four few-shot settings (i.e. one-shot, two-shot, five-shot, and ten-shot) demonstrate that BvSP outperforms strong baselines and has significant gains in other public datasets.

Evaluation on $\mathtt{FSQP}$

Under the same few-shot settings, BvSP demonstrates superior performance compared to MvP. These results validate the effectiveness of BvSP in providing a broader view of templates.

Evaluation on Other Datasets

BvSP continues to outperform all baselines in the full-shot setting. It is worth noting that BvSP exhibits superiority over MvP in both the top-3 and top-15. This underscores that BvSP not only excels in few-shot scenarios but also proves beneficial in full-shot scenarios.

🚀 Quick Start

⚙️ Setup

conda create -n bvsp python=3.8
conda activate bvsp
torch==1.11.0
transformers==4.14.1
pytorch_lightning==0.8.1
numpy==1.21.2

⚡️ Training

To obtain the training model:

Pre-Train models:

bash scripts/pretrain.sh

Train in few-shot settings

bash scripts/run.sh

🤖 Core Code

Jensen-Shannon (JS) divergence

BvSP/js.py

Soft Prompting

Data processing

BvSP/data_utils.py: line 271

Model init and freeze parameters

BvSP/model.py: line 117~127

Model forward

BvSP/model.py: line 263

byinhao/BvSP