This repo our all-in-one Grid Tagging Scheme model inspired by the ISCAS participation in Task 10: "Structured Sentiment Analysis" in SemEval 2022.

Our team revisited the results from the SemEval-2022 Task 10: Structured Sentiment Analysis. We leveraged a Grid Tagging Scheme (GTS) which extracted target, holder, expression, and polarity. We adapted a pre-existing pipeline solution, which consisted of several steps to extract information and transformed into a single-step model that extracts all aspects of the sentiment. The proposed model demonstrated compelling performance when compared against the more complex and resource intensive model it was initially based on.

Prepare all the backbones this repo depends on through executing the following script:

bash ./all_in_one.sh

Train the GTS-based Extraction subsystem:

# Monolingual training
bash ./scripts/train_GTS_*.sh $gpu_num # train individual models

The trained models and evaluation outputs will be saved at ./src/GTS/saved_models/extract/

Predict:

bash ./scripts/predict_GTS.sh $gpu_num # Co-extraction subsystem predict

We would like to thank the ISCAS team for providing an excellent repository which served as a foundation for our study.