Fig 1.
Training from scratch.
Fig 2.
Pre-training and fine-tuning.
Fig 3.
Prompt learning.
Fig 4.
EMSA prompt engineering.
Table 1.
GubaSenti dataset statistics.
Table 2.
10 examples of GubaSenti dataset.
Fig 5.
An example of 1-shot standard prompt template.
Table 3.
Comparison of pre-training and fine-tuning models and prompt engineering on GubaSenti sentiment analysis on the accuracy, precision, recall and F1-score (%).
Fig 6.
Prompt templates with varied combinations of English and Chinese.
Table 4.
Comparison between 3 prompt templates on GubaSenti sentiment analysis on the accuracy, precision, recall, and F1-score (%).
Fig 7.
All of the content used to construct the template is generated by GPT-4 API and the correctness and effectiveness of the generated content have been verified manually. (The English translation is only an English version of the EMSA prompt template and it is not relevant to the experiment.)
Table 5.
Comparison of different prompting approaches.
Table 6.
Overview of baseline models for sentiment analysis comparison.
Fig 8.
Accuracy of multiple GPT models on GubaSenti with different number of templates.
Table 7.
Performance comparison on GubaSenti dataset.
Fig 9.
Confusion matrices for Gubasenti.
Table 8.
Performance comparison on SST-2 dataset.
Table 9.
Performance comparison on SST-5 dataset.
Fig 10.
Confusion matrices for SST-2.
Fig 11.
Confusion matrices for SST-5.
Table 10.
Ablation Study Results on GubaSenti Dataset.
Table 11.
Ablation study results on SST-5 dataset.
Table 12.
Major error categories and their distribution.
Table 13.
Examples of successful implicit sentiment analysis.
Table 14.
Examples of successful context integration.
Table 15.
Examples of sarcasm-related errors.
Table 16.
Examples of mixed sentiment errors.