Target Word Transferred Language Model for tackling ABSA (WordTransABSA)
⭐ ⭐ ⭐ ⭐ ⭐
Word Transferred Language Model for Aspect-based Sentiment Analysis]{WordTransABSA: A Novel BERT-based Aspect-based Sentiment Analysis Approach by Utilizing ``Masked Language Model'' to Predict Affective Tokens for Target Words 😃
[][Model Architecture of WordTransABSA]
> In order to alleviate the limitations of mainstream fine-tuning methods, we propose Target Word Transferred ABSA (WordTransABSA). WordTransABSA subverts the conventional criterion of Transformerbased fine-tuning methods by utilizing the entire parameters in Transformer, to fully exploit the prior knowledge in Transformer.
> Compared with prompt learning, which requiresconstructing extra pre-defined templates with specific slots to better invoke the Pre-train LMs, WordTransABSA only needs sentiment-related pivot tokens to obtain the sentiment polarity-related affective tokens on the position of an aspect term.
> In the specific WordTransABSA implementation, we explore different discover measures to search high quality sentiment-related pivot tokens and try additional transferred word optimization strategies to stimulate the semantic understanding potential of PLMs better.
> The extensive experiments under the data-sufficient scenario (full-data supervised learning) and data-scarce scenario (few-shot learning) validate the superiority and effectiveness of the WordTransABSA, suggesting that regressing to the Transformer pretraining paradigm is a better solution for some specific scenarios like ABSA.
The ABSA experimental performance statistics (% Accuracy) between the WordsTransABSAs and the SOTA baselines.
[
][comparisons]
The performance statistics (% Accuracy) of our WordTransABSA variants and the SOTA baselines under the different few-shot learning settings. All results are averaged over 10 runs to maintain the experimental authenticity.
[
][fewshot_comparisons]
😎 🙏 😇 💗 💗 💗 😝