p06 / Conference
DUTH at SemEval-2023 Task 9: An Ensemble Approach for Twitter Intimacy Analysis
SemEval-2023, 1225-1230 · 2023
In plain language
The DUTH system estimates the intimacy expressed in multilingual tweets for SemEval-2023 Task 9.
Research context
It shows when compact downstream models can be a useful alternative to fine-tuning a larger language model.
Approach
Transformer-generated text embeddings feed three multilayer perceptrons and a CatBoost model, combined into an ensemble.
Findings
In these experiments, preprocessing gives no clear learning benefit and transformer fine-tuning does not outperform embeddings with simpler models. Ensembling improves regression accuracy.
Summary based on abstract.