Oliveira, B. S. N., & family=Silva. Studying the Dependence of Embedding Representations on the Target of NLP Tasks.
|
 |
|
Abstract
|
This work investigates the dependencies and correlations between embeddings and NLP tasks, and investigates how to initially verify if the embedDings' dimensions depend on the final task. In many human languages, linguistic units represent text structure. Vector semantics is used in NLP to represent these units, known as embeddings. Evaluating the learned representations is crucial for identifying critical differences between the diverse existing embedding models in task-specific selection. However, the evaluation process is complex, with two approaches: intrinsic and extrinsic. While useful, aggregated evaluations often lack consistency due to result misalignment. This work investigates the dependencies and correlations between embeddings and NLP tasks. The goal is how to initially verify if the embeddings' dimensions (i.e., features) depend on the final task. The study then explores two research questions and presents findings from experiments.
|
