CARAT
Cheng Pengが第一著者,Web3系の例のZhejiang University
概要
入力がクロスモーダルな感情認識における研究
SoTAをタッっせいした
Intro
This paper addresses the MMER task that identifies multiple emotions from the multi-modal inputs.
Related works
There are mainly two types of methods
aggregation-based: concatenation
alignment-based: neglect the specificity of each modality
hybrid-method: fusion of the above
Proposed method
By using shuffling among modality dimensions, the proposed method implemented the shuffling among modalities and samples.
Created new co-occurrences among positive samples and negative samples.
Experiment
Preparing a lot of datasets such as BR, LP, MulT
感想
上がった数値は一部であり，非常にインクリメンタルな印象に思えた
手法自体のアイデアは面白いと思った．