Core Concepts
The proposed SoftMCL introduces valence ratings as soft-label supervision for contrastive learning to fine-grained measure the sentiment similarities between samples, and performs contrastive learning on both word- and sentence-level to enhance the model's ability to learn affective information.
Abstract
The paper proposes a soft momentum contrastive learning (SoftMCL) approach for fine-grained sentiment-aware pre-training. The key highlights are:
Instead of using hard labels of sentiment polarities, the method introduces valence ratings as soft-label supervision for contrastive learning to fine-grained measure the sentiment similarities between samples.
SoftMCL is conducted on both the word- and sentence-level to enhance the model's ability to learn affective information.
A momentum queue is introduced to expand the contrastive samples, allowing storing and involving more negatives to overcome the limitations of hardware platforms.
Extensive experiments on four different sentiment-related tasks demonstrate the effectiveness of the proposed SoftMCL method, outperforming other sentiment-aware pre-training approaches.
The ablation study shows the importance of the word-level, sentence-level, and momentum contrastive learning components in the proposed framework.
The paper also analyzes the impact of different hyperparameters, such as balance coefficient, temperature, momentum coefficient, and queue size, on the performance of SoftMCL.
Stats
The battery life is long.
It takes a long time to focus.
The scope of his book is ambitious.
The government's decisions to begin the ambitious plans which cost a lot.
Quotes
"The pre-training for language models captures general language understanding but fails to distinguish the affective impact of a particular context to a specific word."
"Learning word sentiment cannot help the model understand the sentiment intention of the whole sentence. Since the expressed sentiment of a sentence is not simply the sum of the polarities or the intensity of its constituent words."