.png){kind=icon}
.png){kind=icon}
Return
Interactive capsule network for implicit sentiment analysis
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-022-03584-3.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Existing sentiment analysis models mainly rely on evident emotive words within phrases. When the apparent emotional words within phrases are eliminated, the performance of these models will inevitably decrease. The implicit communication of emotion without the use of explicit emotional phrases is highly widespread in several cultures. As a result, a classification model is required to learn the link between contexts and the emotions they trigger in an automatic way. Based on whether the sentence should be segmented at the keyword position, existing methods apply either segmented or nonsegmented approaches. When emotional words are removed from a sentence, the nonsegmented approaches may lose syntactic information. To address these issues, an interactive iapsule network was proposed in this paper to extend the segmented approach. Taking the keyword as the segmented position, the network initializes two BERT models from a pretrained checkpoint with shared parameters as the encoder to process both contexts separately. By using both interactive attention and the capsule network with a dynamic routing algorithm, the model can automatically learn the insightful relationship between the former and the latter contexts. After fusing the former and latter context features, the interactive capsule network leverages both local and global attention to complete the sentiment analysis task. Experimental results on both English and Chinese corpora show that the proposed interactive attention model achieves a better performance than existing methods during implicit sentiment analysis tasks. In addition, the proposed model outperformed the top 3 models on WASSA-2018 implicit English shared tasks.
Keywords:
Implicit sentiment analysis
Interactive attention
Capsule network
Natural language processing