Research on Short Video Hotspot Classification Based on LDA Feature Fusion and Improved BiLSTM

Abstract

Short video hot spot classification is a fundamental method to grasp the focus of consumers and improve the effectiveness of video marketing. The limitations of traditional short text classification are sparse content as well as inconspicuous feature extraction. To solve the problems above, this paper proposes a short video hot spot classification model combining latent dirichlet allocation (LDA) feature fusion and improved bi-directional long short-term memory (BiLSTM), namely the LDA-BiLSTM-self-attention (LBSA) model, to carry out the study of hot spot classification that targets Carya cathayensis walnut short video review data under the TikTok platform. Firstly, the LDA topic model was used to expand the topic features of the Word2Vec word vector, which was then fused and input into the BiLSTM model to learn the text features. Afterwards, the self-attention mechanism was employed to endow different weights to the output information of BiLSTM in accordance with the importance, to enhance the precision of feature extraction and complete the hot spot classification of review data. Experimental results show that the precision of the proposed LBSA model reached 91.52%, which is significantly improved compared with the traditional model in terms of precision and F1 value.

Keywords

• feature fusion
• self-attention
• BiLSTM
• hotspot classification