Abstract: 【Purpose/significance】Topic identification has always been a research hotspot in the field of public opinion, nowadays
there are abundant research findings. Existing work mostly uses the traditional bag-of-words model, LDA or word-vector model to rep?
resent public opinion information, which can only symbolize tokens uniquely and traditional model needs to segment words, which may
affect the identification effect due to segmentation error, data sparsity and out-of-vocabulary.【Method/process】This paper proposes atopic identification model of network public opinion based on multi-sample Bidirectional Encoder Representations from Transformers
and text does not need to be segmented before training. The method of combining head text with tail is used for truncation for the
long-text case. Embedding features are extracted from word, segment and position dimensions. The public opinion is represented by
self-attention mechanism. Fine-tuning learning rate and multi-sample dropout are used to enhance the generalization and improve
identification effect.【Result/conclusion】Results show that the proposed model performs well in the task of public opinion topic classi?
fication, and can identify topic accurately without text segmentation.【Innovation/limitation】The innovation of this article is the con?
struction of a new network topic identification model, however, the limitation lies in the complexity of the algorithm, ways to optimize
parameters is the focus of next research.