Abstract: 【Purpose/significance】With the development of social networks and news media, the breeding and spreading of a large
amount of false information has caused serious social problems. Current research mainly relies on the collection of rumors after their
occurrence and dissemination characteristics for identification. In order to detect rumours more accurately at an early stage, this paper
proposes a rumour identification model that incorporates deep semantic knowledge.【Method/process】This paper extracts complex fea? tures of the deep structure of public opinion information by using Transformer and Multi-head attention, and fuses document structure and contextual semantic knowledge representation to improve the accuracy of identifying false public opinion information at an early stage to prevent the spread of rumours in time.【Result/conclusion】Through training and identification experiments on real datasets from various platforms, this paper has improved the accuracy rate by at least 5.6% and at most 24.6% compared with the existing base? line methods. The results show that this model can improve the accuracy of rumour identification by fact-checking early rumour texts to stop the spread of rumours at an early stage.【Innovation/limitation】The rumour recognition model in this paper further combines the semantic knowledge feature representation of public opinion text based on BERT-Base, which can effectively improve the accu? racy of early rumour recognition, but the personalized features of rumour spreaders such as social labels and behavioural information have not been considered yet, and further research is needed on how to integrate more spreaders' features