Abstract: Aiming at the problems that there were  the lack of explicit sentiment clues, mixed sentiment features, polysemy features  and context dependence features in implicit sentiment analysis, we  proposed a Chinese implicit sentiment analysis method based on hierarchical knowledge enhancement and sememe knowledge. We first introduced a sentiment pre-training model based on  bidirectional encoder representation technology of converters  to enhance the ability of  sentiment clue recognition. Then 
we handled mixed sentiment features through character-level information acquisition, region moving box learning, global information learning, and multi-pooling operations. At the same time, we combined sememe knowledge and density matrix, utilized the HowNet knowledge base to allevite polysemy issues, and integrated  with bidirectional long short-term memory network 
features to tackle context dependence features. Experimental results show that the proposed method performs excellently in terms of effectiveness, superiority, and generalizability, providing a valuable technical path for Chinese implicit sentiment analysis and helping to improve sentiment understanding and decision-making support capabilities in scenarios such as social media and user reviews.

Key words: implicit sentiment analysis, external knowledge, knowledge enhancement, SentiBERT model, hierarchical knowledge, sememe knowledge