An Optimized Method for Sentiment Analysis of Short Video Comments in University Reputation Monitoring Based on Domain Dictionary and Prompt Learning

With the rapid growth of short video platforms in campus related communication, universities are more and more exposed in the risk of public opinion reflected in the comment section of the online. This paper proposes a lightweight optimization method integrated a domain specific dictionary with Prompt Learning. Using the comment related to Xi’an Jiaotong-Liverpool University on the Bilibili platform which is publicly available, the paper constructs a dataset of 503 manually annotated short video comment and a domain dictionary who contains 127 entries. Then the paper builds upon a bert-base-chinese fine tuning baseline and three Prompt template variant are designed and evaluated. Experimental results show that the domain dictionary injected Prompt method (Prompt-V2) keep the overall accuracy of 85.15% and improve the negative comment recall from 63.64% to 72.73%, which reduce the risk of missed negative sentiment detection in the university reputation monitoring scenario. The proposed approach offers a practical and effective optimization path for the low resource, highly domain specific sentiment analysis task. In the future work, the paper may expand the scale of the dataset, enrich the domain dictionary dynamically and further explore more advance prompt design strategy to improve the generalization and the overall classification performance.