Keyword extraction methods mainly rely on keyword data statistics and linguistic rules but when dealing with social media data these methods are often difficult to achieve the desired results. With the rise of artificial intelligence and machine learning technology keyword extraction methods based on deep learning have received widespread attention. We propose a keyword extraction method for social media topics based on multi-source information fusion. We construct a decision layer fusion model process the keyword data of social media topics construct a word map model based on the co-occurrence relationship of candidate words under a fixed window size calculate the initial score of candidate words extract the average information entropy features of candidate words and filter out the more important candidate words through hierarchical features on the basis of the initial score of the candidate words and filter out the more frequent but unimportant words through the average information entropy features. Based on the initial scores of candidate words more important candidate words can be filtered out by hierarchical features and the unimportant words with many occurrences can be filtered out by average information esntropy features and finally the keywords can be outputted. The results of the experiment indicate that this paper has the capability to extract topic-related keywords from social media texts more efficiently leading to an enhancement in recall rate.