In order to solve the problem that existing expert recommendation algorithms ignore the influence of the emotion expression contained in user comments on expert expertise representation, which leads to low recommendation accuracy, a deep semantic enhancement algorithm for expert recommendation based on bidirectional encoder representations from transformers-multi-head attention(BERT-MHA) was proposed. The algorithm was based on the pre-trained BERT model, incorporating the MHA mechanism to automatically adjust the emotional attention weights of user reviews towards questions the experts have answered before, capturing dynamic expertise representations. These representations were then combined with static expertise to semantically enhance the expert feature texts, representing the experts' comprehensive expertise. The algorithm identified the characteristics of user questions through the attention mechanism and employed a multi-layer perceptron to model the nonlinear interactions between the experts' comprehensive expertise and user questions, predicting the matching degree of recommended experts. Comparative experiments were conducted using data from haodf.com, with various parameter configurations and algorithms. The results indicate that the proposed algorithm significantly outperforms other algorithms under accuracy(ACC) and area under curve(AUC) indexes, effectively enhancing the accuracy of expert recommendations in online Q & A communities.