Ethnic folk dances possess significant cultural value and require documentation and preservation. This article begins by recognizing the distinctive movement characteristics found in ethnic folk dances. It then collects skeletal motion data of the human body while executing typical movements in ethnic folk dances using Kinect V2. Two primary features, namely angle and relative distance, were extracted. Deep learning was combined with the attention mechanism to design a three-layer BiLSTM-attention method. Experiments were conducted using the typical movement feature set of ethnic folk dance and the MSRAction3D dataset. It was found that the three-layer BiLSTM method exhibited superior performance when compared to other configurations of BiLSTM layer. Additionally, the results derived from the BiLSTM model surpassed those achieved with RNN or LSTM models. Furthermore, the inclusion of the attention layer led to a noteworthy 0.0234 increase in the ACC value compared to models without it. The processed features demonstrated enhanced performance compared to the raw skeletal motion data. ACC values exceeding 0.95 were achieved for the recognition of typical movement features in various types of ethic folk dances. Notably, the ACC value of the three-layer BiLSTM method for the MSR-Action3D dataset was 0.9767, which was superior to the other methods. These outcomes validate the robustness of the methodology presented in this paper for recognizing typical movement features in folk dance and suggest its potential for practical applications.