Cross-prompt English composition automatic scoring method integrating CNN, LSTM, and attention mechanism

This study proposes a cross-prompt English composition automatic scoring method integrating CNN, LSTM, and attention mechanisms (AMs) to enhance generalisation and accuracy.CNN extracts local vocabulary and syntactic features, LSTM captures long-distance semantic dependencies, and the attention mechanism focuses on key scoring information.The final score is generated through a fully connected layer.Experimental results show a Pearson correlation of 0.94 with human scores and an MAE of 0.26, indicating strong correlation and low error.In cross-prompt environmental theme evaluation, the MAE and RMSE were 0.34 and 0.40, with a throughput of 82.11 articles per second and a response delay of 12.09 ms.The proposed method demonstrates high accuracy, efficiency, and adaptability, providing effective technical support for practical automatic English composition scoring.