This paper systematically compares Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) on a wide range of natural language processing (NLP) tasks to provide guidance for selecting the appropriate deep neural network architecture. The study covers tasks such as sentiment classification, textual entailment, answer selection, question-relation matching, path query answering, and part-of-speech tagging. Key findings include: 1. **Complementary Information**: CNNs and RNNs provide complementary information for text classification tasks. The performance of each architecture depends on how semantically understanding the entire sequence is important. 2. **Hyperparameter Sensitivity**: Learning rate changes have a relatively smooth impact on performance, while changes in hidden size and batch size result in significant fluctuations. The experiments support the following observations: - **Text Classification (TextC)**: RNNs outperform CNNs in tasks where the sentiment is determined by the entire sentence or long-range semantic dependencies. - **Sequence Order (SeqOrder)**: RNNs are well-suited for encoding order information, outperforming CNNs in tasks like path query answering. - **Context Dependency (ContextDep)**: CNNs outperform one-directional RNNs in part-of-speech tagging but lag behind bi-directional RNNs. The study also highlights the importance of optimizing hyperparameters such as hidden size and batch size for achieving good performance in both CNNs and RNNs. Overall, the results suggest that the choice between CNNs and RNNs should be based on the specific requirements of the task, particularly the need to understand global or long-range semantics.This paper systematically compares Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) on a wide range of natural language processing (NLP) tasks to provide guidance for selecting the appropriate deep neural network architecture. The study covers tasks such as sentiment classification, textual entailment, answer selection, question-relation matching, path query answering, and part-of-speech tagging. Key findings include: 1. **Complementary Information**: CNNs and RNNs provide complementary information for text classification tasks. The performance of each architecture depends on how semantically understanding the entire sequence is important. 2. **Hyperparameter Sensitivity**: Learning rate changes have a relatively smooth impact on performance, while changes in hidden size and batch size result in significant fluctuations. The experiments support the following observations: - **Text Classification (TextC)**: RNNs outperform CNNs in tasks where the sentiment is determined by the entire sentence or long-range semantic dependencies. - **Sequence Order (SeqOrder)**: RNNs are well-suited for encoding order information, outperforming CNNs in tasks like path query answering. - **Context Dependency (ContextDep)**: CNNs outperform one-directional RNNs in part-of-speech tagging but lag behind bi-directional RNNs. The study also highlights the importance of optimizing hyperparameters such as hidden size and batch size for achieving good performance in both CNNs and RNNs. Overall, the results suggest that the choice between CNNs and RNNs should be based on the specific requirements of the task, particularly the need to understand global or long-range semantics.