This paper introduces Neural Responding Machine (NRM), a neural network-based response generator for Short-Text Conversation (STC). NRM uses an encoder-decoder framework with recurrent neural networks (RNN) for both encoding and decoding. It is trained on a large dataset of one-round conversation data from Weibo, achieving high performance in generating grammatically correct and contextually appropriate responses. NRM outperforms state-of-the-art methods, including retrieval-based and SMT-based models. STC involves generating a response to a short text (post) based on a single round of conversation. Previous methods for STC include retrieval-based methods, which rely on pre-existing responses, and SMT-based methods, which treat response generation as a translation problem. However, these methods have limitations, such as difficulty in customization and semantic equivalence issues. NRM uses a probabilistic model to generate responses, with the encoder summarizing the post into a vector representation and the decoder generating responses based on this. The model employs attention mechanisms to dynamically select relevant parts of the input for response generation. Three variants of NRM are proposed: NRM-glo (global encoding), NRM-loc (local encoding), and NRM-hyb (hybrid encoding combining global and local information). Experiments show that NRM outperforms retrieval-based and SMT-based methods, with NRM-hyb achieving the best performance. Human evaluation indicates that NRM generates fluent and semantically relevant responses, with over 60% of responses labeled as suitable or neutral. NRM-hyb, in particular, outperforms other models, demonstrating the effectiveness of combining global and local representations. The results suggest that NRM is a promising approach for STC, capable of generating diverse and contextually appropriate responses. Future work includes incorporating user intention or sentiment as an external signal to generate responses with specific goals.This paper introduces Neural Responding Machine (NRM), a neural network-based response generator for Short-Text Conversation (STC). NRM uses an encoder-decoder framework with recurrent neural networks (RNN) for both encoding and decoding. It is trained on a large dataset of one-round conversation data from Weibo, achieving high performance in generating grammatically correct and contextually appropriate responses. NRM outperforms state-of-the-art methods, including retrieval-based and SMT-based models. STC involves generating a response to a short text (post) based on a single round of conversation. Previous methods for STC include retrieval-based methods, which rely on pre-existing responses, and SMT-based methods, which treat response generation as a translation problem. However, these methods have limitations, such as difficulty in customization and semantic equivalence issues. NRM uses a probabilistic model to generate responses, with the encoder summarizing the post into a vector representation and the decoder generating responses based on this. The model employs attention mechanisms to dynamically select relevant parts of the input for response generation. Three variants of NRM are proposed: NRM-glo (global encoding), NRM-loc (local encoding), and NRM-hyb (hybrid encoding combining global and local information). Experiments show that NRM outperforms retrieval-based and SMT-based methods, with NRM-hyb achieving the best performance. Human evaluation indicates that NRM generates fluent and semantically relevant responses, with over 60% of responses labeled as suitable or neutral. NRM-hyb, in particular, outperforms other models, demonstrating the effectiveness of combining global and local representations. The results suggest that NRM is a promising approach for STC, capable of generating diverse and contextually appropriate responses. Future work includes incorporating user intention or sentiment as an external signal to generate responses with specific goals.