This paper introduces a diversity-promoting objective function for neural conversation models. Traditional sequence-to-sequence (SEQ2SEQ) models, which maximize the likelihood of outputs given inputs, tend to generate safe, generic responses like "I don't know." The authors propose using Maximum Mutual Information (MMI) as the objective function instead, which measures mutual dependence between inputs and outputs. Experimental results show that MMI models produce more diverse, interesting, and appropriate responses, leading to substantial gains in BLEU scores on two conversational datasets and in human evaluations. The paper discusses the limitations of traditional SEQ2SEQ models in generating diverse responses and proposes MMI as a solution. It outlines practical strategies for using MMI in neural generation models, including the use of a hyperparameter λ to control the penalty for generic responses. The authors also address challenges in applying MMI, such as the difficulty of decoding using the MMI objective directly, and propose solutions like using a modified language model and reranking N-best lists. The paper presents the architecture of SEQ2SEQ models, including the use of LSTM networks for encoding and decoding. It describes the MMI criterion, which aims to balance the likelihood of responses given inputs with the likelihood of inputs given responses. The authors also discuss practical considerations for implementing MMI, including the use of a modified language model and reranking N-best lists. The paper presents experiments on two datasets: the Twitter Conversation Triple Dataset and the OpenSubtitles dataset. Results show that MMI models outperform baseline SEQ2SEQ models in terms of BLEU scores and diversity. Qualitative evaluations with human judges also support the effectiveness of MMI models in generating more diverse and interesting responses. The paper concludes that the use of MMI as an objective function improves the diversity and quality of responses generated by neural conversation models. It also highlights the importance of considering both the likelihood of responses given inputs and the likelihood of inputs given responses in generating diverse, conversationally interesting outputs. The implications of this work extend beyond conversational response generation to other neural generation tasks.This paper introduces a diversity-promoting objective function for neural conversation models. Traditional sequence-to-sequence (SEQ2SEQ) models, which maximize the likelihood of outputs given inputs, tend to generate safe, generic responses like "I don't know." The authors propose using Maximum Mutual Information (MMI) as the objective function instead, which measures mutual dependence between inputs and outputs. Experimental results show that MMI models produce more diverse, interesting, and appropriate responses, leading to substantial gains in BLEU scores on two conversational datasets and in human evaluations. The paper discusses the limitations of traditional SEQ2SEQ models in generating diverse responses and proposes MMI as a solution. It outlines practical strategies for using MMI in neural generation models, including the use of a hyperparameter λ to control the penalty for generic responses. The authors also address challenges in applying MMI, such as the difficulty of decoding using the MMI objective directly, and propose solutions like using a modified language model and reranking N-best lists. The paper presents the architecture of SEQ2SEQ models, including the use of LSTM networks for encoding and decoding. It describes the MMI criterion, which aims to balance the likelihood of responses given inputs with the likelihood of inputs given responses. The authors also discuss practical considerations for implementing MMI, including the use of a modified language model and reranking N-best lists. The paper presents experiments on two datasets: the Twitter Conversation Triple Dataset and the OpenSubtitles dataset. Results show that MMI models outperform baseline SEQ2SEQ models in terms of BLEU scores and diversity. Qualitative evaluations with human judges also support the effectiveness of MMI models in generating more diverse and interesting responses. The paper concludes that the use of MMI as an objective function improves the diversity and quality of responses generated by neural conversation models. It also highlights the importance of considering both the likelihood of responses given inputs and the likelihood of inputs given responses in generating diverse, conversationally interesting outputs. The implications of this work extend beyond conversational response generation to other neural generation tasks.