The paper investigates the construction of open-domain, conversational dialogue systems using generative models based on large dialogue corpora. The authors extend the hierarchical recurrent encoder-decoder (HRED) neural network to the dialogue domain, demonstrating its competitiveness with state-of-the-art neural language models and back-off n-gram models. They explore the limitations of this approach and show how performance can be improved by bootstrapping learning from a larger question-answer pair corpus and pretrained word embeddings. The study focuses on non-goal-driven systems, which can be deployed for tasks like language learning or entertainment, and can also serve as user simulators for goal-driven systems. The authors introduce the MovieTriples dataset, based on movie scripts, to evaluate their models. Experimental results show that the HRED model outperforms both n-gram models and baseline neural network models, with significant gains achieved through bootstrapping. The paper concludes by discussing future work, including the need for larger dialogue datasets and the exploration of other speech acts.The paper investigates the construction of open-domain, conversational dialogue systems using generative models based on large dialogue corpora. The authors extend the hierarchical recurrent encoder-decoder (HRED) neural network to the dialogue domain, demonstrating its competitiveness with state-of-the-art neural language models and back-off n-gram models. They explore the limitations of this approach and show how performance can be improved by bootstrapping learning from a larger question-answer pair corpus and pretrained word embeddings. The study focuses on non-goal-driven systems, which can be deployed for tasks like language learning or entertainment, and can also serve as user simulators for goal-driven systems. The authors introduce the MovieTriples dataset, based on movie scripts, to evaluate their models. Experimental results show that the HRED model outperforms both n-gram models and baseline neural network models, with significant gains achieved through bootstrapping. The paper concludes by discussing future work, including the need for larger dialogue datasets and the exploration of other speech acts.