Exploring Seq2Seq For Generating Human-Like Responses on Internet Forums Shayan Sadigh Computer Science Student University of California, Santa Barbara Santa Barbara, CA 93106 shayan@cs.ucsb.edu Abstract In this paper I explore bleeding-edge techniques for generating human-like posts on internet forums and social networks with neural networks. In particular, I frame the problem of responding to internet posts as a translation problem for seq2seq and test the performance of character level and word level sequence generation on this problem. Finally, I propose a novel method for generating sentences where the model directly generates word vectors rather than pick the next most likely token from a dictionary of tokens. 1 Introduction 1.1 Motivation I introduce two problems that this work touches upon. Problem 1: Starting up a new internet forum Social networks and internet forums must attract users in order to generate content and ad revenue for their owners. Paradoxically, users are attracted to forums which are already generating activity. This raises the question: How do new internet forums and social networks ever attract their first users? According to Reddit cofounder Steve Huffman 1 , Reddit solved this problem by creating fake accounts and "seeding" the forums with their own discussions. This created the illusion of an active forum and attracted real users over time. When Reddit eventually had enough users to generate content without the help of fake accounts, they were abandoned. Clearly Reddit’s technique proved effective but it has its drawbacks - mainly that it requires a lot of manual labor. Forum owners must manage many fake accounts and take care to make original and believable posts under numerous personas. Alternatively, forum owners could generate content through crowdsourcing services such as Mechanical Turk, but this comes at a monetary cost. Problem 2: Turing test The Turing test is a test of whether a machine exhibits intelligent behavior indistinguishable from that of a human’s. In the standard interpretation of the Turing test an evaluator writes questions to two players, a human and a machine, who each send a response to the question to the evaluator. The evaluator does not know the identity of the players and must determine whether the response was written by a human or a machine. The test is considered passed if the evaluator’s decisions are no better than random guessing. 1 See Motherboard article.