ilyasu123
commented
8 years ago Scraping: add a comment in the instructions to obey robots.txt or whatever request that the website has with respect to scraping. I heard that reddit has exposed a few terabytes of reddit comments for download -- you should advice to get that. In fact, I believe that your problem will have a much greater chance of being solved if you preprocess the data yourself!
Step 3 cannot be done with today's technology. The network needs to output text. Text is discrete. So unless you implement a policy gradient algorithm (which is hard but probably doable), this step will just be impossible. I suggest that you request to train a more conventional language model, like the one in https://arxiv.org/abs/1602.02410. I also recommend that you do not include a critic network.
Likewise, I recommend providing clarity regarding step 4. The easiest approach is to use Mechanical Turk, which is not free. One possibility is to create a twitter account.
For this task, training a language model is a much more promising approach than using the stochastic neural network.
I recommend removing the requirement to keep track of previous jokes. To summarize: the goal should be to train the biggest baddest LSTM on jokes.
To improve the LSTM and to make it more interesting, it would be nice if for each joke, you could extract a number of keywords that describe the topic of the joke, or better yet, genre, and train the LSTM to maximize the likelihood of the joke given its topic and genre. If you preprocess the data yourself and put it online someplace, the chance that someone will solve this problem will increase substantially.
iaroslav-ai
Create a deep neural network capable of generating funny jokes.