This is an attempt to create a funny bot trained on a corpus of jokes. The goal of the project is to draw samples which are actually funny, from a character level LSTM model.

Read my blog post for detailed analysis and results.

This project also got coverage in TED ideas: https://ideas.ted.com/can-you-teach-a-computer-to-be-funny/

• Short Jokes: It contains 231,657 short jokes and oneliners. For the language model, csv file of the dataset is processed and written to a seperate text file data/shortjokes.txt with utils/csv_to_text.py.

• F.R.I.E.N.D.S: As a fun task, transcripts of all the episodes of TV Series F.R.I.E.N.D.S are compiled into a single text file of 4.79MB (/data/friends.txt) using utils/friends.py script. The intent is to generate funny text similar to the dialogues in the series. The script is ad-hoc as of now, so contributions are welcome.

• Python - Preprocessing the dataset.
• Torch - Language model is written in Torch.

Navigate to /src/ folder and run the following commands:

python scripts/preprocess.py --input_txt ../data/shortjokes.txt  --output_h5 my_data.h5  --output_json my_data.json
th train.lua -input_h5 my_data.h5 -input_json my_data.json -model_type lstm -num_layers 3 -rnn_size 512

This will start the training session of 50 epochs on jokes dataset and checkpoints are saved in src/cv/ folder every 1000 iterations with names like cv/checkpoint_1000.t7.

To sample data with 2000 characters from the trained checkpoint (say after 3000 iterations), run the following command:

th sample.lua -checkpoint cv/checkpoint_3000.t7 -length 2000

In case of any errors, missing dependencies or more info, refer to torch-rnn.

• Data compiled from utils/friends.py contains many extra headers, which were manually removed. It would be great if this task could be automated.
• It has been attempted to keep the dataset as clean as possible. So, relevant additions to the jokes dataset are welcome.