Humor generation and classification is one the hardest problem in the area of computational Natural Language Understanding. Even humans fail at being funny and recognizing humor. In this project, we attempt to create a joke generator using a large pre-trained language model (GPT2). Further, we create jokes classifier by fine-tuning pre-trained (BERT) to classify the generated jokes and attempt to understand what distinguish joke sentence(s) from non joke sentence(s). Qualitative analysis reveals that the classifier model has specific internal attention patterns while classifying joke sentences which is absent when classifying normal sentences.

For Non-joke sentence

For Joke sentences, there's a visible 'X' pattern which validates the setup-punchline structure.

Detailed view of the Joke attention pattern -

1. Python: 3.7
2. pytorch: 1.2.0
3. Cuda 10.1.243

Some experiments were directly carried out in Google Colab

Pretrained models are from https://github.com/huggingface/transformers

GTP2/fine_tuning.py is mostly the huggingface's run_lm_finetuning.py except for the Dataset Class.

python fine_tuning.py \
	--output_dir output \
	--model_type gpt2 \
	--model_name_or_path distilgpt2 \
	--do_train \
	--train_data_file short_jokes_even_shorter.csv \
	--per_gpu_train_batch_size 5 \
	--save_steps 1000 \
	--num_train_epochs 10

python run_generation.py \
    --model_type=gpt2 \
    --model_name_or_path=output/ \
    --top_k 50 \
    --top_p 1.0 \
    --temperature 0.3 \
    --prompt "Why did jonh call the cops?"

An improved version is in the jupyter notebook along with the rest of the code for generation, analysis and classification.

Jupyter notebook is run on Google Collab, any extra package requried required is being included in the notebook itself.