Text Generation
The past, the present, and the future
Richard Rohla | Laurens ten Cate | Alex Kyalo | Gianluca Sabbatucci | Asier Sarasua | Kamal Nandan | Veronique Wang
WTF is text generation
NLG means generating natural language from a model
Very challenging computational task:
- grammatical complexity
- ambiguity
NLG aims to find an underlying 'language model' from which we can sample to generate text
Initially this was very heuristic based
Probabilistic language models
Probabilistic language models assign a probability for the next part of the sequence.
Conditional Frequency Distribution model
Recurrent Neural Networks
Alex Graves (2013) - Generating Sequences with RNNs
- Good for sequence generation
- Processes sequences one step at the time
- Finds internal representation of sequence instead of exact matches
- Less repetitive generation
- Char vs Word
The future (GANs)
Generative Adverserial Networks
- Discreteness of text data causes problems with back-propagation of error
- SeqGan -- policy gradients + MCM
- We switched to VAEs
VAE
S. Bowman (2016) - Generating Sentences from a Continuous Space
Variational Auto-Encoder
Auto-encoders encode data in a lower n-dimensional space which allows decoding to original data
VAE encodes data not in discrete points but as soft-ellipsoidal regions in this n-dimensional space
Demo
Three models:
- CHAR LSTM
- WORD LSTM
- VAE
nlp2
By laurenstc
