Controlling Linguistic Style Aspects in Neural Language Generation
“We explore a simple neural natural-language generation (NNLG) framework that allows for high-level control on the generated content (similar to previous work) as well as control over multiple stylistic properties of the generated text. We show that we can indeed achieve control over each of the individual properties.
As most recent efforts, our model (section 3) is based on a conditioned language model, in which the generated text is conditioned on a context vector. 1 In our case, the context vector encodes a set of desired properties that we want to be present in the generated text.2 At training time, we work in a fully supervised setup, in which each sentence is labeled with a set of linguistic properties we want to condition on. These are encoded into the context vector, and the model is trained to generate the sentence based on them. At test time, we can set the values of the individual properties to get the desired response. As we show in section 6.3, the model generalizes fairly well, allowing the generation of text with property combinations that were not seen during training.
The main challenge we face is thus obtaining the needed annotations for training time. In section 4 we show how such annotations can be obtained from meta-data or using specialized text based heuristics.
Recent work (Hu et al., 2017) tackles a similar goal to ours. They propose a novel generative model combining variational auto-encoders and holistic attribute discriminators, in order to achieve individual control on different aspects of the generated text. Their experiments condition on two aspects of the text (sentiment and tense), and train and evaluate on sentences of up to 16 words. In contrast, we propose a much simpler model and focus on its application in a realistic setting: we use all naturally occurring sentence lengths, and generate text according to two content-based parameters (sentiment score and topic) and four stylistic parameters (the length of the text, whether it is descriptive, whether it is written in a personal voice, and whether it is written in professional style). Our model is based on a well-established technology - conditioned language models that are based on Long Short-Term Memory (LSTM), which was proven as strong and effective sequence model.
We perform an extensive evaluation, and verify that the model indeed learns to associate the different parameters with the correct aspects of the text, and is in many cases able to generate sentences that correspond to the requested parameter values. We also show that conditioning on the given properties in a conditioned language model indeed achieve better perplexity scores compared to an unconditioned language model trained on the entire dataset, and also compared to unconditioned models that are trained on subsets of the data that correspond to a particular conditioning set. Finally, we show that the model is able to generalize, i.e., to generate sentences for combinations that were not observed in training.”