I have always been very curious about ways to widen the impact of natural language processing on the world. What would be an NLP application that would change the ways of millions of people in a positive way? The first answers coming to mind are obvious: machine translation, speech recognition, or maybe natural language understanding?
However, what about text generation? Let me explain why I am so interested in it: writing, and especially, creative writing is one of the most advanced (and to a large extent mysterious) activities of the human brain. The mechanics of the creative writing process is not well studied yet. However, there have been some interesting attempts to understand it here and here. Inviting a machine to model the creative writing process sounds less mysterious if considered in terms of commonly used machine learning concepts. From this perspective, creativity is considered nothing more than just a loss function within an end-to-end modeling framework, optimizing the writing style, scope, and meaning of writing for a particular metric (for example, conversion rate).
We live during an amazing time in history, where AI is changing the world around us at tremendous speed. There are some impressive attempts to generate copywriting for commercial users (look at Copy.AI, FlowWrite, or Headlime, for example) based on the groundbreaking achievements in the field of pre-trained NLP models (look at GPT-3 and T5 models). However, the majority of these products address creative writing, focusing on quite big problems: email writing, marketing copy creation, or web content generation.
We have decided to take a look at this problem from a different perspective: by narrowing down the scope of the product and finding the niche, which would benefit from the automation of text writing most. Our preliminary conversations with potential technology users brought us to the field of digital marketing and even more concrete to the Google Ads writing. There were two reasons for our choice: a huge potential market and naturally limited content with often well-predefined context.
Over the course of the past months, we have run hundreds of experiments trying various configurations of different pre-trained models, and we think we have found a set of parameters and technologies which allow us to generate creative Google Ad descriptions staying to the point of the offered service or product and being almost endlessly creative at the same time. Our product also involves humans in the loop who shortlist ads they like most and block phrases that are not supposed to appear in the ads (name of competitors, for example).
If you are curious about the WriteWith.AI product, you are more than welcome to give it a try. Please be our guests.