Phrasee AI: Deep Dive

When you see a marketing message as a subject line in your inbox, a push notification, or Facebook ad, millions of neurons in your brain are firing at the same time. You make a split second, subconscious decision on whether or not to open the email, message, or ad.

Phrasee Infographic

Human or Machine?

There are many factors that your brain considers. Simple approaches, such as those based primarily on sentiment analysis or zero-sum emotion tagging, aren’t able to model this complexity. That is where Phrasee AI systems come in.

  • talking

    Phrasee uses an end-to-end deep learning model that automatically identifies thousands of key properties of the raw text. Some of these correspond to familiar concepts such as ‘emotions’ and ‘sentiments’. However, many more subtle language aspect are also encapsulated.

  • brain

    Not all of these parameters can be easily understood and explained, much like you can’t always explain the thought process behind why you made a split-second decision. Phrasee’s Chief Scientist, and architect of the Phrasee method, Neil Yager, PhD, found a way around this.

  • money

    Alongside our team of computational linguists (yes, that’s a real job title), Phrasee has developed a fully automated solution for producing marketing copy that is human sounding, brand compliant, and optimised to get you more impressions, opens, clicks, and conversions.

Natural Language Generation (NLG)

For language generation

Phrasee is at the forefront of NLG technology, producing more brand-compliant, human-sounding subject lines than any copywriter on the planet. Phrasee uses a combination of context-sensitive grammars and recurrent neural networks to ensure that your brand’s voice is captured, that the language sounds natural and fluent, and that it’s unique to you, and no one else.

End-To-End Deep Learning

For performance prediction

Phrasee uses an end-to-end deep learning engine to optimize results on an on-going basis. And this matters. A lot. We use multi-layer neural networks to learn from unstructured text. We don’t need humans tagging individual words and phrases with various sentiments or emotions. That approach has many disadvantages. It limits the language available for you to use, is susceptible to human error and subjectivity, and results in every brand using the same stock language. Furthermore, there is no guarantee that these binary features are the actual drivers of user engagement.

Phrasee’s system builds its own internal representation of language. This multi-dimensional plane considers thousands of interdependent “features” or individual linguistic elements that the model identifies as being important. It’s things like semantics. It’s things like structure. It’s things like meaning. Let’s not forget about punctuation! It’s lots of things. Far too many things for a human to keep track of.

That’s just way too much data for a human to process alone! But, luckily, this is EXACTLY what machines are good at – making sense of large amounts of messy data. Deep learning engines are much better at this than humans are. And that’s why forward-thinking marketers entrust Phrasee with their marketing language generation.

Want to find out more about Phrasee?