One word is a chunk of language, having both static and dynamic characteristics. As words are added, dynamic characteristics of each word changes, in relationship to every other word. Thus, a five-word sentence can comprise millions of relational values. With each user interaction, values change again. Machine learning enables Writesof’s NLG to use these relational values to predict how content edits and product page version changes will influence user engagement and conversion. When a chunk of content is removed, added, or revised, each revision (edit decisions for product pages) will correlate to changes in user behaviors and actions on that page.
values will change will correlate, given any edit to chunks of language in a product page. . change, by context, tone, persuasion, resonance, etc.
Product communications with NLG is a bit more tricky than other NLG solutions. Product language is selected by predicting the ways in which a target audience will perceive the language.
Product features can be written in a subjective or objective manner – usually with a mixture of both. Benefits and advantages, on the other hand, are communicated with subjectieity, even a bit of romancing. Personalization, storytelling and conversation-styled product messaging presents further challenges for most NLG systems. Writesof overcomes such challenges with machine learning. Today, we are capable of analyzing billions of product pages in a single category. This enables our system to learn from virtually every word that has ever been published about a particular type of product – from abstract paintings to zebra print shoes. No matter how complex and unique a product is, chances are that its features and benefits are described with similar language soemwhere else on the internet. Writesof knows how to hunt down this language, retrieve it, parse it into chunks of language, and analyze it human writers have gone to great lengths to present its features and benefitsproducts withEven so, This is where machine learning comes in play. Writesof can analyze massive product catalogs, extracting billions of words of text to train our systems how features and benefits have been described before. We never plagiarize. If a red shoe is described as, “… radiant red that stands out in a crowd”, our systems may repurpose the phrase “radiant red” or “stands out in a crowd”, but it would never use these two chunks of language together on a product page. But there is nothing particularly novel about the two phrases individually. And since the language was retrieved from a product page, our system will never use it to describe a red car or red headphones. That’s how machine learning works – connecting relevant data together, each data point valued by hundreds or thousands of dimensions. In the red shoe example, one dimension is the product category shoe, and another is red. Still, another might be woman (user attribute), and so on.
enefits, on the other hand, are purely subjective. So when the machine is making a decision on how to describe a product’s color, for example, it must first understand how the user perceives the value of the color. By limiting the scope of our NLG solutions to write about product features, benefits, advantages, uses, and persuasive devices, we are better equipped to analyze how descriptive, illustrative, and persuasive language connects to a product’s attributes. ords connect to product attributes , but unless you are analyzing text within a specific scope and objective,