Personalized Message Generation With Natural Language Generation Personas

Each unqiue customer is telling you how they want to be spoken to. Writesof Natural Language Generation (NLG) systems are comprised of one hundred unique NLG Personas – brand-tuned writing machines that create content by listening to users and understanding how different users prefer to engage with content.  Each Persona has a unique communication style and writing ability, restricted by factors such as vocabulary, reading level, community dialogue awareness, etc.

Before an NLG Persona is deployed to go to work for a client, it must first be tuned to write product messages in a brand’s unique voice.  Not every Persona is suitable to write content for a particular brand or to a particular audience, but each unique Persona can be cloned and re-tuned to communicate with its own unique style, vocabulary, and capabilities.  Your NLG Persona must be brand-tuned to communicate in a particular voice and house style.

The brand tuning process analyzes brand communications inputs, for example, text and metadata from 20,000 unique product pages on a brand’s website.  First, we parse and tag each page’s text and metadata by exact location.  Word location data enables our system to understand how all of the different words on a page are used together, by relationship and proximity.  It also enables it to understand a message’s purpose, by content section.  After our systems ordinally maps all of the text, it begins classifying each chunk of language (words, n-grams, phrases, sentences) by their respective qualitative and quantitative linquistic characteristics.  This enables our systems to understand what type of language is used, where it is used, and how it is used, grammatically and stylistically.

In the next stage of language processing, our systems process user profile and behavioral data, each user’s interaction with different parts of text on a page.  Since the system understands the different forms, functions, and communication objectives of each chunk of language on a page, it can begin analyzing how different users engage, interact, and respond with different chunks of language.  It learns based on weighted user engagement actions and outcomes (feedback) that signal the system as to the quality (resonance, value, response) of the message, piece by piece.  In this way, our NLG learns to communicate more effectively in a similar manner to how humans learn – via communication feedback.  Unlike a human communication-feedback, our system is commanded to modify its use of language in a manner that improves communication with the user, to meet a specified objective – better user engagement and increased conversion rates.

Currently, Writesof’s NLG Personas are suitable for product markters who want to produce and edit high volumes of performance-driven product sales communications –  from social media posts, to promotional emails, to fully-enriched product descriptions.

Writesof’s Personalized Message Generation™ (PMG) technology is a type of dynamic Natural Language Generation (NLG) technology that creates user-tailored product messages that are systematically composed by probabalistic ability to to result in increased engagement and response for a particular user or user group.  An NLG Persona is “trained” to write in a particular brand’s voice by analyzing brand-sourced text, such as website product pages, meta data, promotional emails, and reviews.  It can also be trained by analyzing external messaging sources, such as social media group (community) dialogue, compeitor messaging, and product articles.  It can learn exactly how a brand communicates with users and how users interact with brand or product-category messages.  For example, it can learn how a brand communicates to users by analyzing 20,000 product descriptions on a large e-commerce site.  As long as the brand our author-source contains, at minimum, millions of words of text, Writesof NLG Personas can use the text as a foundation for language modeling, constrained by brand voice and communication style.

Our PMG system is designed to learn how unique users engage with unique messages.  Each user interacts with a product sales message in a unique manner.  PMG analyzes multi-dimensional qualities and behaviors of users and calculates intersecting correlation values between user qualities, language qualities, and behavioral actions.  We utilize advanced computational linguistics systems built in a neural network, capable of predicting language’s effect on a particular user or user group.

PMG analyzes language according to its predicted communication value with a paricular user or user group.  We capture user data correlations with language data, which teaches our systems to understand how changing different parts of a message will probabalistically result in changes in user behavior.  Product messages are systematically composed by predicting message engagement outcomes with different user types.  Our PMG systems can also analyze the many different ways that users engage with with language along the customer journey.  The systems learn from user feedback data to semi-autonomously produce and edit messages, with each unique message being adpated for specific communication to a unique user or user group.

Dimensions of PMG
  1. User – unique page visit by a person or a bot – profile and behavioral data
  2. Language – every word and phrase combination, usable by a brand.
  3. Message – language forms and functions, with user-communication objectives
  4. Engagement – unique user’s interaction with any part of a page or message
  5. Brand – word and language usage preferences and limitations – voice, style
  6. Persona – all possible language usage and vocabulary of a style-tuned NLG system

NLP Persona Adaptive and Expansive Language Capabilities

A brand-tuned NLG Persona has the ability to expand its vocabulary and language usage by probabalistically discovering similar langauge forms and communication styles from a variety of sources, including our own internal linguistics libraries, and also external sources such as competitor sites, social media, and product articles.  Our internal systems enable us to filter through language forms and classify chunks of language using hundreds of linguistic criteria.  This enables us to efficiently execute human oversight in the Persona tuning process.

Our systems utilize user data to influence PMG decision-making.  The system can probabalistically determine which version of a product page a user should land on.  It can analyze how unique users engage with content along the customer journey, especially on the landing page.  It also analzes user data to determine what types of language is most likely to create resonance, engagement, and result in conversion, at the individual user level.

An NLG Persona learns to write like a human by analyzing source text and classifying it by linguistic attributes such as diction, grammar, style devices, and sentence structure. With PMG technology, the machine Persona learns how a human writer communicates about different types of products,  product features, benefits, advantages, and uses.  The brand-tuned Persona can actually predict how a human writer would communicate a product sales message by understanding the various ways a human writer informs and persuades customers, with respect to product attributes.

The Persona automates the organization of language compositions by creating adaptations of dynamic language models.  It creates and tests new product page variants and learns which types of language performs best with certain user types on a probability-performance basis.  The PMG system makes decisions on word choice and message mechanics by probabalistically determining which language usage is most probable to achieve the message objective, which is benchmarked by engagement and conversion metrics.

As the system analyzes thousands, or preferably hundreds of thousands of user engagement sessions, it learns how different types of users engage with different dimensions of a message.  As it analyzes increasing numbers of user-message interactions, it begins to improve its predictive capabilities in understanding different types of users and their communication preferences – language that compels user interest and conversion.

When the Persona makes good decisions, it is rewarded with additional creative and experimental language editing freedoms.  When it makes poor decisions, it is restricted to more conservative language edits, at times reverting back to baseline (benchmarked) language choices.  If a product page performs poorly after reverting to a conservative message state, it may be sent to an oversight queue, where human editors can make adjustments to the message.

Each human-implemented edit actually makes our system smarter, as each edit is inputted into the Persona’s language library where it is tagged, valued, and classified for future use.  All language sources are tracked, at all times, with our systems.  We use language data sources (by page) to find correlations in message value by source, message style, vocabulary, etc.

The PMG system can continuously analyze how each word connects to other words, to form a message – by writing mechanics, sentence structures and grammars, word choice, communication styles, message tone, and communication objectives, etc.  Language usage (how words connect to form a message) is determined by a message’s linguistic properties and values, for example, how often it is used in a particular social media group.  More importantly, language values are determined by user engagement and conversion values –  as millions of value correlations for a single product sales message – this is dependent on user data dimensions, data quality, and data volume.   Each user will engage differently with different parts of a message.  The PMG probabalisitcally determines how to compose a message by probabalistically predicting which components of a message, pieced together in a certain order, will result in increased user engagement and sales conversion.

Each of our NLG Personas can create uniquely styled narratives aimed at different audience types, on a wide rage of subjects and topic objectives.  This means that one unique NLG Persona can create product descriptions for shoes, soccer balls or dog food.  Since Writesof currently offers over 100 unique NLG Personas, we can help product marketers select and test different Personas for specific product categories, or A/B test multiple Personas with a product category.

Each NLG Persona has a unique set of writing skills and qualities.  For example, one Persona may be trained with influence from one or more social media groups, learning how users in each group communicate, what words they use, and in what context.  Another Persona may be trained to only write content at a 4th grade reading level.  Still, another may be trained to communicate to a 55+ audience in Houston, Texas.  As marketers improve their ability to classify unqiue user data, they also improve their ability to personalize the customer experience.

An NLG Persona will have a particular vocabulary (restrictive lexicon), able to produce messages at a particular reading-level, with a particular style of communication – tone, persuasion, angles, sentence structure, etc.

Unlike human writers, each Persona is capable of producing content 24 hours per day, 7 days a week, capable of generating thousands of new or versioned product descriptions every day.  It can also edit content or change content distribution decisions in near-real-time, computing live feed user engagement data with millions of respective text strings.   This enables Writesof to deliver a personalized shopping experience for each and every unique user, with near-real-time content edits and updates and user-targeted product page version-dependent redirects.

Our pseronalized messaging technology utilizes consumer behavior data signals to autonomously train each NLG Persona individually.  A Persona communicates with each user in a unique manner.  With user-personalized messages that can be delivered to a target user, in real-time, your organization can publish millions of 100% unique, goal-oriented narratives that are strategically descriptive and persuasive.  In a matter of hours, our system can launch as many as 100 brand-tuned NLG Personas, each delivering highly differentiated product messaging, as many as dozens of product page versions for each SKU.

Each Persona is trained to model a human’s writing style, learning about what types of language describes certain products from published product descriptions, ads, articles, and social media.  Our systems analyzes language by word-string relationships and word locations within content.  It analyzes language usage on a probabalisitc determination model by examining word usage cohesion by frequency.  by word cohesion, with respect to subject attributes and communication objectives, by each language string in a page (document) section.  It tends to be intelligent enough to ignore language strings that are written by different authors unless they happen to work well together.  Thus, our technology can actually identify and predict language mechanics and form from two distinct authors.  It does this by analyzing linguistic parameters, with respect to product attributes and benefits, with respect to the communication objective, which is always determined by audience/user characteristics.  The user is the foundation for which the NLG Persona makes decisions in the language generation process.

Writesof is an author-tuned NLG System that uses audience data to make systematic language decisions.  Audience data is the fuel of any type of cognitive marketing automation technology.  User data makes it possible for a machine to learn how to write, tell a story, and stay on message.  With AI, our NLG can adapt to audience feedback (user engagement and behavior), creating hundreds, even thousands of unique product narratives, targeting individual users.   Like other Natural Language Generation systems, ours can run repetitive writing tasks on autopilot, empowering your organization with communications and messaging, on demand.  Publish first.  Edit first.  Adjust first.  Differentiate first – for each unique user visit.

Face it.  Your customer data is rich with insights, but putting user and audience data into action is not exactly a seamless process.  At best, it empowers your marketing team with more precise targeting capabilities, but it mainly only helps at the human-decicion level.  Most advanced marketing automation solutions have not yet been able to deliver real-time customer-facing product pages that adapt to the user’s needs.  If you are one of the few organizations that are personalizing customer experience with automation (such as display ads), you are a step ahead of the competition.  But the landing page experience is still static.  Your landing pages don’t yet match your targeting efforts and potential cusotmers bounce because they don’t feel the connection.  Meanwhile, your audience is becoming more differentiated, more picky, and a faster-moving target that never lacks in demand for content that speaks to them and informs on a personalized level.

Writesof works with both structured and unstructured data sources. We also provide partner clients with supplemental subject matter and audience-level data, including product language data from category-compeitors.  We use such data internally for system training.   Our NLP systems can learn from demographic, social, and behavioral data, which trains our system on linguistic characteristics that tend to perform well with particular users and user groups.

Prospective customers are activity searching for informative product information.  Most want to be fully informed, compelled, inspired, and assured – in differing ways – before they decide to purchase. As a Writesof partner, you can add substantial value to your internal and third-party customer data, creating personalized product messages that capture your audience’s attention.  Customers will remember high quality content.  They will buy more and they will come back for more.  They are more likely to become brand-loyal fans, promoting your brand and products on social media.  Communications has always been the best path to relationship building – this is especially true with online product communications.

Today, NLG can produce narratives that learn from your audience and adapt to different customer behaviors and interests. Software that interprets language data is only as useful as its ability to interpret language relationships to user profile and behavior data.  If not, the data is simply an indicator of language performance with a single audience.  This has very little value, or at least it will in the near future, as message personalization becomes more standard practice in online retail.

A single message delivered to different users will have differing values of linguistic properties.  Each word is valued, in relationship to other words on the page, with respect to user engagement and profile data.  In fact, each word will have a different value, depending on the user that sees (or doesn’t see) the word.  A user doesn’t have to see the word on a page for the word to have positive value.  A word’s value can be generated by a number of psychoanalytical factors.  Some words are even subliminally consumed or do not directly resonate with the user, but still add value, as a function or perceived feature of the message.  Example:  A segment of users show better engagement and conversion resonse when they view technical jargon on a page, even if they don’t read it; another segment is turned off by the technical jargon.  Both segments didn’t read every word of jargon, but one tended to respond more positively to it.

With AI, our computational lingusitics systems can easily find billions of correlations between the message and the audience. It would take a human writer years, even with a team of data scientists to process, even high-level insights generated from billions of data correlations.  Word relationships are not binary.  They are multivariate.  A single word on one product page has a relationship to a different word on a different product page, even if it’s on a different website.  Conisder the buyer’s journey, or a product category’s / substitute product’s product page, ad, social content footprint.  Now imagine putting all of this data into a multi-demensional spreadsheet, where a single word can be associated with hundreds of attribute variables, built out by different audience / user dimensions.  This technology gap can only be bridged with AI and machine learning, systems that not only interpret the lingusitc data and user data, but also takes on a personal identify of a human writer or team of writers, within brand voice language constaints.  And if your production schedule is time sensitive or market-response sensitive, a human writer would need to continuously adapt to an increasingly diversified and ever-changing audience, analyzing incoming customer data while simultaneously producing new messages that meet audience needs.