Dynamic NLG (natural language generation) is a technology that uses language templates that dynamically output text by matching topic features and values with context strings. Topic features and values are variable data that describe some measurable or pseudo-measurable aspects of a topic. Context strings are the various words, phrases, and sentence structures (strings of text) that add communication clarity to topic features.
Dynamic NLG is used to generate sports articles, financial reports, statistical analyses, and performance evaluations. Associated Press, Yahoo Sports, and Forbes are just a few examples of media outlets that use Dynamic NLG to write articles. Organizations also use it for internal communications to staff, generating statistical and performance reports and memos.
Since the technology automates the narrative generation process, media articles and internal business reports can be generated and published in near-real-time, triggered by data events or updated feeds. One example of this is stock data and public company financial reporting. Systems can process this type of standardized data and generate high quality stories that describe a company’s financials, in comparison to other companies in the industry, including historical analysis. Standardized text – text that describes a change in data – e.g. “the stock price went up”, “the team scored more touchdowns” – there are only a few ways to describe changes in this type of data.
The more advanced Dynamic NLG technologies can extract deeper and more enriched dimensions of language to describe data. Even with standard structured data, Dynamic NLG can extract communicable insights, patterns, and trends – the hidden features of data that tells a more unique and human-quality narrative. Quantitative data can be transformed into qualitative data and binary assertions that add dimensional qualities to the communication produced by the NLG.
A simple example of this can be demonstrated with weather report generation. If the weather has been in a cold trend (+/- prior cold days in succession), and tomorrow is expected to be warm (tomorrows expected low/high temperature increase), Danamic NLG is intelligent enough to open a weather artcicle with, “Leave your jackets at home because tomorrow’s weather is going to be gourgeous”. In this simple example, numerical data was utilized by the NLG system to determine the appropriate text output.
With more advanced Dynamic NLG systems, as many as hundreds of data points can influence word selection and sentences that describe changes in data. This type of qualitative reasoning with quantitative data is, however, limited to extracting mostly metric-based data. With product sales communications, language must be emotionally appealing and written with subjectivity. This is where Writesof has focused our efforts, Dynamic NLG systems that can process subjective data, such as color, texture, style, use, and emotional appeal.
Writesof’s innovative brand of NLG was built on a foundation of Dynamic NLG, but we added subjective writing capabilites that empower writers and marketing professionals to artistically communicate subjective features, benefits, uses, and advantages of products. Also, each product message that our NLG produces is driven by user data, which means that the NLG is more concerned about probabilities of using langauge that influences customer experience, buying decisions and customer relationships.
With advanced machine learning capabilities, Writesof is leading this new shift in NLG whereby communication (engagement) signals from users are received in near-real-time. These data signlals are what drives our NLG’s decisioning – probabilities of language resulting in a positive customer experience, by user. Language is selected and composed, piece by piece, and structured into phrases, sentences, and paragraphs, for the user.
Product pages are an ideal place to collect user behavior data with respect to communications. Social media is also a great source of user data, particularly for discovering shared interests and behaviors among micro-segmented user groups at the community level. We can calculate probabilities and patterns of how a community (e.g., social group) engages with certain types of products or categories and how they engage and respond to page content.
Online shoppers are brutally honest, revealing incredible amounts of data each time they visit a new product page. Each time a prospective or repeat customer visits a page, they tell you what they like and don’t like about your content. Each page’s conversion rate can shift, but even these baseline outcomes can be accounted for, given a change in page content – user experience.
Online retailers know where users come from, what they clicked on to get there, and how they engage at several stages along their customer journey. When the shopper lands on a page, every second of engagement and action reveals details about their on-page experience. In aggregate, this and other types of data extracted along the customer journey is extremely precious to online retailers.
As a Writesof partner, product marketers can use this and other data to analyze every word of text of a product page, by performance – even A/B testing thousands of new product pages per day. This enables our NLG to produce high volumes of content, with each customer’s engagement powering our systems to write better content – truly a continuous improvement model for producing and editing micro-targeted product messages.
NLG systems typically employ a small percentage of a language’s total vocabulary. For example, in English, 170,000 words are used – not a huge variable set. Even large retail websites with over 100,000 SKUs may use as little as 8,000 unique words, less than 5% of the English language. Subjects, topics, objectives, and context, by product category, is substantially reduced to a small brand lexicon. Product communications are further linguistically restricted by brand voice, message tone, author writing style, and communication objectives, to name a few examples. At Writesof, we analyze messages by over 100 variables – linguistics / message attributes and variables, with each word’s location being valued according to its position in the message.
In an NLG system, words are just variables that connect to other words to form text strings, which connect to larger text strings. Given the right message inputs, structuring each string of text together can generate a 10-word advertisement, or a 500-word product page. Each text string is restricted by rules, such as grammar and context rules. Language can also be defined or restricted by region, state, city, community. People communicate differently in New York City than they do in Albany, New York. Restricting language usage by, for example, culture and community (two of many dimensions of a user group) helps markters create messages that speak to customers in a more personalized manner.
NLG systems are able to utilize the input variables (words) and language rules to output fully enriched composititions. Adding user data to the inputs of NLG, creates a new dimension of decisioning that values performance qualities of language. How well does the message resonate? Does it read well with the user? Does it have communication appeal? How does the user engage with page content? What is their response to the message?
Our entire NLG system is built to match audience features with message objectives, with the variables of a message systematically structured into an order that exceeds benchmarked message objectives. If the message objective is to convert more customers, the system determines what “pieces” and “qualities” of messages have a tendency to correlate with higher conversion rates. With the ability to publish dozens of product description versions, we can also A/B test the versions and validate results with control-variable tests. This science only works well with thousands of unique products and preferably hundreds of thousands of unqiue page visits, total, per day.
user engagement signals that correlate with higher conversion rates. By anatransforming user data into connected and words and phrases which are ordered into product messages directed at specific users, with a specific response objective. When response objectives exceed performance benchmarks (e.g., conversion rate), the system is “rewarded” with additional editing lattitude. When performance declines, the system modifies (autonomously edits) language to a state that is most probable to exceed performance benchmarks. The system is more liberal with language editing when page performance improves and more risk averse (tendency to revert to prior language use) when page performance declines.
the type of language used in describing business reports, finance articles, and sports stories are quite similar, in that they describe measurable events that occur on a timeline.
Writing about physical products that need to make an emotional connection with a consumer is an entirely different story. Business, finance, and sports, as for written content goes, tend to use highly standardized language. There are only so many ways a writer can say, “he hit a home run in the 5th inning.” and even the most rudimentary natural language technology can handle all of the variants without oversight from a human writer. That just about sums up the most advanced NLG systems
Time. Data. Resources. Your audience is sitting out there waiting to be informed, compelled, and inspired.
Today, NLG can produce narratives that learn from your audience and adapt, on message, to shifts in behaviors and interests.
Software that interprets data is only as useful as the writer’s ability to uncover insights and construct narratives that consider the myriad of audience engagement metrics that assign valuation to the overall message, as well as each component of a message.
Writesof can easily interpret the KPIs that correlate with every aspect of a narrative, comprehensively, in real-time. Every message delivered to an audience has value. Was there a response? What was the response duration? What was the resulting action by the respondent? How did respondent behavior compare with prior responses?
But for a writer to look at billions of correlations on a spreadsheet or dashboard would be an impossibility. This technology gap can only be bridged with AI that not only interprets the data, but also takes on the role of the writer. Otherwise, you would need to sift through mountains of data, one narrative at a time, to formulate an optimal messaging strategy. And if your messaging is time sensitive or market sensitive, you would need to continuously adapt to the audience, analyzing incoming data feeds while simultaneously formulating new messages that meet audience needs. Enter Writesof NLG.
The software is programmed to reward high value edits with more experimentation options and restrict editing when it makes a poor decision. Human users can easily override edits or revert changes. Our technology performs best where baseline product pages exist, since it can easily compare key performance indicators from the human-written copy compared with AC’s copy.