Gajus - Fotolia
When Grammarly's writing platform was first released in 2009, it was a relatively rudimentary program. The software could check writing samples for spelling and common grammatical mistakes, perform a simple scan for instances of plagiarism, and offer synonym suggestions.
A decade later, Grammarly offers a range of advanced services, including alerting users to passive writing or faulty parallelism. Powered by a variety of Grammarly AI and machine learning algorithms, the platform is continually evolving as its millions of users submit writing samples and feedback.
The San Francisco-based vendor recently on May 2 released an update for Grammarly Premium members that adds a new feature that scans for consistency in a piece of writing; it can detect multiple variants or styles within the same document.
This Q&A with Ayan Mandal, vice president of product, and Joel Tetreault, director of research, explores that update, how the Grammarly AI algorithms work, and what the company is planning next.
How do Grammarly AI algorithms use a person's writing samples to help improve their writing?
Ayan Mandal: Grammarly's algorithms are trained on a huge volume of text -- much more than any single person could generate. We do rely on our users collectively to help us determine how useful our suggestions are.
For example, if many users consistently ignore a particular suggestion, that tells us that they don't see that suggestion as an improvement. Over time, our team can make adjustments to our algorithms to make the suggestion more consistently helpful to more people.
Often, what's really helpful for people is tailoring feedback to specific situations. We have several ways of doing that. For example, in the Grammarly Editor users can specify whether they want to sound formal or informal. Another example would be that when someone is writing on a website like LinkedIn, the Grammarly extension automatically provides writing suggestions that are most appropriate for business writing.
We also send our users a weekly progress report email to give them an idea of how their writing compares to that of other Grammarly users. We focus on mechanical accuracy, vocabulary, and writing productivity. Users can also see the top three issues Grammarly flagged in their writing over the past week, as well as links to articles about many of those issues.
How have the Grammarly AI algorithms changed or improved over the last few years, and how have linguistic researchers helped to train them?
Joel Tetreault: Our algorithms are constantly changing and improving with the help of our research team. Grammarly's computational linguists and deep learning scientists and engineers design algorithms that learn the rules and hidden patterns of good writing by analyzing millions of sentences. Our AI analyzes each sentence and looks for ways to offer better suggestions, whether it's verb tense, word choice, or sentence structure.
Ayan MandalVice president of product, Grammarly
Anytime you hit "accept" or "ignore" on a suggestion, Grammarly gets a little bit smarter. Our team is constantly reviewing suggestions with high ignore rates to make necessary adjustments to improve Grammarly and make it more helpful to users. These nuanced adjustments go far beyond just spelling and grammar. We believe clear and compelling communication must take into account wordiness, vagueness, poor word choice, gnarly sentence structure, and even plagiarism.
Over the years, we've added writing checks to address all of these critical pieces to effective communication, and this new consistency check is the latest installment to help users catch inconsistencies in their writing to sound polished and professional.
How do the Grammarly goal settings, which include experimental features such as emotion and intent, work?
Mandal: Our goal settings customize the type of feedback we give our users. For instance, when a user selects "Creative" as a domain, Grammarly won't flag sentence fragments or other more creative stylistic flourishes. If a user selects "Academic," we detect more strict issues such as "who" vs. "whom" or the use of personal pronouns, which may be frowned upon in formal academic writing.
The goals marked experimental don't currently affect the feedback we provide to users.
Generally, what is the writing level of most users? Do they tend to improve with help from the platform and the Grammarly AI algorithms?
Mandal: The skill level of Grammarly users is self-reported. The majority of our users consider themselves intermediate writers.
Can you explain the new consistency checker in the May 2 update? What other updates can we expect this year?
Mandal: Grammarly's new feature will help by calling out inconsistencies in spelling, punctuation and formatting throughout a piece of text. For example, there is no real difference between 9 a.m. and 9 AM, but writing in a consistent manner will help your document look more polished and professional.
When you're writing in the Grammarly Editor, Grammarly Premium's new consistency checks will alert you when they detect multiple variants or styles within the same document. They'll even ask you which style you want to use and allow you to apply it throughout your document with a single click.
We have a number of product feature launches coming up in 2019, all of which will help users with the clarity and effectiveness of their writing.
Editor's note: This interview has been edited for conciseness and clarity. Also, news writer Mark Labbe has a paid Grammarly Premium account.