The language correction tool Grammarly receives text content of all kinds every day. In the interview, Grammarly’s Head of ML and NLP Products, Timo Mertens, explains how the AI tool deals with content that intentionally breaks the rules of correct language, but also how to handle text with questionable content.
Timo Mertens is Head of ML and NLP Products at Grammarly.
Language thrives on playing with it. Even if there are currently no concrete plans, you indicated in the first interview that you would expand your tool to more languages than just English – German in particular invites you to experiment. How does Grammarly not want to flexibly put a stop to the poetic texts?
That’s a good question. Of course, some writing advice applies to just about any situation, such as correct spelling or punctuation. But some other characteristics of the writing, such as the style or the level of formality, depend heavily on the context and the situation – and even more on the goal of your message.
Grammarly has a feature that allows users to set goals for their writing, allowing them to choose the audience, level of formality, and range of text to get tailored writing suggestions. There are six options to choose from: general, academic, business, email, creative, and casual. Each of them affects the type of suggestions that the user receives.
Regarding your questions about less formal and creative writing styles, the user can choose either the casual or the creative mode for their text, depending on the objective. The casual style is intended for informal texts, such as personal communications. The program then ignores most style issues. For example, abbreviations, the passive voice, informal pronouns, and so on are not emphasized. The creative focus offers the most revealing style. It detects grammatical, punctuation, and spelling mistakes, but also leaves leeway for those who want to intentionally bend grammar rules to achieve specific linguistic effects. In the creative field, sentence fragments, missing subjects or verbs, wordy sentences, colloquialisms, incomplete comparisons, and so on are not flagged.
Grammarly is writing assistance software. This means that it helps users to optimize their own writing in English in every respect. The goal is to enhance human capabilities, not to replace humans. Therefore, the program’s proposals are always just options to consider – not enforced regulations.
Controversial terms sometimes throw language corrections quite out of step. How do you deal with sensitive terms that some find offensive and others find completely normal?
As a company providing AI-enabled product offerings, we know we have a responsibility to develop a product that does not reinforce or perpetuate prejudice. We have in-house teams and structured workflows focused on evaluating the inclusivity and sensitivity of our software. These teams are working to improve how we support responsible and inclusive communication. At the same time, we work with team members across the organization to respond to any issues that arise.
Here are some examples of this development: Grammarly researched possible gender bias in the autocorrect feature before launch; to avoid xenophobic alternatives, we propose official naming of COVID-19 instead; we’ve also developed a number of suggestions for using inclusive LGBTQIA+ language, including support for the singular “they”.
Our work on the topic “Sensitivity” begins with the annotation of the data. For example, one area we focus on is finding ways to mitigate potential bias or stereotypes through strict annotation of the datasets. Also, in the pre-release phase of new features, we conduct thorough assessments to closely investigate potential issues before features are released to users. And once published, we listen carefully to our users and respond with understanding to any issues that arise. We have a team that works cross-functionally to assess issues and develop solutions that align with our values. Direct exchange with our users helps us to continuously improve how we can support conscientious and inclusive communication.
An important factor is that all Grammarly suggestions are accompanied by explanations of the grammar rules or language exercises that influence the recommendation. This allows users to learn the more they use Grammarly. Most importantly, users always have the final say, can weigh the options and decide whether or not to apply a suggestion.
Not every text that you want to have corrected is directly intended for the eyes of others – some contain confidential company information, others maybe embarrassing love letters. Where do you get the data from the uploaded files and where is it processed?
User trust is the basis for conscientious product development. It is a great responsibility to create a product that over 30 million people rely on every day. For this reason, data security is at the heart of our company and our product. We have not in the past, nor will we in the future, sell or rent user data – period. We also do not pass on any information to third parties so that they can advertise to our customers. Everything we do is done with confidence in bringing value to our users. In order to create meaningful value with our product, we must nurture and sustain an ecosystem built on trust; so much so that it’s built into our business model: Grammarly makes money when customers use our world-class communication tool and decide to purchase one of our paid product offerings (like Grammarly Business or Grammarly Premium). So, as an AI company, we take our responsibility as stewards of data very seriously.
We practice “privacy by design”: For example, Grammarly only checks the text requested by users. Verification of text is not performed in sensitive fields (e.g. in credit card forms, password fields, URL fields or similar where private information is provided). Our users can also see all data associated with their accounts.
Because users trust the program with their words and data, we have safeguards in place to ensure their data is safe. We use a combination of technical, physical and logical safeguards to ensure the safety of user data. This includes encryption, secure network configuration and data transmission, restricted access for team members, and other measures.
We have always worked with extremely high safety standards. A few years ago, as we were growing our business within professional teams, we realized that we needed to strive for the certifications and credentials to carry that commitment externally. Since we already had the appropriate basis, we were able to secure several certifications in a very short time. Our security practices are backed by enterprise-level certifications and compliance, such as the SOC 2 (Type 2) report and ISO certifications for security, privacy, availability and confidentiality. These awards have helped us grow our B2B customer base to 50,000 teams, including companies like Siemens, Zoom, Okta and Expensify.
Finally, at Grammarly, one of our five core values is ethical. We always hold ourselves to a high standard of doing what is right, even when no one is looking.
Mr Mertens, thank you very much for your answers. Last week we spoke to Grammarly about the technical implementation of language correction.
In the “Three Questions and Answers” series, iX wants to get to the heart of today’s IT challenges – whether it’s the user’s point of view in front of the PC, the manager’s point of view or the everyday life of an administrator. Do you have suggestions from your daily practice or that of your users? Whose tips on which topic would you like to read in a nutshell? Then please write to us or leave a comment in the forum.
(jvo)
