This finding had implications for the feature we were developing. But when an editor suggests rewriting a sentence to sound less aggressive, for instance, the writer may disagree: “I don’t think that sounds aggressive at all.” Grammarly offers these suggestions just as an expert writing coach would-but still recognizes that they’re inherently more subjective. When a proofreader catches a typo, the writer almost always breathes a sigh of gratitude and never gives it another thought. One just has to look at the world of professional editing for an analogy. Users accept these more “advanced” or “nuanced” suggestions less frequently.įrom a user perspective, accepting fewer suggestions for clarity, structure, and tone is quite understandable. But Grammarly’s writing assistant goes much further than these areas to suggest improvements for clarity, structure, and tone, among other areas.
When the suggestions are about grammar, spelling, and punctuation-those rules of language concerning writing mechanics-they are accepted at a high rate. Helpfully, there is a pattern that we can see when we look at the data. Grammarly’s confidence may be equally high across all of these suggestions, but user acceptance varies. and no.Īs it turns out, there are certain suggestions that people tend to always accept, and others that are more likely to require further consideration, depending on the writer’s specific need. Would this solution solve the problem our Premium users were facing? Yes. As we started asking ourselves what the minimum viable product (MVP) solution might look like, we could see that the simplest approach would be to just bundle all of Grammarly’s suggestions under one big Accept button. The problem was clear: Grammarly Premium users were spending too much time making clicks to accept high numbers of suggestions one by one. Though this feature presents a special case in which disparate pieces fell into place, in many ways this is also characteristic of our process at Grammarly-it’s an apt illustration of how we typically go from idea to launch and all the different teams that contribute along the way. From the natural language processing (NLP) model to the front-end and backend, this feature was essentially bootstrapped from various side projects that ultimately fit together in a cohesive way.
We went through several iterations of the design, and had to carefully consider trade-offs to arrive at a feature that balances convenience and control in the right ways.
0 kommentar(er)
