By Holly Zheng, '22
From the red squiggly lines on Microsoft Word and AutoCorrect on our phones to the list of synonym suggestions by Grammarly, assistant writing apps have gained a significant market by helping many to improve the quality of their emails, essays, and work reports. These highly intelligent personal editors not only take away our worries of misspelling and punctuation mistakes, but also try to give us word suggestions so that the sentence is more concise. The technology behind these devices is a fascinating application of natural language processing and machine learning. Grammarly, one of the most well-known assistant writing app, recently published a blog post where the company explained a new training model to detect and fix run-on sentences. 
“How you correct run-on sentences it’s not as easy as it seems” was the title of the research paper, a humorous choice because the title itself was a run-on sentence. Run-on sentences, or ungrammatical sentences with multiple verbs, are one of the most common mistakes made in English. Compared to words and phrases, run-on sentences, along with other mistakes on the sentence level, are more challenging to detect. They often contain complicated “long-distance dependencies” where a word at the end of a sentence may depend on another word at the beginning. Researchers at Grammarly aimed to improve the accuracy of their editing algorithm in detecting and correcting run-on sentences.
The two models that researchers at Grammarly used in their training were Conditional Random Field (CRF) and Sequence to Sequence (Seq2Seq). CRF treated a sentence as sequences of spaces between tokens that were labeled to indicate whether a period should be inserted in a certain space, whereas Seq2Seq considered an input sentence as one whole entity. Due to lack of labeled run-on sentences in the existing database for grammatical error correction, researchers artificially generated run-on sentences by paring two grammatical sentences and removing the punctuation that separates them.
The result indicated that the CRF model scored highest in precision of inserting a period. This finding was not surprising to the researchers, because the CRF model incorporated syntactic features, rules that govern sentence structure. The Seq2Seq had a higher score on recalling, which means it was better at generalizing in various sentences.
The improvement in fixing run-on sentences only represents one of the many steps in Grammarly’s development. The increasing efficiency of similar AI algorithms has made them extremely beneficial in professional and academic settings, but a question that people raised when Google Translate launched still rings an alarming bell now. When Google’s instantaneous translator appeared, many wondered, “Why should we learn foreign languages any more if we have Google Translate?” Similarly, now if we have a highly intelligent assistant editor already in our keyboard, why should we care about practicing good writing?
Grammarly seems to explain its rationale behind their AI approach to writing enhancement in a very romantic way. “At Grammarly, our goal is to make it possible for everyone to be heard and understood. But there’s no single set of rules that will work for every person in every situation. The beauty of AI is that it can combine many types of information and adjust to the situation at hand,” states a blog post on Grammarly’s official website.  This appealing sentiment of creating space for self-expression is a major factor behind the popularity of writing assistance devices. Still, we should keep in mind that the process of learning how to edit is meaningful in itself. The assistance we receive from Grammarly and other devices should not replace this learning process.
A test passage on Grammarly's advertisement
 Zheng, J., Napoles, C., Tetreault, J., Omelianchuk, K. “How do you correct run-on sentences it’s not as easy as it seems” Proceedings of the 2018 EMNLP Workshop W-NUT: The 4th Workshop on Noisy User-generated Text, pages 33–38, Brussels, Belgium, Nov 1, 2018 [cited: November 26, 2018] http://aclweb.org/anthology/W18-6105
 Grammarly Blog, “Grammarly Spotlight: How We Use AI to Enhance Your Writing” August 14, 2018 [cited: November 26, 2018] https://www.grammarly.com/blog/how-grammarly-uses-ai/