Analyzing ChatGPT’s ability as a grammar fixer

Post by Stephen Horowitz, Professor of Legal English

I recently tried a simple yet potentially helpful ChatGPT activity with my LLM students to (a) build individual grammar awareness, (b) build a better understanding of the benefits and limitations of using ChatGPT to fix one’s grammar, and (c) gain a better understanding of what happens grammatically when ChatGPT is asked to fix grammar.

The Process:

1. As part of the Legal English II course (which teaches US case reading and analysis via a series of Supreme Court decisions about Miranda rights to students in Georgetown Law’s 2-Year LLM program), my students were required to write an essentially IRAC-style answer in response to a fact pattern under timed conditions.
2. Afterwards, as an assignment, I asked my students to input their essay into ChatGPT with the instruction to “Please fix any language issues in this essay:“
3. Students then had to compare the two versions of their essay and write a short analysis or commentary on what they noticed, what ChatGPT did/didn’t do well, how they felt about it, etc. I told students to either put the two versions in a table so they could compare the language side by side, or they could do a use the redline/track changes function to show the differences.
4. I next reviewed the students’ submissions myself. And I then invited two Georgetown Legal English colleagues with PhDs in applied linguistics–Prof. Julie Lake and Prof. Heather Weger–to review the student submissions and then have a group discussion about what we noticed.
5. Upon additional consideration (and inspired by a suggestion from Jack Kenigsberg, a former Hunter MA TESOL classmate), I took one paragraph from one student’s essay and fed it into ChatGPT with the instruction: “Fix any grammar errors in the quoted text. For each change you make, explain why you made the change.” And after it provided its answer, I clicked “Regenerate response” to create a second response to see what (if anything) came out different a second time.

The Takeaways:

The main takeaways by my students, my colleagues and myself were:

1. Natural sounding: ChatGPT very much takes the accent off one’s writing and makes it sound more natural and cohesive. The thing to be careful of is that it sometimes inadvertently changes intended meanings.
2. Overediting? ChatGPT sometimes changes sentences or phrases that seem fine into another form. The problem is that for students who are not native English speakers, it’s hard to know if it really needed changing or not. Sometimes it did need changing because there’s some nuance of the words or grammar that the student couldn’t recognize. Sometimes what the student wrote was completely fine. The solution, however, may be for students to simply ask ChatGPT to provide an explanation for each change it makes in order to better discern which changes to focus on and which ones they might disregard.
3. ChatGPT’s reasons for making changes: Based on its explanations, at least with regard to the one paragraph I re-submitted, fell into the following categories:
   • Make the sentence grammatically correct/Doesn’t make grammatical sense: These two explanations were applied when it changed “protections of due process” to “protections for due process.” In one sense, ChatGPT is correct in identifying it as a grammatical error because “for” is the correct preposition in this phrase. However, prepositions are in many cases fairly arbitrary and often can even be substituted without changing a meaning. In this case, it’s not inconceivable that “of” could be used in this situation. But the fact of the matter is it’s just not the way we would generally say it. In that regard, to me this feels more like an issue of collocation (i.e., co-location)–that is, words that just tend to go together (e.g., how prima facie is usually followed by the words evidence or case.) And that makes even more sense in light of what we know about ChatGPT’s “thinking” process, which is based on predicting the words likely to be used in a given situation. Outside of the obvious uses of prepositions (i.e., the cat being on the house as opposed to in the house), the ways that prepositions combine with various nouns, verbs, etc. are often essentially collocations. There’s no grammatical or semantic explanation as to why we might use one preposition versus another. (Or if there is, it’s too abstract or metaphorical to explain well.) We just know they’re right because as a native speaker we’re just used to hearing them paired that way. This is not, I think, much different from ChatGPT’s thought process.
   • Gender-neutral: In a number of instances, ChatGPT changed a singular pronoun to a plural pronoun. For example, it changed, “….the right of a person not to incriminate oneself” to “….the right of a person not to incriminate themselves.” I originally thought this was just ChatGPT having a preference for the more informal, spoken form of language in which it is acceptable to switch to the plural. However, when I asked for an explanation, ChatGPT unequivocally stated the reason as a preference for gender-neutral pronouns.
   • Make it more accurate/More common phrase: It provided this explanation in response to changing “the right of silence” to “the right to remain silent.” I think both of these explanations make sense. And I think they point back to ChatGPT’s reliance on collocation as a primary source of determining what is and isn’t correct.
   • To clarify that it refers to the suspect’s statements/To use a more appropriate term: ChatGPT seems to like to hang its hat on clarity or making things more clear, and it seems to like terms like more accurate, appropriate or common as interchangeable phrases to convey similar meaning. In this instance, it took the grammatically acceptable phrase, “….statements would be inadmissible in the trial….” and changed it to “any statements made by the suspect would be inadmissible in trial.” To me, this is where the line gets blurry between helping with grammar and the innate desire of all humans (and apparently AI chatbots) to change what someone else has written. This is also an instance where a student genuinely would not know if they made a significant error or not.
4. How you phrase your request to ChatGPT (and AI tools in general) can be extremely important: In this instance, I told my students to use the exact instruction: “Please fix any language issues in this essay.” The second time around I used the instruction “Please correct any grammar errors.” Unfortunately, I still don’t know how big a difference there was as ChatGPT seems to always generate something a little different even in response to the exact same question. To really figure out if there’s a difference, I imagine I would need to ask it one way a bunch of times, and then the other way a bunch more times. And then I could compare and contrast. But for my LLM students’ learning purposes and for my purposes, this fine of a distinction probably doesn’t really matter.

A couple grammar observations:

This investigation is by no means thorough or scientific. It was intended as a first pass at just seeing what kinds of things we notice or jump out. With that in mind, here are a few more specific grammar items that popped out to me and others I spoke to:

1. Nominalization goes both ways: There are instances where ChatGPT revises a phrase to use nominalization. And instances where it takes a nominalized phrase and de-nominalizes it. (Thanks to former Hunter MA TESOL classmate Anne Koenig for this observation!)

However, while the above descriptions sounds fancy and linguistic, I feel like in the end it ChatGPT isn’t thinking about any of that and is ultimately like the proverbial senior law partner or editor who can’t resist the urge to simply change what someone else has written. And I think this ties back to the idea of collocation and ChatGPT just preferring the statistically more common or higher-probability phrasing of an idea.

Original	ChatGPT	Comment
The issue in this case is whether Aaron’s statements would be inadmissible in the trial.	The issue at hand is the admissibility of Aaron’s statements in the trial.	would be admissible –> admissibility
The 5th Amendment of the Constitution provides protections of due process and the right of a person not to incriminate oneself.	The 5th Amendment of the Constitution provides due process and protects the right of a person not to incriminate themselves.	protections of due process –> protects the right of a person

2. Connecting language: A number of students commented–and I’d noticed as well–that ChatGPT does a great job of incorporating cohesive devices to make its writing feel more cohesive. As cohesion is often a challenge for students yet often has a je ne sais quoi element to it, comparing one’s essay against ChatGPT’s “fixed” version provides a great opportunity for students to notice the range of words, phrases, and syntax options for improving cohesion in one’s writing. The fact that the students could often feel and identify the stronger cohesion in ChatGPT’s writing suggests that repeated use of ChatGPT for this kind of comparison combined with active identification and noticing of cohesive devices can potentially help students absorb the use of these cohesive devices–both consciously and incidentally–into their own writing.

By the way, for more on cohesion in US legal writing, and a great primer on cohesion in general, I strongly encourage reading Elizabeth R. Baldwin’s Beyond Contrastive Rhetoric: Helping International Lawyers Use Cohesive Devices in U.S. Legal Writing, 26 Fla. J. Int’l L. 399 (2014), https://digitalcommons.law.uw.edu/faculty-articles/225

Final takeaways:

What I really liked about the assignment is that it succeeded in helping my students build grammatical awareness. Also, I think my students greatly appreciated a way to get detailed language feedback. In my class, while I provide language feedback to the extent it connects to the communicative purposes related to legal analysis and writing, it’s not a grammar class and I don’t have the bandwidth to provide detailed language feedback on every aspect of the students’ writing. (A statement I’m assuming many in the legal English/legal writing field likely relate to.) Additionally, I believe there’s a certain amount of value in receiving the feedback in a manner removed from the power dynamic associated with the student-teacher relationship, i.e., not from an authority figure who has the power to decide your academic fate, but from a neutral, non-judgmental chatbot. And I think it also opens up the potential for more focused conversations and questions between student and teacher about language and grammar, as it helps narrow down areas of student concern from the perspective of each student.

And what I liked about the investigation (in addition to getting a chance to nerd out on grammar with some colleagues) is that it helped me and my students and colleagues get a little better sense of the benefits and limitations of ChatGPT as a tool as well as some ideas on how to refine my use of it with students in the future.

Reactions? Feedback? Other experimental attempts to analyze ChatGPT as a grammar fixer?

If any readers have reactions to this post, or if you’ve attempted to analyze ChatGPT’s grammar correction abilities and patterns, or if you decide to try anything along these lines, or if you’ve heard of anyone else whose done something along these lines, I’d love to hear about it. Please feel free to share in the comments or to contact me directly.

****************************

BONUS CONTENT!

I ran this entire blog post through ChatGPT with the following instruction: “Please make suggestions for fixing any language issues in the following text and provide an explanation for each suggested change:“

Here’s ChatGPT’s response:

Here are some suggestions for fixing language issues in the text:

1. In the first sentence, replace “build individual grammar awareness” with “increase individual grammar awareness.” The latter phrase is more commonly used and better conveys the intended meaning.
2. In the second sentence, change “the benefits and limitations” to “the advantages and limitations.” While “benefits” is technically correct, “advantages” is more commonly used and sounds more natural.
3. In the third sentence, replace “what happens grammatically” with “what occurs grammatically.” “Occurs” is a more precise and appropriate word to use here.
4. In the fourth sentence, add “that” after “course.” This makes the sentence easier to understand.
5. In the fifth sentence, add “under timed conditions” after “fact pattern.” This provides more detail about the assignment.
6. In the sixth sentence, change “input their essay” to “input their essays.” The subject is plural and requires a plural verb.
7. In the eighth sentence, change “either put the two versions” to “either put both versions.” This is a clearer way of expressing the idea.
8. In the ninth sentence, change “myself” to “myself and.” This clarifies that the two professors were also involved in the review process.
9. In the eleventh sentence, remove the word “if” before “anything.” The sentence reads more smoothly without it.
10. In the twelfth sentence, replace “fell into” with “were categorized as.” This is a more precise way of expressing the idea.
11. In the thirteenth sentence, add “that” after “reasons.” This clarifies the meaning of the sentence.
12. In the fourteenth sentence, change “person not to incriminate themselves” to “person’s right not to incriminate themselves.” This makes the sentence clearer and more precise.
13. In the fifteenth sentence, replace “it provided this explanation in response to changing” with “it changed.” This is a clearer way of expressing the idea.
14. In the sixteenth sentence, remove “And I think” at the beginning of the sentence. The sentence is clearer without it.