What is a rhetorical situation?
During one of our demos at CCCC, someone asked me, ‘can WriteLab understand the rhetorical situation of a text?’ My simple answer was ‘no,’ but, after the demo ended, I reflected on the question and tried to articulate a better answer. It evaded me, although I had studied Rhetoric (capital R intended) with Jeffrey Walker, one of the great Classical rhetoricians of our time. After some thought, I concluded the following: a rhetorical situation is not only what the author intends, but also what the reader perceives. It is all the cues that exist, the variety of patterns, that I may perceive and filter through my experiences as a reader, as a writer, as someone new to or familiar with the topic, as a speaker of natural language, and as a human being.
But cues are the bread and butter of WriteLab. They inform and activate the machine learning, the user experience, and, most importantly, the pedagogy of the software. So, even though my answer about the current state of the software is still ‘no,’ why should WriteLab not account for the rhetorical situation of a text?
I understood rhetoric first as a series of figures, schemes, and tropes. One intellectual exercise I too much enjoyed in college, even to the detriment of my personal relationships, was reading Sister Miriam Joseph’s Shakespeare’s Use of the Arts of Language. I teach this book in freshman comp. When I encountered it, I pulled out my copy of Shakespeare’s Complete Works and, for each rhetorical figure, I tagged all the examples that Joseph cites, and I wrote the name of the figure in the margin. After this, I went back through the Comedy of Errors, the first play in my edition. I started marking examples of these figures, examples that Joseph does not cite. Among the epistrophe, the asyndeton, and the anadiplosis, I wanted to find correlations, which would tell me that having found figure A I would likely find figure B, and, more strongly, which might indicate that A strengthens B or makes it rhetorically more effective. This was before I knew anything about programming. I was studying Ancient Greek and had once again, despite a series of numerical disasters in high school, become enchanted with mathematics.
I wanted a way to break language into its smallest rhetorical units, so that I could, in a manner deliberate and poised, add gestures to my prose. I wanted architecture: an outline or a draft that would become an edifice. I started to build a model for such a rhetorical architecture when I read and applied Joseph’s techniques. One of the problems that ensued was that I actually did find correlations between specific rhetorical figures and larger rhetorical strategies, and with these correlations came a lot of complexity.
I published an article in 2011 about Shakespeare’s complex chiasmus. What I had found was that, in several cases, Shakespeare embedded multiple chiasmi to structure speeches and arguments. But even though I felt I had rigorously employed my method, it wouldn’t scale well. For instance, I wouldn’t be able to see how many other authors over time had such a correlation, unless I tagged and counted everything by hand. I needed a way to replicate this tagging, so that I could understand and ultimately utilize ever more complex structures.
I studied formal logic, in the hope that I could derive formulae for rhetorical figures. Rhetorical figures are far too subtle and complex, however, to fit into the clean, binary models of the formal logic that I was learning. One of the shortcomings of formal methods is that they are often deterministic and reductive. I remembered that my calculus teacher said on the first day of class that to estimate situations in real life we needed to settle for reasonable approximations rather than definitive numbers and absolute answers. We approximate, and, based on a seeming infinitude of variables, we form our best guess. I needed not a perfect function, but an approximation, a way of estimating the space under the curve. This seemed totally appropriate, because we do not prove things when we assess a rhetorical situation; neither do we solve things. So I looked around for applications of calculus and statistics that might aid me in this cause.
Shortly after the article came out, I became interested in massive open online courses (MOOCs). I was watching hours of Khan Academy a day, and it wasn’t long before I discovered videos on machine learning. At first, machine learning seemed too technical and foreign, and I absorbed powerful but imprecise impressions of it. What I felt impressed upon me was that there was a way to have algorithms experience things and make decisions based on those experiences. Those experiences have two components: the raw data and the algorithm’s way of processing that data. I discovered a way to engage with this concept when I realized that one application of machine learning is in natural language processing. In NLP, I found that I could build predictive models of language and approximate some of the elusive figures of rhetoric. I found algorithms that I believe can be appropriate for assessing rhetorical situations, those algorithms that estimate probabilities, approximate, converge, and form their best guess. But the work is great, and we must assess only one rhetorical component at a time, breaking a larger rhetorical situation into many smaller features and approximating each of those. When we break down something complex in this way, we must take care not to lose the sense of the whole. Even though a rhetorical situation is more than the sum of its parts, it is important that we manage this complexity as writers or rhetoricians, but especially if we are teachers.
Although I’ve only taught five semesters of writing, I have struggled to find ways to break down literary and rhetorical complexity into a form that my students will not only understand, but also apply. At the end of the day, we do not teach to honor monoliths of classic prose, but to empower our students, to give them their literary and linguistic inheritance with which they can create texts that reflect the integrity of their own intelligence and experience. It’s true that my students are not writing with complex chiasmi, but they perceive the effects of such devices, even as they perceive rhythm and melody in the songs on their iphones.
One of the things that I had wanted as a student, greedy as I was for feedback, was comments on the sound of my prose. I wanted to know about the meter of my lines, the cadences, and the flow, in the hope that I could control my prose and apply new techniques where I saw fit. But now that I teach, I see that giving such feedback is immensely time-consuming. In fact, given the time constraints of a semester and the number of students in a class, it’s not humanly possible to comment on everything we’d like to. This impossibility, however, doesn’t lessen a student’s desire to know more about her writing. In developing WriteLab, we cannot promise that we will be able to identify or comment on all the features that are meaningful to a student or to a teacher. But what we can promise with the help of NLP is that what we can identify reliably, we will identify immediately and at scale. And we will increment what we can identify, making it possible for teachers to focus on more interesting and engaging features of students’ writing, whether it’s their use of chiasmus or subtle aspects of their rhetorical situation.