Filling in the blanks Pt. I - The difference between a cloze passage and a legal argument
What's in an argument?
Hello again everyone! It's Friday again, and time for another issue of this newsletter. I am aiming to produce an issue every week.
This week, I write about a writing task that most, if not all, of us would have come across: the cloze passage. Either you have filled up such a passage yourself, or have a child that has to do such exercises.
Credit: University of Minnesota
Looks familiar?
At the risk of being reductive, the cloze task is what large language models learn. Take large corpora of text, mask certain words, and make the AI model guess what the missing word is. It learns from its mistakes - quite similar to how we learn from cloze tasks.
That's where the similarities end.
Let's use an example.
Bill is a _____, because he eats only plant-based foods.
Is Bill a vegan, or a herbivore, say, a cow?
Arguably, there isn't enough context. If Bill were a cow, then herbivore is the right word. But what if Bill were actually a human, and the writer deliberately used herbivore nonetheless? Do we interpret that as humour, or condescension from someone who cannot imagine such a diet?
Even if an AI model were to correctly guess that these were the two available options, how would it “understand” humour, condescension, or the differences in how we refer to humans and animals?
The answer is that it would build a statistical model of text patterns. When referring to an animal, use “it”. Out of X instances, sometimes the animal is referred to using gender articles. In such cases, this may denote some affection. And so on and so forth.
So much for cloze tasks. What about legal arguments?
Now, obviously, not all legal argumentation is blank filling. But many routine cases are. Many routine cases can be seen as cloze tasks. Say, for example, the issue is one of arguing what the appropriate harm or culpability is of a particular accused person. Do the facts reveal high or low culpability? Do they underline that the harm was high or low?
One may argue that that is a matter of interpretation and argumentation, but let's take a step back. Legal argumentation then becomes an argument about what should go into the harm and culpability brackets. Might we have similar brackets for each argument as a whole? And I think the answer is yes. That's what models like GPT do when prompted. And where perhaps disruption to what junior lawyers will be required to do will be greatest.
Which brings us to the next question, are there forms of legal argumentation where the foregoing simple rubric does not apply to? Or that sit at the penumbra?
Stay tuned.