Live data from Hacker News

The Algorithm Didn’t Like My Essay

nytimes.com

41–50 of 59 posts

Re: The Algorithm Didn’t Like My Essay

#41
Myself and a colleague entered this competition. We came 9th. We were doing this for fun, and aren't experts in the domain, but I think our score was within 'diminishing returns' of the better teams.

There are a couple of things to realize about this challenge:

I wouldn't conceptualise the challenge as trying to find features of good essays. It is more about trying to find features that are predictive of the essay being good.

This is a subtle but important distinction. One example is that the length of essay was hugely predictive of the score the essay would get - longer meant better.

Is a longer essay really a better one? No. But, at the level the students were at, it just so happened that the students who were able to write better, also were able to write longer essays.

While you get good accuracy using techniques like this, its debatable how useful or robust this general approach is - because you aren't really measuring the quality of the essay, so much as you are trying to find features that just happen to be predictive of the quality. Certainly, it would seem fairly easy for future students to game, if such a system was deployed.

This isn't a general attack on machine learning competitions - but I wonder, if for situations that are in some sense adversarial like this (in that future students would have an incentive to game the system), whether some sort of iterated challenge would be better? After a couple of rounds of grading and attempting to game the grading, we'd probably have a more accurate assessment of how a system would work in practice.

There is another important feature of this essay grading challenge, that should be taken into consideration. There were 8 sets of essays, each on a different topic. So, for example, essay set 4 might have had the topic 'Write an essay on what you feel about technology in schools'. To improve accuracy, competitors could (and I would guess most of the better teams did) build separate models for each individual essay-set/topic. This then increased the accuracy of, say, a bag of words approach - if an essay-set 1 essay mentioned the word 'Internet' then maybe that was predictive of a good essay-set-1 essay, even though the inclusion of 'Internet' would not be predictive of essay quality, across all student essays.

Its important to remember this when thinking about the success of the algorithms. The essay grading algorithms were not necessarily general purpose, and could be fitted to each individual essay topic.

Which is fine, as long as we realize it. The fact that it was so easy to surpass inter annotator agreement (how predictive one of the human graders scoring was of the other human graders scoring) was interesting. Its just important to realize the limits of the machine learning contest setup.

I would guess that accuracy would down on essays of older, more advanced students, or in an adversarial situation where there was an incentive to game the system.

Re: The Algorithm Didn’t Like My Essay

#42
post #41

Myself and a colleague entered this competition. We came 9th. We were doing this for fun, and aren't experts in the domain, but I think our score was within 'diminishing returns' of the better teams. There are a couple of things to realize about this challenge: I wouldn't conceptualise the challenge as trying to find features of good essays. It is more about trying to find features that are predictive of the essay be…

While you get good accuracy using techniques like this, its debatable how useful or robust this general approach is - because you aren't really measuring the quality of the essay, so much as you are trying to find features that just happen to be predictive of the quality. Certainly, it would seem fairly easy for future students to game, if such a system was deployed.

I'm not able to dig up the name, but there's a named effect in statistics (especially social-science statistics) describing exactly that. When you find a correlate of a desired outcome that has predictive value, a common result if you then set the correlate as a metric is that a substantial part of the correlation and predictive value quickly disappears, because you've now given people incentives to effectively arbitrage the proxy measure. You've said, I'm going to treat easy-to-measure property A as a proxy for what-I-really-want property B. Now there is a market incentive to find the cheapest possible way to maximize property A, which often ends up being via loopholes that do not maximize property B. A heuristic explanation is that proxies that are easier to measure than the "real" thing are also easier to optimize than the real thing. At the very least, your original statistics aren't valid anymore, because you measured in the context where people were not explicitly trying to optimize for A, but now they are doing so, so you need to re-measure to check if this changed the data.

Re: The Algorithm Didn’t Like My Essay

#43

Earlier quoted context omitted.

You may be overestimating the sophistication of these algorithms. At least for this training set, my algorithm rewarded the length of the essay most of all (something like 65% of the total prediction). The only other significant factors were misspellings and prevalence of certain parts of speech. That model matched the accuracy of human graders and several commercial essay grading packages. Students reverse-engineeri…

currently, if a student writes nonsense, there's a fairly significant chance that they will be caught and penalised. a human can detect nonsense in three minutes. in contrast, i suspect algorithmic approaches can be gamed more easily because they don't adapt in the same way. they're not solving the hard ai problem; they're grading essays (currently) written for a human reviewer. for example, what happens if a child l…

No it's incorrect. Compare "these truths... that all men are created equal, (...)" with "these kittens... that all kittens are created hungry". The that in the second sentence is wrong.

Re: The Algorithm Didn’t Like My Essay

#44
post #28

Earlier quoted context omitted.

The problem with "more money" is that the educational system seems to have a sort of hull speed when it comes to adding money. I'd guess that the same or less is being spent on actual education today and the excesses, whatever they may be, are simply siphoned off by the parasites in the system. It doesn't seem to matter how much $ per student gets added, teachers still end up buying classroom supplies with their own…

University education has been going up at 10%/year for decades now. Way above inflation. It is clear that it is going on an internal dynamic until those dynamics finally conflict with reality. That collision will be..painful. Given the crushing levels of student debt in this country it may happen sooner rather than later. If I were in charge, I'd institute a simple change. Improve financial aid. Beef up the federal l…

> University education has been going up at 10%/year for decades now.

Accounting for inflation and number of students, university budgets have not been going up 10%/year; at some universities, the real per-student cost of higher education has actually been declining. What's been going up at 10%/year is the tuition "sticker price" for students who receive no scholarships or need-based aid. But this is offset by two contrary trends: 1) a decrease in the percentage of students who are actually paying the sticker price, especially at private universities; and 2) large decreases in state funding for public universities.

Consider the University of California system (all figures below in 2012 dollars). In 1990, it spent $21,000 per student (dividing its total budget by its total enrollment). Today it spends $16,500 per student--- a decrease in the cost of education of about 25%. Tuition has nonetheless gone up, because the state-funded portion of its budget has declined even faster: from $16,000 per student in 1990 to $9,500 per student today, and possibly to $8,500 per student in the coming year. As a result, the student-funded portion has risen from $5,000 to $7,000 and soon $8,000 per student on average (with a sticker price a bit over double that).

If you want the UC system to be able to run on tuitions that were sufficient in 1990, you'd have to also return state funding to what it was in 1990...

Re: The Algorithm Didn’t Like My Essay

#45
post #41

Myself and a colleague entered this competition. We came 9th. We were doing this for fun, and aren't experts in the domain, but I think our score was within 'diminishing returns' of the better teams. There are a couple of things to realize about this challenge: I wouldn't conceptualise the challenge as trying to find features of good essays. It is more about trying to find features that are predictive of the essay be…

Is a longer essay really a better one? No. But, at the level the students were at, it just so happened that the students who were able to write better, also were able to write longer essays.

It could also be the case that length is one of the features your human graders are using to grade essays. I.e., it might really be causal, rather than merely correlated.

In my (anecdotal) experience, teachers certainly do this. While in college I developed the skill of utilizing excessively long and verbose language while elucidating simple points simply to incrementally increase the length of essays [1].

Luckily a great prof in grad school (thanks Joel) beat this bad habit out of me.

[1] In college I learned to pad my essays with verbose language.

Re: The Algorithm Didn’t Like My Essay

#46
As long as we're training algorithms to recognize correlates of "high-quality writing" rather than high-quality writing itself, why not use as many predictive features as possible? I'll bet parental income and education level, average home price in the school district, and the percentage of students at the school receiving free or reduced-price lunches, are incredibly correlated with writing quality.

Re: The Algorithm Didn’t Like My Essay

#47
post #34

Earlier quoted context omitted.

"Art" is not even a question - these tests typically apply to high school kids, and you're lucky if bare-bones basic communication is possible there. Most of these kids are straight-up idiots, and I say this having spent several years teaching them. I'd guess that the algorithm "map number of Microsoft Word reported spelling errors from (0 -> 800) to (20 -> 400)" does a pretty damn good job of guessing students' SAT…

I'm sorry, but this is absurd, and it makes me very angry. If I got a whiff of this attitude from you as my teacher, you would not get an ounce of my respect as a student. Why should I spend time trying to produce good work to impress you if you've already decided that what I produce will be shit? If you take the ridiculous view that my attention to grammatical detail, essay length, and variety of vocabulary used are…

...you've already decided that what I produce will be shit...

What he said: "Most of these kids are straight-up idiots..."

He hasn't already decided, he just came up with a prior.

Re: The Algorithm Didn’t Like My Essay

#48
post #41

Myself and a colleague entered this competition. We came 9th. We were doing this for fun, and aren't experts in the domain, but I think our score was within 'diminishing returns' of the better teams. There are a couple of things to realize about this challenge: I wouldn't conceptualise the challenge as trying to find features of good essays. It is more about trying to find features that are predictive of the essay be…

While you get good accuracy using techniques like this, its debatable how useful or robust this general approach is - because you aren't really measuring the quality of the essay, so much as you are trying to find features that just happen to be predictive of the quality. Certainly, it would seem fairly easy for future students to game, if such a system was deployed. I'm not able to dig up the name, but there's a nam…

Goodhart's law?

Re: The Algorithm Didn’t Like My Essay

#49
My instinct is that most algorithms to test this may end up optimizing for essay length and word complexity, rather than being able to assess the content, however I imagine this is also how a lot of teachers grade, given that this article states that they receive 3 minutes of attention each.

Re: The Algorithm Didn’t Like My Essay

#50
post #48

Earlier quoted context omitted.

While you get good accuracy using techniques like this, its debatable how useful or robust this general approach is - because you aren't really measuring the quality of the essay, so much as you are trying to find features that just happen to be predictive of the quality. Certainly, it would seem fairly easy for future students to game, if such a system was deployed. I'm not able to dig up the name, but there's a nam…

Goodhart's law?

Aha, almost it; I was thinking of the very similar Campbell's law, which your mention of Goodhart's law led me to. Somehow no combination of search terms got me to either of those when I was trying to come up with the name, though...
Post reply on HN