The Algorithm Didn’t Like My Essay

http://www.nytimes.com/2012/06/10/business/essay-grading-software-as-teachers-aide-digital-domain.html

by aaronbrethorst • 14 years ago

99 54 14 years ago

kenrikm2• 14 years ago

If it's ok for them to use an algorithm to grade essays then it would be ok for students to use an algorithm to write the papers no? How is it different from using a calculator? it's a tool right? mathematics used to only be only pen and paper. (I know it says it would still be graded by people, I'm making a point)

Writing is an artform and as such should be treated and graded as one. Sometimes you need to use language that is grammatically incorrect to get a point across and software is not nuanced enough to pick up on this. I can see this helping to turn out students that are good grammaticians but terrible writers - optimizing for the wrong outcome.

imd• 14 years ago

The evaluations are on "how well algorithms submitted by professional data scientists and amateur statistics wizards could predict the scores assigned by human graders." As long as some humans are always included as a control, this shouldn't be too big of a problem.

jacktoole13• 14 years ago

I upvoted you for the interesting line of thought, although I disagree with the premise that acceptability of grading algorithmically implies acceptability of writing algorithmically.

In CS, we often grade assignments automatically, although we expect students to write the assignments by hand. Although there are many classes where this would be a bad idea, there are some where it works quite well - if the problems are strictly specified, grading an assignment with what are effectively unit tests is reasonable.

Grading and doing an assignment are different classes of problems - roughly mapping to decision problems and search problems. Just because the grading is computationally tractable or acceptable does not necessarily imply that the act of writing the assignment is.

I agree with your second paragraph by and large. There is a time and place for automatic grading, and most writing should not fall into that category.

_delirium5• 14 years ago

One difference with CS algorithmic grading is that it's much more specific and causally connected to outcomes. It's not looking for statistical correlates of the input text and a notion of "good program", but is actually finding things like compile errors and failed testcases. And, crucially, it can also report those to the student: your program is wrong because when fed 5 it gives an output of 23 when the output should be 25.

It's probably possible to grade student programs like this essay-grading system works: instead of feeding the programs to a compiler and running against test-cases, just turn the source code into a bag-of-words representation and feed it to a statistical classifier. There are almost certainly some statistical correlates between the source code of programs and whether they're good/bad submissions. But the results would be both less reliable and less informative as feedback.

majormajor• 14 years ago

For the standardized test discussed in the article, it says the human graders gave three minutes, on average, to reading each essay. These tests are not wanting art. Writing one of these essays is no more art than writing a basic linked list implementation. It's not concerned with how good a writer a kid is, but that the kid can write at all.

Edit: basically, the sheer scale of the grading required for a standardized test means that it's never going to be a real test of writing skills. It wouldn't surprise me if overly skilled writing is actually a negative even with the human grader—I could easily see a rushed grader misreading something clever and giving it a low score.

(The wisdom or effectiveness of these tests in the first place is a whole different discussion.)

bermanoid2• 14 years ago

"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 Writing scores. Add in a factor for essay length, unique word cound, and grammar errors, and you'd probably get pretty damn close to perfect, within a reasonable statistical tolerance.

Children don't produce art, almost all of them produce garbage, right up to the top 1%. Make no mistake, all that standardized tests aim to measure is how stinky the shit that they produce is. I'm fine with missing a bit of brilliance at the top; those students will do just fine anyways. It's ranking the relative crappitude of the middle that's important here, and I think the standard (if crappy) grading procedures do a pretty good job here.

numlocked2• 14 years ago

(disclaimer - I work at Kaggle)

I would recommend downloading the dataset[1] and taking a look at some of the essays. There is a tendency for an educated audience to (justifiably) worry about exactly what you've articulated, but when you read through the essays, you get an appreciation of how valuable a tool like this could be. These are elementary and middle school essays rife with rudimentary mistakes (which are at times extremely charming) and generally not filled with subtle arguments prone to misinterpretation.

If algorithms can help every seventh grader in the country write a solid 5 paragraph essay, that's progress.

[1] http://www.kaggle.com/c/asap-aes/data

gammarator3• 14 years ago

I participated in the Kaggle competition, but I'm less optimistic about the pedagogic value of these systems. A black box algorithm won't explain to a student what their rudimentary mistakes are.

I wrote about this more extensively here: http://bellm.org/blog/2012/04/23/arguing-to-the-algorithm-ma...

Jach• 14 years ago

Will this help the nearly-illiterate masses? I think so. Will it hurt those who are one standard deviation or more above the norm? I think that's a problem, too. The solution then is to use the algorithm on the first set and more individual-oriented grading methods for the other set. Effort should be spent finding those in the other set and that isn't restricted to a single course or assignment. This challenges the assumption that students are equal and deserve equal grading, but I don't really see another way forward that doesn't hurt any of the best, mediocre, or poor writers.

rwallace• 14 years ago

> then it would be ok for students to use an algorithm to write the papers no?

If a student used an essay-writing algorithm he designed himself, I would certainly consider that okay.

alextp• 14 years ago

Giving students close access to a program that evaluates essays many times would be the worst possible thing to do. People are generally very good at adapting to systems, and the algorithm would break down pretty quickly as people find out which factors it favors and which ones it penalises (while human graders are arguably less strict in their criteria and can hence recognize these attempts at writing to the letter).

biot2• 14 years ago

This inevitably leads to an arms race, with the students and software getting better and better. At some point, sufficiently advanced gaming of essay scores is indistinguishable from students who naturally write well by following the rules which lead to good writing. As long as creativity and originality aren't sacrificed, is it really a problem?

CapitalistCartr• 14 years ago

It does not necessarily lead to an arms race. Once the students learn to write to the gradind software, the schools can claim the students are getting better, and point to an 'objective, unbiased' measuring stick. Everyone wins except for society, i.e. sll of us.

gammarator3• 14 years ago

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-engineering comparable algorithms won't necessarily have to write well to score well.

bermanoid2• 14 years ago

If students en masse merely ran their shitty essays through spell and grammar check, the quality of even college-level essays would be improved by a huge margin.

For 99% of the high school students I've worked with, their essays would be better if they had to pass Microsoft's spelling and grammar checks before they were submitted. The exceptions were all exceptional enough that they knew that they no longer needed to listen to those automated filters.

mingpan• 14 years ago

I think this says more about the assignment and the human grading procedures than anything else. Even the supposedly creative sections in standardized tests are surprisingly narrow in what they consider to be good.

_wp3r• 14 years ago

Exactly what I was thinking. The ability to reliably predict human scores, says more about the terrible way we "standardize" creative writing, than about the quality of the algorithms.

If the grading curve of an essay is a checklist how is this finding a surprise?

InclinedPlane2• 14 years ago

We keep paying more and more per student for public education. And what do we get in return? Robo-grading and prison-like treatment of students (closed campuses, zero tolerance taken to ridiculous extent). And again we come around to the notion that if only we'd spend more everything would get better, which has worked so well every time it's happened in the past.

The increasing reliance on automated tools and standardized tests is just one of many symptoms of the increasing detachment between teachers and students. We have to stop doubling down on the prussian style educational system, it stopped returning dividends decades ago. More so we have to stop acting as though we can just solve our educational problems by throwing more money at it. We need to empower good teachers and parents, we need to encourage and support students. Right now all of the incentives in the system run counter to those things.

adrianN• 14 years ago

But this article advocates using automated grading to reduce costs and free up teacher time for better things.

InclinedPlane2• 14 years ago

Cost is orthogonal to educational quality. As I said, this isn't an example of reducing costs so much as an example of the continual dehumanization of the education system. Any cost savings will just be gobbled up by the entrenched interests in the system, bureaucrats, administrators, etc. And any efficiencies making it easier to treat students like cattle and push "education" more and more toward useless rote memorization (teaching to the test) and busywork. These are not good things.

noonespecial• 14 years ago

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 money. The system just has more $180k "administrators" now than it used to.

btilly• 14 years ago

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 loan guarantees. Only make either available at schools whose tuition was in the bottom 70% when you applied. With that incentive, I think that a lot of schools would find ways to reduce tuitions towards what (inflation adjusted) was more than sufficient several decades ago.

dhx• 14 years ago

How well did the algorithms detect insightful analysis, deep understanding beyond the immediate subject matter, factual correctness, salience and an ability to write to a specific audience?

When reading articles on the Internet I look out for superficial factors including:

‣ Misuse of U+0022 typewriter double quotation marks in place of U+201C and U+201D double quotation marks

‣ Misuse of U+002D hyphens in place of U+2013 en dashes and U+2014 em dashes

‣ Poor understanding of comma and semicolon usage

‣ Choice of typeface, use of ligatures and micro-typographical features, amount of leading used…

‣ Misuse of their, there, they’re, where, were, we’re, affect, effect…

‣ RAS syndrome (sīc)

These factors can assist with analysis of writing. An author can generate credibility in certain contexts by using triangular Unicode bullets instead of ASCII asterisks. Use of lengthy and complex words in place of simple English can also be mistaken for credibility. Analysis of these techniques will show that the writer is often attempting to belittle readers. Documentaries utilise similar tactics that the untrained eye will likely mistake for credibility. Soft filter effects, use of bookcase backdrops for expert interviews and silence have profound manipulative effects on viewers.

Grading algorithms will likely reward use of manipulative and belittling writing techniques and penalise honest superficial mistakes. I would rather read the honest opinion from an author that misuses their and there than have to carefully analyse the carefully crafted writings of a dishonest linguist.

As an aside, I recommend bookmarking The Browser[1] (Writing Worth Reading). It would be interesting to see the results of algorithmic essay grading against this curated collection of articles from across the Internet.

[1] http://thebrowser.com/

DougBTX• 14 years ago

How well did the algorithms detect insightful analysis, deep understanding beyond the immediate subject matter, factual correctness, salience and an ability to write to a specific audience?

I'm far from well informed, but my understanding of standardised tests is that the standard specifies the algorithm, which already ignores your good points above to achieve standardised grading. All that really changes is whether a human or a robot executes the algorithm, the human insight has already been squeezed out of the system.

SagelyGuru• 14 years ago

At best, these algorithms can only pick on obvious spelling, grammar, punctuation and style errors. They have no concept of semantics. Therefore the students will have to polish the form and abandon the content.

To prove this point, I suggest setting up a testing competition, this time for algorithms to generate meaningless 'essays'. I expect their easy success obtaining full marks.

So, what will be the educational outcome of this? In three simple words: more dumbing down.

Is this another example of the obsession of officials everywhere with the pedestrian form at the expense of original content? I think it tells us a lot about their own limited mentality.

ericd3• 14 years ago

This article is coming to the wrong conclusion from the facts given. That the grades given by these algorithms matched the grades given by human essay reviewers on standardized tests almost exactly doesn't mean that our algorithms are great and that they should be rolled out to normal school environments to help students learn to write. It means that our method of grading essays on standardized tests is absolute garbage, because the algorithms can't judge semantic meaning or depth of insight. We're basically grading people on how prepared they are for writing About.com content farm articles.

PaulHoule• 14 years ago

I was told by the head search engineer for a well-known job board site that they'd trained a bag-of-words classifier to find badly written resumes that would benefit from professional editing.

They'd found it was one of the easiest text classification problems that had ever been tried. With a very small training set they'd attained off-the-charts accuracys.

_delirium5• 14 years ago

I wonder how the state-of-the-art in algorithmic essay grading compares to the state-of-the-art in scoring the quality of online content. Does Google have content-quality-estimating algorithms that could improve algorithmic essay grading?

keithpeter• 14 years ago

"And humans are not necessarily ideal graders: they provide an average of only three minutes of attention per essay, Ms. Chow says. "

Well, I'm hardly surprised that an algorithm can duplicate that kind of superficial marking of an essay!

Good heavens, what are you doing to your children in the US?

feral2• 14 years ago

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.

_delirium5• 14 years ago

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.

gwern• 14 years ago

Goodhart's law?

yummyfajitas2• 14 years ago

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.

feral2• 14 years ago

Its possible. More generally, its also possible the human graders were doing a bad job; the ML system can only learn 'essay quality' to the extent that the training data reflects it.

However, the kaggle supplied 'straw-man' benchmark, which worked solely based on the count of characters and words in the essay, had an score of .647 with the training data. (The score metric used isnt trivial to interpret - it was 'Weighted Mean Quadratic Weighted Kappa' - but for reference the best entries had a score of ~.8 at the end)

The score of .647, just using length, is quite high. For length to have this powerful a causal predictive effect, the human graders would have to be weighting for length, as a feature, very heavily.

I can't rule that out; but I think its highly likely a major component of the predictive effect of length was correlative, rather than causal.

fiatmoney• 14 years ago

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.

reubensutton• 14 years ago

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.

The Algorithm Didn’t Like My Essay

Related Stories