Hacker News Reader

Comment by _delirium | Hacker News Reader

I work in AI, and was very surprised when I heard about this (a few years ago). I don't think anyone who works in the area thinks the tech is ready for this kind of deployment. There is research on the subject [1], and NLP systems can do better than baseline methods, but the error rates are still pretty high.

A thing you quickly find if you try to download off-the-shelf NLP tools and apply them to anything is how little is reliable at all, unless you can constrain the domain. Even basic topic identification only works with low error rates when constrained to something like NYT stories, or PubMed abstracts, not arbitrary text by arbitrary writers. And I would bet ETS is using worse tech than research state-of-the-art.

[1] e.g. https://www.aclweb.org/anthology/P15-1053

View on HN

You've noticed though that the AI con is on. This damages your work as people get burned and will bring about the second "AI winter"

People making big decisions with a lot of money around computing know nothing about it and are marks for con-artists. Think big consulting firms selling to senior public servants in washington. "For a successful technology reality must take precedence of public relations." But reality just gets in the way when conning a mark for a successful snake oil sale, right?

Call it out, publically, cite your credentials. Encourage colleagues, your competition and everyone with a clue to pour scorn on whoever is selling this evil, toxic waste as drinkable.

Second? We’re heading to #3 — fully cyclical

Hmmm. I also work in AI, in fact professionally in information retrieval and NLP. I disagree strongly with what you say. Basic topic summarization and keyword / named entity extraction on unstructured sources of text works reasonably well. It’s easy to modify BERT and GPT on smaller problems, language classification is borderline totally solved by extremely easy to train neural network models.

I still agree that automatic essay grading is beyond the reach of SOTA NLP models today, but youmake it sound like virtually nothing can be done in a production-grade manner that solves real world unconstrained NLP problems. This is manifestly false.

It's completely possible I'm not fully up on recent progress, especially since a bunch of stuff seems to have moved in the past 6 months. But I haven't seen any general models that can solve open-domain problems, without specifically retraining on each domain. Do you have any pointers? E.g. a single pretrained BERT model that can reliably extract topics from: tweets, paragraphs from 19th-century novels, mathematics journal articles, and Wikipedia articles? All the systems with very low error rates that I know of target one specific domain. The last time I looked into sentiment analysis (a year or so ago), it wasn't even that great on many individual domains, e.g. it would get tripped up by sentences from novels that used "negative" keywords in a humorous or ironic way.

In production problems that I work on, we don’t even really use things from within the past year. These problems are just incredibly well-solved with fairly vanilla LSTM networks from 2-3 years ago. Enough so that while it’s probably premature for fully automated essay grading, it’s not _crazy_ to make a product from models trained to solve this problem.

I have a grant where were are doing just that. Implementing more or less SOTA research using fairly vanilla LSTM networks from 2-3 years ago (primarily Taghipour & Ng) to provide low stakes feedback to students on their essays in one of our teaching tools at Purdue. It’s based on research using the Kaggle ASAP database and we have found it to be pretty accurate across a variety of domains in early testing. Though some essay prompts seem to do better with CNNs vs. RNNs. I doubt many of the systems in TFA are based on LSTMs or neural nets at all. They are probably doing regression on hand-crafted features.

Very interesting. Are there any meta-analyses / reviews that summarize progress in this area? Would it be possible to share your grant proposal -- I'd be curious to get an idea of what is being attempted.

It's an internal grant and I'm not sure I'd be allowed to share it. We are adding AES to our peer-review app. Currently as an additional "grader" to the peer reviews since that's what the PI requested. Since the tool allows unlimited submissions until the review date, I hope to add it as a "pre-flight" estimate to give students a chance to get a rough prediction of the score they will receive and a metric they can use as they revise until the due date.

I'm not aware of any meta-analyses myself. I have been keeping up with the ASAP competition and various attempts to improve on the initial systems for a number of years. The two papers I believe are having the most success are [1] and [2]. [3] seems promising for balancing the opposing forces of high accuracy for true positives and the risk of false positives via adversarially crafted inputs.

I'm also vaguely aware of research happening around extracting features from neural nets. I'd love to be able to help students understand why the system is predicting a particular score.

[1] https://www.aclweb.org/anthology/D16-1193 [2] https://arxiv.org/pdf/1606.04289.pdf [3] https://arxiv.org/pdf/1804.06898.pdf

Reply on news.ycombinator.com