4 ms·

> For example, given state of the art in current benchmarks on language understanding, if those benchmarks really did measure language understanding, we could a

by nl 6y ago

> For example, given state of the art in current benchmarks on language understanding, if those benchmarks really did measure language understanding, we could all be AIs debating the paper that itself could have been written by AI. Suffice it to say, this is not likely given the current level of "understanding" in modern systems. I'd note that I said "clear progress on language understanding". > we could all be AIs debating the paper if those benchmarks really did measure language understanding in terms of the level of proficiency between the researchers, then it's no surprise that the results are very different. The next chart shows the results of the most common tests for language comprehension. If we look at the results of the different tests, then we see that the average proficiency is very high (12.9 out of 20). If we look at the results of the tests for other things, then the average proficiency is very low. Again, the charts are very similar to the results of the two separate results of the two separate tests for language comprehension. So, the question is, what does this tell us about the impact of language learning on literacy? I'll give you what I think is the relevant point. I think the key point here is that language learning is the process of learning new things. So, there are some things that are learning in the language. The text here was generated by a GPT-2 based model (aitextgen). I think there is a fair argument that it isn't far off the average level of discourse on HN.

YeGoblynQueenne 6y ago
I think you're being unnecessarily mean, rather than "fair". Anyway I'll take the opportunity to point out that one reason why language understanding benchmarks are not adequate for the task is that it's very difficult to know what a system "understands" (if anything) just by looking at its output. It is also very difficult to evaluate language _generation_ tasks accurately. Finally, I have no idea how you could convince me that the above was not written by a human, rather than generated by a language model, if I didn't believe you. Which is to say, there are no good ways to know this kind of thing for sure. At the end of the day, we have a bunch of metrics that don't measure what we want them to measure and a bunch of systems that don't learn what we want them to learn, and that are very good at gaming any metric we throw at them. The end result is a lot of uncertainty regarding true capabilities of those systems. And when careful scholarship is turned to the analysis of those systems' results, it tends to find that they're not as good as the metrics suggest. I'm repeating the point made by the article, but I agree with it very much. Edit: the paper I link to above proposes a new benchmark called HANS (Heuristic Analysis for NLI Systems) that tries to correct for learning shortcuts in language models. It finds that (state of the art language model) BERT performs dismally on that benchmark. That's one datum. The other is that I haven't so far seen results on HANS reported in papers or benchmark aggregators etc. Language modelling work likes to cite results on e.g. GLUE, SUPERGLUE, etc, but these are exactly the kinds of benchmarks that are full of loopholes for the current approaches to exploit (even though they're supposed to not be).
- nl 6y ago
  In the HANS paper I'd point out this bit: "we retrained each model on the MNLI training set augmented with a dataset structured exactly like HANS (i.e. using the same thirty subcases) but containing no specific examples that appeared in HANS.... In general, the models trained on the augmented MNLI performed very well on HANS" I think it's great to find examples where language models break. But it doesn't look like they have found a fundamental problem here - just a weakness that can be fixed with "just" engineering. I do agree there are things we don't know how to do yet. I think Chollet's "On the Measure of Intelligence" paper[1] is the best writing I've seen on this, which shows specific things that are currently hard for computers to do and give a route towards general intelligence. [1] https://arxiv.org/abs/1911.01547 https://arxiv.org/abs/1911.01547

by nl 6y ago

YeGoblynQueenne 6y ago

I think you're being unnecessarily mean, rather than "fair". Anyway I'll take the opportunity to point out that one reason why language understanding benchmarks are not adequate for the task is that it's very difficult to know what a system "understands" (if anything) just by looking at its output. It is also very difficult to evaluate language _generation_ tasks accurately. Finally, I have no idea how you could convince me that the above was not written by a human, rather than generated by a language model, if I didn't believe you. Which is to say, there are no good ways to know this kind of thing for sure. At the end of the day, we have a bunch of metrics that don't measure what we want them to measure and a bunch of systems that don't learn what we want them to learn, and that are very good at gaming any metric we throw at them. The end result is a lot of uncertainty regarding true capabilities of those systems. And when careful scholarship is turned to the analysis of those systems' results, it tends to find that they're not as good as the metrics suggest. I'm repeating the point made by the article, but I agree with it very much. Edit: the paper I link to above proposes a new benchmark called HANS (Heuristic Analysis for NLI Systems) that tries to correct for learning shortcuts in language models. It finds that (state of the art language model) BERT performs dismally on that benchmark. That's one datum. The other is that I haven't so far seen results on HANS reported in papers or benchmark aggregators etc. Language modelling work likes to cite results on e.g. GLUE, SUPERGLUE, etc, but these are exactly the kinds of benchmarks that are full of loopholes for the current approaches to exploit (even though they're supposed to not be).

nl 6y ago
In the HANS paper I'd point out this bit: "we retrained each model on the MNLI training set augmented with a dataset structured exactly like HANS (i.e. using the same thirty subcases) but containing no specific examples that appeared in HANS.... In general, the models trained on the augmented MNLI performed very well on HANS" I think it's great to find examples where language models break. But it doesn't look like they have found a fundamental problem here - just a weakness that can be fixed with "just" engineering. I do agree there are things we don't know how to do yet. I think Chollet's "On the Measure of Intelligence" paper[1] is the best writing I've seen on this, which shows specific things that are currently hard for computers to do and give a route towards general intelligence. [1] https://arxiv.org/abs/1911.01547 https://arxiv.org/abs/1911.01547