14 ms·
ChatGPT outperforms crowd-workers for text-annotation tasks
- karmasimida 4y agoIt should comes with no surprise to anyone. Text classification is an easy task. ChatGPT is definitely overqualified to perform this task.
- halfjoking 4y agoWe wanted to give you the job ChatGPT, but you’re overqualified. We’re going to give the compute to davinci. But I need that extra compute to train my kids. :(
- halfeatenscone 4y ago"Classify the following strings according to whether they represent Python programs which terminate..."
- karmasimida 4y agoHuman can't do it either
- Tepix 4y agoIt's weird, i would have expected the mturks' tasks to be of equal quality because they have outsourced their tasks to ChatGPT...
- meltedcapacitor 4y agoCheaper for operators of swarms of mturks to use simpler but less effective machine learning algos, which are good enough for this type of task? For things like answering surveys of lazy social science researchers, I wouldn't even bother and just use rand(3).
- danpalmer 4y agoAt my previous job we had a human review stage in a data pipeline. 5-10 people at an outsourcing company in Bangladesh would review things via a simple web interface we provided for them. There were ~10 factors they were reviewing, all fixed options (no free text), but varying from 5 to 500 options per factor. It was all based on a few text fields and around 5 images. On the surface of it, I'd expect ChatGPT to do very well at this. It's simple text and images, not many options and theoretically very limited context. However the more I think about it the less sure I am. Firstly these weren't crowd-sourced reviews, they were trained reviewers, paid hourly not per review. Incentives were definitely in favour of the long term business relationship. Then there was the training doc, we maintained a vast disambiguation doc used to resolve things that were vague or could be interpreted multiple ways, this was constantly being revised. All necessary context should have been in that but it wasn't and reviewers definitely found patterns that worked and didn't. Lastly the reviewers were in a Slack channel where they would ask questions to their manager on our side, and while this might have only been ~1% of tasks, it was an important process. So maybe you could point ChatGPT at it and let it run, but the oversight process we had would still be necessary. The disambiguation doc would have been too long for ChatGPT's context at the moment, but that will likely change in the near future. Would the workflow be to keep tweaking the prompt to add special case after special case? How do you scale "do this, but not that, but add this, but..." in prompting, and would ChatGPT become as confused as a human after enough of that – I expect so given that it's only a language model and that's not effective communication.
- kiratp 4y agoLangChain + a vector DB for embedding [0] sections of your doc would solve the problem today. You could also have failsafes that trigger human oversight based on confidence levels or other factors. [0] https://python.langchain.com/en/latest/modules/indexes/getting_started.html https://python.langchain.com/en/latest/modules/indexes/getti...
- danpalmer 4y agoIt looks like for it to work like this the doc would need to be divisible into obvious units that could be retrieved by some key/terms, but in our case it was much looser than that. There were no fixed rules (otherwise we'd have written business logic), it was guidance that really assumed the whole doc as context, and human judgement of the relative weightings of parts of the doc. The failsafes based on confidence levels is an interesting idea though, that is very possible. I suspect we'd have started out with manually reviewing all decisions and slowly backed it off as the prompts got better.
- dougb5 4y ago"ChatGPT’s intercoder agreement exceeds that of both crowd-workers and trained annotators for all tasks." -- wait, what? However good ChatGPT is at approximating trained annotators for these Twitter tasks, it's an algorithm, so the level of simulated "inter-annotator agreement" is in the authors' control (in the case of GPT, via the temperature parameter, for which they try just two values, 0.2 and 1.) And why does this paper not make any effort to describe the wide range of annotation tasks for which this kind of simulated annotation is not a good idea -- for example, where you care about the subjective opinions of specific subgroups of people at specific times. And even for the tasks they mention, what about the risks of reinforcing biases by using a model's output to train new models? Good grief this paper is lazy!
- bumbledraven 4y ago> We used the ChatGPT API with the ‘gpt-3.5-turbo’ version to classify the tweets.
- sorz 4y agoIs NLP a solved problem now?
- matusp 4y agoShameless self-promotion, I have recently written a blog about this. ChatGPT actually is usually a little bit worse than older models for these classical NLP tasks. Of course the older models are not zero-shot. https://www.opensamizdat.com/posts/chatgpt_survey/ https://www.opensamizdat.com/posts/chatgpt_survey/
- sorz 4y agoJust watched a talk[0] about natural language understanding research in post-GPT-3 era. Old issues may has been solved, while new topics are coming to this area (quoted from the slides): - Retrieval augmented in-context learning - Better benchmarks - Last mile for productive application - Faithful, human-interoperable explanations [0] https://www.youtube.com/watch?v=-lnHHWRCDGk https://www.youtube.com/watch?v=-lnHHWRCDGk
- est 4y agofor stuff that before 2021-09, mostly.
- ninjin 4y agoObviously not. If we had a general solution to language intelligence we would have artificial intelligence at the level of at least human intelligence – which we do not. Rather, the right question to ask is which language intelligence tasks currently have acceptable performance and under which conditions (text domain, etc.). Clearly this is a much more difficult question and with a lot more nuance to it, even if it is undeniably that things have moved very quickly over the last few years. Skimming the abstract as a senior academic in the area. This looks like preliminary work and a limited investigation for a single (non-standard) task. Thus far from a strong result published at say a top-tier conference or journal. Still, interesting direction and if expanded upon could absolutely be impactful. I should also mention that I am not familiar with the related literature, so it could very much be that there is similar (better?) work out there exploring the same question.
- PostOnce 4y agoBefore you ask, because I was curious, from the paper: "For MTurk, we aimed to select the best available crowd-workers, notably by filtering for workers who are classified as “MTurk Masters” by Amazon, who have an approval rate of over 90%, and who are located in the US."
- swyx 4y agohonestly i wonder if @dang will approve an auto summarizer bot on HN since it helps improves the quality of discussions. finetune on HN comments, anticipate the top few questions, and then answer from the source doc
- theamk 4y agoI hope not, as such bot will inevitably introduce factual errors every once in a while. When people don't read articles others can see that and correct misconceptions. Officialy approved bots will not be able to receive or act on this feedback.
- codetrotter 4y agoDo like the mods in some subreddits have done. 1. The comment with the summary is made like any other comment. (In the case of Reddit, that’s perhaps mainly because the mods have no other option, but the point still stands.) Because of this, other users can downvote the bot comment if it is incorrect, as well as respond to it with corrections. Exactly the same way as you’d interact with a person in the past if their summary comment was incorrect; downvote them and/or correct them. 2. Include a disclaimer stating that the summary may be inaccurate. Encourage people to correct and/or downvote bad summaries in the disclaimer
- 93po 4y agoi think this would better as a browser extension for those who want it, the simplicity of HN is nice
- 8organicbits 4y ago
- jacknews 4y agoSo is this 'AI trains AI better than people can'? And presumably the better-trained AI will also be better again at training. I think I've seen this movie.
- intended 4y agoNo its not. This is High quality training data set made by trained humans. (Ground truth) Chat GPT was given the rules and data set. then asked to evaluate the data. MTurkers were given the rules and data set. Then asked to evaluate the data. GPT was closer to the gold standard/Ground truth. If you want a historical comparison, this is industrialization. Before we could only scale using humans via mturk. Now if you can scale using GPT. However, for ALL of GPT related shenanigans - 1) This is English First. You cannot apply this to under-resourced languages. 2) You need to have the ability to verify the output. ——- I really want to solve 1.
- mistrial9 4y agobut its a super-spell checker in a way.. it does not understand the meaning of the patterns.
- sacrosancty 4y ago[dead]
- GaggiX 4y agoWhat does it mean to really understand the meaning of a pattern? An answer that is not a simple "look at ChatGPT, it fails at this task", since in most cases it is already no problem for GPT-4, and does not really prove that this particular task cannot be solved by a LLM. Also I see, "super-spell checker" is the new "fancy markov chain".
- doctor_eval 4y agoI don’t know how I understand patterns either. So I’m not entirely sure consciousness exists.
- nico 4y ago
- Imnimo 4y agoMy main take away here is that Turkers are terrible at some of these tasks. The "stance" task is, "Classify the tweet as having a positive stance towards Section 230, a negative stance, or a neutral stance.", and the Turkers accuracy was like 20%. Even in its best task, ChatGPT only got 75% accuracy.
- p1necone 4y ago20% means you can beat ChatGPT if you invert the answers!
- quickthrower2 4y agoThere are 3 stances. But 33% is achievable by chance.
- xyzzyz 4y agoOnly if probability distribution of answers is known.
- riku_iki 4y agojust speculating, there is a chance ChatGPT was trained on this dataset in the past, and learned probability distribution..
- raincole 4y agoEh, even if the probability distribution is unknown, a random guess should still have 33% chance to be correct.
- p1necone 4y agoI started explaining why you were wrong and got about 5 words in before I realized you're correct.
- 4y ago
- joshxyz 4y agooh dear, i am foreseeing prompt writing mechanical turks in the future where their instance of ai got plugins tailored for their industry, task, and datasets.
- visarga 4y agoIn my experience GPT labeling is good, but it makes errors, about 10% in my case. Maybe it's better on average than a human, but not perfect for the task. I was doing open-ended schema matching, a hard task because of the vast number of filed names.
- kolinko 4y agoDid you try using multiple AIs for the same task (e.g. GPT & llama), or GPT with two-three prompts?
- visarga 4y agoYes, it contradicts itself if you call it multiple times. So I get to measure ensemble agreement as a proxy for correctness. But what do I do with the examples where the ensemble is not confident? They are too many to label by hand, and a pity to throw out - they could be the most interesting training examples.
- alphabetatheta 4y agoI wonder how much we can extend this to image or video labelling tasks? this could really jeopardize the business moats of lot of data labelling startups and services.
- arbol 4y agoYou can extend it to data labelling tasks and it will do pretty well. https://imgur.com/a/3XfOBq1 https://imgur.com/a/3XfOBq1 This is stable diffusion processing a small, blurry image on my PC and correctly identifying it. It did take 7 minutes but my 1 GPU is no match for the ~250M USD worth of GPUs owned by OpenAI.
- whatever1 4y agoOn average, no doubt chatgpt will be great for annotation. But annotation is mostly needed at the boundaries. At the very edge cases where it is not clear if something is a dog or a shape that looks like a dog. I really doubt that a generic model like chat gpt can really help in these tail cases.
- dx034 4y agoBut it shows here that it performs better than humans who did that in the past. Few annotation tasks were performed by experts. For example, I guess most annotations for animals weren't created by biologists. So tail cases were already handled incorrectly.
- sandGorgon 4y agoCurious here - OpenAI talks a LOT about how RLHF (Reinforcement Learning Through Human Feedback) is core to how GPT is tuned. Including safety. Are we getting to the point where GPT will be tuned by GPT without the need for HF ?
- alophawen 4y agoNo. ChatGPT or any other LLM model requires to be trained on curated input in order to demonstrate anything useful. If you would train it on a stream of hallucinations, you would end up no better than hallucinatory noise. Already fooling some, though.
- sandGorgon 4y agobut isnt that a direct contradiction of this particular paper anyways - that chatgpt anyways outperforms human annotation. so permit me to act as devil's advocate to your statement - prove that (in context of this paper), your hypothesis is still correct.
- alophawen 4y agoWhy would I need to prove anything. The unproved claim is in the paper.
- CGamesPlay 4y agoThe researchers in the paper used human-curated results to classify the accuracy of the GPT results. So it had that human in the loop.
- redox99 4y agoHere it's outperforming because ChatGPT is already good at these tasks (and the MTurks aren't very good, OpenAI labelers are probably better, and a panel of experts much better). To further improve ChatGPT shortcomings (assuming such flaws are because of alignment and not lack of capability of the base model) you need Human labels. Feeding it's own outputs would achieve nothing. However feeding it's outputs can make a non aligned model become aligned (that's what alpaca did with llama+chatgpt).
- winrid 4y agoIt does seem to work pretty well. I'm using it to analyze all US Congress bills: https://govscent.org/bill/USA/118hres190ih https://govscent.org/bill/USA/118hres190ih It extracts the topics and determines how on topic the bill is. Soon we're adding a topic browser and the homepage will have some fun stats :) it's all free.
- Ajedi32 4y agoWow, I didn't realize how many congressional bills are just pointless resolutions with zero legislative impact. Is this list of bills curated in any way? Where are you sourcing it from? A quick scan of https://www.govinfo.gov/app/collection/bills/ https://www.govinfo.gov/app/collection/bills/ seems to turn up bills with a lot more substance. Edit: Upon further investigation it seems like a lot of those are not technically bills, but rather House or Senate resolutions. Explained here: https://www.senate.gov/legislative/common/briefing/leg_laws_acts.htm https://www.senate.gov/legislative/common/briefing/leg_laws_... The govinfo.gov list seems to indicate that such resolutions are pretty common, but significantly less so than bills, so I'm not sure why the front page seems to consist only of resolutions right now.
- winrid 4y agoIt's just everything from https://github.com/unitedstates/congress https://github.com/unitedstates/congress The homepage just shows the most recent data (it auto updates every 6 hours). I plan to add more filters and things to make it more useful. I've only spent a few Sundays on it so far :P It's in Django so very easy to contribute to: https://github.com/winrid/govscent https://github.com/winrid/govscent I also plan to add all bills, all the way back to 1800.
- winrid 4y agoPart of the problem is also the 3.5-turbo token limit. I'm awaiting approval for GPT4 which raises the limit to 8k tokens, should help process some larger bills.
- winrid 4y agoUpdate - long bills are now supported: https://govscent.org/bill/USA/111s570pcs https://govscent.org/bill/USA/111s570pcs
- tasubotadas 4y agoNow you won't have to worry that those people are not paid enough.
- YeGoblynQueenne 4y agoSuppose you have two classifiers, A and B, and some un-annotated data, D. You want to know how good is classifier B at annotating the data, compared to classifier A. One problem is that you don't have the ground truth for D. So you start by annotating D with the labels assinged by a third classifier, C: C(D) → D₁ Having thus established a modicum of "ground truth", ish, you proceed to annotate D with the two classifiers you are comparing, A and B: A(D) → D₂ B(D) → D₃ Then you compare the classifications D₂ and D₃ to D₁, and find that D₃ better approximates D₁ than D₂. What is the result of the experiment? I summarise it as follows: Classifier B better approximates the labelling of D by classifier C than classifier A. Now we can name the three classifiers as they were used in the experiments in the linked article: A: Human annotators. B: ChatGPT C: Human annotators. So the result of the paper is that, plugging in the names: ChatGPT better approximates the labelling of D by Human annotators than Human annotators. And that, is the finding of the paper. Which is clearly absurd and a cause to re-think methodology.
- nmca 4y agoAh yes, ye olde "all humans are equally capable of all tasks" axiom (the paper is correct and the parent comment is wrong, perhaps obviously)
- YeGoblynQueenne 4y agoTo be clear, that's a different criticism of the article's methodology than mine, yes? If you assume that the two sets of human annotators are different, then what are you comparing, exactly? The ability of one group to second-guess the other? There are other issues if you choose to assume that the two groups are fundamentally dissimilar: one group was two grad students, the other a number of Mechanical Turks. You can expect there to be more disagreement between (more than two? I'm not sure) Mechanical Turks than between two grad students (both in political science). Ultimately, my problem with the study is that the labelling they took as ground truth (i.e. that they compared ChatGPT and Mechanical Turks against) is too uncertain, or even subjective, to know for sure what exactly they found out. Edit: Oh, wait, I didn't put that in the comment above. Damn. I thought I had.