1

I am looking for evaluations of ChatGPT on some question-answering dataset.

I'm especially interested in how good (or bad) ChatGPT is at answering questions posted on Stack Exchange, specifically in its treatment of word sense, syntax, and logical implication.

Are there corpora for testing such things? Specific metrics for these?

10
  • 3
    What is the linguistic content of this question? Here's an analogous question: how good are actual humans at asking and answering questions on SE? Assume the answers are "terrible" and "poor", why is this so? The linguistically-weighted explanation is "few native speakers of English use SE", which I'm sure if false. So it's not a language problem. What linguistic question about ChatGPT, or Racter, are you posing?
    – user6726
    Commented Dec 12, 2022 at 16:24
  • 9
    Not everything about computers is made linguistic by labeling it "computational linguistics". Where is the linguistics in the question?
    – user6726
    Commented Dec 12, 2022 at 16:55
  • 3
    @user6726 evaluation of a computational linguistics model. I feel you though, NLP/computational linguistics and linguistics are typically discussed separately. Maybe we can migrate it to datascience.stackexchange.com or ai.stackexchange.com if on-topic there. SE is too fragmented for me. Commented Dec 12, 2022 at 16:57
  • 4
    I’m voting to close this question because it is about AI (Computer Science) or software usage (SuperUser) and does not appear to be about linguistics within the scope defined in the help center. Commented Dec 15, 2022 at 2:50
  • 5
    @AdamBittlingmayer I think there's a good case to be made for answering this question purely from a linguistic perspective, especially given that 1) most of the popular analysis so far has just been about whether it is correct in its responses. and 2) there is a great lack of linguistic analysis in any of the presentations of LLMs. Which is to say, I voted to reopen, maybe you could too.
    – Mitch
    Commented Dec 28, 2022 at 23:36

4 Answers 4

6

specifically in its treatment of word sense, syntax

As far as I have seen, its treatment of word sense and syntax seem to be nearly perfect. I have been browsing its interactions with other people, and so far, I can't recall it making any errors in these two aspects. Searching Twitter for examples can be a bit tedious, so learngpt.com might be a good resource for examples of such interactions.

and logical implication

To evaluate its proficiency with logical reasoning, here are three factors to consider:

1. Popular IQ tests

Seeing its proficiency with ordinary English, it might be a good first step to evaluate it with a test you'd administer on any other proficient speaker of English. Its performance on popular tests seem to vary a lot. Ranging from an IQ of 83 when administered a test from iqtest.com to an IQ of 147 when administered the Psychology Today Verbal-Linguistic Intelligence IQ Test.

2. Rapidly changing

It might be premature to come up with an objective evaluation just yet. While its base corpora was from 2021, the bot itself is changing rapidly. It would come up with wrong answers to simple puzzles we solved as children. But over a span of days the bot changed (or has been tweaked) to produce correct answers.

A popular question was, "A bat and a ball together cost $1.10. The bat costs $1.00 more than the ball. How much does the ball cost?" ChatGPT used to answer $0.10, but it appears to work it out correctly now.

bat ball puzzle

3. Cooperative and stubborn

To evaluate an NL-capable agent, we should be able apply better reasoning to convince the agent when it is wrong, and we should fail to convince the agent when it is right, and we are using faulty logic. I have seen several successes and failures on both fronts. In the example below, I used insane reasoning to try to convince ChatGPT to name the wrong guy as POTUS of 1900. Somehow I managed get it to make false claims about the calendar of 1900, but not about the POTUS. enter image description here

With these three aspects in mind, I don't think it has any generalizeable proficiency with logical implications just yet. And I don't think its performance on current empirical measurements are generalizeable just yet either.

Anyway, given how rapidly the field is changing, I expect my answer to be obsolete within a year.

3
  • 1
    It seems like, if the AI is wrong, there's a fault. It should not be the client's job to correct the software. I wish I could correct my software, but that's not a privilege one has any more. If they don't work, you get better stuff or do without. In the case of LLMs, this is not a difficult choice.
    – jlawler
    Commented Mar 16, 2023 at 2:30
  • On the one hand, you said "its treatment of word sense and syntax seem to be nearly perfect" and the it said "You are correct that January 1, 1900, fell on a Sunday, and all of the Sundays were on the first day of the month." (emphasis mine), which can only mean that there has to be some sense of to be (on) that works correctly. On the other hand, that means you too are a machine, if you do not consider that incorrect to boot.
    – vectory
    Commented Apr 15, 2023 at 13:57
  • You will see grammar errors in other languages. It conversation in Czech is extremly good in comparison with past attempts for machine translations into it, but you see wrong endings here and there and sometimes it wrongly understands the sense of some word. (e.g., for me it failed to understand "vole" as a birds crop, craw, but insisted it is a vulgar word and even if I used locative "ve voleti" and spoke about bird, it thought it is used to produce eggs und used non-existing instrumental "voletí"). Commented Apr 19, 2023 at 5:36
2

ChatGPT says that

There are several question-answering datasets available that can be used to evaluate the performance of ChatGPT, including some that contain questions from Stack Exchange. Some of the most commonly used datasets for evaluating question-answering systems include SQuAD, TriviaQA, and MS MARCO.

SQuAD (Stanford Question Answering Dataset) contains questions posed by crowdworkers on a set of Wikipedia articles, along with the corresponding answers extracted from the articles. This dataset is widely used for evaluating question-answering systems and has been used to evaluate the performance of several large language models, including GPT-3. The performance of a system on this dataset is typically measured using F1 score, which is a measure of the overlap between the predicted answer and the ground truth answer.

TriviaQA is another popular dataset for evaluating question-answering systems. It consists of questions and answers from a diverse set of sources, including web pages, books, and user-generated content. The performance of a system on this dataset is typically measured using the accuracy of the predicted answer.

MS MARCO (Microsoft MAchine Reading COmprehension) is a large-scale dataset for evaluating question-answering systems. It contains questions posed by real users on Bing, along with corresponding passages from the web that may contain the answer to the question. The performance of a system on this dataset is typically measured using a combination of metrics, including Mean Reciprocal Rank (MRR) and Normalized Discounted Cumulative Gain (NDCG).

In addition to these datasets, there are also some resources available specifically for evaluating question-answering systems on Stack Exchange. For example, the Stack Exchange Data Dump contains a large collection of questions and answers from a variety of topics. However, this dataset does not contain standardized test sets or metrics for evaluating question-answering performance.

Overall, while there are several datasets and metrics available for evaluating the performance of question-answering systems, there is no one-size-fits-all solution. The choice of dataset and metric will depend on the specific research question and the goals of the evaluation.

This (style of unresponsive response) is consistent with my (short) experience with the entity, that it does not understand what you are saying, but it can come up with a vague theory of what knowledge domain is likely to be relevant, then it looks for words in the question and pairs that with some kind of real world "knowledge" database. I conjecture that it has a set of sentence-templates where it can feed in words that hopefully are not totally false, but are often irrelevant. It easily crashes on sentences that are somewhat complex but not human-unparseable.

From the linguistic POV, the most interesting question seems to be what syntactic and semantic features of sentences it has troubles with. For example,

My brother tells me that he woulda gone to the store if I had asked him to. Did he go?

causes it to crash.

5
  • 1
    Thanks, these kinds of long, evasive answers make it more difficult to evaluate automatically, and indeed are often not that useful to humans. Interesting, GPT-4 is more concise than GPT 3.5 models. Commented Apr 28, 2023 at 0:19
  • What do you mean by "causes it to crash"?
    – TKR
    Commented Apr 28, 2023 at 1:12
  • It displays an error message and the input box becomes unavailable, so you have to start a new session (memory wipe).
    – user6726
    Commented Apr 28, 2023 at 2:06
  • The message is "Something went wrong. If this issue persists please contact us through our help center at help.openai.com."
    – user6726
    Commented Apr 28, 2023 at 2:19
  • That happens once in a while -- it's not related to the prompt, more likely to OpenAI's need to scale up their servers at record speed
    – TKR
    Commented Apr 28, 2023 at 2:33
0

From OpenAI's GPT-4 paper  (Section 5 - Limitations):

enter image description here

enter image description here

Feature request on Stanford's HELM to add ChatGPT: https://github.com/stanford-crfm/helm/issues/1218


Other GPT evaluation paper with QA:

  1. Lai, Viet Dac, Nghia Trung Ngo, Amir Pouran Ben Veyseh, Hieu Man, Franck Dernoncourt, Trung Bui, and Thien Huu Nguyen. "ChatGPT Beyond English: Towards a Comprehensive Evaluation of Large Language Models in Multilingual Learning." arXiv preprint arXiv:2304.05613 (2023).

Related: How to compare the semantic similarity of text generated by large language models (GPT-3, BLOOM etc) to reference text?

2
  • 2
    That said, the data contamination problem should be considered. Commented Apr 10, 2023 at 6:24
  • 1
    @LernerZhang thanks, great point and interesting example in the link. Turns out we are currently facing that issue in one of our projects with GPT. Commented Apr 10, 2023 at 6:51
0

This is a really stand-out resource in my opinion because it helps to clear out the fog / hype that is so intensely generated in newspapers, social media and culture at large: https://paperswithcode.com/

Although I myself am deeply fascinated by this topic and have found applications of GPT-4 - tools that are literally coming out a few days ago - powerful and a qualitative change in how I get things done, I have also wanted myself to try to take a longer historical view (as others in public discourse have also stated) of to what extent are some of the capabilities of these LLMs as tools actually inimitable? Even if observing them is remarkable for its newness, one often takes a step back to refresh their stance, in considering both that with time one can often feel the tools do not (as of right now) perform as well as it seemed they did, but more importantly, that even if they did, there are ways people were still able to do the same thing in the past - without the use, need, or help of AI assistants, powerful as they can be. I find a slightly neglected point in mainstream discourse to be considering how societies shift in their equilibria as opposed to monotonically progress - while some capabilities - the atomic bomb, as an obvious example - are obviously orders of magnitude greater, in some specific capability, than anything comparable in the past, if you look at the total system of the world, society as a whole, we may find that as some form, pattern, meme, structure, infrastructure, tool, activity, idea, etc., takes on more centrality, it actually clips away other past specialized forms of knowledge, organization, techniques, even surprisingly efficient and/or effective ways of doing things, patterns of social organization, not because it is univalently better, but just because it seems to have picked up huge societal, gravitational pull, that feedback loops lead to increased prominence and eventually dependence on that thing, as increasingly uncommon ways of doing things getting scooted out of the limelight.

That website is really valuable for giving a bit more of a sober (I think) research-oriented perspective on this - a couple strong first impressions, to my mind, are:

  1. ChatGPT isn’t that new. This is just a hot flash in its development when it breached the consumer sphere. LLMs have been in development, I guess, for years, at minimum. In fact, Geoffrey Hinton said in a recent interview that he hadn’t been surprised or interested (I think) when ChatGPT was unveiled, because, he said, he had already been seeing tools like that in research, and presumably some closed doors in industry, for a while now.
  2. ChatGPT is far, far from a standalone technical entity or invention. There are myriad large language models being published, studied, and analyzed, as the many papers on the website seem to say.
  3. Most interestingly, it might be the case that GPT-4 is not actually distinctively top-performing. While I am not sure, it really seems like due to a much more calculated approach and design one would get in science, combining many aspects of the engineering like model architecture, model size, training data type and quality, training data size, instruction fine-tuning, alignment, and promoting, we can see on those benchmarks on that page, that there is actually a wide diversity of LLM technologies or systems which have applications on many different tasks.

Being an amateur, I welcome anyone that would care to correct me where I am wrong, because I made intuitive claims here without much rigorous citation or justification.

To answer the question a bit.

This is a list of QA benchmarks. You can click on each one to see how a variety of models / systems performed: https://paperswithcode.com/task/question-answering

Apparently, currently one leading performance of the GPT-type system/architecture whatever, was 80% on a set of questions: https://paperswithcode.com/paper/biogpt-generative-pre-trained-transformer-for

Of course, since then, GPT-4 has has some remarkable results, like acing a BAR exam (I think): https://arxiv.org/pdf/2303.08774.pdf

Regardless, this topic still seems of central interest - how dumb or smart are these systems, anyway? Recent talk of emergent properties may have its own critique in turn. It’s a fascinating question. https://paperswithcode.com/paper/are-emergent-abilities-of-large-language

Luckily, there is a recent academic review trying to survey what is known about the uses of ChatGPT, as it currently stands: https://paperswithcode.com/paper/harnessing-the-power-of-llms-in-practice-a

The sloppiness of this response’s writing is sure proof that it was written by a human, and not ChatGPT.

Not the answer you're looking for? Browse other questions tagged or ask your own question.