Are transformer-based language models overfitting on the paraphrase identification task? What tools overcome this?

Question

I've been working on a sentence transformation task that involves paraphrase identification as a critical step: if we are confident enough that the state of the program (a sentence repeatedly modified) has become a paraphrase of a target sentence, stop transforming. The overall goal is actually to study potential reasoning in predictive models that can generate language prior to a target sentence. The approach is just one specific way of reaching that goal. Nevertheless, I've become interested in the paraphrase identification task itself, as it's received some boost from language models recently.

The problem I run into is when I manipulate sentences from examples or datasets. For example, in this HuggingFace example, if I negate either sequence or change the subject to Bloomberg, I still get a majority "is paraphrase" prediction. I started going through many examples in the MSRPC training set and negating one sentence in a positive example or making one sentence in a negative example a paraphrase of the other, especially when doing so would be a few word edit. I found to my surprise that various language models, like bert-base-cased-finetuned-mrpc and textattack/roberta-base-MRPC, don't change their confidences much on these sorts of changes. It's surprising as these models claim an f1 score of 0.918+. The dataset is clearly missing a focus on negative examples and small perturbative examples.

My question is, are there datasets, techniques, or models that deal well when given small edits? I know that this is an extremely generic question, much more than is typically asked on StackOverflow, but my concern is in finding practical tools. If there is a theoretical technique, then it might not be suitable as I'm in the category of "available tools define your approach" rather than vice-versa. So I hope that the community would have a recommendation on this.

Answer 1

Short answer to the question: yes, they are overfitting. Most of the important NLP data sets are not actually well-crafted enough to test what they claim to test, and instead test the ability of the model to find subtle (and not-so-subtle) patterns in the data.

The best tool I know for creating data sets that help deal with this is Checklist. The corresponding paper, "Beyond Accuracy: Behavioral Testing of NLP models with CheckList" is very readable and goes into depth on this type of issue. They have a very relevant table... but need some terms:

We prompt users to evaluate each capability with three different test types (when possible): Minimum Functionality tests, Invariance, and Directional Expectation tests... A Minimum Functionality test (MFT), is a collection of simple examples (and labels) to check a behavior within a capability. MFTs are similar to creating small and focused testing datasets, and are particularly useful for detecting when models use shortcuts to handle complex inputs without actually mastering the capability.

...An Invariance test (INV) is when we apply label-preserving perturbations to inputs and expect the model prediction to remain the same.

A Directional Expectation test (DIR) is similar, except that the label is expected to change in a certain way. For example, we expect that sentiment will not become more positive if we add “You are lame.” to the end of tweets directed at an airline (Figure 1C).

Answer 2

I haven't been actively involved in NLG for long, so this answer will be a bit more anecdotal than SO's algorithms would like. Starting with the fact that in my corner of Europe, the general sentiment toward peer review requirements for any kind of NLG project are higher by several orders of magnitude compared to other sciences - and likely not without reason or tensor thereof.

This makes funding a bigger challenge, so wherever you are, I wish you luck on that front. I'm not sure of how big of a deal this site is in the niche, but Ehud Reiter's Blog is where I would start looking into your tooling ideas.

Maybe even reach out to them/him personally, because I can't think of another source that has an academic background and a strong propensity for practical applications of NLG, at least based on the kind of content they've been putting out over the years.

Your background, environment/funding, and seniority level/control you have over the project will eventually compose your vector decision for you. I's just how it goes on the bleeding edge of anything. What I will add, though, is not to limit yourself to a single language or technology in this phase because of those precise reasons you've mentioned. I'd recommend the same in terms of potential open source involvement but if your profile information is accurate, that probably won't happen, no matter what you do and accomplish.

But yeah, in the grand scheme of things, your question is far from too broad, in my view. It identifies a rather unmistakable problem pattern that not all branches of science are as lackadaisical to approach as NLG-adjacent fields seem to be right now. In that regard, it's not broad enough and will need to be promulgated far and wide before community-driven tooling will give you serious options on a micro level.

Blasphemy, sure, but the performance is already stacked against you As for the question potentially being too broad, I'd posit it is not broad enough, so long as we collectively remain in a "oh, I was waiting for you to start doing something about it" phase.

P.S. I'd eliminate any Rust and ECMAScript alternatives prior to looking into Python, blapshemous as this might sound to a 2021 data scientist . Some might ARight nowccounting forr the ridicule this would receive xou sltrsfx hsbr s fszs drz zhsz s mrnzsl rcrtvidr, sz lrsdz due to performance reasons.