We conduct experiments in machine translation and text generation and demonstrate that this yields tighter prediction sets with better coverage! Check the paper for more analyses and discussions 🙂 [4/4]
However, token-level generation violates the i.i.d. assumption. We show that by building on the non-exchangeable CP by Barber et al., we can construct token-level prediction sets with guarantees by retrieving information about similar tokens from a datastore! [3/4]
Conformal prediction provides statistical guarantees about the correct prediction being contained in a prediction set with a pre-defined probability in expectation. Such a property is appealing for NLG, where we would like to avoid sampling implausible tokens. [2/4]
Interested in how to apply Conformal Prediction to NLG? Check out our work on "Non-Exchangeable Conformal Language Generation with Nearest Neighbors" with Chryssa Zerva and
@andre_t_martins accepted to EACL2024 Findings! [1/4] 🧵
https://browse.arxiv.org/abs/2402.00707
Non-Exchangeable Conformal Language Generation with Nearest Neighbors
Quantifying uncertainty in automatically generated text is important for letting humans check potential hallucinations and making systems more reliable. Conformal prediction is an attractive framework to provide predictions imbued with statistical guarantees, however, its application to text generation is challenging since any i.i.d. assumptions are not realistic. In this paper, we bridge this gap by leveraging recent results on non-exchangeable conformal prediction, which still ensures bounds on coverage. The result, non-exchangeable conformal nucleus sampling, is a novel extension of the conformal prediction framework to generation based on nearest neighbors. Our method can be used post-hoc for an arbitrary model without extra training and supplies token-level, calibrated prediction sets equipped with statistical guarantees. Experiments in machine translation and language modeling show encouraging results in generation quality. By also producing tighter prediction sets with good coverage, we thus give a more theoretically principled way to perform sampling with conformal guarantees.
The survey comprehensively summarizes and compares approaches for classification and regression tasks, as well as containing all the necessary derivations, common datasets, references to code, an in-depth discussion and more! (3/3)
We survey the literature on a class of models which parameterize prior and posterior distributions instead of likelihoods. By modelling distributions over distributions, we can now estimate the distributional uncertainty of the model as well, all in a single forward pass! (2/3)
I am happy to announce that our work about Evidential Deep Learning methods for Uncertainty Quantification with @ch_hardmeier
and @jesfrellsen
got accepted at
#TMLR! 🥳 (1/3) 🧵
https://openreview.net/forum?id=xqS8k9E75c
Prior and Posterior Networks: A Survey on Evidential Deep Learning...
Popular approaches for quantifying predictive uncertainty in deep neural networks often involve distributions over weights or multiple models, for instance via Markov Chain sampling, ensembling, or...
🤖 💬 ❓ nlp-uncertainty-zoo — nlp-uncertainty-zoo 1.0.1 documentation
(4/5) To give an quick intro into the usage, you can find a demo notebook here:
https://github.com/Kaleidophon/nlp-uncertainty-zoo/blob/main/demo.ipynb
nlp-uncertainty-zoo/demo.ipynb at main · Kaleidophon/nlp-uncertainty-zoo
Model zoo for different kinds of uncertainty quantification methods used in Natural Language Processing, implemented in PyTorch. - nlp-uncertainty-zoo/demo.ipynb at main · Kaleidophon/nlp-uncertain...
(3/5) This version also simplified the evaluation of calibration and uncertainty properties of your model:
