2022

Impact of Pretraining Term Frequencies on Few-Shot Reasoning

Razeghi, Yasaman, Logan IV, Robert L., Gardner, Matt et al.

Understand

Pretrained Language Models (LMs) have demonstrated ability to perform numerical reasoning by extrapolating from a few examples in few-shot settings.

However, the extent to which this extrapolation relies on robust reasoning is unclear.
In this paper, we investigate how well these models reason with terms that are less frequent in the pretraining data.
In particular, we examine the correlations between the model performance on test instances and the frequency of terms from those instances in the pretraining data.

Built on

Gender bias in coreference resolution
Rudinger, R., Naradowsky, J., Leonard, B., and Van Durme, B · 2002
Earlier work this paper cites.
Exploiting machine learning to subvert your spam filter
Nelson, B., Barreno, M., Chi, F. J., Joseph, A. D., Rubinstein, B. I. P., Saini, U., Sutton, C., Tygar, J. D., and Xia, K · 2008
Earlier work this paper cites.
Mental models and human reasoning
Johnson-Laird, P. N · 2010
Earlier work this paper cites.
A thorough examination of the CNN/Daily Mail reading comprehension task
Chen, D., Bolton, J., and Manning, C. D · 2016
Earlier work this paper cites.
Annotation artifacts in natural language inference data
Gururangan, S., Swayamdipta, S., Levy, O., Schwartz, R., Bowman, S., and Smith, N. A · 2017
Earlier work this paper cites.
Adversarial examples for evaluating reading comprehension systems
Jia, R. and Liang, P · 2017
Earlier work this paper cites.
Numeracy for language models: Evaluating and improving their ability to predict numbers
Spithourakis, G. and Riedel, S · 2018
Earlier work this paper cites.
What makes reading comprehension questions easier?
Sugawara, S., Inui, K., Sekine, S., and Aizawa, A · 2018
Earlier work this paper cites.
The secret sharer: Evaluating and testing unintended memorization in neural networks
Carlini, N., Liu, C., Erlingsson, U., Kos, J., and Song, D · 2019
Earlier work this paper cites.
Commonsense knowledge mining from pretrained models
Davison, J., Feldman, J., and Rush, A · 2019
Earlier work this paper cites.
Openwebtext corpus
Gokaslan, A. and Cohen, V · 2019
Earlier work this paper cites.
We need to talk about standard splits
Gorman, K. and Bedrick, S · 2019
Earlier work this paper cites.
Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference
McCoy, T., Pavlick, E., and Linzen, T · 2019
Earlier work this paper cites.
Language models as knowledge bases?
Petroni, F., Rocktäschel, T., Riedel, S., Lewis, P., Bakhtin, A., Wu, Y., and Miller, A · 2019
Earlier work this paper cites.
Language models are unsupervised multitask learners
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., and Sutskever, I · 2019
Earlier work this paper cites.
What’s in a name? Reducing bias in bios without access to protected attributes
Romanov, A., De-Arteaga, M., Wallach, H., Chayes, J., Borgs, C., Chouldechova, A., Geyik, S., Kenthapadi, K., Rumshisky, A., and Kalai, A · 2019
Earlier work this paper cites.
Analysing mathematical reasoning abilities of neural models
Saxton, D., Grefenstette, E., Hill, F., and Kohli, P · 2019
Earlier work this paper cites.
Auditing data provenance in text-generation models
Song, C. and Shmatikov, V · 2019
Earlier work this paper cites.

Similar

Do NLP models know numbers? probing numeracy in embeddings
Wallace, E., Wang, Y., Li, S., Singh, S., and Gardner, M · 2019
Cited alongside, same era.
Language models are few-shot learners
Original
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al · 2020
Cited alongside, same era.
Poison attacks against text datasets with conditional adversarially regularized autoencoder
Chan, A., Tay, Y., Ong, Y.-S., and Zhang, A · 2020
Cited alongside, same era.
Does Learning Require Memorization? A Short Tale about a Long Tail , pp. 954–959
Feldman, V · 2020
Cited alongside, same era.
What neural networks memorize and why: Discovering the long tail via influence estimation
Feldman, V. and Zhang, C · 2020
Cited alongside, same era.
The Pile: An 800gb dataset of diverse text for language modeling
Original
Gao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N., Presser, S., and Leahy, C · 2020
Cited alongside, same era.
RealToxicityPrompts: Evaluating neural toxic degeneration in language models
Gehman, S., Gururangan, S., Sap, M., Choi, Y., and Smith, N. A · 2020
Cited alongside, same era.
Injecting numerical reasoning skills into language models
Geva, M., Gupta, A., and Berant, J · 2020
Cited alongside, same era.
End-to-end bias mitigation by modelling biases in corpora
Karimi Mahabadi, R., Belinkov, Y., and Henderson, J · 2020
Cited alongside, same era.
Are pretrained language models symbolic reasoners over knowledge?
Kassner, N., Krojer, B., and Schütze, H · 2020
Cited alongside, same era.
Question and answer test-train overlap in open-domain question answering datasets
Original
Lewis, P., Stenetorp, P., and Riedel, S · 2020
Cited alongside, same era.
Birds have four legs?! NumerSense: Probing Numerical Commonsense Knowledge of Pre-Trained Language Models
Lin, B. Y., Lee, S., Khanna, R., and Ren, X · 2020
Cited alongside, same era.

Then

GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow, March 2021
Black, S., Gao, L., Wang, P., Leahy, C., and Biderman, S · 2021
Later among the works it cites.
On the opportunities and risks of foundation models
Original
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al · 2021
Later among the works it cites.
Evaluating entity disambiguation and the role of popularity in retrieval-based NLP
Chen, A., Gudipati, P., Longpre, S., Ling, X., and Singh, S · 2021
Later among the works it cites.
Documenting large webtext corpora: A case study on the colossal clean crawled corpus
Dodge, J., Sap, M., Marasović, A., Agnew, W., Ilharco, G., Groeneveld, D., Mitchell, M., and Gardner, M · 2021
Later among the works it cites.
Memorization vs. generalization : Quantifying data leakage in NLP performance evaluation
Elangovan, A., He, J., and Verspoor, K · 2021
Later among the works it cites.
Competency problems: On finding and removing artifacts in language data
Gardner, M., Merrill, W., Dodge, J., Peters, M., Ross, A., Singh, S., and Smith, N. A · 2021
Later among the works it cites.
Datasheets for datasets, 2021
Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J. W., Wallach, H., au2, H. D. I., and Crawford, K · 2021
Later among the works it cites.
Question and answer test-train overlap in open-domain question answering datasets
Lewis, P., Stenetorp, P., and Riedel, S · 2021
Later among the works it cites.
What’s in the box? an analysis of undesirable content in the Common Crawl corpus
Luccioni, A. and Viviano, J · 2021
Later among the works it cites.
Are nlp models really able to solve simple math word problems?, 2021
Patel, A., Bhattamishra, S., and Goyal, N · 2021
Later among the works it cites.
Masked language modeling and the distributional hypothesis: Order word matters pre-training for little
Sinha, K., Jia, R., Hupkes, D., Pineau, J., Williams, A., and Kiela, D · 2021
Later among the works it cites.
We need to talk about random splits
Søgaard, A., Ebert, S., Bastings, J., and Filippova, K · 2021
Later among the works it cites.
Concealed data poisoning attacks on NLP models
Wallace, E., Zhao, T., Feng, S., and Singh, S · 2021
Later among the works it cites.
Mesh-Transformer-JAX: Model-Parallel Implementation of Transformer Language Model with JAX
Wang, B · 2021
Later among the works it cites.
Frequency effects on syntactic rule learning in transformers
Wei, J., Garrette, D., Linzen, T., and Pavlick, E · 2021
Later among the works it cites.
Counterfactual memorization in neural language models
Original
Zhang, C., Ippolito, D., Lee, K., Jagielski, M., Tramèr, F., and Carlini, N · 2021
Later among the works it cites.
Rethinking the role of demonstrations: What makes in-context learning work?
Original
Min, S., Lyu, X., Holtzman, A., Artetxe, M., Lewis, M., Hajishirzi, H., and Zettlemoyer, L · 2022
Closest in time.
Hypothesis only baselines in natural language inference
Poliak, A., Naradowsky, J., Haldar, A., Rudinger, R., and Van Durme, B · 2023
Closest in time.

Beyond the bibliography

alphaXiv searches the wider corpus for related work and actual follow-ups.

Open on alphaXiv

alphaXiv is searching for related work…

Reading the bibliography…