Avoiding Inference Heuristics in Few-shot Prompt-based Finetuning

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9063–9074 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 9063 Avoiding Inference Heuristics in Few-shot Prompt-based Finetuning Prasetya Ajie Utama†‡ , Nafise Sadat Moosavi‡, Victor Sanh♣, Iryna Gurevych‡ †Research Training Group AIPHES ‡UKP Lab, Technische Universität Darmstadt ♣Hugging Face, Brooklyn, USA ‡https://www.ukp.tu-darmstadt.de utama@ukp.tu-darmstadt.de Abstract Recent prompt-based approaches allow pre- trained language models to achieve strong per- formances on few-shot finetuning by reformu- lating downstream tasks as a language mod- eling problem. In this work, we demon- strate that, despite its advantages on low data regimes, finetuned prompt-based models for sentence pair classification tasks still suffer from a common pitfall of adopting inference heuristics based on lexical overlap, e.g., mod- els incorrectly assuming a sentence pair is of the same meaning because they consist of the same set of words. Interestingly, we find that this particular inference heuristic is sig- nificantly less present in the zero-shot evalu- ation of the prompt-based model, indicating how finetuning can be destructive to useful knowledge learned during the pretraining. We then show that adding a regularization that pre- serves pretraining weights is effective in mit- igating this destructive tendency of few-shot finetuning. Our evaluation on three datasets demonstrates promising improvements on the three corresponding challenge datasets used to diagnose the inference heuristics.1 1 Introduction Prompt-based finetuning has emerged as a promis- ing paradigm to adapt Pretrained Language Models (PLM) for downstream tasks with limited number of labeled examples (Schick and Schütze, 2021a; Radford et al., 2019). This approach reformulates downstream task instances as a language modeling input,2 allowing PLMs to make non-trivial task- specific predictions even in zero-shot settings. This in turn, provides a good initialization point for data efficient finetuning (Gao et al., 2021), resulting in 1The code is available at https://github.com/ UKPLab/emnlp2021-prompt-ft-heuristics 2E.g., appending a cloze prompt “It was [MASK]” to a sentiment prediction input sentence “Delicious food!”, and obtaining the sentiment label by comparing the probabilities assigned to the words “great” and “terrible”. a strong advantage on low data regimes where the standard finetuning paradigm struggles. However, the success of this prompting approach has only been shown using common held-out evaluations, which often conceal certain undesirable behaviors of models (Niven and Kao, 2019). One such behavior commonly reported in down- stream models is characterized by their preference to use surface features over general linguistic in- formation (Warstadt et al., 2020). In the Natu- ral Language Inference (NLI) task, McCoy et al. (2019) documented that models preferentially use the lexical overlap feature between sentence pairs to blindly predict that one sentence entails the other. Despite models’ high in-distribution performance, they often fail on counterexamples of this inference heuristic, e.g., they predict that “the cat chased the mouse” entails “the mouse chased the cat”. At the same time, there is a mounting evidence that pre-training on large text corpora extracts rich linguistic information (Hewitt and Manning, 2019; Tenney et al., 2019). However, based on recent studies, standard finetuned models often overlook …

Content truncated to 500 words