Selective-LAMA: Selective Prediction for Confidence-Aware Evaluation of Language Models

Findings of the Association for Computational Linguistics: EACL 2023, pages 2017–2028 May 2-6, 2023 ©2023 Association for Computational Linguistics Selective-LAMA: Selective Prediction for Confidence-Aware Evaluation of Language Models Hiyori Yoshikawa†‡ Naoaki Okazaki† †Tokyo Institute of Technology, Japan ‡Fujitsu Limited, Japan {hiyori.yoshikawa@nlp., okazaki@}c.titech.ac.jp Abstract Recent studies have suggested that neural lan- guage models learn and store a large amount of facts and commonsense knowledge from train- ing data. The ability of language models to restore such knowledge is often evaluated via zero-shot cloze-style QA tasks. However, such evaluations rely only on prediction accuracy without punishing the systems for their mis- takes, e.g., simply guessing or hallucinating likely answers. Selective prediction is a more informative evaluation framework that takes the confidence of predictions into account. Under the selective prediction setting, a model is eval- uated not only by the number of correct predic- tions, but also by the ability to filter out dubious predictions by estimating the confidence of in- dividual predictions. Such confidence-aware evaluation is crucial for determining whether to trust zero-shot predictions of language models. In this paper, we apply the selective predic- tion setting to an existing benchmark, LAMA probe, and conduct extensive experiments with recent neural language models and different confidence functions. We empirically show that our Selective-LAMA evaluation is more robust to the effect of simple guesses than the conventional accuracy-based evaluation. Our evaluation reveals the importance of the choice of confidence functions by showing that sim- ply relying on token probabilities is not always the best choice. Further analysis shows that various confidence functions exhibit different preferences over predicted tokens for a given context. 1 Introduction Recently, knowledge stored in pre-trained lan- guage models has been intensively investigated. Many studies have suggested that language mod- els trained on a large amount of textual corpora, such as BERT (Devlin et al., 2019) and GPT (Rad- ford et al., 2019; Brown et al., 2020), store both linguistic knowledge (Warstadt et al., 2019; Mi- aschi et al., 2020) and factual and commonsense knowledge (Bosselut et al., 2019; Roberts et al., 2020) during training. However, this knowledge is embedded in the parameters of these language models and thus is difficult to interpret, in contrast to symbolic knowledge bases, which allows us to inspect and edit stored facts explicitly. Petroni et al. (2019) proposed a benchmark task, the LAMA probe, that aims at evaluating the amount of relational knowledge, such as common- sense knowledge and facts, which is stored in a language model. In LAMA probe, a relational fact is converted into a cloze statement (query) and then given to a language model as a fill-in-the-blank question. If the language model fills in the blank with the correct answer, the model is considered to possess “knowledge” of the relation. According to Petroni et al.’s experiments, the BERT language model (Devlin et al., 2019) has a comparable perfor- mance to a supervised relation extraction baseline, with precision ranging from 10.5 to 32.3 depending on the dataset type. However, in many applications, we …

Content truncated to 500 words