Code accompanying the ACL paper A Better Way to Do Masked Language Model Scoring by Carina Kauf & Anna Ivanova.

Code adapts the minicons implementation of Salazar et al. (2020)'s PLL metric.

Estimating the log-likelihood of a given sentence under an autoregressive language model is straightforward: one can simply apply the chain rule and sum the log-likelihood values for each successive token. However, for masked language models, there is no direct way to estimate the log-likelihood of a sentence. To ad- dress this issue, Salazar et al. (2020) propose to estimate sentence pseudo-log-likelihood (PLL) scores, computed by successively masking each sentence token, retrieving its score using the rest of the sentence as context, and summing the resulting values. Here, we demonstrate that the original PLL method yields inflated scores for out-of-vocabulary words and propose an adapted metric, in which we mask not only the target token, but also all within-word tokens to the right of the target. We show that our adapted metric (PLL-word-l2r) outperforms both the original PLL metric and a PLL metric in which all within-word tokens are masked. In particular, it better satisfies theoretical desiderata and better correlates with scores from autoregressive models. Finally, we show that the choice of metric affects even tightly controlled, minimal pair evaluation benchmarks (such as BLiMP), underscoring the importance of selecting an appropriate scoring metric for evaluating MLM properties

We improve on the widely used pseudo-log-likelihood sentence score estimation method for masked language models from Salazar et al. (2020) by solving the score inflation problem for multi-token words the original PLL metric introduces, without overly penalizing multi-token words. The new metric also fulfills several theoretical desiderata of sentence/word scoring methods better than alternative PLL metrics and is more cognitively plausible:

git clone --recursive https://github.com/carina-kauf/better-mlm-scoring.git
cd better-mlm-scoring
pip install -r requirements.txt 

Scoring the bert-base-cased model on the EventsAdapt dataset using the PLL-word-l2r metric:

cd better-mlm-scoring-analyses

python dataset_scoring.py --dataset EventsAdapt \
                          --model bert-base-cased \
                          --which_masking within_word_l2r

Masking options are original, within_word_l2r, within_word_mlm, global_l2r

If you use this work, please cite:

@misc{kauf2023better,
      title={A Better Way to Do Masked Language Model Scoring}, 
      author={Carina Kauf and Anna Ivanova},
      year={2023},
      eprint={2305.10588},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

• adapted_minicons : A stripped version of the minicons package in which we adapt the MaskedLMScorer class in the scorer module.
• better-mlm-scoring-analyses : Data and scripts used to analyze the novel masked language model scoring metrics

• master : Code accompanying the paper A Better Way to Do Masked Language Model Scoring by Carina Kauf & Anna Ivanova
• pr-branch-into-minicons : Code used to add the optimized masked-language modeling metric (PLL-word-l2r) to the minicons library maintained by Kanishka Misra. This repo is a wrapper around the transformers library from HuggingFace 🤗 and supports functionalities for (i) extracting word representations from contextualized word embeddings and (ii) scoring sequences using language model scoring techniques.