# About fill-mask pipeline with \[mask\] made up of multiple tokens

**URL:** <https://discuss.huggingface.co/t/about-fill-mask-pipeline-with-mask-made-up-of-multiple-tokens/37557>\
**Category:** 🤗Transformers\
**Created:** [April 24, 2023, 8:13am UTC](https://discuss.huggingface.co/t/about-fill-mask-pipeline-with-mask-made-up-of-multiple-tokens/37557 "2023-04-24T08:13:10Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![iamseokhyun](https://avatars.discourse-cdn.com/v4/letter/i/ecd19e/32.png) [@iamseokhyun](https://discuss.huggingface.co/u/iamseokhyun)\
**Post date:** [April 24, 2023, 8:13am UTC](https://discuss.huggingface.co/t/about-fill-mask-pipeline-with-mask-made-up-of-multiple-tokens/37557/1 "2023-04-24T08:13:10Z")

</div>

For the bert-base-uncased tokenizer, it tokenizes a “rectangle” into two tokens. How does the fill-mask pipeline work when a word like rectangle in a sentence is whole-word masked?

For example, if the sentence “My car is rectangle shape” is the original sentence, and “My car is [MASK] shape” is given as an input, can the entire word rectangle be restored? If not, what is the general approach to recovering whole words?

I have to perform a task to determine which shape word has the highest score among triangle, circle, and rectangle in that [MASK].