BabyBERTa-2

BabyBERTA

Overview

BabyBERTa is a light-weight version of RoBERTa trained on 5M words of American-English child-directed input. It is intended for language acquisition research, on a single desktop with a single GPU - no high-performance computing infrastructure needed.

The three provided models are randomly selected from 10 that were trained and reported in the paper.

Loading the tokenizer

BabyBERTa was trained with add_prefix_space=True, so it will not work properly with the tokenizer defaults. For instance, to load the tokenizer for BabyBERTa-1, load it as follows:

tokenizer = RobertaTokenizerFast.from_pretrained("phueb/BabyBERTa-1",
                                                 add_prefix_space=True)

Hyper-Parameters

See the paper for details. All provided models were trained for 400K steps with a batch size of 16. Importantly, BabyBERTa never predicts unmasked tokens during training - unmask_prob is set to zero.

Performance

BabyBerta was developed for learning grammatical knowledge from child-directed input. Its grammatical knowledge was evaluated using the Zorro test suite. The best model achieves an overall accuracy of 80.3, comparable to RoBERTa-base, which achieves an overall accuracy of 82.6 on the latest version of Zorro (as of October, 2021). Both values differ slightly from those reported in the CoNLL 2021 paper. There are two reasons for this:

1. Performance of RoBERTa-base is slightly larger because the authors previously lower-cased all words in Zorro before evaluation. Lower-casing of proper nouns is detrimental to RoBERTa-base because RoBERTa-base has likely been trained on proper nouns that are primarily title-cased. In contrast, because BabyBERTa is not case-sensitive, its performance is not influenced by this change.
2. The latest version of Zorro no longer contains ambiguous content words such as "Spanish" which can be both a noun and an adjective. this resulted in a small reduction in the performance of BabyBERTa.

Overall Accuracy on Zorro:

Additional Information

This model was trained by Philip Huebner, currently at the UIUC Language and Learning Lab.

More info can be found here.