DrewG/AAVE_PoS-Tagger

bert-finetuned-AAVE-PoS

This model is a version of bert-base-cased fine-tuned on a dataset of African American Vernacular English (AAVE) which was published alongside Jørgensen et al. 2016. It achieves the following results on the evaluation set:

• Loss: 0.2582
• Precision: 0.8632
• Recall: 0.8730
• F1: 0.8681
• Accuracy: 0.9356

Model description

More information needed

Intended uses & limitations

This model is intended to help close the gap in part-of-speech tagging performance between Standard American English (SAE) and African American English (AAVE) which differ liguistically in many well-documented ways. It was fine-tuned on data gathered from Twitter, and is thus ingrained with what linguists call 'register bias'.

Training and evaluation data

Code hosted at GitHub.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3 (this amount of data overfits on 3+)

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
No log	1.0	223	0.2982	0.8196	0.8350	0.8272	0.9216
No log	2.0	446	0.2625	0.8599	0.8680	0.8640	0.9326
0.4647	3.0	669	0.2582	0.8632	0.8730	0.8681	0.9356

Framework versions

• Transformers 4.29.2
• Pytorch 1.13.1+cpu
• Datasets 2.12.0
• Tokenizers 0.13.3