File size: 2,439 Bytes
023b7aa
 
 
 
 
 
 
 
 
 
 
 
2fba2f8
 
 
023b7aa
 
 
 
 
2fba2f8
023b7aa
fdaee21
023b7aa
 
 
 
 
 
 
 
 
 
 
 
 
f289ddb
023b7aa
 
 
5e4fefd
023b7aa
 
 
 
 
 
 
 
 
 
 
 
f220d74
023b7aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2fba2f8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
---
license: apache-2.0
tags:
- generated_from_trainer
metrics:
- precision
- recall
- f1
- accuracy
model-index:
- name: bert-finetuned-ner
  results: []
language:
- en
pipeline_tag: token-classification
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

# bert-finetuned-AAVE-PoS

This model is a version of [bert-base-cased](https://huggingface.co/bert-base-cased) fine-tuned on a [dataset](https://bitbucket.org/soegaard/aave-pos16/src/master/data) of African American Vernacular English (AAVE) which was published alongside [Jørgensen et al. 2016](https://aclanthology.org/N16-1130.pdf).
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](http://www.johnrickford.com/portals/45/documents/papers/Rickford-1999e-Phonological-and-Grammatical-Features-of-AAVE.pdf) 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](https://github.com/DrewGalbraith/AAE-PoS/tree/main).

## 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