-
Notifications
You must be signed in to change notification settings - Fork 0
/
Abstract.txt
73 lines (47 loc) · 3.83 KB
/
Abstract.txt
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
Class Project: Explore the biases (Gender, unintended, etc.) in the knowledge distilled model (TinyBert) of Bert and potential ways to mitigate them.
UIN:
Problem description and motivation:
Although transformer based models like ELMO, Bert, and OpenGPT have pushed the boundaries of NLP across various language tasks,
they suffer from significant amount of gender and unintended biases. When these large models are knowledge distilled into smaller
models they propagate the biases, along with the learnings, they have into the distilled student models. As studied in [1], the
distilled versions have biases exacerbated in them compared to the corresponding teacher or source model. Our objective here is to
explore bias, for example gender bias, unintended bias, in a smaller knowledge distilled version of Bert model, called TinyBert.
Understand what contributes to the bias and explore the methods that could potentially help mitigate them.
Dataset and Evaluation Plan:
Following are the datasets that we might be using:
1) Wikipedia dataset - https://huggingface.co/datasets/wikipedia#20220301en
2) BookCorpus dataset - https://huggingface.co/datasets/bookcorpusopen
3) IMdB sentiment dataset - http://ai.stanford.edu/~amaas/data/sentiment/
4) Automatic Misogyny Identification dataset - https://github.com/MIND-Lab/ami2018 - this requires filling google form
5) MLMA hate speech dataset - https://github.com/HKUST-KnowComp/MLMA_hate_speech
6) SEAT related synthetic dataset - https://github.com/W4ngatang/sent-bias
Note: Depending on the amount of time available, we may not use all the datasets mentioned above or we may even use some extra
dataset to substantiate the analysis.
Following are the metrics we use to measure gender bias and other unintended bias:
1. SEAT : Sentence Encoder Association Test. This is a test that used to measure bias in sentence encoders.
2. F1 scores and AUC (when testing the model backbone on a classification downstream task)
Note: Depending on the kind of analysis we moving forwad, we may add other metrics for the analysis.
Planned Tentative Approach (Summary):
1) Understand and process the datasets and the evaulation criterian math(SEAT especially).
2) Configure the TinyBert or re-train one from scratch using knowledge distillation.
3) Understand what kind of biases it contain and quantify them.
4) Explore bias mitigation methods onto the TinyBert and re-train it accordingly.
5) Finally compare the performance of debiased TinyBert model with non-debiased TinyBert model.
Refernces:
[1] https://aclanthology.org/2022.gebnlp-1.27.pdf
https://huggingface.co/huawei-noah/TinyBERT_General_4L_312D/tree/main
https://github.com/yinmingjun/TinyBERT
Training - handle pipeline for training from scratch and explore mitigation pipelines too.
https://arxiv.org/pdf/1909.10351.pdf
https://aclanthology.org/2022.gebnlp-1.27.pdf
https://aclanthology.org/2022.findings-acl.88.pdf
testing - they can take off the shelf tinybert and use it to create the pipeline for testing. Referencing IMDb, SEAT, MISOGYNY and MLMN
https://aclanthology.org/2022.gebnlp-1.20.pdf - IMDB related links are posted in this paper, testing related stuff is there in this paper
that we can expolate for our project
https://aclanthology.org/2022.gebnlp-1.27.pdf - SEAT related papers are posted in this paper, testing related stuff is there in this paper
that we can expolate for our project
https://aclanthology.org/2022.findings-acl.88.pdf - you can see information and link regarding the the misogyny dataset in this papar, testing related stuff is there in this paper
that we can expolate for our project
Notes:
1) Abstract understanding and then splitting, in the coming week. 22 Sept - 25 sept
2) Wrap it up becore thanks giving