jasonshaoshun / amsal

code for "Erasure of Unaligned Attributes from Neural Representations"

Erasure of Unaligned Attributes from Neural Representations

This repository contains the code for the experiments and algorithm from the paper Erasure of Unaligned Attributes from Neural Representations (appears at TACL 2023).

Introduction

We present the Assignment-Maximization Spectral Attribute removaL (AMSAL) algorithm, which aims at removing information from neural representations when the information to be erased is implicit rather than directly being aligned to each input example.

It first loops between two Alignment and Maximization steps, during which it finds an alignment (A) based on existing projections and then projects the representations and guarded attributes into a joint space based on an existing alignment (M). After these two steps are iteratively repeated and an alignment is identified, the algorithm takes another step to erase information from the input representations based on the projections identified.

Experimental Setting and Datasets

We use the Dataset and Experimental Settings from the paper

Paper	Github Link	Notes
INLP	View	Null it out: guarding protected attributes by iterative nullspsace projection
Bias-Bench	View	An Empirical Survey of the Effectiveness of Debiasing Techniques for Pre-trained Language Models
vulgartwitter	View	Expressively vulgar: The socio-dynamics of vulgarity and its effects on sentiment analysis in social media

Datasets Download

The external datasets required by the projects above

Project	Dataset	Download Link	Notes	Download Directory
Bias-Bench	Wikipedia-2.5	Download	English Wikipedia dump used for SentenceDebias and INLP. (Section 4.3)	data/text
INLP	Word Embedding	Download	GloVe Word Embeddings for Most Gendered Word (Section 4.1)	data/word-embedding
INLP	BiasBios	Download	BiasBios dataset for fair profession prediction (Section 4.1)	data/biography

Each dataset should be downloaded to the specified path, relative to the root directory of the project.

Datasets Size

This dataset and model size

Section	Experiment	Model	Dataset size	Model Size
4.1	Word Embedding Debiasing	GloVe	18M	1 M
4.2	BiasBios	BertModel	340M	432 M
~	~	FastText	134M	164 M
4.3	BiasBench	BertModel + ALBERT + RoBERTa + GPT-2	>> 1G (Provided by BiasBench)	257M
4.4	Multiple-Guarded Attribute Sentiment	BertModel	55M	76 M
4.5	Method Limitation (sentiment analysis on twitter)	DeepMoji	244M	332 M

Experiments

Suppose each dataset are downloaded or created to the specified path.

Word Embedding Debiasing (Section 4.1 in the paper)

Move the script file under the folder of script_glove to the root of matlab. Then call the script file to run AMSAL with specified hyperparameters, or call another script file to queue the experimental scripts automatically by the following code.

./src/matlab/matscript.sh

Please note: you need to change the start index and end index in matscript.sh, depends on how many matlab jobs you want to submit at once.

Check the spreadsheet created under the folder data, named full.xlsx and epoch.xlsx.

BiasBios Experiments (Section 4.2 in the paper)

AM Step

Move the script files under the folder of biography to the root of matlab, then run AMSAL with hyperparameters specified in the scripts by calling the scripts one by one or automatically queueu the jobs by calling the following code.

./src/matlab/matscript.sh

Check the result in the spreadsheet created under the folder of data, named full.xlsx and epoch.xlsx.

Please also check the script files under the folder of BertModel_PARTIAL and FastTextModel_PARTIAL to run the partially supervised assignment of AMSAL.

Removal Step and Downstream Tasks Evaluation

Run the script files to debias the neural representations of the biographies on genders, and perform the profession classifications.

./src/assignment/tpr-gap_biasbios.sh

Export the results to spreadsheet and overleaf code

./src/assignment/export_tpr-gap_biography.sh

The tables will be stored in the folder of tables and the overleaf code will be stored in the folder of biography.

BiasBench Experiments (Section 4.3 in the paper)

Create Dataset

./src/bias-bench/batch_jobs/create_data.sh

Alignment and Maximization step

Move the script files under the folder of biasbench to the root of matlab, then run AMSAL with hyperparameters specified in the scripts by calling the scripts one by one or automatically queueu the jobs by calling the following code.

./src/matlab/matscript.sh

Check the result in the spreadsheet created under the folder of data, named full.xlsx and epoch.xlsx.

Removal Step

# Compute the projection matrix by SAL
./src/bias-bench/batch_jobs/sal_eigenvector.sh

# Compute the projection matrix by INLP
./src/bias-bench/batch_jobs/inlp_projection_matrix.sh

Downstream Tasks

Task 1 Stereoset (section 4.3 in the paper)

# Evaluates non-debiased models against StereoSet
./src/bias-bench/batch_jobs/stereoset.sh

# Evaluates debiased models against StereoSet
./src/bias-bench/batch_jobs/stereoset_debias.sh

# Summary the results
./src/bias-bench/batch_jobs/stereoset_evaluation.sh

# Export to LaTeX code
./src/bias-bench/batch_jobs/export_stereoset.sh

Please check the spreadsheet and results json file under the folder of stereoset, also the LaTex code stored in the folder of tables.

Task 2 CrowS-Pairs (section 4.3 in the paper)

# Evaluates non-debiased models against CrowS-Pairs
./src/bias-bench/batch_jobs/crows.sh

# Evaluates debiased models against CrowS-Pairs
./src/bias-bench/batch_jobs/crows_debias.sh

# Export to LaTeX code
./src/bias-bench/batch_jobs/export_crows.sh

Please check the results saved in the folder of CrowS-Pairs and LaTex code stored in the folder of tables.

Task 3 SEAT (Appendix B)

# Evaluates non-debiased models against SEAT
./src/bias-bench/batch_jobs/seat.sh

# Evaluates debiased models against SEAT
./src/bias-bench/batch_jobs/seat_debias.sh

# Export to LaTeX code
./src/bias-bench/batch_jobs/export_seat.sh

Please check the results saved in the folder of SEAT and LaTex code stored in the folder of tables.

Task 4 GLUE (Appendix B)

Very Complex, will be explained in a separate page

Twitter Sentiment with Multiple Guarded Attributes Experiments (Section 4.4 in the paper)

Create Dataset

# Save the Twitter Sentiment (Political) datasets in npz and matlab readable format
# Twitter dataset with original labelling
./src/assignment/create_twitter_dataset_.sh

# Twitter dataset labelled by word overlap
./src/assignment/create_twitter_dataset_labelled_by_overlap_.sh

Deepmoji Experiments (Section 4.5 in the paper)

Deepmoji experiments consider the stereotypes of race in the twitters.

AM Step

Move the script files under the folder of deepmoji to the root of matlab, then run AMSAL with hyperparameters specified in the scripts by calling the scripts one by one or automatically queueu the jobs by calling the following code.

./src/matlab/matscript.sh

Check the result in the spreadsheet created under the folder of data, named full.xlsx and epoch.xlsx.

Please also check the script files under the folders of R5S5, R5S8 and R8S5 to run the partially supervised assignment of AMSAL.

Removal Step and Downstream Tasks Evaluation

Run the script files to debias the neural representations of the biographies on genders, and perform the profession classifications.

./src/assignment/tpr-gap_deepmoji.sh

Export the results to spreadsheet and overleaf code

./src/assignment/export_tpr-gap_table-deepmoji.sh

The tables will be stored in the folder of tables and the overleaf code will be stored in the folder of deepmoji.