US 11,657,236 B2
Tunable bias reduction pipeline
Aonghus McGovern, Dublin (IE); Abhishek Khanna, Dublin (IE); Rebekah Murphy, Dublin (IE); Steve Cooper, Dublin (IE); and Xin Zuo, Dublin (IE)
Assigned to Accenture Global Solutions Limited, Dublin (IE)
Filed by Accenture Global Solutions Limited, Dublin (IE)
Filed on May 26, 2020, as Appl. No. 16/883,673.
Application 16/883,673 is a continuation of application No. 16/423,067, filed on May 27, 2019, granted, now 10,671,942.
Claims priority of provisional application 62/812,005, filed on Feb. 28, 2019.
Prior Publication US 2020/0285999 A1, Sep. 10, 2020
This patent is subject to a terminal disclaimer.
Int. Cl. G06N 20/00 (2019.01); G06F 40/20 (2020.01); G06F 18/22 (2023.01); G06F 18/24 (2023.01)
CPC G06N 20/00 (2019.01) [G06F 18/22 (2023.01); G06F 18/24 (2023.01); G06F 40/20 (2020.01)] 20 Claims
OG exemplary drawing
 
1. A system for reducing bias in an artificial intelligence model, the system comprising:
a processor and a computer readable medium, the computer readable medium comprising instructions executable by the processor to:
receive a word embedding model generated based on a corpus of words, the word embedding model comprising word vectors representative of the corpus of words;
determine a bias definition vector based on the word embedding model, the bias definition vector being defined along a bias axis representative of a bias type in the word embedding model;
receive a bias classification criteria, the bias classification criteria comprising logic to group the word vectors based on a distance measurement from the bias definition vector;
identify, in the word embedding model, a first group of word vectors and a second group of word vectors based on the bias classification criteria and the bias definition vector, the first group of word vectors being representative of a first bias category for the bias type and the second group of word vectors being representative of a second bias category for the bias type;
adjust a first bias threshold to decrease the quantity of word vectors in the first group of word vectors and adjust a second bias threshold to increase the quantity of word vectors in the second group of word vectors;
generate a word embedding model where the word vectors from the first group and the word vectors from the second group are weighted;
adjust, based on the adjusted first bias threshold and the adjusted second bias threshold, the respective word vectors included in the first group of word vectors and the second group of word vectors;
determine that a debias criteria is satisfied based on the adjusted quantity of word vectors in the first group of word vectors and the adjusted quantity of word vectors in the second group of word vectors; and
generate a debiased artificial intelligence model, the debiased artificial intelligence model comprising metrics representative of words, wherein metrics for words associated with the first group of word vectors and metrics for words associated with the second group of word vectors are weighted based on a non-zero penalization factor.