Unconscious bias detection

Patent metadata

Field	Value
Publication number	US-10839153-B2
Application number	US-201715604549-A
Country	US
Kind code	B2
Filing date	May 24, 2017
Priority date	May 24, 2017
Publication date	Nov 17, 2020
Grant date	Nov 17, 2020

Abstract

The discussion relates to unconscious bias detection. One example can detect potential bias words in a text sample and suggest alternative phrasing to eliminate the potential bias words from the text sample.

First claim

The invention claimed is:

1. A system, comprising:
storage storing computer-executable instructions for:
obtaining text that includes words arranged as phrases or sentences;
identifying unconscious bias candidate words from the text by determining that a gender-specific pro-form does not have a corresponding expression in the text that the gender-specific pro-form refers to;
selecting a bias score for an individual unconscious bias candidate word from a table of known bias words and associated bias severities for the known bias words;
determining a particular class from an ontological classification that the individual unconscious bias candidate word belongs to, the ontological classification including a plurality of classes and acceptable bias thresholds associated with the plurality of classes;
determining a particular acceptable bias threshold associated with the determined particular class to which the individual unconscious bias candidate word belongs;
in an instance when the bias score exceeds a flagging threshold and also exceeds the particular acceptable bias threshold, flagging the individual unconscious bias candidate word and suggesting one or more non-biased words that are gender-neutral to replace the individual unconscious bias candidate word; and
presenting the flagged individual unconscious bias candidate word and the one or more non-biased words on a graphical user interface by displaying, in a first subset of the graphical user interface, a portion of the text containing the flagged individual unconscious bias candidate word and highlighting the flagged individual unconscious bias candidate word, and further displaying, in a second subset of the graphical user interface, a listing of suggested alternative language choices using the one or more non-biased words in association with the flagged individual unconscious bias candidate word, wherein selection of one of the suggested alternative language choices automatically replaces the flagged individual unconscious bias candidate word in the text; and,

a processor that executes the computer-executable instructions.

2. The system of claim 1 , embodied on a user device.

3. The system of claim 1 , embodied on cloud-based resources.

4. The system of claim 1 , wherein highlighting the flagged individual unconscious bias candidate word includes bolding the flagged individual unconscious bias candidate word.

5. The system of claim 1 , wherein highlighting the flagged individual unconscious bias candidate word includes underlining the flagged individual unconscious bias candidate word.

6. The system of claim 1 , wherein the listing of suggested alternative language choices is displayed in association with the flagged individual unconscious bias candidate word by being displayed in proximity to the flagged individual unconscious bias candidate word.

7. At least one computer-readable storage medium having instructions stored thereon that, when executed by a processor of a computing device, cause the computing device to perform acts, comprising:
identifying a potential bias word that indicates a gender bias and relationships between the potential bias word and other words of user text based at least in part on determining that the potential bias word is a gender-specific pro-form that does not correspond to an expression in the other words of the user text that the gender-specific pro-form refers to;
determining a bias score of the potential bias word by accessing a listing of bias words and associated bias scores;
determining a particular class and a particular acceptable bias threshold from an ontological classification that the potential bias word belongs to, the ontological classification including a plurality of classes and acceptable bias thresholds associated with the plurality of classes;
indicating when the bias score of the potential bias word exceeds the particular acceptable bias threshold by displaying, on a first subset portion of a graphical user interface, the potential bias word being highlighted and further displaying, on a second subset portion of the graphical user interface, a non-biased word alternative in association with the potential bias word, the non-biased word alternative being a gender-neutral word alternative; and
replacing the potential bias word with the non-biased word alternative in the user text upon receiving a selection of the displayed non-biased word alternative via the graphical user interface of the computing device.

8. The at least one computer-readable storage medium of claim 7 , wherein the determining the bias score comprises comparing the potential bias word to the listing of bias words.

9. The at least one computer-readable storage medium of claim 7 , wherein the highlighting comprises bolding or underlining the potential bias word in the first subset portion of the graphical user interface.

10. The at least one computer-readable storage medium of claim 7 , wherein the indicating comprises providing the non-biased word alternative to a user that generated the user text.

11. The at least one computer-readable storage medium of claim 7 , wherein the indicating comprises providing an explanation to a user that generated the user text in a third subset portion of the graphical user interface.

12. The at least one computer-readable storage medium of claim 7 , wherein the indicating comprises providing an explanation about the potential bias word in a third subset portion of the graphical user interface, and providing the non-biased word alternative to a user that generated the user text in the second subset portion of the graphical user interface.

13. The at least one computer-readable storage medium of claim 7 , wherein the determining the bias score comprises comparing the potential bias word to a static listing of known bias words and a dynamic listing of known bias words.

14. The at least one computer-readable storage medium of claim 7 , wherein the acts further comprise:
removing one or more stop words from the user text; or
stemming the potential bias word.

15. The at least one computer-readable storage medium of claim 7 , wherein the acts further comprise:
using a natural language processing module to extract syntactic features from the user text,
wherein the syntactic features include parts of speech and/or co-references.

16. A method, comprising:
identifying a gender-specific pro-form that does not have a corresponding expression in a text sample that the gender-specific pro-form refers to;
determining a bias score for the gender-specific pro-form, the bias score indicating a severity of bias;
determining a particular class from an ontological classification that the gender-specific pro-form belongs to, the ontological classification including a plurality of classes and bias thresholds associated with the plurality of classes;
flagging the gender-specific pro-form as an unconscious bias word in response to determining that the bias score for the gender-specific pro-form exceeds a particular bias threshold associated with the particular class to which the gender-specific pro-form belongs;
displaying a graphical user interface having first and second subset portions, the first subset portion displaying the unconscious bias word highlighted, and the second subset portion displaying a listing of suggested alternative phrasing to eliminate the unconscious bias word from the text sample, the listing of suggested alternative phrasing being displayed in association with the unconscious bias word, wherein the alternative phrasing is a gender-neutral phrasing that can be selected via

Assignees

• Microsoft Technology Licensing Llc

Inventors

• Gaur Naveen
• Kulkarni Priyanka Subhash
• Menezes Erika Debra
• Zhang Wenlei

Classifications

• G06F40/35

  Discourse or dialogue representation · CPC title

• G06F40/253 Primary

  Grammatical analysis; Style critique · CPC title

• G06F40/211

  Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars · CPC title

• G06F40/289

  Phrasal analysis, e.g. finite state techniques or chunking · CPC title