Google's sister company announces the "Perspective" tool to detect hate speech, and what is the result of the expert's accuracy verification?

"Jigsaw", a sister company of Google under the umbrella of Alphabet Group, is a tool that automatically detects hate speech on the Internet by machine learning "PerspectiveWe announced. Although it is an API that can prevent diffusion of hate speech by implementing in the comment section on the website etc., I know that comments that are not many hate speech will be regarded as "harmful comments".

Google's attempt to police trolling and hate on the internet does not work yet - Quartz
https://qz.com/918640/alphabets-hate-fighting-ai-doesnt-understand-hate-yet/

Jigsaw's "Perspective"Official websiteThere is a comment input column on the top, so that anyone can easily try out the "hazard" of written contents. Many experts have tried this function, and it is said that harmless words such as "hate is bad." "Garbage truck (garbage truck)" are considered harmful hate speech It has been reported.


For example, in the following example, while "Racism is bad." Is judged as "harmful comment with a probability of 60%", "Racism is good." (Racial discrimination "The principle is a good thing)" is judged as "a harmful comment with a probability of 35%", and words that recommend racial discrimination show low harmfulness.

Content neutral algorithms will never help usPic.twitter.com/2x7BUdYEyw

- guillotines for (@ jbradfield)February 23, 2017

Being a Google former engineerMIT Tech ReviewReporter David Auerbach himself inputs various comments to Perspective and posts a list with the results to Facebook. Words such as "Donald Trump is a meretricious buffoon" (Donald · Tramp is an invalid dog) and "few muslims are a terrorist threat (several Muslims are a threat of terrorism)" indicate correctly high hazard On the other hand, the word 'I fucking love you man. Happy birthday' (I love you, birthday Happy birthday) 'shows the harmful side by side with them, so it is difficult to distinguish "F word" etc. Appearance. However, if Perspective 's judgment is incorrect with the website verification tool, you can send feedback to correct the mistake.

The Jigsaw development team is aware of this problem and emphasizes that Perspective is an experimental tool for "alpha version" which is not ready for large-scale deployment. Jigsaw has already worked with Wikipedia and The New York Times and uses a training set that collected hundreds of comments and judged harmful but released an unfinished state Perspective The reason is to obtain real learning data of Perspective which is a machine learning tool.

CJ Adams, product manager for Jigsaw said, "As more perspectives are used, more and more verbatim samples continue to gather and that users around the world improve their ability to detect hate speech by correcting mistakes I am expecting it. "