Debiasing word embeddings | talking about word embedding and deviation removal ＃ yyds dry goods inventory ＃

This article is my note sharing , The content mainly comes from teacher Wu Enda's in-depth learning course . ^[1]

The existence of stereotypes

word embedding It has a very important impact on the generalization of our model , Therefore, we should also ensure that they are not affected by unexpected forms of bias . Such as sexism , Racial discrimination , Religious discrimination and so on .

Of course, I think the word hint is a little serious , Here we can understand it as stereotype .

Take a chestnut ：

My father is a doctor , My mother is _______ .

My father is a company employee , My mother is _______ .

Boys like _______ . Girls like _______ .

The first empty one, of course, is likely to be “ The nurse ”. The second empty answer is likely to be “ housewife ”. The third empty answer is likely to be “ The transformers ”. The fourth empty answer is likely to be “ Barbie doll ”.

What is this ？ This is the so-called gender stereotype . These stereotypes are related to socio-economic status .

Learning algorithms are not stereotyped , But the words written by human beings are stereotyped . and Word embedding Can “ very good ” Learn these stereotypes .

So we need to modify the learning algorithm as much as possible , Minimize or idealize , Eliminate these unexpected types of bias .

Over many decades, over many centuries,I think humanity has made progress in reducing these types of bias. And I think maybe fortunately AI, I think we actually have better ideas for quickly reducing the bias in AI than for quickly reducing the bias in the human race. Although I think we are by no means done for AI as well, and there is still a lot of research and hard work to be done to reduce these types of biases in our learning of learning algorithms.

---

Eliminate word embedding stereotypes

With the aid of arXiv:1607.06520 ^[2] Methods .

It is mainly divided into the following three steps ：

1. Identify bias direction.
2. Neutralize: For every word that is not definitional, project to get rid of bias.
3. Equalize pairs.

Suppose now we have a good student word embedding.

Or continue our previous style . It uses 300 Dimension characteristics , Then we map it to a two-dimensional plane . The distribution of these words on the plane is shown in the figure .

1. Find a way

To find out the main direction of stereotype between two words , This method we talked about earlier word embedding I mentioned the feature once . Is to subtract two vectors to get the main dimension of their difference .

$e_{doctor}-e_{nurse}$

$e_{boy}-e_{girl}$

$e_{he}-e_{she}$

$e_{grandmother}-e_{grandfather}$

After subtracting the above, you will find that their differences are mainly in gender In this dimension .

Then make a for the above Average .

We can get the following result ：

We can find out the main direction of our stereotype bias . Then you can also find a direction that is not related to a particular bias .

Be careful ： In this case , We think our bias is in the direction of “gender” It's a one-dimensional space , And the other irrelevant direction is 299 The subspace of dimension . This is simplified compared with the original paper . Specifically, you can read the references provided at the end of the article .

2. Neutralization treatment

There is this word, which is clearly gender differentiated , But some words should exist fairly without gender distinction .

Gender specific words , such as grandmother and grandfather, There is no gender distinction , such as nurse,doctor. For this kind of words, we should neutralize them , That is, reduce the horizontal distance in the direction of bias .

3. Balancing

The second step is to deal with words that are gender neutral . What's wrong with gender specific words .

We can clearly see from the above figure . about nurse The word , It is associated with girl The distance is significantly longer than boy A more recent . So if the text is generated , mention nurse, appear girl Will be more likely . So we need to balance the distance through calculation .

After calculation, translate it , It's a gender neutral word. It's . The distance between words with gender distinction is equal .

---

1.  AI Official website of China - The world's leading online AI education 、 Practice platform (deeplearningai.net) ︎

2.  [1607.06520v1] Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings (arxiv.org) ︎