• Corpus ID: 1704893

Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings

@inproceedings{Bolukbasi2016ManIT,
  title={Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings},
  author={Tolga Bolukbasi and Kai-Wei Chang and James Y. Zou and Venkatesh Saligrama and Adam Tauman Kalai},
  booktitle={NIPS},
  year={2016}
}
The blind application of machine learning runs the risk of amplifying biases present in data. [] Key Method Geometrically, gender bias is first shown to be captured by a direction in the word embedding. Second, gender neutral words are shown to be linearly separable from gender definition words in the word embedding.

Figures and Tables from this paper

Replicating ’Man is to Computer Programmer as Woman is to Homemakers? Debiasing Word Embeddings’

This work replicates, testing, and visualizes a methodology to remove gender stereotypes from word embeddings to affirm the results reported by the original researchers and test their model’s robustness.

Fair Is Better than Sensational: Man Is to Doctor as Woman Is to Doctor

A series of clarifications are provided that should put well-known, and potentially new analogies into the right perspective, which might have yielded a distorted picture of bias in word embeddings.

Understanding the Origins of Bias in Word Embeddings

Given a word embedding trained on a corpus, this work develops a technique for understanding the origins of bias in word embeddings that identifies how perturbing the corpus will affect the bias of the resulting embedding.

What are the Biases in My Word Embedding?

An algorithm for enumerating biases in word embeddings that outputs a number of Word Embedding Association Tests (WEATs) that capture various biases present in the data, which makes it easier to identify biases against intersectional groups, which depend on combinations of sensitive features.

Bias in word embeddings

A new technique for bias detection for gendered languages is developed and used to compare bias in embeddings trained on Wikipedia and on political social media data, and it is proved that existing biases are transferred to further machine learning models.

Nurse is Closer to Woman than Surgeon? Mitigating Gender-Biased Proximities in Word Embeddings

RAN-Debias is proposed, a novel gender debiasing methodology that not only eliminates the bias present in a word vector but also alters the spatial distribution of its neighboring vectors, achieving a bias-free setting while maintaining minimal semantic offset.

Impact of Gender Debiased Word Embeddings in Language Modeling

This paper studies how an state-of-the-art recurrent neural language model behaves when trained on data, which under-represents females, using pre-trained standard and debiased word embeddings, to address gender, race and social biases.

A Source-Criticism Debiasing Method for GloVe Embeddings

A simple yet effective method for debiasing GloVe word embeddings which works by incorporating explicit information about training set bias rather than removing biased data outright, and reduces the effect size on Word Embedding Association Test (WEAT) sets without sacrificing training data or TOP-1 performance.

Word embeddings are biased. But whose bias are they reflecting?

From Curriculum Vitae parsing to web search and recommendation systems, Word2Vec and other word embedding techniques have an increasing presence in everyday interactions in human society. Biases,

Humpty Dumpty: Controlling Word Meanings via Corpus Poisoning

This work develops an explicit expression over corpus features that serves as a proxy for distance between words and establishes a causative relationship between its values and embedding distances, and shows how the attacker can generate linguistically likely corpus modifications, thus fooling defenses that attempt to filter implausible sentences from the corpus using a language model.
...

References

SHOWING 1-10 OF 47 REFERENCES

Linguistic Regularities in Continuous Space Word Representations

The vector-space word representations that are implicitly learned by the input-layer weights are found to be surprisingly good at capturing syntactic and semantic regularities in language, and that each relationship is characterized by a relation-specific vector offset.

Linguistic Regularities in Sparse and Explicit Word Representations

It is demonstrated that analogy recovery is not restricted to neural word embeddings, and that a similar amount of relational similarities can be recovered from traditional distributional word representations.

Deep Recursive Neural Networks for Compositionality in Language

The results show that deep RNNs outperform associated shallow counterparts that employ the same number of parameters and outperforms previous baselines on the sentiment analysis task, including a multiplicative RNN variant as well as the recently introduced paragraph vectors.

Distributed Representations of Words and Phrases and their Compositionality

This paper presents a simple method for finding phrases in text, and shows that learning good vector representations for millions of phrases is possible and describes a simple alternative to the hierarchical softmax called negative sampling.

Improving Document Ranking with Dual Word Embeddings

This paper investigates the popular neural word embedding method Word2vec as a source of evidence in document ranking and proposes the proposed Dual Embedding Space Model (DESM), which provides evidence that a document is about a query term.

Automated Experiments on Ad Privacy Settings

AdFisher, an automated tool that explores how user behaviors, Google's ads, and Ad Settings interact, finds that the Ad Settings was opaque about some features of a user’s profile, that it does provide some choice on advertisements, and that these choices can lead to seemingly discriminatory ads.

Certifying and Removing Disparate Impact

This work links disparate impact to a measure of classification accuracy that while known, has received relatively little attention and proposes a test for disparate impact based on how well the protected class can be predicted from the other attributes.

Semi-supervised Question Retrieval with Gated Convolutions

This paper designs a recurrent and convolutional model (gated convolution) to effectively map questions to their semantic representations and demonstrates that the model yields substantial gains over a standard IR baseline and various neural network architectures (including CNNs, LSTMs and GRUs).

It's a Man's Wikipedia? Assessing Gender Inequality in an Online Encyclopedia

This paper presents and applies a computational method for assessing gender bias on Wikipedia along multiple dimensions and finds that while women on Wikipedia are covered and featured well in many Wikipedia language editions, the way women are portrayed starkly differs from the way men are portrayed.

Efficient Estimation of Word Representations in Vector Space

Two novel model architectures for computing continuous vector representations of words from very large data sets are proposed and it is shown that these vectors provide state-of-the-art performance on the authors' test set for measuring syntactic and semantic word similarities.