“Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks”の学習報告

We survey 146 papers analyzing "bias" in NLP systems, finding that their motivations are often vague, inconsistent, and lacking in normative reasoning, despite the fact that analyzing "bias" is an inherently normative process. We further find that these papers' proposed quantitative techniques for measuring or mitigating "bias" are poorly matched to their motivations and do not engage with the relevant literature outside of NLP. Based on these findings, we describe the beginnings of a path forward by proposing three recommendations that should guide work analyzing "bias" in NLP systems. These recommendations rest on a greater recognition of the relationships between language and social hierarchies, encouraging researchers and practitioners to articulate their conceptualizations of "bias"---i.e., what kinds of system behaviors are harmful, in what ways, to whom, and why, as well as the normative reasoning underlying these statements---and to center work around the lived experiences of members of communities affected by NLP systems, while interrogating and reimagining the power relations between technologists and such communities.

Language (Technology) is Power: A Critical Survey of "Bias" in NLP

Исследование влияния пола и психологических характеристик автора на количественные параметры его текста с использованием программы Linguistic Inquiry and Word Count

We survey 146 papers analyzing "bias" in NLP systems, finding that their motivations are often vague, inconsistent, and lacking in normative reasoning, despite the fact that analyzing "bias" is an inherently normative process We further find that these papers' proposed quantitative techniques for measuring or mitigating "bias" are poorly matched to their motivations and do not engage with the relevant literature outside of NLP Based on these findings, we describe the beginnings of a path forward by proposing three recommendations that should guide work analyzing "bias" in NLP systems These recommendations rest on a greater recognition of the relationships between language and social hierarchies, encouraging researchers and practitioners to articulate their conceptualizations of "bias"---ie, what kinds of system behaviors are harmful, in what ways, to whom, and why, as well as the normative reasoning underlying these statements---and to center work around the lived experiences of members of communities affected by NLP systems, while interrogating and reimagining the power relations between technologists and such communities

/pdf/language-technology-is-power-a-critical-survey-of-bias-in-40xzysfmb0.pdf

Gender stereotypes are manifest in most of the world’s languages and are consequently propagated or amplified by NLP systems. Although research has focused on mitigating gender stereotypes in English, the approaches that are commonly employed produce ungrammatical sentences in morphologically rich languages. We present a novel approach for converting between masculine-inflected and feminine-inflected sentences in such languages. For Spanish and Hebrew, our approach achieves F1 scores of 82% and 73% at the level of tags and accuracies of 90% and 87% at the level of forms. By evaluating our approach using four different languages, we show that, on average, it reduces gender stereotyping by a factor of 2.5 without any sacrifice to grammaticality.

Counterfactual Data Augmentation for Mitigating Gender Stereotypes in Languages with Rich Morphology

Many modern Artificial Intelligence (AI) systems make use of data embeddings, particularly in the domain of Natural Language Processing (NLP). These embeddings are learnt from data that has been gathered “from the wild” and have been found to contain unwanted biases. In this paper we make three contributions towards measuring, understanding and removing this problem. We present a rigorous way to measure some of these biases, based on the use of word lists created for social psychology applications; we observe how gender bias in occupations reflects actual gender bias in the same occupations in the real world; and finally we demonstrate how a simple projection can significantly reduce the effects of embedding bias. All this is part of an ongoing effort to understand how trust can be built into AI systems.

/pdf/biased-embeddings-from-wild-data-measuring-understanding-and-19kp1qicnx.pdf

Biased Embeddings from Wild Data: Measuring, Understanding and Removing

In Western societies, the stereotype prevails that pink is for girls and blue is for boys. A third possible gendered colour is red. While liked by women, it represents power, stereotypically a masculine characteristic. Empirical studies confirmed such gendered connotations when testing colour-emotion associations or colour preferences in males and females. Furthermore, empirical studies demonstrated that pink is a positive colour, blue is mainly a positive colour, and red is both a positive and a negative colour. Here, we assessed if the same valence and gender connotations appear in widely available written texts (Wikipedia and newswire articles). Using a word embedding method (GloVe), we extracted gender and valence biases for blue, pink, and red, as well as for the remaining basic colour terms from a large English-language corpus containing six billion words. We found and confirmed that pink was biased towards femininity and positivity, and blue was biased towards positivity. We found no strong gender bias for blue, and no strong gender or valence biases for red. For the remaining colour terms, we only found that green, white, and brown were positively biased. Our finding on pink shows that writers of widely available English texts use this colour term to convey femininity. This gendered communication reinforces the notion that results from research studies find their analogue in real word phenomena. Other findings were either consistent or inconsistent with results from research studies. We argue that widely available written texts have biases on their own, because they have been filtered according to context, time, and what is appropriate to be reported.

/pdf/english-colour-terms-carry-gender-and-valence-biases-a-13xjsqh0jg.pdf

English colour terms carry gender and valence biases: A corpus study using word embeddings.

Many modern Artificial Intelligence (AI) systems make use of data embeddings, particularly in the domain of Natural Language Processing (NLP). These embeddings are learnt from data that has been gathered "from the wild" and have been found to contain unwanted biases. In this paper we make three contributions towards measuring, understanding and removing this problem. We present a rigorous way to measure some of these biases, based on the use of word lists created for social psychology applications; we observe how gender bias in occupations reflects actual gender bias in the same occupations in the real world; and finally we demonstrate how a simple projection can significantly reduce the effects of embedding bias. All this is part of an ongoing effort to understand how trust can be built into AI systems.

Word Embeddings are used widely in multiple Natural Language Processing (NLP) applications. They are coordinates associated with each word in a dictionary, inferred from statistical properties of these words in a large corpus. In this paper we introduce the notion of "concept" as a list of words that have shared semantic content. We use this notion to analyse the learnability of certain concepts, defined as the capability of a classifier to recognise unseen members of a concept after training on a random subset of it. We first use this method to measure the learnability of concepts on pretrained word embeddings. We then develop a statistical analysis of concept learnability, based on hypothesis testing and ROC curves, in order to compare the relative merits of various embedding algorithms using a fixed corpora and hyper parameters. We find that all embedding methods capture the semantic content of those word lists, but fastText performs better than the others.

On the Learnability of Concepts: With Applications to Comparing Word Embedding Algorithms

Word Embeddings are used widely in multiple Natural Language Processing (NLP) applications They are coordinates associated with each word in a dictionary, inferred from statistical properties of these words in a large corpus In this paper we introduce the notion of “concept” as a list of words that have shared semantic content We use this notion to analyse the learnability of certain concepts, defined as the capability of a classifier to recognise unseen members of a concept after training on a random subset of it We first use this method to measure the learnability of concepts on pretrained word embeddings We then develop a statistical analysis of concept learnability, based on hypothesis testing and ROC curves, in order to compare the relative merits of various embedding algorithms using a fixed corpora and hyper parameters We find that all embedding methods capture the semantic content of those word lists, but fastText performs better than the others

/pdf/on-the-learnability-of-concepts-23ap94wynm.pdf

Adam Sutton

Papers

Biased Embeddings from Wild Data: Measuring, Understanding and Removing

English colour terms carry gender and valence biases: A corpus study using word embeddings.

Biased Embeddings from Wild Data: Measuring, Understanding and Removing

On the Learnability of Concepts: With Applications to Comparing Word Embedding Algorithms

On the Learnability of Concepts