Dataset Information

Wide range screening of algorithmic bias in word embedding models using large sentiment lexicons reveals underreported bias types.

ABSTRACT: Concerns about gender bias in word embedding models have captured substantial attention in the algorithmic bias research literature. Other bias types however have received lesser amounts of scrutiny. This work describes a large-scale analysis of sentiment associations in popular word embedding models along the lines of gender and ethnicity but also along the less frequently studied dimensions of socioeconomic status, age, physical appearance, sexual orientation, religious sentiment and political leanings. Consistent with previous scholarly literature, this work has found systemic bias against given names popular among African-Americans in most embedding models examined. Gender bias in embedding models however appears to be multifaceted and often reversed in polarity to what has been regularly reported. Interestingly, using the common operationalization of the term bias in the fairness literature, novel types of so far unreported bias types in word embedding models have also been identified. Specifically, the popular embedding models analyzed here display negative biases against middle and working-class socioeconomic status, male children, senior citizens, plain physical appearance and intellectual phenomena such as Islamic religious faith, non-religiosity and conservative political orientation. Reasons for the paradoxical underreporting of these bias types in the relevant literature are probably manifold but widely held blind spots when searching for algorithmic bias and a lack of widespread technical jargon to unambiguously describe a variety of algorithmic associations could conceivably be playing a role. The causal origins for the multiplicity of loaded associations attached to distinct demographic groups within embedding models are often unclear but the heterogeneity of said associations and their potential multifactorial roots raises doubts about the validity of grouping them all under the umbrella term bias. Richer and more fine-grained terminology as well as a more comprehensive exploration of the bias landscape could help the fairness epistemic community to characterize and neutralize algorithmic discrimination more efficiently.

SUBMITTER: Rozado D

PROVIDER: S-EPMC7173861 | biostudies-literature | 2020

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Wide range screening of algorithmic bias in word embedding models using large sentiment lexicons reveals underreported bias types.

Rozado David D

PloS one 20200421 4

Concerns about gender bias in word embedding models have captured substantial attention in the algorithmic bias research literature. Other bias types however have received lesser amounts of scrutiny. This work describes a large-scale analysis of sentiment associations in popular word embedding models along the lines of gender and ethnicity but also along the less frequently studied dimensions of socioeconomic status, age, physical appearance, sexual orientation, religious sentiment and political ...[more]

PMID: 32315320

Similar Datasets

Project description:BackgroundLanguage in nonmedical data sets is known to transmit human-like biases when used in natural language processing (NLP) algorithms that can reinforce disparities. It is unclear if NLP algorithms of medical notes could lead to similar transmissions of biases.Research questionCan we identify implicit bias in clinical notes, and are biases stable across time and geography?Study design and methodsTo determine whether different racial and ethnic descriptors are similar contextually to stigmatizing language in ICU notes and whether these relationships are stable across time and geography, we identified notes on critically ill adults admitted to the University of California, San Francisco (UCSF), from 2012 through 2022 and to Beth Israel Deaconess Hospital (BIDMC) from 2001 through 2012. Because word meaning is derived largely from context, we trained unsupervised word-embedding algorithms to measure the similarity (cosine similarity) quantitatively of the context between a racial or ethnic descriptor (eg, African-American) and a stigmatizing target word (eg, nonco-operative) or group of words (violence, passivity, noncompliance, nonadherence).ResultsIn UCSF notes, Black descriptors were less likely to be similar contextually to violent words compared with White descriptors. Contrastingly, in BIDMC notes, Black descriptors were more likely to be similar contextually to violent words compared with White descriptors. The UCSF data set also showed that Black descriptors were more similar contextually to passivity and noncompliance words compared with Latinx descriptors.InterpretationImplicit bias is identifiable in ICU notes. Racial and ethnic group descriptors carry different contextual relationships to stigmatizing words, depending on when and where notes were written. Because NLP models seem able to transmit implicit bias from training data, use of NLP algorithms in clinical prediction could reinforce disparities. Active debiasing strategies may be necessary to achieve algorithmic fairness when using language models in clinical research.

Project description:BackgroundBy the end of 2022, more than 100 million people were infected with COVID-19 in the United States, and the cumulative death rate in rural areas (383.5/100,000) was much higher than in urban areas (280.1/100,000). As the pandemic spread, people used social media platforms to express their opinions and concerns about COVID-19-related topics.ObjectiveThis study aimed to (1) identify the primary COVID-19-related topics in the contiguous United States communicated over Twitter and (2) compare the sentiments urban and rural users expressed about these topics.MethodsWe collected tweets containing geolocation data from May 2020 to January 2022 in the contiguous United States. We relied on the tweets' geolocations to determine if their authors were in an urban or rural setting. We trained multiple word2vec models with several corpora of tweets based on geospatial and timing information. Using a word2vec model built on all tweets, we identified hashtags relevant to COVID-19 and performed hashtag clustering to obtain related topics. We then ran an inference analysis for urban and rural sentiments with respect to the topics based on the similarity between topic hashtags and opinion adjectives in the corresponding urban and rural word2vec models. Finally, we analyzed the temporal trend in sentiments using monthly word2vec models.ResultsWe created a corpus of 407 million tweets, 350 million (86%) of which were posted by users in urban areas, while 18 million (4.4%) were posted by users in rural areas. There were 2666 hashtags related to COVID-19, which clustered into 20 topics. Rural users expressed stronger negative sentiments than urban users about COVID-19 prevention strategies and vaccination (P<.001). Moreover, there was a clear political divide in the perception of politicians by urban and rural users; these users communicated stronger negative sentiments about Republican and Democratic politicians, respectively (P<.001). Regarding misinformation and conspiracy theories, urban users exhibited stronger negative sentiments about the "covidiots" and "China virus" topics, while rural users exhibited stronger negative sentiments about the "Dr. Fauci" and "plandemic" topics. Finally, we observed that urban users' sentiments about the economy appeared to transition from negative to positive in late 2021, which was in line with the US economic recovery.ConclusionsThis study demonstrates there is a statistically significant difference in the sentiments of urban and rural Twitter users regarding a wide range of COVID-19-related topics. This suggests that social media can be relied upon to monitor public sentiment during pandemics in disparate types of regions. This may assist in the geographically targeted deployment of epidemic prevention and management efforts.

Project description:BackgroundSuicide is a serious public health issue, accounting for 1.4% of all deaths worldwide. Current risk assessment tools are reported as performing little better than chance in predicting suicide. New methods for studying dynamic features in electronic health records (EHRs) are being increasingly explored. One avenue of research involves using sentiment analysis to examine clinicians' subjective judgments when reporting on patients. Several recent studies have used general-purpose sentiment analysis tools to automatically identify negative and positive words within EHRs to test correlations between sentiment extracted from the texts and specific medical outcomes (eg, risk of suicide or in-hospital mortality). However, little attention has been paid to analyzing the specific words identified by general-purpose sentiment lexicons when applied to EHR corpora.ObjectiveThis study aims to quantitatively and qualitatively evaluate the coverage of six general-purpose sentiment lexicons against a corpus of EHR texts to ascertain the extent to which such lexical resources are fit for use in suicide risk assessment.MethodsThe data for this study were a corpus of 198,451 EHR texts made up of two subcorpora drawn from a 1:4 case-control study comparing clinical notes written over the period leading up to a suicide attempt (cases, n=2913) with those not preceding such an attempt (controls, n=14,727). We calculated word frequency distributions within each subcorpus to identify representative keywords for both the case and control subcorpora. We quantified the relative coverage of the 6 lexicons with respect to this list of representative keywords in terms of weighted precision, recall, and F score.ResultsThe six lexicons achieved reasonable precision (0.53-0.68) but very low recall (0.04-0.36). Many of the most representative keywords in the suicide-related (case) subcorpus were not identified by any of the lexicons. The sentiment-bearing status of these keywords for this use case is thus doubtful.ConclusionsOur findings indicate that these 6 sentiment lexicons are not optimal for use in suicide risk assessment. We propose a set of guidelines for the creation of more suitable lexical resources for distinguishing suicide-related from non-suicide-related EHR texts.

Dataset Information

Wide range screening of algorithmic bias in word embedding models using large sentiment lexicons reveals underreported bias types.

Publications

Wide range screening of algorithmic bias in word embedding models using large sentiment lexicons reveals underreported bias types.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets