When building natural language processing (NLP) systems, we frequently require an objective baseline for human emotion, sentiment, or semantic meaning. To achieve this, researchers traditionally turn to foundational lexical resources: dictionaries (like WordNet) and emotion lexicons (like the NRC Emotion Lexicon). Because these resources are curated by experts or derived from structured linguistic relationships, they are often treated as "gold standards" or neutral anchors capable of cleaning up noisy, biased web data.
However, a deeper look reveals a fundamental paradox: lexical resources are human artifacts. They inherit the demographic biases of their annotators, the historical prejudices embedded in definitions, and structural imbalances in how language is expanded. Instead of serving as neutral baselines, they frequently encode and amplify systemic biases.
The Paradox of Dictionary-Based Debiasing
If a word embedding model learns a biased association from web text—such as linking specific genders or races to distinct career paths—a intuitive mitigation strategy is to anchor the model to a dictionary. Since dictionary definitions aim to define words objectively, forcing word vectors to align with definitions should theoretically strip away extraneous social stereotypes.
Yet, research shows this assumption fails in practice. Kaneko et al. and An et al. both explored using dictionary definitions as neutral reference points to debias word embeddings. In Dictionary-based Debiasing of Pre-trained Word Embeddings, Kaneko et al. demonstrated that this process is strictly limited by the inherent cultural and gender biases present within lexicographic definitions themselves. If the dictionary definition contains a stereotypical association, the debiasing process inadvertently reinforces it.
In fact, Lauscher et al. discovered in Are We Consistently Biased? Multidimensional Analysis of Biases in Distributional Word Vectors that utilizing formal dictionary definitions to refine word vectors can paradoxically increase measurable bias within the resulting embeddings.
Why does this happen? Dictionaries reflect the historical and societal context of the language they document. This mechanism is made clear by Bartl et al. in Inferring Gender: A Scalable Methodology for Gender Detection with Online Lexical Databases. When attempting to infer lexical gender from definitions, neutral terms are frequently misclassified because the definitions reflect historical gender distributions (e.g., associating certain professions or roles with a specific gender). As a result, the "objective" resource acts as a repository for historical skew.
Demographic and Structural Skews in Sentiment Lexicons
A parallel issue compromises sentiment and emotion lexicons, which map words to emotional valences. These resources are typically built through two methods: human crowdsourcing or automated expansion from a small set of seed words. Both pathways introduce systemic bias.
Crowdsourcing and Annotator Demographics
Human-curated lexicons reflect the demographic makeup of their creators. In Best Practices in the Creation and Use of Emotion Lexicons, Mohammad argues that bias is an intrinsic characteristic of emotion lexicons. Because emotion perception is deeply tied to socio-cultural identity, a failure to document annotator demographics and actively incorporate diverse viewpoints means the lexicon will project a narrow, non-universal emotional framework.
When these demographic perspectives are unbalanced, the resulting resource can perpetuate harmful stereotypes. For instance, Zad et al. identified significant societal biases within word-emotion assignments in Hell Hath No Fury? Correcting Bias in the NRC Emotion Lexicon, showing that foundational resources require active, semi-automatic correction to prevent unfair downstream outcomes.
Even semi-automated generation methods fail to escape demographic alignment. Guerini et al. demonstrated in Sentiment Analysis: How to Derive Prior Polarities from SentiWordNet that automatically generated sentiment resources can align significantly closer with male human ratings than female ones. This indicates that the underlying seed data or structural assumptions favored a specific demographic's emotional perception.
The Geometry of Thesaurus Expansion
When lexicons are expanded computationally using structural relationships like thesaurus links, they run into a different, mathematical form of bias. In Generating High-Coverage Semantic Orientation Lexicons From Overtly Marked Words and a Thesaurus, Mohammad et al. showed that expanding sentiment seeds through a thesaurus inherently yields a much larger volume of positive entries than negative ones. This occurs because human language is naturally skewed toward positive expression (the Pollyanna hypothesis), meaning the graph structure of a thesaurus is fundamentally asymmetric.
Downstream Impact: Sentiment, Regard, and Framing
When models inherit these biased lexical foundations, the distortions propagate directly into how they evaluate real-world text, affecting fairness in downstream applications like code-mixed emotion recognition (Niță et al.).
To diagnose these biases, researchers have had to shift how they measure model outputs. As Udagawa et al. argue in Bias Analysis and Mitigation through Protected Attribute Detection and Regard Classification, simply counting how often a protected demographic attribute is mentioned is misleading. Instead, we must measure regard—the sentiment valence directed toward that attribute.
This distinction between frequency and valence is exactly how ideological framing operates in text. In a multi-dimensional analysis of online social networks, Coletto et al. (2016, 2017) showed that the choice of terminology when describing the same group of people (e.g., choosing "refugees" to evoke humanitarian empathy versus "migrants" to minimize it) serves as a strong proxy for underlying sentiment and political framing.
[Linguistic Choice] ---> [Altered Sentiment/Regard] ---> [Biased Framing]
e.g., "Refugees" Positive/Empathetic Humanitarian
e.g., "Migrants" Negative/Neutral Security Threat
This lexicon-driven framing is highly effective at encoding bias. Bourahouat et al. leveraged this mechanism in The Guidelines Specialists at FIGNEWS 2024 Shared Task to map how media narratives use specific word choices to either demonize groups or evoke sympathy. Similarly, Ferrer et al. demonstrated in Discovering and Categorising Language Biases in Reddit that analyzing the sentiment polarity of community-specific terms can automatically expose localized stereotypes.
Ultimately, when large language models are trained on these uncorrected text distributions, they internalize these exact dynamics. By training models on specific internet sub-communities, Haller et al. created OpinionGPT (OpinionGPT: Modelling Explicit Biases in Instruction-Tuned LLMs) to make these latent perspectives explicit. Their evaluation confirmed that different demographic and political training pipelines directly alter the proportion of negative regard a model holds toward specific groups, demonstrating that the systemic biases originally embedded in our basic lexicons continue to shape the behavior of modern AI.
Key Foundational Work
Sentiment Analysis: How to Derive Prior Polarities from SentiWordNet (2013) — contributed 3 concept elements to this direction.
Generating High-Coverage Semantic Orientation Lexicons From Overtly Marked Words and a Thesaurus (2009) — contributed 1 concept elements to this direction.