Linguistic Relativity Theory: Sapir-Whorf Hypothesis Explained

This section presents the hypothesis that language structure influences cognition, distinguishing weak linguistic relativity (language influences thought) from strong linguistic determinism (language determines thought). Examples include Native American languages lacking abstract concepts requiring detailed description, and Northern peoples having many color terms for snow differentiation. Early experiments by Brown and Lenneberg found weak correlations between naming ease and recognition accuracy, leading to methodological refinements emphasizing communicative context. This theoretical framework sets up subsequent research on how language shapes cognitive organization and perception across cultures.

Linguistic relativity, also known as the Sapir-Whorf Hypothesis, proposes that the language or languages one speaks shapes the way one thinks and perceives the world; this hypothesis exists in two versions—a strong version suggesting language determines thought, and a weak version suggesting language merely influences thought—with research demonstrating that languages differ significantly in how they encode spatial relationships (such as prepositions in English versus verb-framed constructions in Spanish), categorization systems (like classifiers in Navajo and Guugu Yimithirr), lexicalization of concepts (such as snow terms in Inuit), color terminology (with Berlin and Kay's implicational hierarchy), and directional systems (geocentric versus geographic), all of which constrain speakers' attention to and conceptualization of different aspects of reality.

Contemporary cognitive science distinguishes between strong and weak versions of linguistic relativity. The strong version claims that language determines thought so completely that speakers of different languages think fundamentally differently—a position largely rejected due to overwhelming evidence of universal cognitive foundations. The weak version accepts that while humans share a common foundation of thought, small variations exist at the margins, sometimes having significant effects at both perceptual and conceptual levels. This nuanced view acknowledges that language can influence thinking without determining it absolutely, representing the current consensus in cognitive science.

Dani speakers with only two color terms could learn English color terms, refuting strong linguistic relativity. Research in many domains has refuted the strong hypothesis, but new research supports the weaker version. The debate was sidelined after Chomsky's innateness hypothesis but revived after the cognitive revolution. Cognitive linguistics and bilingualism research have brought new insights into the language-cognition interface.

Multilingualism functions like binocular vision, expanding cognitive capabilities by integrating multiple linguistic worldviews; research demonstrates that bilingual individuals develop enhanced perceptual discrimination (such as distinguishing more shades of blue), reduced susceptibility to gender stereotypes, flexible temporal reasoning, improved economic decision-making, and significantly better protection against cognitive decline in old age, with multilinguals experiencing approximately five additional years of healthy brain function compared to monolinguals.

A French psychoanalyst said that having two languages is to have two souls. The word gives us our worldview, and the more languages you speak, the more diverse your worldview becomes. The more you can see through the programming and the veil of illusion created by a particular culture. People who have access to different cultures and languages often carry deep wisdom because they can blend different sets of coding. This suggests that multilingualism provides cognitive flexibility and broader perspectives on human experience.

This segment presents research findings on the cognitive and social benefits of bilingualism for children. Managing two languages requires the brain to constantly inhibit one language while using the other, improving attentional control and executive function. Research shows bilingual individuals make faster decisions between options and demonstrate greater cognitive flexibility. Exposure to multiple grammatical systems helps children understand that there are multiple valid ways to encode reality, fostering tolerance for diversity. Code-switching, far from being random, follows specific grammatical rules, demonstrating that bilinguals maintain separate but interconnected language systems. The professor argues that multilingualism represents the human norm rather than an exception, contrasting this with the modern emphasis on 'one nation, one language' that emerged in the 19th century with nation-state formation.

Research from the University of Edinburgh shows that being bilingual can enhance cognitive abilities, including improved general intelligence, better reading comprehension, and enhanced decision-making skills, while also potentially slowing brain aging and delaying dementia; these benefits apply regardless of when a second language is learned, even after age 18, though bilingual individuals may experience slightly slower word retrieval due to the brain activating both languages simultaneously.

Conceptual metaphor theory, developed by linguists George Lakoff and Mark Johnson, explains that abstract concepts are understood through concrete, sensory domains; for example, emotional states are conceptualized as temperatures (hot/cold), knowledge is understood as vision (seeing/blindness), and relationships are framed as journeys (crossroads/travel), demonstrating that our entire conceptual system is fundamentally metaphorical rather than purely literal.

Conceptual metaphors are pervasive frameworks that structure how humans think about abstract concepts. George Lakoff and Mark Johnson's 1980 work established that metaphor is not merely rhetorical but fundamental to ordinary thought, particularly for abstract concepts. Abstract domains like love relationships are structured through multiple metaphorical mappings simultaneously—relationships can be understood as journeys, natural forces, buildings, or containers. Primary metaphors, rooted in recurring embodied correlations (big is important, up is good, warmth is affection), form the foundation for all conceptual metaphors. These primary metaphors explain why abstract concepts can be understood through embodied experiences, as they emerge from the most salient correlations in human bodily experience.

George Lakoff and Mark Johnson argue that the essence of metaphor is to understand and live one type of thing in terms of another type of thing. Metaphor is not decorative but part of conceptual systems that highlight some aspects and hide others. Conceptual metaphors can be described at the neural level as mappings between conceptual domains where structures of the source domain are reused to reason about the target domain. These mappings are typically partial and learned automatically without consciousness. In 'Where Mathematics Comes From', they argue that many mathematical ideas use conceptual metaphors to make available inferential structures, and concepts like actual infinity would be fundamentally metaphorical constructions distinct from potential infinity understood as an endless process.

George Lakoff and Mark Johnson challenge the traditional view of metaphor as merely poetic or rhetorical. They argue that metaphor impregnates not just language but also thought and action. Our ordinary conceptual system, through which we think and act, is fundamentally metaphorical in nature. Metaphors are fruits of our knowledge about immediate reality, particularly for abstract concepts like peace or intelligence that resist direct expression. They produce knowledge by providing forms through which we learn about realities before us. Metaphors also connect cognition to action, as we do not merely contemplate the world but engage in dialogue with it, acting upon it and with those around us.

The Conceptual Metaphor Theory, proposed by George Lakoff and Johnson, argues that metaphor is not merely a literary device but a fundamental cognitive mechanism that structures human thought and understanding. According to this theory, we comprehend abstract concepts (target domains) by mapping them onto concrete experiences (source domains), such as understanding 'time' through the domain of 'money' or 'discussion' through the domain of 'war'. This conceptual mapping occurs unconsciously in everyday language and thought, making metaphor a ubiquitous and necessary tool for human cognition rather than an ornamental linguistic feature.

Language models can perform Chain-of-Thought reasoning in languages that occupy very small fractions of their training data (e.g., Swahili, Bengali, Telugu, Thai). Despite these languages representing less than 0.01% of pre-training data, models perform surprisingly well on math word problems in these languages. This suggests that reasoning abilities may be somewhat language-independent, with models deriving reasoning structure separately from language-specific patterns.

The overwhelming majority of language-based AI tools focus on English and perform much better for languages like English and more globally dominant languages because smaller languages have far less available training data. For GPT-3, 93% of its training data was in English. LLMs combine languages with each other, mixing them into the same model so they all learn from each other. This means the model might begin to make certain patterns or vocabulary in Polish, for example, look more like the patterns of grammar and vocabulary in English because of the imbalance of training data. Additionally, models will reflect any ideological biases implicit in their training data, such as those related to race or gender.

Lang is virtual - it's everywhere but not in the pieces of the expression. AI training data is 500 million paroles (individual utterances), and by looking at this carefully, humans can infer the structure of meaning (lang). AI gets to the thing that linguists are trying to study - the virtual totality of language - by training on massive amounts of parole.

LLMs trained primarily on English data (99% of Llama's training data) perform better in English than in other languages. Even when asked questions in French, translating to English, generating the response in English, then translating back often yields better results. This bias occurs because most training data and evaluation benchmarks are in English.