From Text to Thought: How Analyzing Language Can Advance Psychological Science. Joshua Conrad Jackson et al. Perspectives on Psychological Science, October 4, 2021. https://doi.org/10.1177/17456916211004899
Abstract: Humans have been using language for millennia but have only just begun to scratch the surface of what natural language can reveal about the mind. Here we propose that language offers a unique window into psychology. After briefly summarizing the legacy of language analyses in psychological science, we show how methodological advances have made these analyses more feasible and insightful than ever before. In particular, we describe how two forms of language analysis—natural-language processing and comparative linguistics—are contributing to how we understand topics as diverse as emotion, creativity, and religion and overcoming obstacles related to statistical power and culturally diverse samples. We summarize resources for learning both of these methods and highlight the best way to combine language analysis with more traditional psychological paradigms. Applying language analysis to large-scale and cross-cultural datasets promises to provide major breakthroughs in psychological science.
Keywords: natural-language processing, comparative linguistics, historical linguistics, psycholinguistics, cultural evolution, emotion, religion, creativity
Psychological science still has work to do before researchers can master NLP and comparative linguistic methods. We dedicate the rest of this article to illustrating how that might happen. First, we present Figure 4, which is a visual flowchart illustrating how the language-analysis methods discussed in this article can be employed to address psychological questions. We then summarize three case studies that demonstrate how NLP and comparative linguistics can yield new insights and increase the scale and diversity of study into three psychological constructs that have been notoriously difficult to study—emotion, religion, and creativity. In these sections, we highlight research that has used language analysis to address new questions or solve long-standing debates or that has used language-analysis methods to increase the scale or cultural diversity of research in these fields. This work illustrates the utility of language analysis for asking enduring psychological questions and foreshadows the potential of these tools to address psychological constructs across social, cultural, cognitive, clinical, and developmental psychology.
Fig. 4. A flowchart of different language-analysis methods and the kinds of questions they are best suited to answer. Orange boxes represent methods from comparative linguistics, and gray boxes represent methods from NLP. Black boxes approximate the questions that may guide researchers toward these methods. Concepts are defined here as the meaning associated with words. This is meant as a general guide for researchers interested in language analysis, and there is some overlap in classifications. For example, word embeddings can show how language conveys moods and attitudes, and colexification can sometimes uncover evolutionary dynamics.
Emotion
Questions and debates about the nature of human emotion have existed since the earliest days of psychological science (Darwin, 1872/1998; James, 1884; Spencer, 1894; Wundt, 1897) and are relevant to psychological questions pertinent to social, clinical, and developmental psychology. Language-analysis methods have already increased the scope of this long-standing field and generated original methods of addressing old debates.
One of the most enduring debates about emotions concern whether emotions are universal, inborn categories that possess little variation around the world or are socially learned categories that vary in their experience and conceptualization across cultures (Cowen & Keltner, 2020; Ekman & Friesen, 1971; Izard, 2013; Plutchik, 1991; Lindquist et al., 2012; Mesquita et al., 2016; Russell, 2003). We recently addressed this question by means of a comparative-linguistics approach using colexifications (Jackson, Watts, et al., 2019). This analysis allowed us to increase the scale and generalizability over previous field studies of cross-cultural differences in emotion that had relied on smaller sample sizes and two-culture comparisons (Bryant & Barrett, 2008; Ekman & Friesen, 1971; Gendron et al., 2014, 2015, 2020).
In our study, we computationally aggregated thousands of word-lists and translation dictionaries into a large database named “CLICS” (https://clics.clld.org/), and we used this database to examine colexification patterns of 24 emotion concepts across 2,474 languages. We constructed networks of colexification in which nodes represented concepts (e.g., “anger”) and edges represented colexifications (instances in which people had named two concepts with the same word), and then compared emotion colexification networks across language families. In contrast to Youn and colleagues (2016), who found universal colexification patterns involving concepts such as “sun” and “sky,” we found wide cultural variation in the colexification of emotion concepts such as “love” and “fear.” In fact, clusters of emotion colexification varied more than three times as much as the clustering patterns of colors—our set of control concepts—across language families (see Fig. 5). For example, “anxiety” was perceived as similar to “fear” among Tai-Kadai languages, but was more related to “grief” in Austroasiatic languages, suggesting that speakers of these language may conceptualize anxiety differently.
Fig. 5. The colexification structure of emotion concepts for all languages (top left) and for five individual language families in Jackson and colleagues (2019) analysis of emotion. Nodes are emotion concepts, and links between concepts represent the likelihood that these concepts will be colexified in a language. Color indicates semantic community, which refers to clusters of emotions that are similar in meaning. From Jackson, J. C., Watts, J., Henry, T. R., List, J. M., Forkel, R., Mucha, P. J., Greenhill, S., Gray, R. D., & Lindquist, K. A. (2019). Emotion semantics show both cultural variation and universal structure. Science, 366(6472), 1517–1522. https://doi.org/10.1126/science.aaw8160. Reprinted with permission from AAAS.
The variability in emotion meaning that we observed was associated with the geographic proximity of language families, suggesting that the meaning of emotion may be transmitted through historical patterns of contact (e.g., warfare, trade) and common ancestry. We also found that emotions universally clustered together on the basis of their hedonic valence (whether or not they were pleasant to experience) and to a lesser extent, by their physiological activation (whether or not they involved high levels of physiological arousal), suggesting valence and physiological activation might be biologically based factors that provide “minimal” universality to the meaning of emotion. In sum, this study used an unprecedented sample of cultures to yield new insights into the structure and cultural variation of human emotion.
A different set of language-analysis studies involving NLP are improving how psychologists measure emotion and track it over time and across social networks. For example, in a study of unprecedented historical scale, Morin and Acerbi (2017) used sentiment analysis to examine English fiction from 1800 to 2000 to assess whether the expression of emotion had changed systematically over time. They found a decrease in positive (but not negative) emotions conveyed in language over history in three separate corpora of text. This change could not be explained by changing writer demographics (e.g., age and gender), vocabulary size, or genre (fiction vs. nonfiction), raising the possibility that something about emotion or its expression has itself changed over time.
Other studies have also used language analysis to track faster emotional dynamics, such as measuring the emotional qualities of social-media posts (Roberts et al., 2012; Yu & Wang, 2015) and testing whether the emotions of one person are likely to rapidly spread via language throughout that person’s social network. Such studies have shown experimentally that emotional sentiment conveyed by language on social-media websites (e.g., Facebook) is more likely to make individuals who view that language express similar emotions (Kramer et al., 2014). Correlational studies find that social-media information with high emotional content is more likely to be shared than information with low emotional content (Brady et al., 2017). These studies show how affect can spread across many social-media users in a short period of time.
Religion
The science of religion has a rich legacy equal to that of the psychology of emotion; many psychological studies have addressed questions about the social value and historical development of religion. Language analysis has recently begun answering both kinds of questions with a scope and ecological validity that was not possible with traditional methods.
NLP analyses have shed light on the positive and negative ways that religion affects happiness and intergroup relations. Some social theorists view religion as a primarily positive force because it reinforces social connections and promotes well-being (Brooks, 2007). On the other hand, “New Atheism” suggests that religion has a more negative effect on psychology by narrowing people’s worldviews and homogenizing the beliefs of religious adherents (Dawkins & Ward, 2006; Hitchens, 2008). Evidence for this debate has been mixed because of methodological challenges. For example, religious people frequently report more well-being than atheists in large national surveys, but they also show more social-desirability bias (Gervais & Norenzayan, 2012), which makes their self-reports less reliable.
NLP analyses are able to overcome these social-desirability limitations and have begun to show ecologically valid evidence that religion is linked to well-being. For example, Ritter et al. (2014) conducted a sentiment analysis of 16,000 users on Twitter and found that Christians expressed more positive emotion, less negative emotion, and more social connectedness than nonreligious users. Wallace et al. (2019) conducted a creative analysis of obituaries, finding that people whose obituaries mentioned religion had lived significantly longer than people whose obituaries did not mention religion, even controlling for demographic information.
Other NLP research has called the New Atheist proposition of religious worldview homogeneity into question. For example, Watts and colleagues (2020) analyzed the explanations that Christian and nonreligious participants generated to explain a wide range of supernatural and natural phenomena and estimated the overlap of these explanations as a measure of worldview homogeneity. If religion does indeed homogenize adherents’ worldviews, one would expect that religious people’s explanations would share greater overlap than nonreligious people’s explanations. Watts and colleagues (2020) used a text analysis approach known as Jaccard distances, which was able to estimate the similarity between participants’ explanations of the world using overlapping key words, and test whether religious people offered more homogeneous explanations than did nonreligious people. Using this algorithm, the researchers found that religious people’s explanations of supernatural phenomena were more homogeneous than nonreligious people’s explanations, but their explanations of natural phenomena (e.g., the prevalence of parasites) were more diverse than were nonreligious explanations, probably because they drew on supernatural as well as scientific concepts when explaining the natural world.
Comparative linguistics has mostly contributed to questions about how religion has developed over time across cultures. Many of these analyses have focused on the “supernatural monitoring hypothesis”: that watchful and punitive gods contributed to the evolution of social groups by increasing in-group prosociality and fostering large-scale cooperation (Johnson, 2016; Norenzayan et al., 2016). This idea is nearly a century old, arguably dating back to Durkheim (1912/2008), but most tests of the hypothesis have been correlational, and there is an ongoing debate about whether societies with large-scale cooperation tend to adopt moralistic religions or societies that adopt moralistic religions tend to be more cooperative (Whitehouse et al., 2019).
Researchers using comparative-linguistics methods recently addressed these debates by focusing on the development of religion in the Pacific Islands, where linguistic analyses have mapped out cultural phylogenies that can then be repurposed for cross-cultural research (R. D. Gray et al., 2009). Using these phylogenetic trees and implementing a method known as Pagel’s discrete (Pagel, 1999), Watts and colleagues (2015) inferred the probability that ancestor cultures had high levels of political complexity (indicating large-scale cooperation), the probability that they believed in supernatural punishment, and the probability that they worshiped moralizing high gods. Their results showed partial support for both sides of the debate about religion and cooperation. Broad supernatural punishment (e.g., punishment for violating taboos) tended to precede and facilitate political complexity. However, belief in watchful and punitive high gods (e.g., the Christian God) tended to occur only when societies were already politically complex.
Phylogenetic analyses have also shed light on the darker side of religious evolution, such as ritualized human sacrifice practices, which were common across the ancient world. According to the social-control hypothesis, ritual human sacrifice was used as a tool to help build and maintain social inequalities by demonstrating the power of leaders and instilling fear among subjugates. Yet evidence in support of this theory was based largely on individual case studies showing that higher classes often orchestrated ritual sacrifices (Carrasco, 1999; Turner & Turner, 1999). Watts and colleagues (2016) tested this prediction by examining patterns of ritual human sacrifice and social inequality across 93 Pacific societies that had been mapped onto an established language phylogeny (R. D. Gray et al., 2009). They found evidence that ritual human sacrifice often preceded, facilitated, and helped to sustain social inequalities, supporting the social-control hypothesis.
Creativity
Compared with the psychology of emotion and religion, that of creativity has a shorter history in psychology. Most psychologists agree that creativity contributes to personal feelings of self-fulfillment and societal innovation (Pratt & Jeffcutt, 2009; Wright & Walton, 2003), but the field is still exploring the best ways to measure creativity as a psychological construct. More than a dozen creativity-measurement paradigms exist in psychology. One such measure asks participants to name multiple uses for common household items such as article clips and bricks (Guilford, 1950), whereas others require participants to think of creative marketing schemes (Lucas & Nordgren, 2015) or draw an alien from another planet (Ward, 1994). In each paradigm, responses are qualitatively scored on creativity by trained research assistants. Although these tasks are themselves quite creative, the coding process can be onerous, and it can take months to obtain creativity ratings for a small behavioral study. Because these measures require custom tasks and laboratory settings, they are also rarely suitable for analyzing real-world creative behavior.
Language analysis has only recently been applied to study creativity, but NLP techniques are already advancing the measurement of creativity with paradigms that can be applied to both individuals in a small study as well as millions of people around the world. One such paradigm is “forward flow” (K. Gray et al., 2019). Forward flow asks people to free associate concepts, much like classic psychoanalysis methods. But rather than qualitatively deconstructing these free associations, forward flow uses word embeddings to quantitatively analyze the extent that present thoughts diverge from past thoughts. For example, because “dog” and “cat” are frequently used together in large corpora, “dog” → “cat” would not represent as much divergence as “dog” → “fortress,” which are less frequently used together. Forward flow correlates with higher creativity scores on validated behavioral tasks such as the multiple uses task, and creative professionals such as actors, performance majors, and entrepreneurs score highly on forward flow (K. Gray et al., 2019). Forward flow in celebrities’ social-media posts can even predict their creative achievement (K. Gray et al., 2019). Forward flow may represent a rich and low-cost measure that could help capture creativity across people and societies.
Other NLP analyses have captured creativity in terms of divergences from normative language (e.g., Kuznetsova et al., 2013). Much like an unorthodox-looking alien, unorthodox patterns of language can signal creativity. However, it can be difficult to distinguish nonnormative and creative language (e.g., “metal to the pedal,” which is a reformulation of “pedal to the metal”) from nonnormative and nonsensical language (e.g., “the metal pedal to”). Berger and Packard (2018) developed a potential solution to this problem in a study of the music industry and used this method to test how creativity related to a product’s success. Their approach first used topic modeling to develop words that frequently appeared in different genres of music. For instance, words about bodies and movement were often featured in dance songs, whereas words about women and cars were often featured in country music songs. The study next quantified each song from the sample on its typicality according to how much it used language typical of its genre. Analyzing these trends found that songs that broke from tradition and featured atypical language performed better than songs featuring more typical language, offering some evidence that people prefer creative cultural products.
Recent language-analysis studies have already made a considerable impact on the study of creativity and show the potential of NLP for capturing and quantifying variability in creativity across people and products. Although no comparative-linguistics research has examined creativity, this subfield also has great potential for examining whether creativity varies in its structure across cultures and how creativity has evolved across history. Some historical analyses suggest that creativity has been highest during periods of societal looseness—periods with less rigid social norms and more openness (Jackson, Gelfand, et al., 2019). But this research was done on American culture, and it is not clear whether these findings would generalize around the world.
Posts
