Pasar al contenido principal

Resultados relacionados

Resultados totales (Incluyendo duplicados): 18
Encontrada(s) 1 página(s)

Repositorio Digital de la UPF
Publicación Artículo científico (JournalArticle)

The common behavior effect in norm learning: when frequent observations override the behavior of the majority

  • Woiczyk, Thomas Karl Alfred
  • Hosseini, Rahil
  • Le Mens, Gaël
Prior research suggests that descriptive norms correspond to what most people do-the "behavior of the majority." We examine norm perception in situations where the behavior of the majority differs from the most frequently observed behavior-the "common behavior." In environments where individuals learn descriptive norms through repeated observations of a reference group, we propose that perceived norms align more closely with the common behavior than with the behavior of the majority. Consequently, individuals are more likely to follow the common behavior, even when it differs from what most people do. We argue that this 'common behavior effect' arises from a combination of two factors: the structure of the information environments in which the behavior of the majority and the common behavior differ, and imperfect source memory of the observed behaviors. We provide evidence for the basic phenomenon and test two moderators in four studies reported in the body of the article and two ancillary studies reported in the appendix. These findings are important for our understanding of social norms, because they challenge the assumption that norms simply reflect the behavior of the majority. They also cast light on phenomena such as pluralistic ignorance, majority illusions in online and offline environments or the spread of misinformation on social media. Finally, they have practical implications for how to shape norms in organizations., This work was sup-ported by the European Commission (grants 772268 and MCIN/AEI/10.13039/501100011033 to G.L.M.), the Agencia Estatal de Investigación, Spain (grants PID2019-105249GB-I00 and PID2022-137908NB-I00 to G.L.M.; PID2023-152226NB-I00 to T.K.A.W; PID2022-140026NB-I00 to R.H), the Fundación BBVA, Spain (Grant G999088Q to G.L.M.), the Barcelona School of Economics (grant CEX2019-000915-S to G.L.M.), the Instituto Nacional de Ciberseguridad (Grant INCIBE-CYBERTHREAT C115/23 to R.H.) and the ICREA Academia (grant to G.L.M). Guillem Pros Rius provided excellent research assistance with programming for Study 2.
Repositorio Digital de la UPF
Publicación Artículo científico (JournalArticle)

How politicians learn from citizens' feedback: the case of gender on Twitter

  • Schöll, Nikolas
  • Gallego Dobón, Aina
  • Le Mens, Gaël
Includes supplementary materials: online appendix; replication file., This article studies how politicians react to feedback from citizens on social media. We use a reinforcement-learning framework to model how politicians respond to citizens’ positive feedback by increasing attention to better received issues and allow feedback to vary depending on politicians’ gender. To test the model, we collect 1.5 million tweets published by Spanish MPs over 3 years, identify gender-issue tweets using a deep-learning algorithm (BERT) and measure feedback using retweets and likes. We find that citizens provide more positive feedback to female politicians for writing about gender, and that this contributes to their specialization in gender issues. The analysis of mechanisms suggests that female politicians receive more positive feedback because they are treated differently by citizens. To conclude, we discuss implications for representation, misperceptions, and polarization., The research leading to these results has received financial support from the project “Local politicians: selection and performance (LEADERS)” (CSO2016-79569-P) to Aina Gallego and from the grants AEI/FEDER UE-PSI201675353, PID2019-105249GB-I00/AEI/10.13039/501100011033, #RYC-2014-15035 to Gaël Le Mens, funded by the Spanish Ministry for the Economy, Industry and Competitiveness. It also benefited from funding from ERC Consolidator Grant #772268 and BBVA Foundation Grant G999088Q to Gaël Le Mens.
Repositorio Digital de la UPF
Publicación Artículo científico (JournalArticle)

Social media feedback and extreme opinion expression

  • Konovalova, Elizaveta
  • Le Mens, Gaël
  • Schöll, Nikolas
On popular social media platforms such as Twitter, Facebook, Instagram, or Tiktok, the quantitative feedback received by content producers is asymmetric: counts of positive reactions such as 'likes,' or 'retweets,' are easily observed but similar counts of negative reactions are not directly available. We study how this design feature of social media platforms affects the expression of extreme opinions. Using simulations of a learning model, we compare two feedback environments that differ in terms of the availability of negative reaction counts. We find that expressed opinions are generally more extreme when negative reaction counts are not available than when they are. We rely on analyses of Twitter data and several online experiments to provide empirical support for key model assumptions and test model predictions. Our findings suggest that a simple design change might limit, under certain conditions, the expression of extreme opinions on social media., G. Le Mens benefited from financial support from ERC Consolidator Grant \#772268 from the European Commission, an ICREA Academia grant, the Fundación BBVA Grant G999088Q, and the Agencia Estatal de Investigación (AEI), through grant PID2019-105249GB-I00/AEI/10.13039/501100011033. G. Le Mens also acknowledge financial support from the Severo Ochoa Programme for Centres of Excellence in R\&D (Barcelona School of Economics CEX2019-000915-S), funded by MCIN/AEI/10.13039/501100011033.
Repositorio Digital de la UPF
Publicación Artículo científico (JournalArticle)

Evaluating categories from experience: the simple averaging heuristic

  • Woiczyk, Thomas Karl Alfred
  • Le Mens, Gaël
Includes experimental data and R scripts used to create the analyses, figures, and tables of the experiments., We analyze how people form evaluative judgments about categories based on their experiences with category members. Prior research suggests that such evaluative judgments depend on some experience average, but is unclear about the specific kind of average. We hypothesized that evaluations of categories could be driven either by the ‘simple average’ of experiences with the category or by the ‘member average’ (the average of the evaluations of the category members, where the evaluation of a category member is the average of experiences with this particular member). Understanding whether evaluations of categories are driven by the ‘simple average’ or the ‘member average’ is important in settings where people obtain unbalanced numbers of observations about category members such as when people form opinions about a social group and predominantly interact with just a few members of this group. Across 9 studies (N=1,966), we consistently found that evaluative judgments about categories were better explained by the simple average than by the member average. We call the underlying cognitive strategy the ‘simple averaging heuristic.’ Collected evidence indicates that participants relied on simple averaging even in settings where normative principles required avoiding the use of this cognitive strategy, leading to systematic mistakes. Our findings contribute to several areas of social cognition such as research on redundancy biases, information aggregation, social sampling, and norm perceptions., G. Le Mens benefited from financial support from grants #AEI/FEDER UE-PSI2016-75353, PID2019-105249GB-I00/AEI/10.13039/501100011033 and RYC-2014-15035 from the Spanish Ministerio de Ciencia, Innovacion y Universidades (MCIU) and the Agencia Estatal de Investigacion (AEI), ERC Consolidator Grant #772268 from the European Commission and BBVA Fundation Grant G999088Q. T. Woiczyk benefited from financial support from grant BES-2015-073750 from the Spanish MINECO.
Repositorio Digital de la UPF
Publicación Artículo científico (JournalArticle)

Revisiting the competency trap

  • Denrell, Jerker
  • Le Mens, Gaël
We revisit the competency trap and reexamine when it occurs. We show that a bias against alternatives that improve with practice does not require that learning is myopic in the sense of lacking foresight or failing to explore. The same bias occurs even if learners engage in substantial exploration and have foresight. In fact, we demonstrate that even a rational and foresighted learner, who follow an optimal strategy for balancing exploration and exploitation, will learn to prefer alternatives with initially high payoffs that decrease with practice over alternatives, with identical expected values, that have initially low payoffs that increase with practice. Our results show that a bias against alternatives that improve with practice is due to an asymmetry in error correction rather than to myopic learning. The implication is that a wide range of selection systems, even optimally designed ones, will be biased against late-bloomers., G. Le Mens benefited from financial support from grants \#AEI/FEDER UE-PSI2016-75353 and Ramon y Cajal Fellowship (RYC-2014-15035), from the Spanish MINECO, and ERC Consolidator \#772268 from the European Commission.
Repositorio Digital de la UPF
Publicación Artículo científico (JournalArticle)

Uncovering the semantics of concepts using GPT-4

  • Le Mens, Gaël
  • Kovács, Balázs
  • Hannan, Michael T.
  • Pros Rius, Guillem
Includes supplementary materials for the online appendix., The ability of recent Large Language Models (LLMs) such as GPT-3.5 and GPT-4 to generate human-like texts suggests that social scientists could use these LLMs to construct measures of semantic similarity that match human judgment. In this article, we provide an empirical test of this intuition. We use GPT-4 to construct a measure of typicality—the similarity of a text document to a concept. We evaluate its performance against other model-based typicality measures in terms of the correlation with human typicality ratings. We conduct this comparative analysis in two domains: the typicality of books in literary genres (using an existing dataset of book descriptions) and the typicality of tweets authored by US Congress members in the Democratic and Republican parties (using a novel dataset). The typicality measure produced with GPT-4 meets or exceeds the performance of the previous state-of-the art typicality measure we introduced in a recent paper [G. Le Mens, B. Kovács, M. T. Hannan, G. Pros Rius, Sociol. Sci. 2023, 82–117 (2023)]. It accomplishes this without any training with the research data (it is zero-shot learning). This is a breakthrough because the previous state-of-the-art measure required fine-tuning an LLM on hundreds of thousands of text documents to achieve its performance., G.L.M. and G.P. received financial support from European Research Council (ERC) Consolidator Grant 772268 from the European Commission. G.L.M. also received financial support from an Catalan Institution for Research and Advanced Studies (ICREA) Academia grant, grant PID2019-105249GBI00/AEI/10.13039/501100011033 from the Spanish Ministerio de Ciencia, Innovacion y Universidades and the Agencia Estatal de Investigacion, and the Severo Ochoa Programme for Centres of Excellence in R&D (Barcelona School of Economics CEX2019-000915-S), funded by MCIN/AEI/10.13039/501100011033. B.K. was supported by Yale School of Management. M.T.H. was supported by the Stanford Graduate School of Business. We thank Ido Erev, Tim Sels, Alex Tyulyupo, and the participants in the 2023 Nagymaros Conference for insightful discussions and comments.
Repositorio Digital de la UPF
Publicación Artículo científico (JournalArticle)

An information sampling explanation for the in-group heterogeneity effect

  • Konovalova, Elizaveta
  • Le Mens, Gaël
Includes supplementary material: replication file., People often perceive their in-groups as more heterogeneous than their out-groups. We propose an information sampling explanation for this in-group heterogeneity effect. We note that people frequently obtain larger samples of information about in-groups than about out-groups. Using computer simulations, we show that this asymmetry in sample sizes implies the in-group heterogeneity effect under a wide range of assumptions about how experience affects perceived variability. This is the case even when perceived variability is the outcome of rational information processing, implying that the structure of the environment is sufficient to explain the emergence of the in-group heterogeneity effect. A key assumption of our explanation is that perceived group variability depends on the size of the sample observed about this group. We provide evidence in support for this assumption in two experiments. Our results considerably expand the scope and relevance of a prior sampling explanation proposed by Linville, Fischer, and Salovey (1989). They also complement other explanations that proposed that information about in-groups and out-groups is processed differently., Le Mens benefited from financial support from Southern Denmark University and Grants PSI2013-41909-P, #AEI/FEDER UE-PSI2016-75353, Ramon y Cajal Fellowship (RYC-2014-15035) from the Spanish MINECO, Grant IN[15]_EFG_ECO_2281 from the BBVA Foundation and ERC Consolidator #772268 from the European Commission. E. Konovalova was funded by Spanish MINECO Grant PSI2013-41909-P to G. Le Mens.
Repositorio Digital de la UPF
Publicación Artículo científico (JournalArticle)

Positioning political texts with large language models by asking and averaging

  • Le Mens, Gaël
  • Gallego, Aina
Data de publicació electrònica: 27-01-2025, We use instruction-tuned large language models (LLMs) like GPT-4, Llama 3, MiXtral, or Aya to position political texts within policy and ideological spaces. We ask an LLM where a tweet or a sentence of a political text stands on the focal dimension and take the average of the LLM responses to position political actors such as US Senators, or longer texts such as UK party manifestos or EU policy speeches given in 10 different languages. The correlations between the position estimates obtained with the best LLMs and benchmarks based on text coding by experts, crowdworkers, or roll call votes exceed.90. This approach is generally more accurate than the positions obtained with supervised classifiers trained on large amounts of research data. Using instruction-tuned LLMs to position texts in policy and ideological spaces is fast, cost-efficient, reliable, and reproducible (in the case of open LLMs) even if the texts are short and written in different languages. We conclude with cautionary notes about the need for empirical validation., This research was funded by ERC Consolidator Grant 772268 from the European Commission to G.L.M, ICREA Academia grants to A.G and G.L.M, grants PID2021-123111OB-I00 (A.G.) and PID2022-137908NB-I00 (G.L.M.) funded by MICIN/AEI/10.13039/501100011033 and by 'ERDF/UE A way of making Europe', and the Severo Ochoa Programme for Centres of Excellence in R&D (Barcelona School of Economics CEX2019-000915-S) funded by MCIN/AEI/10.13039/501100011033.
Repositorio Digital de la UPF
Publicación Artículo científico (JournalArticle)

Using machine learning to uncover the semantics of concepts: how well do typicality measures extracted from a BERT text classifier match human judgments of genre typicality?

  • Le Mens, Gaël
  • Kovács, Balázs
  • Hannan, Michael T.
  • Pros Rius, Guillem
Includes data, material, and analysis code for all analyses., Social scientists have long been interested in understanding the extent to which the typicalities of an object in concepts relate to its valuations by social actors. Answering this question has proven to be challenging because precise measurement requires a feature-based description of objects. Yet, such descriptions are frequently unavailable. In this article, we introduce a method to measure typicality based on text data. Our approach involves training a deep-learning text classifier based on the BERT language representation and defining the typicality of an object in a concept in terms of the categorization probability produced by the trained classifier. Model training allows for the construction of a feature space adapted to the categorization task and of a mapping between feature combination and typicality that gives more weight to feature dimensions that matter more for categorization. We validate the approach by comparing the BERT-based typicality measure of book descriptions in literary genres with average human typicality ratings. The obtained correlation is higher than 0.85. Comparisons with other typicality measures used in prior research show that our BERT-based measure better reflects human typicality judgments., Pros received financial support from ERC Consolidator Grant #772268 from the European Commission. G. Le Mens also received financial support from grant PID2019-105249GB-I00/AEI/10.13039/501100011033 from the Spanish Ministerio de Ciencia, Innovacion y Universidades (MCIU) and the Agencia Estatal de Investigacion (AEI) and from the BBVA Foundation Grant G999088Q.
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

Using machine learning to uncover the semantics of concepts: how well do typicality measures extracted from a BERT text classifier match human judgments of genre typicality?

  • Le Mens, Gaël
  • Kovács, Balázs
  • Hannan, Michael T.
  • Pros Rius, Guillem
Includes data, material, and analysis code for all analyses., Social scientists have long been interested in understanding the extent to which the typicalities of an object in concepts relate to its valuations by social actors. Answering this question has proven to be challenging because precise measurement requires a feature-based description of objects. Yet, such descriptions are frequently unavailable. In this article, we introduce a method to measure typicality based on text data. Our approach involves training a deep-learning text classifier based on the BERT language representation and defining the typicality of an object in a concept in terms of the categorization probability produced by the trained classifier. Model training allows for the construction of a feature space adapted to the categorization task and of a mapping between feature combination and typicality that gives more weight to feature dimensions that matter more for categorization. We validate the approach by comparing the BERT-based typicality measure of book descriptions in literary genres with average human typicality ratings. The obtained correlation is higher than 0.85. Comparisons with other typicality measures used in prior research show that our BERT-based measure better reflects human typicality judgments., Pros received financial support from ERC Consolidator Grant #772268 from the European Commission. G. Le Mens also received financial support from grant PID2019-105249GB-I00/AEI/10.13039/501100011033 from the Spanish Ministerio de Ciencia, Innovacion y Universidades (MCIU) and the Agencia Estatal de Investigacion (AEI) and from the BBVA Foundation Grant G999088Q.
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

Revisiting the competency trap

  • Denrell, Jerker
  • Le Mens, Gaël
We revisit the competency trap and reexamine when it occurs. We show that a bias against alternatives that improve with practice does not require that learning is myopic in the sense of lacking foresight or failing to explore. The same bias occurs even if learners engage in substantial exploration and have foresight. In fact, we demonstrate that even a rational and foresighted learner, who follow an optimal strategy for balancing exploration and exploitation, will learn to prefer alternatives with initially high payoffs that decrease with practice over alternatives, with identical expected values, that have initially low payoffs that increase with practice. Our results show that a bias against alternatives that improve with practice is due to an asymmetry in error correction rather than to myopic learning. The implication is that a wide range of selection systems, even optimally designed ones, will be biased against late-bloomers., G. Le Mens benefited from financial support from grants \#AEI/FEDER UE-PSI2016-75353 and Ramon y Cajal Fellowship (RYC-2014-15035), from the Spanish MINECO, and ERC Consolidator \#772268 from the European Commission.
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

Evaluating categories from experience: the simple averaging heuristic

  • Woiczyk, Thomas Karl Alfred
  • Le Mens, Gaël
Includes experimental data and R scripts used to create the analyses, figures, and tables of the experiments., We analyze how people form evaluative judgments about categories based on their experiences with category members. Prior research suggests that such evaluative judgments depend on some experience average, but is unclear about the specific kind of average. We hypothesized that evaluations of categories could be driven either by the ‘simple average’ of experiences with the category or by the ‘member average’ (the average of the evaluations of the category members, where the evaluation of a category member is the average of experiences with this particular member). Understanding whether evaluations of categories are driven by the ‘simple average’ or the ‘member average’ is important in settings where people obtain unbalanced numbers of observations about category members such as when people form opinions about a social group and predominantly interact with just a few members of this group. Across 9 studies (N=1,966), we consistently found that evaluative judgments about categories were better explained by the simple average than by the member average. We call the underlying cognitive strategy the ‘simple averaging heuristic.’ Collected evidence indicates that participants relied on simple averaging even in settings where normative principles required avoiding the use of this cognitive strategy, leading to systematic mistakes. Our findings contribute to several areas of social cognition such as research on redundancy biases, information aggregation, social sampling, and norm perceptions., G. Le Mens benefited from financial support from grants #AEI/FEDER UE-PSI2016-75353, PID2019-105249GB-I00/AEI/10.13039/501100011033 and RYC-2014-15035 from the Spanish Ministerio de Ciencia, Innovacion y Universidades (MCIU) and the Agencia Estatal de Investigacion (AEI), ERC Consolidator Grant #772268 from the European Commission and BBVA Fundation Grant G999088Q. T. Woiczyk benefited from financial support from grant BES-2015-073750 from the Spanish MINECO.
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

Uncovering the semantics of concepts using GPT-4

  • Le Mens, Gaël
  • Kovács, Balázs
  • Hannan, Michael T.
  • Pros Rius, Guillem
Includes supplementary materials for the online appendix., The ability of recent Large Language Models (LLMs) such as GPT-3.5 and GPT-4 to generate human-like texts suggests that social scientists could use these LLMs to construct measures of semantic similarity that match human judgment. In this article, we provide an empirical test of this intuition. We use GPT-4 to construct a measure of typicality—the similarity of a text document to a concept. We evaluate its performance against other model-based typicality measures in terms of the correlation with human typicality ratings. We conduct this comparative analysis in two domains: the typicality of books in literary genres (using an existing dataset of book descriptions) and the typicality of tweets authored by US Congress members in the Democratic and Republican parties (using a novel dataset). The typicality measure produced with GPT-4 meets or exceeds the performance of the previous state-of-the art typicality measure we introduced in a recent paper [G. Le Mens, B. Kovács, M. T. Hannan, G. Pros Rius, Sociol. Sci. 2023, 82–117 (2023)]. It accomplishes this without any training with the research data (it is zero-shot learning). This is a breakthrough because the previous state-of-the-art measure required fine-tuning an LLM on hundreds of thousands of text documents to achieve its performance., G.L.M. and G.P. received financial support from European Research Council (ERC) Consolidator Grant 772268 from the European Commission. G.L.M. also received financial support from an Catalan Institution for Research and Advanced Studies (ICREA) Academia grant, grant PID2019-105249GBI00/AEI/10.13039/501100011033 from the Spanish Ministerio de Ciencia, Innovacion y Universidades and the Agencia Estatal de Investigacion, and the Severo Ochoa Programme for Centres of Excellence in R&D (Barcelona School of Economics CEX2019-000915-S), funded by MCIN/AEI/10.13039/501100011033. B.K. was supported by Yale School of Management. M.T.H. was supported by the Stanford Graduate School of Business. We thank Ido Erev, Tim Sels, Alex Tyulyupo, and the participants in the 2023 Nagymaros Conference for insightful discussions and comments.
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

An information sampling explanation for the in-group heterogeneity effect

  • Konovalova, Elizaveta
  • Le Mens, Gaël
Includes supplementary material: replication file., People often perceive their in-groups as more heterogeneous than their out-groups. We propose an information sampling explanation for this in-group heterogeneity effect. We note that people frequently obtain larger samples of information about in-groups than about out-groups. Using computer simulations, we show that this asymmetry in sample sizes implies the in-group heterogeneity effect under a wide range of assumptions about how experience affects perceived variability. This is the case even when perceived variability is the outcome of rational information processing, implying that the structure of the environment is sufficient to explain the emergence of the in-group heterogeneity effect. A key assumption of our explanation is that perceived group variability depends on the size of the sample observed about this group. We provide evidence in support for this assumption in two experiments. Our results considerably expand the scope and relevance of a prior sampling explanation proposed by Linville, Fischer, and Salovey (1989). They also complement other explanations that proposed that information about in-groups and out-groups is processed differently., Le Mens benefited from financial support from Southern Denmark University and Grants PSI2013-41909-P, #AEI/FEDER UE-PSI2016-75353, Ramon y Cajal Fellowship (RYC-2014-15035) from the Spanish MINECO, Grant IN[15]_EFG_ECO_2281 from the BBVA Foundation and ERC Consolidator #772268 from the European Commission. E. Konovalova was funded by Spanish MINECO Grant PSI2013-41909-P to G. Le Mens.
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

The common behavior effect in norm learning: when frequent observations override the behavior of the majority

  • Woiczyk, Thomas Karl Alfred
  • Hosseini, Rahil
  • Le Mens, Gaël
Prior research suggests that descriptive norms correspond to what most people do-the "behavior of the majority." We examine norm perception in situations where the behavior of the majority differs from the most frequently observed behavior-the "common behavior." In environments where individuals learn descriptive norms through repeated observations of a reference group, we propose that perceived norms align more closely with the common behavior than with the behavior of the majority. Consequently, individuals are more likely to follow the common behavior, even when it differs from what most people do. We argue that this 'common behavior effect' arises from a combination of two factors: the structure of the information environments in which the behavior of the majority and the common behavior differ, and imperfect source memory of the observed behaviors. We provide evidence for the basic phenomenon and test two moderators in four studies reported in the body of the article and two ancillary studies reported in the appendix. These findings are important for our understanding of social norms, because they challenge the assumption that norms simply reflect the behavior of the majority. They also cast light on phenomena such as pluralistic ignorance, majority illusions in online and offline environments or the spread of misinformation on social media. Finally, they have practical implications for how to shape norms in organizations., This work was sup-ported by the European Commission (grants 772268 and MCIN/AEI/10.13039/501100011033 to G.L.M.), the Agencia Estatal de Investigación, Spain (grants PID2019-105249GB-I00 and PID2022-137908NB-I00 to G.L.M.; PID2023-152226NB-I00 to T.K.A.W; PID2022-140026NB-I00 to R.H), the Fundación BBVA, Spain (Grant G999088Q to G.L.M.), the Barcelona School of Economics (grant CEX2019-000915-S to G.L.M.), the Instituto Nacional de Ciberseguridad (Grant INCIBE-CYBERTHREAT C115/23 to R.H.) and the ICREA Academia (grant to G.L.M). Guillem Pros Rius provided excellent research assistance with programming for Study 2.
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

How politicians learn from citizens' feedback: the case of gender on Twitter

  • Schöll, Nikolas
  • Gallego Dobón, Aina
  • Le Mens, Gaël
Includes supplementary materials: online appendix; replication file., This article studies how politicians react to feedback from citizens on social media. We use a reinforcement-learning framework to model how politicians respond to citizens’ positive feedback by increasing attention to better received issues and allow feedback to vary depending on politicians’ gender. To test the model, we collect 1.5 million tweets published by Spanish MPs over 3 years, identify gender-issue tweets using a deep-learning algorithm (BERT) and measure feedback using retweets and likes. We find that citizens provide more positive feedback to female politicians for writing about gender, and that this contributes to their specialization in gender issues. The analysis of mechanisms suggests that female politicians receive more positive feedback because they are treated differently by citizens. To conclude, we discuss implications for representation, misperceptions, and polarization., The research leading to these results has received financial support from the project “Local politicians: selection and performance (LEADERS)” (CSO2016-79569-P) to Aina Gallego and from the grants AEI/FEDER UE-PSI201675353, PID2019-105249GB-I00/AEI/10.13039/501100011033, #RYC-2014-15035 to Gaël Le Mens, funded by the Spanish Ministry for the Economy, Industry and Competitiveness. It also benefited from funding from ERC Consolidator Grant #772268 and BBVA Foundation Grant G999088Q to Gaël Le Mens.
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

Positioning political texts with large language models by asking and averaging

  • Le Mens, Gaël
  • Gallego, Aina
Data de publicació electrònica: 27-01-2025, We use instruction-tuned large language models (LLMs) like GPT-4, Llama 3, MiXtral, or Aya to position political texts within policy and ideological spaces. We ask an LLM where a tweet or a sentence of a political text stands on the focal dimension and take the average of the LLM responses to position political actors such as US Senators, or longer texts such as UK party manifestos or EU policy speeches given in 10 different languages. The correlations between the position estimates obtained with the best LLMs and benchmarks based on text coding by experts, crowdworkers, or roll call votes exceed.90. This approach is generally more accurate than the positions obtained with supervised classifiers trained on large amounts of research data. Using instruction-tuned LLMs to position texts in policy and ideological spaces is fast, cost-efficient, reliable, and reproducible (in the case of open LLMs) even if the texts are short and written in different languages. We conclude with cautionary notes about the need for empirical validation., This research was funded by ERC Consolidator Grant 772268 from the European Commission to G.L.M, ICREA Academia grants to A.G and G.L.M, grants PID2021-123111OB-I00 (A.G.) and PID2022-137908NB-I00 (G.L.M.) funded by MICIN/AEI/10.13039/501100011033 and by 'ERDF/UE A way of making Europe', and the Severo Ochoa Programme for Centres of Excellence in R&D (Barcelona School of Economics CEX2019-000915-S) funded by MCIN/AEI/10.13039/501100011033.
Recercat. Dipósit de la Recerca de Catalunya
Publicación Artículo científico (JournalArticle)

Social media feedback and extreme opinion expression

  • Konovalova, Elizaveta
  • Le Mens, Gaël
  • Schöll, Nikolas
On popular social media platforms such as Twitter, Facebook, Instagram, or Tiktok, the quantitative feedback received by content producers is asymmetric: counts of positive reactions such as 'likes,' or 'retweets,' are easily observed but similar counts of negative reactions are not directly available. We study how this design feature of social media platforms affects the expression of extreme opinions. Using simulations of a learning model, we compare two feedback environments that differ in terms of the availability of negative reaction counts. We find that expressed opinions are generally more extreme when negative reaction counts are not available than when they are. We rely on analyses of Twitter data and several online experiments to provide empirical support for key model assumptions and test model predictions. Our findings suggest that a simple design change might limit, under certain conditions, the expression of extreme opinions on social media., G. Le Mens benefited from financial support from ERC Consolidator Grant \#772268 from the European Commission, an ICREA Academia grant, the Fundación BBVA Grant G999088Q, and the Agencia Estatal de Investigación (AEI), through grant PID2019-105249GB-I00/AEI/10.13039/501100011033. G. Le Mens also acknowledge financial support from the Severo Ochoa Programme for Centres of Excellence in R\&D (Barcelona School of Economics CEX2019-000915-S), funded by MCIN/AEI/10.13039/501100011033.