This project includes data and code for the following paper: **Schmidt, H., Tran, S., Medaglia, J.D., Ulichney, V., & Helion, C. (2022). Conversational linguistic features predict social network learning. *PsyArXiv https://doi.org/10.31234/osf.io/fn4my*** Additional information can be found on GitHub at: https://github.com/hschmidt12/NLP-Social-Network-Learning. ---------- ## Abstract ## Whether it is the first day of school or a new job, individuals often find themselves in situations where they must quickly and accurately learn novel social networks. However, the mechanisms through which this learning occurs remain unclear. We posit that individuals use linguistic features of conversations to identify the valence and strength of social relationships. Across three studies using a naturalistic behavioral task (57 adults; 34,735 observations), we employ novel person- and stimulus-focused approaches to investigate social network learning success, examine how distinct linguistic features predict network learning, and explore the association between semantic similarity and relationship formation. We found that participants learned similar network structures, linguistic features uniquely predicted network learning, and greater semantic similarity was associated with perceived friendship formation. These findings suggest that naturalistic conversational content is both a potential mechanism of social network learning and a promising avenue for future research on social relational inference. ---------- ## Data ## The `data` folder contains all de-identified data using in this paper. - `dat_survivor.csv` contains raw participant-level responses - `df-clean.csv` contains all raw participant-level responses merged with NLP linguistic scores (similarity, sentiment, clout) - `df-ling.csv` contains raw participant-level responses merged with NLP linguistic scores, but only contains trials that have corresponding linguistic data - `friends-dyads.csv` contains data on how often each contestant pair (dyad) was chosen as friends - `friends_rivals_dyads.csv` contains data on how often each dyad was chosen as friends and rivals - `LIWC_clout.csv` contains output from LIWC with clout scores per sentence of contestant dialogue - `rival-dyads.csv` contains data on how often each dyad was chosen as rivals - `similarity_choices.csv` contains data on participant-level responses merged with similarity scores - `Similarity-data.csv` contains reformatted similarity scores for all episode clips (created from the clip-specific similarity matrices) - `survivor-split-sentences.csv` contains all contestant dialogue from the episode with each row containing one sentence of dialogue; each row is labeled with the word count in that sentence - `survivor-text.csv` contains raw contestant dialogue from the episode with each row containing one sentence of dialogue Within the `data` folder, there is a subfolder called `clips` that contains the clip-specific semantic similarity matrices (6 .csv files). ---------- ## Scripts ## All scripts associated with this paper are included in the `scripts` folder. These scripts are formatted as R Markdowns and each script has a corresponding .html file. 1. `Universal-Sentence-Encoder-Similarity-Scores` - Calculates the semantic similarity between each contestant's dialogue per episode clip and saves a matrix of scores. 2. `Data-Cleaning-and-Setup` - Reads in raw response data and formats it for NLP analyses 3. `Study1-Network-Discernment` - Runs five analyses to assess how participants learned the social network of the episode 4. `Study2-GLM-Linguistic-Features` - Fits three separate beta generalized linear mixed effect models to assess to what extent each linguistic feature predicts relationship judgments 5. `Study3-Similarity-Over-Time` - Fits a multilevel linear regression to assess how relationship judgments at the time of evaluation predict earlier levels of semantic similarity 6. `Supplemental` - Contains supplementary analysis for Study 2. ---------- Please direct any comments or questions to Helen Schmidt at helen_schmidt@temple.edu. Please feel free to use any of these scripts, but unfortunately I am unable to provide support if you adapt them to your own data. If you notice a bug or error, please tell me!