You Sound Like an Evil Young Man: A Distributional Semantic Analysis of Systematic Form-meaning Associations for Polarity, Gender, and Age in Fictional Characters' Names

We detail a successful attempt in modeling associations about the age, gender, and polarity of fictional characters based on their names alone. We started by collecting ratings through an online survey on a sample of annotated names from young adult, children, and fan fiction stories. We collected ratings over three semantic differentials (gender: male - female; age: old - young; polarity: evil - good) using a slider bar. First, we show that participants tend to agree with authors: names judged to better suit female/young/evil characters tend to be assigned to female/young/evil characters in the original stories. We then show that, in a series of computational studies, we can predict participants' ratings on the three attributes using a distributional semantic model which derives representations for both lexical and sub-lexical patterns. This attempt was successful for all names, including made-up ones, and using both a supervised and an unsupervised approach. The prediction supported by distributed representations is much better than that afforded by symbolic features such as letters and phonological features, also when accounting for the complexity of the feature spaces. Our results show that people interpret both known and novel names relying on lexical and sub-lexical patterns, which suggests the availability of systematic form-meaning mappings in everyday language use. This further lends credit to the hypothesis that language internal statistics can support systematic form-meaning associations which apply to both known and novel lexical items.