Your spouse needs professional help: Determining the Contextual Appropriateness of Messages through Modeling Social Relationships

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10994–11013 July 9-14, 2023 ©2023 Association for Computational Linguistics Your spouse needs professional help: Determining the Contextual Appropriateness of Messages through Modeling Social Relationships David Jurgens3 University of Michigan jurgens@umich.edu Agrima Seth3 University of Michigan agrima@umich.edu Jackson Sargent⋆ University of Michigan jacsarge@umich.edu Athena Aghighi⋆ University of California, Davis aaghighi@ucdavis.edu Michael Geraci⋆ University of Buffalo megeraci@buffalo.edu Abstract Understanding interpersonal communication requires, in part, understanding the social con- text and norms in which a message is said. However, current methods for identifying of- fensive content in such communication largely operate independent of context, with only a few approaches considering community norms or prior conversation as context. Here, we introduce a new approach to identifying in- appropriate communication by explicitly mod- eling the social relationship between the in- dividuals. We introduce a new dataset of contextually-situated judgments of appropri- ateness and show that large language mod- els can readily incorporate relationship infor- mation to accurately identify appropriateness in a given context. Using data from online conversations and movie dialogues, we pro- vide insight into how the relationships them- selves function as implicit norms and quan- tify the degree to which context-sensitivity is needed in different conversation settings. Further, we also demonstrate that contextual- appropriateness judgments are predictive of other social factors expressed in language such as condescension and politeness. 1 Introduction Interpersonal communication relies on shared ex- pectations of the norms of communication (Hymes et al., 1972). Some of these norms are widely shared across social contexts, e.g., racial epithets are taboo, enabling NLP models to readily iden- tify certain forms of offensive language (Fortuna and Nunes, 2018). Yet, not all norms are widely shared; the same message said in two different social contexts may have different levels of accept- ability (Figure 1). While NLP has recognized the role of social context as important (Hovy and Yang, 2021; Sheth et al., 2022), few works have directly 3These authors contributed equally to this work ⋆These authors contributed equally to this work “You look beautiful today” “You look like you need some help” “When can I see you again?” grandpa to grandchild police officer to person teacher to a student student to a teacher romantic partner coworker Figure 1: The same message can be appropriate or not depending on the social context in which it is said. incorporated this context into modeling whether messages violate social norms. Here, we explicitly model relationships as the social context in which a message is said in order to assess whether the message is appropriate. NLP models have grown more sophisticated in modeling the social norms needed to identify offen- sive content. Prior work has shown the benefits of modeling context (Menini et al., 2021), such as the demographics of annotators and readers (Sap et al., 2019; Akhtar et al., 2021) and the online commu- nity in which a message is said (Chandrasekharan et al., 2018; Park et al., 2021). However, these …

Content truncated to 500 words