Detecting age-related linguistic patterns in dialogue Toward adaptive conversational systems

This work explores an important dimension of variation in the language used by dialogue participants: their age. While previous work showed differences at various linguistic levels between age groups when experimenting with written discourse data (e.g., blog posts), previous work on dialogue has largely been limited to acoustic information related to voice and prosody. Detecting fine-grained linguistic properties of human dialogues is of crucial importance for developing AI-based conversational systems which are able to adapt to their human interlocutors. We therefore investigate whether, and to what extent, current text-based NLP models can detect such linguistic differences, and what the features driving their predictions are. We show that models achieve a fairly good performance on age-group prediction, though the task appears to be more challenging compared to discourse. Through in-depth analysis of the best models' errors and the most predictive cues, we show that, in dialogue, differences among age groups mostly concern stylistic and lexical choices. We believe these findings can inform future work on developing controlled generation models for adaptive conversational systems.