Positivity Z-Score

High-Level Intuition

This feature measures how positive an utterance is relative to other messages. This positivity is gauged by comparing each message against all messages in a dataset or within the same conversation.

Citation

Tausczik & Pennebaker (2013)

Implementation

A positivity score is first assigned using RoBERTa; Liu et al., (2019). We then compute two types of z-scores based on this positivity score:

  1. Positivity among all messages (`positivity_zscore_chats`): - This z-score measures how a message’s positivity compares to all other messages in the entire dataset.

  2. Positivity within the same conversation (`positivity_zscore_conversations`): - This z-score measures how a message’s positivity compares to other messages within the same conversation (grouping by the unique conversational identifier).

Interpreting the Feature

The positivity z-scores can be both negative and positive, with no fixed bounds.

  • Negative Score: Indicates that the utterance is less positive compared to other messages (either in the entire dataset or within the same conversation, depending on the reference point of the z-score).

  • Zero Score: Indicates that the utterance has a typical (average) level of positivity compared to other messages.

  • Positive Score: Indicates that the utterance is more positive compared to other messages.