Sentiment (RoBERTa)

High-Level Intuition

This feature measures the positivity of a message using RoBERTa’s generated valence sentiment markers.

Citation

Twitter-roBERTa-base-sentiment model from the Hugging Face Transformers library.

Implementation Basics

The code runs inference on the the Twitter-roBERTa-base-sentiment model to predict how relatively positive, negative, and neutral a message is on a 0-1 scale.

Implementation Notes/Caveats

This feature precomputes these valence ratings in the data preprocessing step and stores them locally; this essentially “caches” the sentiment markers, preventing the case where a user spends extra time regenerating these ratings on subsequent requests.

Interpreting the Feature

This feature returns 3 general sentiment markers: positive_bert, negative_bert, and neutral_bert. Each score ranges from 0-1, and all three scores add up to 1. This feature measures the extent to which a particular utterance aligns with each label, relative to the other labels.

Below is an example output file:

Output File

message	positive_bert	negative_bert	neutral_bert
The idea sounds great!	0.97	0.01	0.02
I disagree, this idea is terrible.	0.02	0.92	0.06
Who’s idea was it?	0.05	0.35	0.60

Related Features

This feature is one of several that measure sentiment. Other sentiment-related features include Negative_Emotion and Positive_Emotion, which are attributes of Politeness/Receptiveness Markers; Textblob Polarity; and LIWC (a relevant column name being positive_affect_lexical_per_100). A z-scored version of the positivity RoBERTa score is also included separately (Positivity Z-Score).