Mimicry (BERT)

High-Level Intuition

This feature measure how much the current utterance “mimics” the previous utterance in a conversation.

Citation

N/A; This is a novel measure that builds on related measures of mimicry/accommodation (see “Related Features”).

Implementation Basics

Using BERT’s Sentence Transfomers model (https://sbert.net/), utterances are represented as multidimensional embeddings. Stepping through each message in a conversation, this feature computes the cosine similarity of the current embedding and previous embedding to determine their degree of mimicry.

Implementation Notes/Caveats

Note that the first utterance in a conversation cannot have a mimicry score, as there is no “previous utterance” to associate it with. In this case, we assign a value of 0 to this utterance.

Interpreting the Feature

This feature generates a score between 0-1 for each utterance in a conversation, with scores closer to 0 representing a more original thought compared with the previous chat (lacking mimicry), while scores near 1 represent a higher degree of mimicry/similarity with the previous chat.

It’s important to note that this score doesn’t measure the overall mimicry of the conversation. As an utterance-level feature, it computes the mimicry only between the selected chat and the previous. If a particular message is only similar to chats exchanged before it’s direct previous chat, therefore, it won’t have a high mimicry score (see below). In the same vein, high mimicry score for an individual chat does not signal that a conversation overall employed high mimicry.

Output File

message	speaker	mimicry_bert
Hi, my name is Shruti!	Speaker A	0
Hey, my name is Nathaniel, but I go by Nate.	Speaker B	0.89
What’s the plan for today?	Speaker A	0.12
My name is Emily.	Speaker C	0.09

Related Features

This toolkit incorporates a host of mimicry-related features, with others including Function Word Accommodation, Content Word Accommodation, and Moving Mimicry. The former two features use a bag-of-words approach to compute mimicry within two discrete categories. Mimicry (BERT) is similar in that it still uses sBERT embeddings to compute similarity, but differs in that it helps reason towards the mimicry discretely between a single utterance and the previous utterance, rather than the overall flow of mimicry throughout a conversation up until a certain point.