![Screen Shot 2022-07-19 at 12.53_edited.jpg](https://static.wixstatic.com/media/c9ce98_0587daa6872f478392045ed0cc6ae49e~mv2.jpg/v1/fill/w_356,h_384,al_c,q_80,usm_0.66_1.00_0.01,enc_auto/Screen%20Shot%202022-07-19%20at%2012_53_edited.jpg)
Cyrano
Improving Couples’ Communication Through Empathy
About
Cyrano is a conversational AI that coaches couples to respond more empathetically to one another in text exchanges. It does so by suggesting a more empathetic response than the one a user is contemplating sending.
The Problem and the Opportunity
It's often said that communication is one of the keys to a good relationship. However, many partners don't know when they might be saying something that might be harsher or less empathetic than they might want it to be. That's particularly true for text exchanges, which lack the context of face-to-face conversations.
​
In a Turkish study of divorce that asked couples more than 50 questions, divorced vs not divorced couples had the most divergent answers to statements related to communication. For example, most non-divorced couples strongly disagreed with the statement “I can insult in our discussions”, while divorced couples tended to agree with it to some degree. That divergence is shown in a chart from that study* below.
![Screen Shot 2022-07-19 at 11.58.57 PM.png](https://static.wixstatic.com/media/c9ce98_bea0ae427cea45f0bd8cf870eead6e7a~mv2.png/v1/fill/w_452,h_285,al_c,q_85,usm_0.66_1.00_0.01,enc_auto/Screen%20Shot%202022-07-19%20at%2011_58_57%20PM.png)
From the study, being quick to anger, insulting, or dismissive are some of the most common traits in divorced couples. This inspired our idea – to help couples communicate more empathetically using machine learning. With millions of couples and billions of interactions, the potential positive impact on peoples' lives is immense.
Demo
See Cyrano in action in an interactive demo. We show four examples: first and second being examples of Cyrano working perfectly as intended. Third and fourth examples still produce coherent sentences but are flawed and demonstrate current limitations of MVP. More testing and iterations will greatly improve performance.
Example Use Case
Currently, our model is a standalone tool. However, we anticipate this will eventually be integrated into a mobile phone's keyboard, as in the following mock up.
![Screen Shot 2022-07-20 at 2.00.02 PM.png](https://static.wixstatic.com/media/c9ce98_8ed16e7ebaca4a9b9ff3832fa7ca9e8d~mv2.png/v1/fill/w_541,h_560,al_c,lg_1,q_85,enc_auto/Screen%20Shot%202022-07-20%20at%202_00_02%20PM.png)
The workflow, in use, is very simple:
-
A user with Cyrano installed types a response to their partner.
-
Cyrano scans the recent text history as well as the user's draft response.
-
If Cyrano scores the draft response too low, it will suggest an alternative that the user can either accept or reject.
Output Requirements
There are three considerations that Cyrano attempts to balance when suggesting an alternative response:
​
-
​Fluency: The words need to ‘seem’ like natural written text
-
Contextual Relevance: The response needs to be relevant to both:
-
The context used by the conversation partner
-
The user’s originally intended response
-
-
Empathy: Account for the emotional ‘context’ of the partner and respond appropriately
​
One of the challenges we face is balancing these factors; we attempt to improve empathy without sacrificing fluency and contextual relevance.
How Cyrano Works:
1. Empathy Evaluation
There has been relatively little research into empathy evaluation and very few empathy datasets. One of the most cited, and the one used for Cyrano, was the result of a paper titled Learning Word Ratings for Empathy and Distress from Document-Level User Responses by Sedoc, Buechel, et al.
​
The dataset contains roughly 10,000 unigrams with empathy scores between 1 (least empathetic) and 7 (most empathetic). This lexicon was created using empathy ratings for comments under news articles that were annotated by psychologists and trained on a feed-forward neural network.
![](https://static.wixstatic.com/media/c9ce98_c37839eca28648699d87a2459de8232c~mv2.jpg/v1/fill/w_270,h_236,al_c,q_80,usm_0.66_1.00_0.01,enc_auto/c9ce98_c37839eca28648699d87a2459de8232c~mv2.jpg)
Sentences are scored based on the mean empathy score of their words. Words not in the lexicon are given the mean score of 3.5, which we consider to be neutral.
2. Alternative Text Generator
Cyrano's alternative text generator/chatbot is adapted from an article titled How to build a State-of-the-Art Conversational AI with Transfer Learning, written by the co-founder of Huggingface. The original model would return a response from GPT that was relative to one of 10,000 Facebook user volunteers' chatlog entries. We replaced those chatlog entries with a partner's initial text to the user. (We refer to the user of Cyrano as the user and their partner as...their partner.)
​
Our altered version simply uses GPT1 to predict the next tokens based on the prior conversation history and the text generated by the user up until this point. A diagram of how proposed alternative responses is below.
![cyrano workflow.png](https://static.wixstatic.com/media/c9ce98_f085cfc3bcc8435bada84d14225b16ea~mv2.png/v1/fill/w_732,h_578,al_c,q_90,usm_0.66_1.00_0.01,enc_auto/cyrano%20workflow.png)
Get in Touch!
Request an interactive demo, get product updates, more technical details, or just say hi!