Overview

mpathic builds AI models for detecting and improving or correcting human behaviors in conversations. Our unique, continuously-evolving models are built using real-world data labeled by mental health and communication experts. Our API exposes these models via REST API.

Key Concepts

The key concepts revolve around utterances, behaviors, and conversations.

Utterance

An utterance is a complete thought. An utterance can span a part of a sentence or multiple parts of a sentence. Generally an utterance is not more than one sentence. The behaviors and conversation requests both take an input, which is analyzed based on utterance.

For example, take the following request to the behaviors endpoint:

{
  "input": "I am unhappy! I'm leaving!",
  "confidence_threshold": 0.5,
  "max_behaviors": 1
}

The number of utterances is two. "I am unhappy!" and "I'm leaving!"

Behaviors

Behaviors contain the behavior name, confidence level, and any tips, definitions, or suggestions (if requested).

Zero to the max number of behaviors requested are returned. Using the above example, one behavior would be returned for "I am unhappy" and one for "I am leaving".

Behaviors are returned based on the confidence threshold provided. If you provide a confidence threshold of 0.5 you will only get behaviors equal to or above that threshold.

Types of Behaviors

Name	Definition
appreciation	Providing positive feedback on another person’s strengths, abilities, or other attributes, such as providing appreciation and encouragement.
autonomy	Acknowledging another person's freedom of choice.
check for understanding	Periodically checking for understanding throughout the course of a conversation.
clarification question	Seeking clarity in a conversation to reduce ambiguity.
collaborative question	Asking questions that increase problem-solving or solicit another person's thoughts.
concern	Expressing worry or concern about another person's actions.
conflict	Language that is harmful to others.
confront	Negative, expert-like responses including disapproval, disagreement, and shame
discovery question	Asking targeted questions meant to solicit details or specific information.
engagement	Showing interest in a conversation.
giving information	Providing information on or education on a topic or personal disclosure.
high curiosity	Asking questions that encourage more information sharing.
instruct	Offering suggestions or recommendations for change
instruct direction	Giving orders, commands, or instruction.
low curiosity	Asking questions that only allow for a limited range of information.
preparation	Providing clear transitions and structure in a conversation.
reflective listening	Making efforts to understand another's points of view.
sympathy	Expressing reassurance and support.

See how behaviors positively and negatively impact metrics.

Conversation

A conversation is split up by talk turns. Each talk turn has an input (which can be zero to many utterances), a talk turn id, and an associated speaker.

For example, take the following request to the conversations endpoint:

{
    "conversation_id": "ExampleConversation",
    "talk_turns": [
        {
            "input": "Are you done with the report yet?",
            "talk_turn_id": "1",
            "speaker": {
                "id": "Mr. Smith",
                "role": "Manager"
            }
        },
        {
            "input": "No, I quit!",
            "talk_turn_id": "2",
            "speaker": {
                "id": "John Doe",
                "role": "Employee"
            }
        }
    ]
}

On the response you will get speaker, role, and conversation analytics along with metrics. As long as there is more than one speaker you will also get synchrony scores.

Analytics

As mentioned, there are three types of analytics returned:

speaker
role
conversation

The speaker and role analytics will split the data by id, in the case of speaker, or role in the case of role. The conversation analytics returns data on the entire conversation. See the examples for how the APIs behave.

Metrics

Three types of metrics (empathy, collaboration, curiosity) are provided on the conversations response per piece of analytical data (speaker/role/conversation) per id or role (Speaker: John Doe Role: Employee).

empathy: Accurate understanding. Empathy is a measure of efforts to understand another person’s point of view.

Positively Impacts	Negatively Impacts
appreciation reflective listening high curiosity	low curiosity instruct confront

collaboration: Sharing power. Actively inviting another person’s thoughts and supporting their autonomy accompanied with less advice giving.

Positively Impacts	Negatively Impacts
reflective listening	instruct

curiosity: Open versus directive and driving the agenda. Encouraging others to share more information.

Positively Impacts	Negatively Impacts
high curiosity	low curiosity

Synchrony scores

A synchrony score is an objective indicator of empathy representing the degree to which two people unconsciously align and synchronize in language style. Synchrony is the only metric not provided per speaker as it is calculated between two individuals.