Researchers at the University of Warsaw Developed a Deep Learning (DL) Dialogue System Called Serena for Mental Health Therapy

One of the largest problems in the world is still that mental health therapy is not widely accessible. According to estimates, 658 million individuals worldwide experience psychological anguish, which has increased by 50% over the past 30 years. However, less than 25% of those suffering from mental health illnesses have ever “seen someone,” and only 35% obtain mental health care. Psychological counseling and therapy are beneficial in treating many problems, including anxiety, depression, obsessive-compulsive disorder, personality disorders, eating disorders, and many more. Though 56% of those going through a mental health crisis said, they handled their issues alone, over 48% of them found that talking to friends was beneficial.

A straightforward yet persistent question may have a solution thanks to deep generative learning (DL) models: how might they make mental health therapy more accessible? According to their hypothesis, a virtual mental health counselor built on generative deep learning models might improve many user profiles’ mental health outcomes. In this article, they will outline the development of a deep-learning conversation system for psychiatric therapy. They must first explore why most individuals cannot or do not want to get mental health therapy to address the issue properly. The most obvious factor is the cost of frequent, in-person counseling, which is the most helpful. Time is a comparable barrier.

People with enough money to pay for high-quality therapy might need more time to devote to the process, which calls for scheduling, commuting, arranging for child care, etc., in addition to the actual sessions. They also fear counseling because of perceived stigma. To address as many of these aspects as possible, they created a DL-based discussion system called Serena, focusing on bridging the gaps left by conventional, in-person therapy. The suggested method is meant to take something other than traditional treatment.

Instead, they see it as 1) a backup plan for those who are unable to participate in traditional therapy due to cost or time constraints; 2) a motivator for getting people comfortable with the idea of talking about their feelings through dialogue, which may lead to them scheduling in-person sessions; 3) a device for determining treatment requirements and monitoring adherence to a virtual counseling model across a broad population, to raise the standard and accessibility of mental health resources worldwide. The model may be used on their website and was deployed using Google Kubernetes Engine (GKE).

Their solution relies on the abstractions provided by the ParlAI platform2 to put up an interactive dialogue model. Their website uses FastAPI3 to retrieve replies from the model via REST API. The model, which runs on a single Nvidia T4 GPU, has to be containerized for deployment using GKE. After interacting with the model for a while, users may complete a survey included in their deployment. Users are asked to score how well the model comprehends their communications and if they find the produced answers interesting and useful. Behaviors Their conversation model demonstrates a clear comprehension of the user’s requests and is capable of replying in a manner that appears sympathetic (an example in figure 1). By posing pertinent questions, the model engages the user in conversation and stimulates further reflection.

Serena’s frequent hallucinations of user knowledge, a well-known issue with transformer-based generative models, are one of its key drawbacks. She may, for instance, assert that she has already met the user or appear to be knowledgeable about their personal history. They are working to address this problem by adding words that suggest these hallucinations to the previously specified exclusion list. Hallucinations are thought to result from data noise, such as information in the output that isn’t present in the input, and they want to investigate a potential remedy to this.

Serena frequently uses questions to respond to the user’s requests, another problem. While this is ideal for including the user in the dialogue, early feedback from test participants suggests that this behavior is seen as unpleasant and maybe even disrespectful. They now use hardcoded procedures to select from candidate answers, which reduces the number of questions that may be created to a minimum. However, this strategy is prone to failure since the candidate list frequently omits answers that are not questions. By carefully balancing the number of questions and statements in the data used for fine-tuning the generative model, they intend to address this problem. The project is live and available for users with various pricing tiers. The good thing is there is a free tier as well.

Check out the Paper and Tool. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 13k+ ML SubReddit, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.