What makes you happy? This question is a central topic in positive psychology — particularly, which behavioral modifications provide people with happiness in a sustainable fashion? To elucidate this information, we’ve developed the Happiness Entailment Recognition prototype. This module analyzes input text describing an event and helps determine repeatable activities that could improve overall well-being.
The Pursuit of Happiness Can Be Hard
Finding out what makes you happy can be surprisingly difficult. People tend to misgauge feelings in the moment. Hindsight, especially from an outside observer, is a much more reliable indicator of what fosters happiness in individuals. For this reason, self-reporting assessment tools have helped the field of psychology make tremendous strides in understanding happiness.
Building on this success, a more recently introduced approach requires subjects to document experiences (often referred to as moments) in short textual descriptions. Research has shown that recognizing “happy moments” and expressing gratitude can cultivate better subjective well-being. Besides this benefit, these concise journal entries also shine a light on the full spectrum of human emotion, from gratefulness and happiness to boredom, anxiety, and even depression.
By examining and interpreting this text, psychologists can theoretically distinguish happy moments from unhappy ones and identify patterns that sustainably contribute to a person’s happiness. Making individually-tailored suggestions based on this information has long been in the domain of mental health professionals. More people than ever before are turning to these experts. Unfortunately, there is a shortage of these specialists around the world; not everyone who would benefit has access to them.
How Do You Define Happiness?
To help address this growing issue, we developed an AI system that can supplement professional help by generating suggestions for well-being. But to begin solving this problem, we must answer one tough question: What is happiness, exactly?
Obviously, this topic is open to vast interpretation. In order to make our solution versatile, we wanted to interpret this question as generally as possible while still maintaining enough specificity to measure our success meaningfully. If given a happy moment, could our system make inferences about how likely another activity is to contribute to an individual’s happiness? And could our system model this distribution over a full spectrum of activities? This is what we call happiness entailment:
Probability(activity X making the subject happy | their recorded happy moment Y)
We hypothesize that various happy moments contain activities that can be extracted and suggested for repetition. Here are the two main criteria we used for suggestibility:
- Repeatability: The activity can be repeated with a reasonable amount of time and effort.
- Sustainability: Repeating the activity would benefit the individual’s short-term well-being without being detrimental in the long term.
Swimming or walking in the park are examples of activities that meet both of these conditions. In contrast, events such as adopting a puppy or buying a new car fail these requirements because their repetition isn’t sustainable.
Introducing the Happiness Entailment Recognition Model
With our two suggestibility rules in tow, an automated system should be able to accomplish two tasks using information from an individual’s short journal entries:
- Discover sustainable suggestions for activities.
- Identify which suggestions make sense for the individual.
For the first step, suggestion candidates can either be manually crafted or automatically filtered from a corpus such as HappyDB, our database of 100,000 crowd-sourced happy moments. Ideally, domain experts such as psychologists would help in creating these sustainable suggestions. They can also be based on knowledge of the users, such as their environment or previously recorded happy moments.
The Happiness Entailment Recognition module takes care of the second task. This work was inspired by the problem of Recognizing Textual Entailment (RTE), a field of Natural Language Understanding (NLU): If given a set of user-reported happy moments and a set of possible suggested activities, which activities would potentially make the user happy?
How the Happiness Entailment Recognition Module Works
The Happiness Entailment Recognition module is a neural network that’s implemented as a dual-encoder (DE) model, a standard for NLU tasks. Its shared encoder takes two inputs:
- A short text describing an event (a happy moment)
- A sustainable candidate suggestion
The module then outputs a prediction of whether the suggestion is likely to be good (entailment) or bad (non-entailment) based on the event description. By considering several of these determinations, we can build a clearer picture of what makes a user happy.
Successfully evaluating if a suggestion would make a user happy depends on both common sense and inference to understand the relationship between a happy moment and the suggestion in question. For example, if a user reports getting a haircut as a happy moment, “paint your nails” may serve as a suitable suggestion, as the model might potentially learn representations of these activities are similar, as they both reflect self-care activities. However, this suggestion becomes irrelevant if the user’s happy moment describes painting a portrait.
Figure: Overview of our system. The system outputs a best followup suggestion based on an input happy moment provided by the user.
To train the Happiness Entailment Recognition module, we created a dataset. In it, each sample contained a happy moment, a sustainable suggestion, and a label for the pair. For the first component, we filtered happy moments from HappyDB. From there, we created clusters based on the verbs from these sentences and crafted 36 suggestions. We then utilized Amazon Mechanical Turk (MTurk) for crowdsourced annotations.
Evaluating the Happiness Entailment Recognition Module
Under these conditions, the Happiness Entailment Recognition module was remarkably accurate at predicting the relationship between happy moments and sustainable suggestions; it achieved an AU-ROC score of 0.770.
To enhance its functionality, we also tried equipping our model with additional classifiers to determine the following psychological features:
- Agency: Was the user in control of this happy moment?
- Sociality: Was the happy moment a social one?
- Concept: Were any of the 15 most commonly occurring concepts in happy moments from HappyDB (e.g., family, career, weather) present?
With these augmentations, the Happiness Entailment Recognition module was able to predict whether a suggestion fits a happy moment with an AU-ROC of 0.831. As a baseline, we tested the performance of the state-of-the-art Textual Entailment model from AllenNLP (trained on the Stanford Natural Language Inference (SNLI) dataset). Our module outperformed this model by more than 48% in happiness entailment recognition.
This notable improvement in performance confirms that these classifiers allow the model to capture psychological perspectives that are crucial for accomplishing the task effectively.
Table: Experimental result for the HER task
Future Work in Happiness Entailment
Future work in this field includes incorporating more sophisticated user information; integrating contextual data regarding factors like location and weather; and collecting more diverse suggestions. All of these options must be implemented in ways that do not compromise user privacy.
Further research into the roles of psychological features such as agency and sociality is also needed. Similarly, it is also imperative to investigate which use-case systems like the Happiness Entailment Recognition module are most applicable. For example, it could aid in redirecting users towards positive coping mechanisms when they are experiencing anxiety or boredom. How people feel about interacting with this type of technology will ultimately determine its utility.
However, it must be noted that this system is no substitute for a mental health professional during times of emotional crisis. The Happiness Entailment Recognition module could never replace a mental health professional; rather, it may be a tool to help the user reflect on the possible activities that she can engage to make him/her happy.
For now, the results of our research are an impressive start. The Happiness Entailment Recognition module demonstrates that the field of positive psychology could profoundly benefit from NLU and automation. These technologies open up numerous opportunities to help solve the mental health expert shortage and improve the well-being of individuals around the world.
Written by Sara Eversen and Megagon Labs