ArguMessage

Objective

Individuals are increasingly recognising the importance of healthy eating and its effects on well-being. However, many find it difficult to eat healthily, leading to negative outcomes such as diabetes and obesity. Personalised messages have previously been shown to impact positive health behaviour, and so may be used to promote healthy eating habits. Researchers have investigated the personalisation of messages by adapting which of the widely used Cialdini principles of persuasion should be applied. The number of Cialdini principles is limited, and so the question arises as to whether the far more detailed and structured logical statements commonly used in everyday dialogue, i.e., argumentation schemes, could be used to provide finer-grained personalisation.

In our previous studies, we manually created and validated messages for each Cialdini principle (which was extremely time-consuming). Since argumentation schemes have a definite structure with easily modifiable variables, it may be easier to automate the process of message creation after the initial validation of message types. In addition, variables can be substituted with alternatives that can help in building a large corpus of messages that can be used by, for example, intelligent healthy eating trainer software.

Work

Our primary research objective is to automate personalised persuasive messages that will be able to sustain behaviour change. This could be achieved by incorporating Cialdini's principles of persuasion and argumentation schemes. Argumentation schemes are rules leading from assumptions to conclusions that are often found in everyday dialogues. In this work, we illustrate the system built based on the mapped argumentation schemes.

Results

The mapping of Cialdini’s principles to the argumentation schemes is summarised below. We developed a message generation system using this mapping as its foundation. Given below is an explanation of one of the argumentation schemes.

Argument from commitment with goal. This scheme states that the proposed “action” supports the “actor” in fulfilling a “goal” they committed to previously. In the domain of healthy eating, this scheme can be used to encourage users to commit to a positive healthy eating “action” backed by their previous “commitment “. The generated message is developed using a message structure created for each argumentation scheme, as demonstrated below for the “argument from commitment with goal” argumentation scheme.

To create automated messages for the argument from commitment with goal scheme, we needed to describe a specific “commitment”, “goal” and “action” for the “actor” who would be the intended subject of the message. Our aim is to crowd-source such messages, and our system, therefore — as shown below — presents a user with a sample message using the message structure and poses questions to instantiate the scheme’s variables.

In this argumentation scheme, we asked three questions:

Q1. What is the goal of the user?

A. The goal of the user is to _________. This provides the input for Goal G.

Q2. What is the user therefore committed to do?

A. The user is committed to _________. This provides the input for Commitment C.

Q3. What specific action contributes to achieve this commitment?

A. The user should ________. This provides the input for Action A.

To instantiate the variables appropriately, the user’s answers are required to be in a verb form. To achieve this, we provided the user with the first part of the answer (e.g., stating that “The goal of the user is to . . . ” for Question 1).

The remaining 13 argumentation schemes and the questions for the users along with the answer structures that we have developed are illustrated in the Appendix.

Using the system: We intend to use the system within a set of user studies. The participant is presented with the summary of the study instructions which states that they are required to generate a total of three messages with three “recipes” (argumentation schemes) by answering some questions that provide the input for generating messages. Next, they are shown the explanation of a “recipe”. This is followed by a set of questions that require a small amount of participant input to generate the message. An example of the completed participant inputs is shown above. Then, the participant presses the ’Create Message’ button, which takes them to the second step which shows the generated message. In this case, the message generated would be “As you want to improve skin texture, you are committed to consume sources rich in Vitamin C and potassium. So you’re also committed to consume fruits such as kiwis and bananas as it helps you to consume sources rich in Vitamin C and potassium”. The system uses template-based natural language generation to produce these messages. Participants provide their level of satisfaction with the message generated on a 5-point Likert scale that ranges from not satisfied to totally satisfied. In addition, they may provide detailed feedback, as input to further improve the system. When the participant presses the ’Submit’ button, they are taken to the next randomly selected recipe. The same process is repeated to generate a set of three messages per participant in total.

Reflection

We will conduct studies with lay people; argumentation scheme experts; and domain experts (e.g., dieticians etc.) to generate a corpus of messages using the developed system and investigate the extent to which the system makes it easy to produce good messages. The system is currently only used to generate individual persuasive messages. These messages could then be used by a dialogue system.

Appendix