Skip to Content
Version 4.1 by Liza Wensink on 2023/03/19 22:08
Hide last authors
Liza Wensink 1.1 1 **Article: **Persuasive Conversational Agent with Persuasion Tactics. [[https:~~/~~/link.springer.com/chapter/10.1007/978-3-642-13226-1_4>>https://link.springer.com/chapter/10.1007/978-3-642-13226-1_4]]
2
3 A number of studies had been done regarding the persuasiveness of conversational agents and how convincing an agent actually might be to a human person. This paper highlights that for a conversational agent to be persuasive and influence a person's behavior they need to be able to adapt to the outcomes of the conversation and the interactions it has with the human, as would a human who wants to be persuasive and convincing.
4
5 This article also contains useful information when it comes to designing a persuasive conversational agent using the Wizard of Oz method. I won't describe it in detail here, but if it is needed there is some information to be taken from the article.
6
7 When it comes to designing a persuasive conversational agent, there are several models that can be used. The general approach is to select the response and rule of replying that is most likely to lead to success.
8
9 **Goal-oriented conversational model. **
Liza Wensink 2.1 10
11 (Quotes from the article)
12 \\- The conversation model can be represented as a state transition tree where a statement is represented as a link to change a state from one to another.
Liza Wensink 1.1 13 - Two different types of states, agent states and user states (the human). 
14 - They are interleaved on a conversation path.
15 - A conversation path represents the flow of conversation between the agent and one or more users and begins with the initial state and terminates with either success or failure.
16 - If the input matches a statement on a link to an agent state, the agent chooses a statement that links the agent state to a user state with the greatest success probability.
17
18 This might not be necessarily a structure we need to implement in its entirety, but some information could definitely be taken from it.
19
Liza Wensink 2.1 20 **Updating conversation model. **
Liza Wensink 1.1 21
Liza Wensink 2.1 22 When updating the above conversation model needs to be updated, it goes according to the following:
23
24 - When input from the user does not match any statement on the stored conversation path, the conversation path is branched and the success probability scores are updated depending on persuasion success/failure. (Once again, maybe not something we will be able to implement but can try to somehow mimic the idea of).
25
Liza Wensink 3.1 26 The article does, however, continue by mentioning that the Wizard approach where the robot is simply controlled by a human in a wizard-like fashion managed to persuade 25/60 users and the conversation agent based on the model only managed 1 out of 10 users. It is necessary to remember that designing a persuasive conversational agent consists out of two important aspects - having the robot follow general human conversational rules, but also applying persuasiveness tactics. I will attempt to clarify these tactics a bit below.
Liza Wensink 2.1 27
Liza Wensink 3.1 28 **The persuasiveness example given in this study entails:**
29
30 "We first show two digital cameras to a customer A and B as shown in Table 1. Camera A has better features about the number of pixels and image stabilizer than camera B, but the price and the weight of A are more than those of B. The purpose of this persuasion is to make the user change his/her choice from the initial one to another one."
31
32 The way the persuasion was designed in this case is according to the following:
33 "Each phase has a goal to achieve such as “Ask which camera he/she prefers?” Hence the process of persuasive conversation can be represented as a sequence of phases. The sequence of phases may change depending on the responses from the user. If the user likes a camera because of the number of pixels, the agent tries to explain that the number of pixels is not important to choose a camera. If the user likes a camera because of its image stabilizer, the agent tries to explain that the image stabilizer is useless if photos are taken only in the day time."
34
35 [[image:attach:flowchart.PNG]]
36
37 From this particular case it is clear that the persuasive strategy is based on the fact that there is a set of expected things the user might bring up (like, a priori assumed aspects that the user might talk about) that the robot will attempt to explain away, or explain why the user does not need to bother about that when choosing the camera. In our case maybe we could also attempt to catch some reasons somebody might not want to go walking for example, and then try to explain away those reasons (once again, just an idea) to try to persuade the user to actually go out on the walk.
38
39
Liza Wensink 4.1 40 **Article**: Applying Psychology of Persuasion to Conversational Agents through Reinforcement Learning: an Exploratory Study. 
41 [[https:~~/~~/ceur-ws.org/Vol-2481/paper27.pdf>>https://ceur-ws.org/Vol-2481/paper27.pdf]]
Liza Wensink 3.1 42
Liza Wensink 4.1 43 This study concerns itself with agents trying to induce a healthier diet into the human they are attempting to persuade, which could be somewhat similar to what we are attempting to do.
44
45 This study mentions: 
46 "Three relevant psychosocial antecedents of behaviour change are the following: Self-Efficacy (the individual perception of being able to eat healthy), Attitude (the individual evaluation of the pros and cons) and Intention Change (the individual willingness of adhering to a healthy diet). These psychosocial dimensions cannot be directly observed and need to be measured as latent variables. To this purpose, questionnaires are used..."
47
48 What was later done during the test was...
49
50 "In a subsequent phase (i.e. message intervention), participants were randomly assigned to one of four groups, each receiving a different type of persuasive message: gain (i.e. positive behavior leads to positive outcomes), non-gain (negative behavior prevents positive outcomes), loss (negative behavior leads to negative outcomes) and non-loss (positive behavior prevents negative outcomes)." Could be something that can be considered during the persuasion stage.
51
52 All this together is maybe be a bit much for us to implement. These questionnaires are quite lengthy and complicated to design and evaluate, since these aspects need to be monitored through latent variables. While we shouldn't and can't implement this in our project currently, it might be good to include as a side point when it comes to designing the complete system.
53
54 Further this article mostly descends into how to translate this different aspects and variables into a Bayesian network and then training the agents using RL, which is not relevant for this course even if it is interesting. Once again, could maybe be mentioned as a side note.
55
56
Liza Wensink 1.1 57