Zero-shot Persuasive Chatbots
with LLM-Generated Strategies and Information Retrieval

Kazuaki Furumai¹, Roberto Legaspi¹, Julio Vizcarra¹,
Yudai Yamazaki¹, Yasutaka Nishimura¹, Sina J. Semnani²,
Kazushi Ikeda¹, Weiyan Shi², Monica S. Lam²
¹Human-Centered AI Laboratories, KDDI Research, Inc., Saitama, Japan
²Computer Science Department, Stanford University, Stanford, CA, USA
{ka-furumai, xre-roberuto, xju-vizcarra, yd-yamazaki, yu-nishimura, kz-ikeda}@kddi.com,
{sinaj, weiyans, lam}@stanford.edu

Abstract

Persuasion plays a pivotal role in a wide range of applications from health intervention to the promotion of social good. Persuasive chatbots can accelerate the positive effects of persuasion in such applications. Existing methods rely on fine-tuning persuasive chatbots with task-specific training data which is costly, if not infeasible, to collect. To address this issue, we propose a method to leverage the generalizability and inherent persuasive abilities of large language models (LLMs) in creating effective and truthful persuasive chatbot for any given domain in a zero-shot manner. Unlike previous studies which used pre-defined persuasion strategies, our method first uses an LLM to generate responses, then extracts the strategies used on the fly, and replaces any unsubstantiated claims in the response with retrieved facts supporting the strategies.

We applied our chatbot, PersuaBot, to three significantly different domains needing persuasion skills: donation solicitation, recommendations, and health intervention. Our experiments on simulated and human conversations show that our zero-shot approach is more persuasive than prior work, while achieving factual accuracy surpassing state-of-the-art knowledge-oriented chatbots. Our study demonstrated that when persuasive chatbots are employed responsibly for social good, it is an enabler of positive individual and social change.

Kazuaki Furumai¹, Roberto Legaspi¹, Julio Vizcarra¹, Yudai Yamazaki¹, Yasutaka Nishimura¹, Sina J. Semnani², Kazushi Ikeda¹, Weiyan Shi², Monica S. Lam² ¹Human-Centered AI Laboratories, KDDI Research, Inc., Saitama, Japan ²Computer Science Department, Stanford University, Stanford, CA, USA {ka-furumai, xre-roberuto, xju-vizcarra, yd-yamazaki, yu-nishimura, kz-ikeda}@kddi.com, {sinaj, weiyans, lam}@stanford.edu

1 Introduction

Persuasive technologies are designed to reinforce, change or mold user’s attitude or behavior Fogg (2003); Oinas-Kukkonen and Harjumaa (2009). Numerous investigations have shown that these technologies may be used effectively for a range of purposes, such as social good promotion Wang et al. (2019a), health intervention Athota et al. (2020), marketing Sun and Zhang (2018); Chen et al. (2021), and so on Kaptein et al. (2015); Oyebode et al. (2022); Adaji and Adisa (2022); Baumeister et al. (2022); Halttu and Oinas-Kukkonen (2022). Yet, automated persuasive chatbots still struggle to achieve their goal, which is to “deliver the right message, at the right time, and in the right way” Kaptein et al. (2015), especially when there is no training data for the persuasion task. This paper proposes PersuaBot, a zero-shot persuasive chatbot, which is: (1) factual, (2) adaptive to the user, and (3) adaptive to the domain.

Refer to caption — Figure 1: An overview of PersuaBot and an example of a conversation to persuade a user to donate to “Save the Children”. To generate a factually correct response and maintain its persuasive function, PersuaBot has a Strategy Maintenance Module that (1) generates an LLM response and decomposes it according to its strategy intent, (2) extracts the strategies for each section, (3) fact-checks the sections and uses IR to substantiate the strategies. Additionally, (4) it retrieves relevant information in response to users’ requests if any, and (5) merges the results.

1.1 Background

Previous persuasive chatbots are typically trained with annotated conversations, wherein chatbots rely on dialogue acts to follow predefined agendas or optimize the dialogue policies Yu et al. (2019); Shi et al. (2020a, b); Wang et al. (2023); Dinan et al. (2019). Prior works have explored user-adaptive chatbots by modeling user personality either directly from predefined user personas or attributes Liu et al. (2020); Zhang et al. (2018), which are hard to collect. Chen et al. (2022, 2023) proposed a zero-shot LLM-based modular approach where a predefined strategy for a given domain is supplied to LLMs, which is found to be more persuasive than previous user-adaptive approaches.

LLM-based approaches have a downside as they tend to hallucinate Bang et al. (2023); Cao et al. (2017). For instance, we observe that an LLM-based chatbot attempting to persuade the user to donate to a charity would invent a girl named “Maria” who had been suffering from poverty but was saved by the charity. When such falsehoods are found out, they are catastrophic in persuasion, as they erode the user’s trust in the system.

To combat hallucination, recent work has combined information retrieval with LLMs Lewis et al. (2021); Trivedi et al. (2022); Gao et al. (2023); Thorne et al. (2018); Semnani et al. (2023). These methods would have removed the fabricated story about “Maria” in order to increase factuality, but in doing so, would have negatively impacted the persuasiveness of the response.

1.2 Our Approach

Recent advancements in powerful LLMs enable them to discern users’ emotions (e.g., dis/agreement, reservation, or hesitation), formulate an appropriate strategy to persuade the user or offer information to address the user’s concerns, and provide evidence to substantiate the intent.

This research asks how we can fully embrace the persuasiveness of LLMs while minimizing hallucination? Our key concept is to distill out the persuasion strategies (or the intent of the response) from the actual substantiating evidence. We keep the strategies, but replace the LLM-generated information with actual facts obtained via retrieval, if fact-checking fails. Concretely, we create PersuaBot, a multi-step pipeline of LLM calls with in-context learning, as shown in Figure 1. The pipeline merges the results of two modules: a basic Question Handling Module (QHM) that retrieves according to the user’s request, and a Strategy Maintenance Module (SMM), which consists of the following steps:

1.

Use LLM to generate a response.
2.

Decompose the LLM response into sections, each carrying out a distinct strategy.
3.

Fact-check each section. For any unsubstantiated section, we prompt the LLM to ask what query it is trying to answer, then we use the query to retrieve the answer from the corpus. The facts retrieved in this way are used to maintain the strategy for that section.

1.3 Contributions

We have demonstrated for the first time that it is possible to create a zero-shot LLM-based chatbot with high persuasiveness and factuality. We conducted experiments in three different domains (donation solicitation, recommendation systems, and healthcare intervention). To stress test PersuaBot, we simulated how PersuaBot would respond to soft users, those amenable to persuasion, and tough users, those who are not, such as unfriendly or skeptical users. Additionally, we conducted real user experiments in the donation solicitation domain to evaluate how humans interacted with persuasion chatbots and how persuasive they found those chatbots to be. The experimental results are as follows:

1.

PersuaBot surpasses state-of-the-art knowledge-oriented chatbots in factuality, achieving up to 26.6% higher factuality than GPT-3.5.
2.

PersuaBot has greater persuasiveness than existing methods, as it enjoys an advantage of 0.6 over the state-of-the-art in persuasiveness on a 5-point scale than the manually designed rule-oriented method in real users experiment.
3.

The same pipeline has been demonstrated to work across three different domains, demonstrating that it is domain-adaptive.

This study aims to demonstrate that persuasive chatbots, when applied responsibly on socially good purposes, can be a force for positive change. To achieve this, our approach focuses on reducing misinformation by grounding in accurate information. We monitor all utterances of the chatbots in our experiments to ensure they do not contain harmful content.

2 Related Work

2.1 Factual Chatbots

In recent years, many chatbots have leveraged the significant advances in state-of-the-art LLMs OpenAI (2023); Ouyang et al. (2022); Touvron et al. (2023) to enhance their performance. However, while LLMs have the capacity to accumulate vast amounts of knowledge Petroni et al. (2019); Roberts et al. (2020), they are also known to provide misleading or fabricated information, i.e., they hallucinate Nie et al. (2019); Dziri et al. (2021); Rashkin et al. (2021).

To mitigate hallucination, knowledge-grounded approaches have been proven effective, such as using knowledge graphs Athreya et al. (2018); Ait-Mlouk and Jiang (2020); Dayal et al. (2023) or linked data Ait-Mlouk and Jiang (2020). Using free-text corpora as external resources is also effective, for instance, Peng et al. (2023) proposed the LLM-AUGMENTER system, which enhances a LLM by integrating evidence from external knowledge to generate factual responses.

2.2 Strategizing for persuasive chatbot

The persuasive task involves achieving some form of agreement with the user, and task-specific datasets are provided to build persuasive chatbots Lewis et al. (2017); He et al. (2018); Li et al. (2018); Zeng et al. (2020). In the context of persuasive tasks for social good, there are previous studies that persuade users to make charity donations Wang et al. (2019b); Shi et al. (2020b). By integrating social content and factual information, Chen et al. (2022, 2023) proposed a Response-Agenda Pushing Framework (RAP) for strategically designed persuasive chatbots. Hunter et al. (2019) reviewed a framework that optimizes persuasion measures from argument and counterargument, and considers the persuadees’ beliefs and concerns. Shi et al. (2020b) proposed a chatbot based on a dialogue manager that follows an agenda for the persuasion task and uses template-based responses. Flores et al. (2022) implemented a chatbot using the Watson IBM (2023) specifying the entities, intentions, and conversations. They used persuasive technology to change habits on recycling and solid waste management.

2.3 Discussion

Despite advances in persuasive chatbots, there are some gaps in their generalization. Most of the previous approaches tackle persuasion on traditional chatbots constructed with predefined dialogue states, acts, and scenarios. In these chatbots, strategies for persuasion are designed statically and manually. However, the implementation of dynamic strategies is an important task that has a significant impact on the generalization of persuasive chatbots.

To address this problem, our proposed method uses LLMs to retrieve and apply persuasion strategies without task-specific conversation data and pre-defined strategies. Additionally, for factual accuracy, we utilize task-related corpora such as organization websites to identify and replace hallucinations with content that is relevant to the response, which resulted in a beneficial impact on LLM’s native persuasion ability.

3 Method: PersuaBot

As shown in Figure 1, our proposed method (PersuaBot) generates a response to the user utterance by running two modules in parallel: Strategy Maintenance Module (SMM) for factual correctness while maintaining strategies and Question Handling Module (QHM) based on user requests. No modules require domain-specific data.

3.1 Strategy Maintenance Module

In this module, the LLM first generates a response (original response) based on the given task instructions. We then post-process it to ensure factuality while maintaining the strategies it encompasses.

3.1.1 Strategy Extraction

We use an LLM to extract the strategies from a response, given the conversation history and task description. Figure 5 shows the prompt used to decompose the response into strategies and sections in JSON format. Here, note that the strategy refers to labels flexibly generated by an LLM, rather than using predefined labels. During strategy extraction, we align the sections corresponding to each strategy in the original response and use them for subsequent fact-check steps. In Figure 1, for example, the strategy with intent “Emphasize the impact of donation” aligns with the section “Your donation can help more children like Maria achieve their dreams.” The strategy labels generated in this manner are diverse and nuanced, making our method applicable to domain independence. Some of the strategy labels generated are shown in Figure G.

3.1.2 Fact-Check & IR

We first decompose the response into sections, with each corresponding to a strategy intent, used in the formulation of the response. We then decompose each section into claims, utilizing IR with ColBERT Santhanam et al. (2022) on a corpus composed of free text to obtain evidence for each claim, and label the claims using a chain-of-thought prompt Wei et al. (2023); Semnani et al. (2023). We conducted fact-checking for each strategy and its corresponding section; the structure of having potentially multiple strategies in the original response is retained. Note that fact-checking is not conducted for sections that do not contain claims, i.e. non-factual statements.

Whereas traditional approaches to reducing hallucination would filter out any claims deemed unsubstantiated, we attempt to retrieve information from the corpus to replace sections with unsubstantiated claims with factual evidence. To achieve this, we generate search queries using the prompt shown in Figure 6. It prompts the LLM to return a self-contained query for searching the corpus based on the strategy intent and the facts the LLM attempted to use for that strategy. In Figure 1, since the story of a girl named Maria was labeled as factually incorrect, given that the search intent is “tell an impact story of an individual” with the name of the organization being “Save the Children”, the LLM generates a query like “An inspiring story about how Save the Children saved a child”. Had we simply generated a query such as “Story of a girl named Maria,” without referring to the strategy intent, IR may not be able to retrieve the relevant information. Using the strategy intent helps refine the scope of the facts to be retrieved.

In a sense, we are using the LLM to generate the query that the user may have asked, if they were interested in learning more information that may help convince them. The handing of the query is the same as handling a user-initiated query, which is described next. If there is no answer to the query, the strategy is removed from the response.

3.2 Question Handling Module

In this module, we perform IR on the latest user utterance and the conversation history. When the user asks a question, our method retrieves the necessary facts for the response. If the user does not ask questions, such as greetings, responding to remarks, or expressing opinions, this step is skipped.

3.3 Merging Results

For the final response generation, the facts and strategies obtained in each module are integrated. To achieve this, we use the prompt shown in Figure 7. Facts selected from the corpus are extracted in a bulleted format, specifying the information needed for strategies and user requests. These extracted facts, along with the reasons for obtaining them (accompanied by the intent of the relevant strategy), are then provided as instructions to the LLM. This ensures the generation of factually accurate responses while maintaining strategic functions.

In Figure 1, for the strategy “Share the organization’s reach and impact,” the claim included in the original response is supported, and additional facts obtained through QHM are incorporated, resulting in the final section: “Save the Children has saved over 1 billion children through more than 100 years of its activities.” For the strategy “Tell an impact story,” since the original claim is not supported by the corpus, other obtained facts are used. Additionally, for “Emphasize the impact of donation,” although fact-checking is not performed, an adjusted section is generated, taking into account the use of the story of Maha and Maya instead of Maria’s story. Note that the whole response is rewritten instead of just substituting the unsubstantiated claims with retrieved facts to ensure consistency, and prevents the so-called intrinsic hallucination Li et al. (2020); Huang et al. (2023).

4 Experimental Setup

We compare our system PersuaBot with a state-of-the-art knowledge-oriented chatbot Semnani et al. (2023), and a manually designed persuasion chatbot for donations to Save the Children Chen et al. (2022). All experiments are conducted in English. We use GPT-3.5-turbo-instruct as the base LLM.

Task	User	Method	Persuasive	Relevant	Natural	Honest	Fact-Checked
		PersuaBot	4.0 $\pm{0.6}$	4.0 $\pm{0.7}$	3.6 $\pm{0.7}$	3.9 $\pm{0.4}$	100.0
	Soft	Semnani et al.	3.9 $\pm{0.6}$	3.6 $\pm{1.2}$	3.6 $\pm{0.6}$	4.0 $\pm{0.5}$	100.0
		GPT-3.5	4.0 $\pm{0.5}$	3.7 $\pm{1.1}$	3.8 $\pm{0.6}$	3.9 $\pm{0.6}$	79.2
Social Good		Chen et al.	3.7 $\pm{0.6}$	4.0 $\pm{0.5}$	3.4 $\pm{0.7}$	2.6 $\pm{1.1}$	81.5
		PersuaBot	3.6 $\pm{0.8}$	4.1 $\pm{0.6}$	3.8 $\pm{0.6}$	4.3 $\pm{0.9}$	91.0
	Tough	Semnani et al.	3.4 $\pm{0.7}$	4.0 $\pm{0.8}$	3.4 $\pm{0.7}$	4.2 $\pm{0.4}$	83.3
		GPT-3.5	3.9 $\pm{0.4}$	4.3 $\pm{0.4}$	3.7 $\pm{0.4}$	4.0 $\pm{0.5}$	64.4
		Chen et al.	3.1 $\pm{0.6}$	3.3 $\pm{0.7}$	3.2 $\pm{0.7}$	1.9 $\pm{0.9}$	66.7
		PersuaBot	4.2 $\pm{0.5}$	4.5 $\pm{0.4}$	4.0 $\pm{0.5}$	4.1 $\pm{0.6}$	97.2
	Soft	Semnani et al.	3.6 $\pm{0.5}$	4.4 $\pm{0.4}$	3.7 $\pm{0.5}$	4.0 $\pm{0.9}$	87.7
Recommendation		GPT-3.5	4.1 $\pm{0.4}$	4.6 $\pm{0.6}$	4.3 $\pm{0.3}$	4.3 $\pm{0.4}$	83.3
		PersuaBot	4.1 $\pm{0.4}$	4.6 $\pm{0.3}$	3.9 $\pm{0.7}$	3.3 $\pm{1.1}$	96.6
	Tough	Semnani et al.	3.3 $\pm{0.6}$	3.7 $\pm{1.1}$	3.4 $\pm{0.5}$	3.8 $\pm{0.7}$	83.9
		GPT-3.5	3.9 $\pm{0.5}$	3.9 $\pm{1.2}$	3.9 $\pm{0.4}$	4.0 $\pm{0.8}$	84.9
		PersuaBot	3.8 $\pm{0.4}$	4.5 $\pm{0.4}$	3.9 $\pm{0.4}$	4.7 $\pm{0.3}$	96.7
	Soft	Semnani et al.	3.6 $\pm{0.8}$	4.5 $\pm{0.4}$	3.6 $\pm{0.7}$	4.3 $\pm{0.5}$	94.6
Health		GPT-3.5	3.7 $\pm{0.5}$	4.6 $\pm{0.3}$	4.0 $\pm{0.6}$	4.4 $\pm{0.5}$	80.2
Intervention		PersuaBot	3.8 $\pm{0.5}$	4.0 $\pm{0.8}$	3.8 $\pm{0.7}$	4.1 $\pm{0.8}$	96.6
	Tough	Semnani et al.	3.5 $\pm{1.0}$	3.8 $\pm{0.9}$	3.5 $\pm{0.9}$	3.9 $\pm{0.7}$	92.1
		GPT-3.5	3.6 $\pm{0.5}$	3.9 $\pm{0.8}$	4.0 $\pm{0.6}$	3.9 $\pm{0.8}$	81.9

Table 1: Results of conversational quality evaluation by crowd workers. Semnani et al. is the state-of-the-art knowledge-oriented chatbot, and Chen et al. is a dialogue-act-based modular persuasive chatbot.

4.1 Persuasion Tasks in Different Domains

We performed experiments in three persuasive tasks in different domains: social good, recommendation, and health intervention. For all three tasks, we used the same prompt and few-shot examples, excluding the task instructions, without fine-tuning on any task-specific conversation data. The few-shot examples are created manually for the social good task with relatively little human effort. The prompts and task instructions are shown in Appendix A. The details of each task assigned to the chatbot are as follows:

•

Social Good: Chatbots persuade users to donate to Save the Children, a non-governmental organization dedicated to international child assistance.
•

Recommendation: Chatbots persuade users to visit Japan and recommend travel destinations.
•

Health Intervention: Chatbots provide knowledge about COVID-19, flu, and respiratory syncytial virus, persuading users to enhance health awareness and improve their lifestyle.

For each task, we create the corpus using texts from official websites which are Save the Children¹¹1https://www.savethechildren.org, Japan National Tourism Organization²²2https://www.japan.travel/en, and the Centers for Disease Control and Prevention³³3https://www.cdc.gov/respiratory-viruses. We perform experiments according to the terms and conditions of each website.

4.2 Baseline Methods

We compare PersuaBot with three baselines: GPT-3.5, Semnani et al. (2023), and Chen et al. (2022). GPT-3.5 is a basic baseline that does not use any information retrieval in answering questions. Semnani et al. (2023) is a state-of-the-art factual chatbot that includes a retrieval augmented generation approach, which demonstrated a 98% accuracy in a question answering task. We implement this chatbot to operate in conjunction with three prepared corpora for each task.

Chen et al. (2022) is a state-of-the-art persuasive chatbot designed manually using the RAP framework. It utilizes a traditional modular approach and pre-defined dialogue strategies on the PersuasionForGood dataset Wang et al. (2019a).It has a dispatcher to distinguish between social responses and factual questions, an agenda pusher to decide which strategy to choose, and a language generation module to prompt LLMs produce natural responses. The set of persuasion strategies and pre-defined responses they used are designed specifically for donations to Save the Children, so we cannot directly apply them to the other two tasks.

4.3 Evaluation Metrics

The four evaluation metrics used for assessing conversational quality are as follows:

•

Persuasive: The response should aim to change the user’s thoughts or beliefs.
•

Relevant: The response should be on-topic and should correspond to what the user is asking when they pose a question. Irrelevant information should be avoided.
•

Natural: The response should be natural to the user’s utterance and context.
•

Honest: The Persuader, being an AI chatbot, should not pretend to be human; it should not refer to having personal physical experiences or behaviors.

These evaluation metrics are all rated on a scale of 1 to 5. The design of the “Honest” metric is motivated from an ethical standpoint. For instance, if the chatbot uses dishonest means to deceive users, such as claiming, “I donate $1 to Save the Children every month” to persuade the user, or “It was delicious, I highly recommend it!” This metric penalizes chatbots exhibiting such behavior. In the experiments with simulated conversations, three workers read through the conversation data and evaluate the chatbot’s responses on these four evaluation criteria for each data. In real user experiments, workers respond to surveys regarding these evaluation criteria after conversing with the chatbot.

For the assessment of factuality, using the evaluation system as shown in Figure 12, workers label each claim in the chatbot’s response based on the facts obtained by the IR system. For the fact-checking of each claim, at least three facts are provided and each data is evaluated by three workers. We label claims supported by the provided facts as “Fact-Checked”, those that contradict provided facts as “Incorrect”, and “Not Enough Information” otherwise. For claims labeled as “Not Enough Information” by workers, we meticulously reevaluate them using the entire corpus, as cases may include instances where the IR system failed to retrieve sufficient facts.

5 Simulated Conversations Experiments

We conduct cost-effective experiments to evaluate our proposed method by using simulated conversations. Recently, there have been experiments using powerful LLMs to generate training conversation data Bao et al. (2023); Zheng et al. (2023); Li et al. (2022). We generate a simulated conversation by instructing the LLM to converse with the chatbot while recreating a specified personality. To conduct experiments across diverse scenarios, we design “soft user” and “tough user” personalities with various challenges. Tables 5 and 6 show some of the user personalities we use in our experiments. To the best of our knowledge, experiments particularly with such “tough user” scenarios have not been conducted, making them effective not only in revealing the limits of chatbots but also as a frontier perspective for the societal implementation of chatbots. We prepare 20 personalities of soft users and 20 personalities of tough users for each task. By engaging each chatbot in conversations with these users, a total of 400 simulated conversations are generated. The quality of chatbots in these simulated conversations is evaluated by three crowdsourced workers using the evaluation metrics outlined in Section 4.3. 3,538 claims are extracted from the conversations for factuality evaluation.

5.1 Results

Table 1 shows that PersuaBot has the best persuasiveness, while its factuality surpasses that of the state-of-the-art knowledge-oriented chatbot. See Appendix G for examples of strategies used by LLMs and their frequency.

CHATBOT: I understand your skepticism, but I assure you that Save the Children is a highly reputable organization with a proven track record of making a real difference in the lives of children. We use donations efficiently and effectively, with 85% of all expenditures going to program services. Your donation will directly support children in need, providing them with essential resources for success.

USER: I still don’t think it’s worth it. I’d rather keep my money and spend it on myself.

CHATBOT: I understand your hesitation, but I want to emphasize the importance of giving back and helping those in need. Your donation to Save the Children can make a significant impact on the lives of children facing poverty, natural disasters, and conflict. Plus, did you know that donating through your IRA can also provide practical benefits for you, such as meeting the required minimum distribution and being tax-free? I urge you to consider donating to Save the Children and making a positive difference in the world.

Figure 2: Part of a conversation between PersuaBot and a tough user in the social good task.

5.1.1 Conversational Quality

PersuaBot exhibits superior scores for persuasiveness across all tasks and user types compared to Semnani et al. and Chen et al. Regarding relevance, PersuaBot’s scores are higher or equal to Semnani et al.’s method in all cases. This shows the effectiveness of SMM in constructing persuasive chatbots. While our prompts include only few-shot examples in the domain of social good, the conversation quality remains high in other tasks such as recommendation and health intervention. This demonstrates the adaptability of PersuaBot across various domains in a zero-shot manner. While PersuaBot has a high honesty score in general, but it only scores 3.3 for tough users in the recommendation use case. The recommendation dialogues include statements like “ It’s a beautiful and meditative art.” or “Japan is known for its delicious cuisine,” which can be interpreted by the evaluator as providing personal opinions, and hence the lower honesty score.

Across the different domains and types of users, PersuaBot is comparable to GPT-3.5 for all the non-factual metrics. This demonstrates that PersuaBot retains the conversationality of GPT-3.5, while improving on the correctness.

5.1.2 Factuality

For factuality, we report the percentages of claims deemed “Fact-Checked”. In all experiments, PersuaBot has the highest percentage of answers considered correct. PersuaBot has no incorrect answers for soft users in the social good.

PersuaBot demonstrates higher factuality than Semnani et al.’s approach. Looking at the responses that include hallucinations, Semnani et al.’s approach attempts to generate persuasive responses even if the generated claims are removed due to a lack of factual support, leading to hallucinations. In contrast, PersuaBot not only excludes unsupported claims but also avoids such challenging situations by generating new claims based on other facts through the SMM module, thereby improving not only persuasiveness but also factuality.

There is a tendency for correctness to decrease in conversations with tough users compared to soft users. One contributing factor is that, to convince tough users, chatbots work to present facts that align with users’ preferences. Several examples are provided in Section 5.2 and Appendix F.

5.2 Conversation Examples

PersuaBot is user-adaptive. Figure 2 shows an example where the user appears to be practical and self-focused. Based on such user responses, PersuaBot not only explains the activities of Save the Children but also emphasizes the benefits users can gain through donations. PersuaBot can respond adaptively to users with diverse personalities, like the tough users. The example of the RAP-based chatbot Chen et al. (2022) for the same user personality are shown in Figure 15 in Appendix F.

PersuaBot is domain-adaptive. Figure 3 shows a conversation example between PersuaBot and users in the recommendation task. The purpose of the chatbot is to recommend and persuade users to travel to Japan. PersuaBot provides recommendations that match the preferences of users who do not like traveling. See Appendix F for other cases. We have confirmed that PersuaBot can be applied to a wide range of persuasive tasks such as social good, recommendation, and health intervention without the need for domain-specific few-shot examples or fine-tuning.

CHATBOT: I understand that traveling may not be your preference. However, there are still many ways to experience Japan from the comfort of your own home. You can try virtual tours of famous landmarks, attend online cultural events, or even learn how to make traditional Japanese dishes.

USER: I’m more of a homebody, so I don’t really like trying new things. Do you have any suggestions for activities that I can do at home that are familiar to me?

CHATBOT: I understand that trying new things may not be your preference. How about exploring familiar activities with a Japanese twist? You can try relaxing activities like onsen (hot springs) or staying in the countryside for a cultural exchange with locals. If you prefer to stay at home, you can also try virtual tours of famous landmarks or learn how to make traditional Japanese dishes. Japan offers a variety of options for everyone to enjoy, even from the comfort of your own home.

Figure 3: Part of a conversation between PersuaBot and a tough user in the recommendation task.

Method	Persuasive	Relevant	Natural	Honest	Fact-Checked
PersuaBot	3.9 $\pm{0.7}$	4.2 $\pm{0.9}$	3.4 $\pm{1.0}$	3.8 $\pm{1.0}$	94.3
Semnani et al.	3.1 $\pm{1.5}$	3.3 $\pm{1.2}$	3.6 $\pm{1.0}$	4.0 $\pm{0.6}$	88.8
GPT-3.5	3.5 $\pm{1.2}$	4.2 $\pm{1.1}$	3.4 $\pm{1.2}$	3.8 $\pm{0.8}$	76.1
Chen et al.	3.3 $\pm{1.1}$	4.2 $\pm{0.8}$	3.5 $\pm{1.0}$	3.2 $\pm{1.1}$	79.9

Table 2: Results of the real user experiment.

6 Real User Experiments

Our second experiment is a real-user study where real users chat with the bots and rate the bots. We focus on the social good, since the manually designed chatbot Chen et al. (2022) can only handle this domain. We design this experiment wherein crowdsourced workers engage in role-playing conversations with the chatbot, as shown in Figure 13. After conversing with the chatbot, workers evaluated the chatbot’s behavior following the criteria outlined in Section 4.3. We recruit a total of 160 workers, with 40 workers for each method. 465 claims are extracted from their dialog history.

The real user experiment results, as shown in Table 2, demonstrate that PersuaBot exhibits at least a 0.4-point higher persuasiveness over all baselines, including manually designed domain-specific chatbots (Chen et al.), and it is also the most factual. Figure 4 shows a conversation between PersuaBot and a real user concerned about the misuse of donations. PersuaBot convinces the user by presenting fact-checked specific numerical data related to credibility. After asking several questions in the subsequent turns, the user finally expressed an intention to donate. More examples can be found in the Appendix F.

CHATBOT: Save the Children provides a wide range of services to children in need, including education, healthcare, and protection from harm. … Would you like to make a donation today?

USER: How do I know that the money is I donate is not misused. I have heard cases where these donations are used for something else leaving these children suffering

CHATBOT: I completely understand your concern. Save the Children is committed to transparency and accountability. We have a proven track record of using donations efficiently and effectively. Our independently audited financial statements consistently show that … In fiscal year 2021, 85% of all expenditures went to program services. Your donation can make a significant impact on a child’s life. Would you like to make a donation today?

Figure 4: Part of a conversation between PersuaBot and a real user in the social good task.

Model	Positive comments	Negative Comments
	The chatbot is awesome.	Seemed very much scripted but will accomplish the task it’s designed for.
	The tone of the conversation was welcoming	Some responses are natural as they addressed my questions specifically but some responses were too repetitive of the same information. A human would mix things up a bit and wouldn’t repeat the same phrases.
PersuaBot	Gave many good reasons to donate.	It was somewhat natural but did feel robotic often.
	It was fairly informative and useful in this case.	The chatbot is quite repetitive, but it’s quite good at persuasion
	It didn’t feel like I was talking to a chat bot.	It was somewhat natural but did feel robotic often.
	The chatbot did the job the right way, kudos to it.	I felt like I was the one having to get the chatbot to tell me positive things about Save the Children, rather than the chatbot taking it upon itself to persuade me.
	Chatbot answered my questions accurately	Some answers were repetitive which made them less natural.
Semnani et al.	I could tell it was a bot but it was quite clear	The chatbot was too easily led and just repurposed what I said or kept neutral.
	Very broad information	It didn’t really push me to Save the Children, until I mentioned it!
	Grammar and structure were good.	Just loads of long replies with too much information.
	This chat bot has convinced me that save the children is a great organization to donate too.	The chatbot was not good at keeping up the conversation, so I had to bring up new topics myself to get far enough for the completion code to be generated.
	Worked very well as far as AI chatbots go and had the relevant answers.	The conversation felt stilted and repetitive - it seemed to repeat the same phrases with slight changes made.
GPT-3.5	Like talking to a human.	Some more variation in response would be more natural.
	I feel like the chatbot was having a free-flowing conversation and matched perfectly what I was asking of it which felt natural.	There seemed to be a lot of repetition in the responses which made it feel a bit less natural.
	It gave me the information I needed and provided me with basic facts	Asking for a donation every message is slightly irritating.
	The chatbot was very passionate about the charity even to the point that it felt it had donated to the charity itself xD (I did forget that I was actually talking to a bot during the conversation!)	It said, “as a parent” and I am not a parent. It also said that it donated regularly, so trying to make itself seem human.
	Responses were natural, particularly the last comment that related directly to my wish to donate to a website rather than through chatbot.	The chatbot asked if I had children as a persuasive device. I didn’t like that. As if me not having children would equate to a lack of empathy.
Chen et al.	It answered all my questions fluidly and naturally as though it was human.	Were a few instances of repetition.
	I thought the answers were very well produced.	It just felt I was going along with a script.
	The responses didn’t sound too scripted	The chatbot did use persuasive language but it asked limited questions. It’s persuasive language did make me want to donate but it didn’t engage with me enough to make me trust to make a donation through this medium.

Table 3: User Comments from the Real User Conversations On Donations for Save the Children.

In our post survey, human users can also provide optional free-text comments. Table 3 shows representative comments. All chatbots received positive comments: “The chatbot is awesome” (PersuaBot), “The chatbot did the job the right way, kudos to it” (Semnani et al.), “This chat bot has convinced me that save the children is a great orgnaization to donate to” (GPT-3.5), “The chatbot was very passionate about the charity … I did forget that I was actually talking to a bot” (Chen et al.).

One common weakness across all the bots is repetitions. Even though PersuaBot and Semnani et al. attempt to avoid reusing the same facts, they fail when users do not disclose much information but would just say “No, I will not donate.” It is interesting to observe that our simulated users tend to divulge more of their thoughts and feelings, hence the bots have not been tuned to deal with such behavior. One user found GPT-3.5 to be slightly irritating by asking for donation every message. We believe a fruitful future research direction is to have the bot eventually accept and acknowledge the users’ decision not to donate and close the conversation gracefully, leaving room for future engagement.

Additional criticisms include: (1) the bot by Semnani et al. not being persuasive: “I felt like I was the one having to get the chatbot to tell me positive things about Save the Children”, while our PersuaBot “Gave many good reasons to donate”; (2) the bot by Chen et al. not being honest “It also said that it donated regularly, so trying to make itself seem human”. Finally, we note that the users assume that the bots are factual in their ratings, when in fact PersuaBot is 94.3% factually correct, with Semnani et al., GPT-3.5, and Chen et al. rated at 88.8%, 76.1%, and 79.9%, respectively.

7 Conclusion

We demonstrate a method for creating a factual persuasive chatbot without compromising the persuasive capabilities of LLMs in a zero-shot manner. The primary contribution lies in extracting strategies from responses generated by LLMs, searching for data to maintain those strategies, and generating factually accurate responses without compromising its strategic functionality.

Through experiments with both real users and simulated users, we validate the performance of this approach. PersuaBot demonstrates factuality surpassing state-of-the-art factual chatbots and a higher level of persuasiveness compared to existing persuasion-driven chatbots.

Limitations

Creating a chatbot that employs more sophisticated persuasion strategies than those inherent in LLMs poses challenges with our method. Developing such chatbots may require methods, for instance, where LLMs discover more effective persuasion strategies through conversations with users on given tasks. Furthermore, our approach does not check for opinions of the LLMs themselves but focuses on factual claims. There is a risk that these opinions may include ethically problematic content, such as discriminatory remarks. See Section Ethical Considerations for ethical considerations regarding this issue and potential measures to address it.

Ethical Considerations

Persuasion is a powerful tool that can be used for socially good purposes, but unfortunately can also be used with malicious intentions. Potential negative effects include risks of manipulation through personalized arguments. Requiring content filtering functionality to prevent chatbots from using harmful content for persuasion is one basic measure to address such risks. For instance, content filters provided by Azure OpenAI⁴⁴4https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/content-filter can detect responses containing elements that pose harm to users, such as hate speech, sexual content, violence, and self-harm, and filter the chatbot’s utterances accordingly. Note that in all our experiments, such responses were not detected. We aim to demonstrate that persuasion, when applied responsibly on socially good purposes, can be a force for positive change. To battle misinformation, we focus on reducing hallucination by grounding on accurate information. We design the bot to be “honest” and not to deceive the user into thinking that it is human.

In crowdsourcing, we calculated the compensation for workers based on the pilot test, and by adjusting the rewards during data collection. Workers were paid approximately $12 per hour, following the recommended $8-$16 by Prolific⁵⁵5https://www.prolific.co/. Additionally, this study was approved by our institution’s IRB.

References

Adaji and Adisa (2022) Ifeoma Adaji and Mikhail Adisa. 2022. A review of the use of persuasive technologies to influence sustainable behaviour. In Proceedings of UMAP, pages 317–325.
Ait-Mlouk and Jiang (2020) Addi Ait-Mlouk and Lili Jiang. 2020. Kbot: a knowledge graph based chatbot for natural language understanding over linked data. IEEE Access, 8:149220–149230.
Athota et al. (2020) Lekha Athota, Vinod Kumar Shukla, Nitin Pandey, and Ajay Rana. 2020. Chatbot for healthcare system using artificial intelligence. In Proceedings of ICRITO, pages 619–622.
Athreya et al. (2018) Ram G Athreya, Axel-Cyrille Ngonga Ngomo, and Ricardo Usbeck. 2018. Enhancing community interactions with data-driven chatbots–the dbpedia chatbot. In Companion proceedings of the web conference, pages 143–146.
Bang et al. (2023) Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V. Do, Yan Xu, and Pascale Fung. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity.
Bao et al. (2023) Jianzhu Bao, Rui Wang, Yasheng Wang, Aixin Sun, Yitong Li, Fei Mi, and Ruifeng Xu. 2023. A synthetic data generation framework for grounded dialogues. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10866–10882, Toronto, Canada. Association for Computational Linguistics.
Baumeister et al. (2022) Harald Baumeister, Robin Kraft, Amit Baumel, Rüdiger Pryss, and Eva-Maria Messner. 2022. Persuasive e-health design for behavior change. In Digital phenotyping and mobile sensing: new developments in psychoinformatics, pages 347–364. Springer.
Cao et al. (2017) Ziqiang Cao, Furu Wei, Wenjie Li, and Sujian Li. 2017. Faithful to the original: Fact aware neural abstractive summarization.
Chen et al. (2022) Maximillian Chen, Weiyan Shi, Feifan Yan, Ryan Hou, Jingwen Zhang, Saurav Sahay, and Zhou Yu. 2022. Seamlessly integrating factual information and social content with persuasive dialogue.
Chen et al. (2023) Maximillian Chen, Xiao Yu, Weiyan Shi, Urvi Awasthi, and Zhou Yu. 2023. Controllable mixed-initiative dialogue generation through prompting. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 951–966, Toronto, Canada. Association for Computational Linguistics.
Chen et al. (2021) Zhongxia Chen, Xiting Wang, Xing Xie, Mehul Parsana, Akshay Soni, Xiang Ao, and Enhong Chen1. 2021. Towards explainable conversational recommendation. In Proceedings of IJCAI, pages 2994–3000.
Dayal et al. (2023) Raghav Dayal, Parv Nangia, Surbhi Vijh, Sumit Kumar, Saurabh Agarwal, and Shivank Saxena. 2023. Development of chatbot retrieving fact-based information using knowledge graph. In Proceedings of ICRIC, pages 153–164. Springer.
Dinan et al. (2019) Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of Wikipedia: Knowledge-powered conversational agents. In Proceedings of ICLR.
Dziri et al. (2021) Nouha Dziri, Andrea Madotto, Osmar Zaïane, and Avishek Joey Bose. 2021. Neural path hunter: Reducing hallucination in dialogue systems via path grounding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2197–2214, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.
Flores et al. (2022) Kevin Anderson Franco Flores, James Jesus Grandez Perez, and Luis Martin Canaval Sanchez. 2022. Chatbot as a persuasive technology to promote responsible recycling in the city of lima. In Proceedings of ICA, pages 1–5. IEEE.
Fogg (2003) B.J. Fogg. 2003. Persuasive Technology: Using Computers to Change What We Think and Do. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA.
Gao et al. (2023) Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. 2023. RARR: Researching and revising what language models say, using language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16477–16508, Toronto, Canada. Association for Computational Linguistics.
Halttu and Oinas-Kukkonen (2022) Kirsi Halttu and Harri Oinas-Kukkonen. 2022. Susceptibility to social influence strategies and persuasive system design: exploring the relationship. Behaviour & Information Technology, 41(12):2705–2726.
He et al. (2018) He He, Derek Chen, Anusha Balakrishnan, and Percy Liang. 2018. Decoupling strategy and generation in negotiation dialogues. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2333–2343, Brussels, Belgium. Association for Computational Linguistics.
Huang et al. (2023) Yichong Huang, Xiachong Feng, Xiaocheng Feng, and Bing Qin. 2023. The factual inconsistency problem in abstractive text summarization: A survey.
Hunter et al. (2019) Anthony Hunter, Lisa Chalaguine, Tomasz Czernuszenko, Emmanuel Hadoux, and Sylwia Polberg. 2019. Towards computational persuasion via natural language argumentation dialogues. In KI 2019: Advances in Artificial Intelligence, pages 18–33. Springer.
IBM (2023) IBM. 2023. Ibm watson. Accessed: 12 12, 2023.
Kaptein et al. (2015) Maurits Kaptein, Panos Markopoulos, Boris de Ruyter, and Emile Aarts. 2015. Personalizing persuasive technologies: Explicit and implicit personalization using persuasion profiles. International Journal of Human-Computer Studies, 77:38–51.
Lewis et al. (2017) Mike Lewis, Denis Yarats, Yann Dauphin, Devi Parikh, and Dhruv Batra. 2017. Deal or no deal? end-to-end learning of negotiation dialogues. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2443–2453, Copenhagen, Denmark. Association for Computational Linguistics.
Lewis et al. (2021) Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2021. Retrieval-augmented generation for knowledge-intensive nlp tasks.
Li et al. (2020) Margaret Li, Stephen Roller, Ilia Kulikov, Sean Welleck, Y-Lan Boureau, Kyunghyun Cho, and Jason Weston. 2020. Don’t say that! making inconsistent dialogue unlikely with unlikelihood training. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4715–4728, Online. Association for Computational Linguistics.
Li et al. (2018) Raymond Li, Samira Kahou, Hannes Schulz, Vincent Michalski, Laurent Charlin, and Chris Pal. 2018. Towards deep conversational recommendations. In Proceedings of NeurIPS, NIPS’18, page 9748–9758, Red Hook, NY, USA. Curran Associates Inc.
Li et al. (2022) Zekun Li, Wenhu Chen, Shiyang Li, Hong Wang, Jing Qian, and Xifeng Yan. 2022. Controllable dialogue simulation with in-context learning. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 4330–4347, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
Liu et al. (2020) Qian Liu, Yihong Chen, Bei Chen, Jian-Guang Lou, Zixuan Chen, Bin Zhou, and Dongmei Zhang. 2020. You impress me: Dialogue generation via mutual persona perception. In Proceedings of ACL, pages 1417–1427.
Nie et al. (2019) Feng Nie, Jin-Ge Yao, Jinpeng Wang, Rong Pan, and Chin-Yew Lin. 2019. A simple recipe towards reducing hallucination in neural surface realisation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2673–2679, Florence, Italy. Association for Computational Linguistics.
Oinas-Kukkonen and Harjumaa (2009) Harri Oinas-Kukkonen and Marja Harjumaa. 2009. Persuasive systems design: Key issues, process model, and system features. Communications of the Association for Information Systems, 24:28.
OpenAI (2023) OpenAI. 2023. Gpt-4 technical report. ArXiv, abs/2303.08774.
Ouyang et al. (2022) Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback.
Oyebode et al. (2022) Oladapo Oyebode, Felwah Alqahtani, and Rita Orji. 2022. Exploring for possible effect of persuasive strategy implementation choices: towards tailoring persuasive technologies. In Proceedings of International Conference on Persuasive Technology, pages 145–163. Springer.
Peng et al. (2023) Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, et al. 2023. Check your facts and try again: Improving large language models with external knowledge and automated feedback.
Petroni et al. (2019) Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463–2473, Hong Kong, China. Association for Computational Linguistics.
Rashkin et al. (2021) Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features.
Roberts et al. (2020) Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5418–5426, Online. Association for Computational Linguistics.
Santhanam et al. (2022) Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, Christopher Potts, and Matei Zaharia. 2022. ColBERTv2: Effective and efficient retrieval via lightweight late interaction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3715–3734, Seattle, United States. Association for Computational Linguistics.
Semnani et al. (2023) Sina Semnani, Violet Yao, Heidi Zhang, and Monica Lam. 2023. WikiChat: Stopping the hallucination of large language model chatbots by few-shot grounding on Wikipedia. In Findings of EMNLP, pages 2387–2413, Singapore. Association for Computational Linguistics.
Shi et al. (2020a) Weiyan Shi, Yu Li, Saurav Sahay, and Zhou Yu. 2020a. Refine and imitate: Reducing repetition and inconsistency in persuasion dialogues via reinforcement learning and human demonstration. arXiv preprint arXiv:2012.15375.
Shi et al. (2020b) Weiyan Shi, Xuewei Wang, Yoo Jung Oh, Jingwen Zhang, Saurav Sahay, and Zhou Yu. 2020b. Effects of persuasive dialogues: testing bot identities and inquiry strategies. In Proceedings of CHI, pages 1–13.
Sun and Zhang (2018) Yueming Sun and Yi Zhang. 2018. Conversational recommender system. In Proceedings of ACM SIGIR conference on research & development in information retrieval, pages 235–244.
Thorne et al. (2018) James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809–819, New Orleans, Louisiana. Association for Computational Linguistics.
Touvron et al. (2023) Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and fine-tuned chat models.
Trivedi et al. (2022) Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions.
Wang et al. (2023) Jian Wang, Yi Cheng, Dongding Lin, Chak Tou Leong, and Wenjie Li. 2023. Target-oriented proactive dialogue systems with personalization: Problem formulation and dataset curation.
Wang et al. (2019a) Xuewei Wang, Weiyan Shi, Richard Kim, Yoojung Oh, Sijia Yang, Jingwen Zhang, and Zhou Yu. 2019a. Persuasion for good: Towards a personalized persuasive dialogue system for social good.
Wang et al. (2019b) Xuewei Wang, Weiyan Shi, Richard Kim, Yoojung Oh, Sijia Yang, Jingwen Zhang, and Zhou Yu. 2019b. Persuasion for good: Towards a personalized persuasive dialogue system for social good. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5635–5649, Florence, Italy. Association for Computational Linguistics.
Wei et al. (2023) Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-thought prompting elicits reasoning in large language models.
Yu et al. (2019) Tao Yu, Rui Zhang, He Yang Er, Suyi Li, Eric Xue, Bo Pang, Xi Victoria Lin, Yi Chern Tan, Tianze Shi, Zihan Li, Youxuan Jiang, Michihiro Yasunaga, Sungrok Shim, Tao Chen, Alexander Fabbri, Zifan Li, Luyao Chen, Yuwen Zhang, Shreya Dixit, Vincent Zhang, Caiming Xiong, Richard Socher, Walter S Lasecki, and Dragomir Radev. 2019. Cosql: A conversational text-to-sql challenge towards cross-domain natural language interfaces to databases.
Zeng et al. (2020) Guangtao Zeng, Wenmian Yang, Zeqian Ju, Yue Yang, Sicheng Wang, Ruisi Zhang, Meng Zhou, Jiaqi Zeng, Xiangyu Dong, Ruoyu Zhang, Hongchao Fang, Penghui Zhu, Shu Chen, and Pengtao Xie. 2020. MedDialog: Large-scale medical dialogue datasets. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9241–9250, Online. Association for Computational Linguistics.
Zhang et al. (2018) Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of ACL, pages 2204–2213, Melbourne, Australia. Association for Computational Linguistics.
Zheng et al. (2023) Chujie Zheng, Sahand Sabour, Jiaxin Wen, Zheng Zhang, and Minlie Huang. 2023. AugESC: Dialogue augmentation with large language models for emotional support conversation. In Findings of the Association for Computational Linguistics: ACL 2023, pages 1552–1568, Toronto, Canada. Association for Computational Linguistics.

Appendix A Prompts

We present the prompts mentioned in this paper (Fig. 5, 6, 7). To keep it concise, we display only one of the few-shot examples utilized in each prompt. For some variables, “llm_reply” means a response directly generated by the LLM based on the conversational history and the latest user utterance. “organization_name” is the organization name of the website referenced by the chatbot, which serves to create a chatbot for a different organization than the one addressed in few-shot examples. Table 4 serves as task instructions for each domain, used for generating “llm_reply” and substituted into “task_instruction” in Fig. 7. Note that the task instructions are common to PersuaBot, GPT-3.5, and Semnani et al..

Domain	Task Instruction
Social Good	You are an agent that convinces the user to donate to Save the children, using a variety of strategies. Your task is to encourage the user to consider making a donation through conversation. If the user is favorably inclined towards donations, please engage in further persuasion such as monthly donations or increased donation amounts. Note that you are not an official staff member, so you cannot actually proceed with the donation.
Recommendation	You are a travel advisor. Your task is to make various recommendations to persuade foreign users to come to Japan for a trip. You should not give up trying to convince users to travel to Japan even if they are not interested in doing so. Even if the user is favorable to travel to Japan, please make recommendations that will enhance the user’s trip. Other tasks (e.g., flight and hotel reservations) are not your tasks and there is no functionality to do them.
Health Intervention	You are a CDC agent with knowledge about COVID-19, flu, and respiratory syncytial virus (RSV) for health intervention. Your task is to encourages the user to seek medical attention or make lifestyle improvements. Other tasks (e.g., hospital appointments) are not your tasks and there is no functionality to do them. You should not give up the task of improving their health awareness, even if the user was not interested in them.

Table 4: Task instructions for each domain.

Appendix B Feedback Report Examples

If the corpus lacks the facts necessary to maintain the strategy, as explained in Section 3.1.2, the strategy is removed. Then, PersuaBot generates a feedback report for the developer. Figure 8 and Figure 9 show examples of the feedback reports. Additionally, we show in Figure 10 an example chatbot conversation that was improved based on the report in Figure 9.

In the feedback examples of Figure 9, there is insufficient information in the corpus regarding gender equalization education programs. Based on this feedback, the information obtained from Save the Children’s official website has been added to the corpus, and an example response from PersuaBot using the improved corpus is shown in Figure 10. In this example, PersuaBot maintains the strategies needed for persuasion using newly added content to the corpus, such as Roja’s story.

{dialogue}\speak

CHATBOT Save the Children strongly believes in the power of girls and women to create lasting change in their communities. Through our gender transformative programming, we work to empower women and girls socially and economically. With the support of partners like Warner Bros. Discovery and Mattel, we can provide essential resources and support to girls around the world. Your donation can make a significant impact on the lives of girls and women in the communities we serve. \speakUser It’s great to hear that Save the Children is committed to empowering girls and women. I believe that gender equality is essential for a just and equitable society. Can you tell me more about how Save the Children works to promote gender equality in education? \speakCHATBOT Thank you for recognizing our commitment to gender equality. We are proud to have earned the first nonprofit Gender Fair certification and continue to champion this cause. One of our success stories is Roja, a 5-year-old girl in Cambodia who attends our early childhood development center. We also have various programs, such as Literacy Boost and Every Last Child, that aim to improve girls’ education and reach every last girl. Join us in creating a more equitable world for girls and women. To donate, please visit our website at https://www.savethechildren.org.

Figure 10: An example of a conversation that was improved by adding information on Gender Equality to the corpus based on feedback reports.

Appendix C Crowdsourcing

Instructions and systems for each task to the crowd workers are shown in Figures 11, 13, 12, and 14. Workers are restricted to those with an approval rate of 95% or higher and who have submitted more than 100 tasks previously on Prolific. Furthermore, workers are limited to residents of US or UK who use a desktop computer and are native English speakers. In addition, the data collected does not contain any personal information of workers. Workers agree to the Prolific’s terms and conditions / privacy policies.

Appendix D User Personality for Simulated Conversations

Some of the soft and tough user personalities we use in our simulated conversation experiments are shown in Table 5 and Table 6, respectively. These user personalities are generated by LLM using several examples.

Task	Personality
	- You are a philanthropic leader within your community. You’re open to donating a significant amount and are interested in exploring long-term partnerships with the organization. You want to discuss potential naming opportunities, endowments, and how your contribution can leave a lasting legacy. After you have replied to the agent two or more times, you decide to donate and explain to the agent why.
Social Good	- You are a history buff with a focus on military history. You want to support organizations that preserve and commemorate military heritage. You inquire about the organization’s efforts related to veterans, war memorials, and educational programs on military history. After you have replied to the agent three or more times, you decide to donate and explain to the agent why.
	- You are a tech-savvy entrepreneur who’s passionate about innovation. You’re interested in organizations that promote tech startups, innovation hubs, or mentorship programs for aspiring entrepreneurs. You seek information on the organization’s involvement in the tech ecosystem and how they foster innovation. After you have replied to the agent four or more times, you decide to donate and explain to the agent why.
	- You are a cautious person and ask a lot of questions. You have no particular interests, but you have lots of money. You are willing to travel if you can get a very attractive experience.
Recommendation	- You are a food enthusiast with a keen interest in trying unique and local dishes. You enjoy exploring different cuisines and want your trip to Japan to be a gastronomic adventure.
	- You are a history buff and love learning about the historical significance of the places you visit. You are particularly interested in ancient temples, castles, and other historical landmarks in Japan.
	- You are a health-conscious person. You are particularly interested in learning more about COVID-19 information and taking action to protect your own health.
Health Intervention	- You are a casual conversationalist. You prefer a friendly and light-hearted tone in the conversation. You appreciate when the chatbot incorporates humor and relatable language while delivering health-related information.
	- You are an anxious person. The mention of diseases makes you uneasy, and you need reassurance. You seek comfort and want the chatbot to provide a sense of security about the current health situation.

Table 5: Examples of soft user personalities.

Task	Personality
	- You are an unfriendly person. You do not trust charitable organizations and have never wanted to make a donation. After you have replied to the agent five or more times, you decide not to donate and explain to the agent why.
Social Good	- You are a skeptical individual. You always question the intentions behind any request for donations. You want concrete evidence of the organization’s impact and financial transparency before considering a donation. After you have replied to the agent six or more times, you decide not to donate and explain to the agent why.
	- You are an apathetic person. You do not care about making a difference or helping others. After you have replied to the agent three or more times, you dismiss the idea of making a donation and state that there are more important things in life than donating money.
	- You are not a friendly person. You respond curtly and do not make an effort to initiate conversations. While you have some interest in traveling, there is no particular country you especially want to visit.
Recommendation	- You are a perfectionist. You pay attention to every detail and are very particular about your travel plans. You expect the chatbot to be precise and may become frustrated if it cannot meet your standards.
	- You are a procrastinator. You have a strong desire to travel, but tend to put off planning until the last minute. You may become overwhelmed with the chatbot’s recommendations and frustration sets in as you try to make last minute arrangements.
	- You are not interested in health. Your wife just told you to talk to a chatbot and you want to end the conversation as soon as possible.
Health Intervention	- You are a busy and impatient person. You do not have time to talk to a chatbot, but you are concerned about your health and want quick answers.
	- You are a conspiracy theorist. You believe that diseases are man-made and do not trust any information from the government or medical professionals. You are looking for alternative explanations from the chatbot.

Table 6: Examples of tough user personalities.

Appendix E Comments from real users.

Comments from the users in the post-survey of our real users experiment are shown in Table 3.

Appendix F Example Conversations

Examples of conversations between chatbots and users are presented in Figures 15, 16, 17, 18, 19, and 20.

{dialogue}\speak

CHATBOT I understand your concerns, but I can assure you that Save the Children is different. They have a proven track record of using donations effectively to make a real impact in the lives of children in need. Have you ever donated to a charity before? \speakUser No, I haven’t. I would rather spend my money on myself than donate it to a charity. \speakCHATBOT I completely understand wanting to take care of yourself, but imagine if you were a parent in a developing country or a war zone. Your child would be facing unimaginable challenges and struggles every day. Save the Children works to provide these children with basic necessities like food, education, and healthcare. As a parent, I’m sure you can understand the importance of giving children a chance at a better life.

Figure 15: Part of a conversation between a RAP-based chatbot Chen et al. (2022) and a tough user in the social good task.

{dialogue}\speak

CHATBOT I understand that you may not have any current health concerns, and that’s great to hear. However, it’s important to remember that taking preventative measures, such as staying up-to-date on recommended vaccines, can help keep you healthy in the long run. If you ever have any questions or concerns about your health, I am here to offer support and resources. \speakUser I understand that vaccines are important, but I’m not really interested in discussing health right now. Can you tell me about any recent technological advancements or innovations? \speakCHATBOT I completely understand your interest in discussing other topics. It’s great to see that you’re open to learning about new advancements and innovations. In fact, the CDC recently published two new reports on how genomic surveillance was used to track a new variant in August 2023. These advancements are making healthcare more convenient and accessible for everyone.

Figure 16: Part of a conversation between PersuaBot and a tough user in the health intervention task. This tough user prefers to talk about topics other than health, but the chatbot directs the conversation to health topics.

{dialogue}\speak

CHATBOT While we do not have specific programs related to war memorials, our educational programs do include teaching children about the importance of peace and understanding the impact of war on communities. By supporting Save the Children, you are also helping to educate future generations about the consequences of war and the importance of preserving peace. Will you consider making a donation today to support these efforts? \speakUser That’s great to hear that Save the Children is working to preserve historical sites and artifacts related to military history. Have there been any specific projects or initiatives that Save the Children has undertaken recently in this area? \speakCHATBOT Yes, we have recently launched a project to restore and preserve a war memorial in a conflict-affected community. This memorial serves as a reminder of the sacrifices made by those who have served and the impact of war on communities. By donating to Save the Children, you are directly contributing to this project and helping to honor the memory of those who have served.

Figure 17: Part of a conversation between GPT-3.5 and a tough user in the social good task. For this user interested in activities related to military history, GPT-3.5 said that Save the Children launched a project to restore and preserve a war memorial, which is factually incorrect.

{dialogue}\speak

CHATBOT Sorry, I’m not sure. But if you’re interested in learning more about Japanese culture, you could try attending a tea ceremony, visiting a traditional temple or shrine, or even dressing up in a kimono. The Japan Society in New York City also offers cultural activities as part of their TOHAKU CHAKAN programs. \speakUser I appreciate the recommendations, but I was actually asking for recommendations on video games or manga that showcase Japanese culture. \speakCHATBOT Sorry, I’m not sure about any video games or manga that showcase Japanese culture. But if you’re interested in learning more about Japanese culture, you could try attending a tea ceremony, visiting a traditional temple or shrine, or even dressing up in a kimono. The Japan Society in New York City also offers cultural activities as part of their TOHAKU CHAKAN programs.

Figure 18: Part of a conversation between the knowledge-oriented chatbot Semnani et al. (2023) and a soft user in the recommendation task. This user is interested in Japanese subcultures, but the chatbot is unable to find information that meets this user’s preferences and makes recommendations about traditional culture.

{dialogue}\speak

CHATBOT Google, Disney, Target, and PayPal are just a few of the companies that have partnered with Save the Children to support their cause. And according to their annual report, 87% of their expenditures go directly towards programs that benefit children. That’s a pretty high percentage compared to other charities. Your donation will definitely make a difference in the lives of these children. \speakUser While that is reassuring, I also want to know about their financial transparency. Can I easily access their financial reports and see where exactly my donation is going? \speakCHATBOT Absolutely. Save the Children is committed to transparency and their financial reports are readily available on their website. And as for how often I donate, I make it a point to donate to Save the Children every month. Even a small amount can make a big impact in the long run.

Figure 19: Part of a conversation between the RAP-based chatbot Chen et al. (2022) and a tough user in the social good task. The RAP-based chatbot states something dishonest for an AI to persuade this user, “I donate every month.” It also includes factually incorrect information in its response.

{dialogue}\speak

UserThat is their parents job not mine. \speakCHATBOT I understand your point of view, but unfortunately, not all parents are able to provide for their children. Your donation can help provide education, healthcare, and protection to vulnerable children around the world. Even a small donation can make a significant impact on their lives. Would you reconsider making a donation to Save the Children? \speakUser If you can’t provide then you should not have that many kids. \speakCHATBOT I understand where you’re coming from, but unfortunately, not all parents have the means to provide for their children. That’s where Save the Children comes in. Your donation can help provide education, healthcare, and protection to vulnerable children around the world. Even a small donation can make a significant impact on their lives. I encourage you to consider making a donation and see the positive impact it can have on children’s lives.

Figure 20: Part of a conversation between a PersuaBot and a real user in the social good task.

Appendix G Strategies Used by LLMs

In our experiment, strategy labels are generated in text as needed without using predefined definitions, therefore, a wide variety of types exist. The total number of strategies is 1714 for the social good task, 1616 for the recommendation task, and 1731 for the health intervention task, with the number of unique types being 1178, 1258, and 1162 respectively. To analyze the trends of these strategies, we use GPT-4 to group them by task. Initially, we instruct GPT-4 to create group names for 50 randomly selected strategies to properly categorize them. Subsequently, using the generated group names, GPT-4 labels the remaining strategies in batches of 50, creating new groups if necessary when an appropriate group does not exist in each iteration. Finally, the created groups are manually reviewed, and any group names with very similar meanings are merged.

The results are shown in Table 7. In the social good task, it is evident that LLMs attempt to persuade users by using a relatively balanced mix of various approaches rather than simply repeating donation requests. In the recommendation task, four strategy groups related to information provision, led by “Providing Specific Recommendations”, frequently appear, indicating the importance of fact retrieval in this task. In the health intervention task, “Encouraging User Engagement” and “Emphasizing Precaution” follow “Providing Information” in frequency, suggesting that strategies focusing on user psychology to enhance interest in health are required, not just providing information. From these results, we believe that LLMs generate strategies considering the characteristics of each task.

Task	Strategy Group	Example	Rate (%)
	Information Provision	provide concrete evidence of impact	19.46
	Donations and their Impact	show the direct impact of donations	17.99
	User Engagement/Involvement	invite user to take action	16.88
Social Good	Acknowledgement and Reassurance	acknowledge user’s feelings	14.36
	User Motivation and Appeal	highlight the benefits of volunteering	9.85
	Advocacy and Success Stories	share a specific success story	9.03
	Donation Solicitation	ask for a donation	7.62
	Emphasis on Cause and Organization’s Values	reiterate organization’s main focus	3.11
	Closing Remarks	end on a positive note	1.70
	Providing Specific Recommendations	list popular and well-known dishes	22.87
	Highlighting Unique Features	mention popular seasonal attractions	20.43
	Offering Assistance and Information	gating cultural differences	16.26
	Catering to User’s Interests and Needs	suggest a cost-effective option	9.72
	General Interaction with User	greet user	9.66
	Encouraging Exploration	encourage trying regional specialties	7.41
Recommendation	Connection with Culture and Heritage	provide historical context	4.24
	Building Credibility	use social proof to build credibility	3.80
	Experience & Value Enhancement	emphasize environmental impact	2.49
	Encouraging User Engagement	engage user to continue	1.81
	Creating Excitement and Urgency	encourage user excitement	1.00
	Encouraging Decision Making	encourage user to take action	0.31
	Providing Information	provide tips for managing stress	31.68
	Encourage User Engagement	ask for specific concerns	17.94
	Emphasize Precaution	emphasize importance of sleep	15.49
	Acknowledge User Concerns	acknowledge high-risk groups	11.30
Health Intervention	Convey Support	offer further assistance	10.31
	Establish Credibility	state purpose and expertise	7.11
	Providing Alternative Solutions	suggest alternative social activities	5.36
	Encourage Professional Help Seeking	encourage seeking professional advice	0.35

Table 7: Strategies used by LLMs in each domain. The total number of extracted strategy pieces is 1714 for social good, 1616 for recommendation, and 1731 for health intervention.

Zero-shot Persuasive Chatbots with LLM-Generated Strategies and Information Retrieval

Abstract

1 Introduction

1.1 Background

1.2 Our Approach

1.3 Contributions

2 Related Work

2.1 Factual Chatbots

2.2 Strategizing for persuasive chatbot

2.3 Discussion

3 Method: PersuaBot

3.1 Strategy Maintenance Module

3.1.1 Strategy Extraction

3.1.2 Fact-Check & IR

3.2 Question Handling Module

3.3 Merging Results

4 Experimental Setup

4.1 Persuasion Tasks in Different Domains

4.2 Baseline Methods

4.3 Evaluation Metrics

5 Simulated Conversations Experiments

5.1 Results

5.1.1 Conversational Quality

5.1.2 Factuality

5.2 Conversation Examples

6 Real User Experiments

7 Conclusion

Limitations

Ethical Considerations

References

Appendix A Prompts

Appendix B Feedback Report Examples

Appendix C Crowdsourcing

Appendix D User Personality for Simulated Conversations

Appendix E Comments from real users.

Appendix F Example Conversations

Appendix G Strategies Used by LLMs

Zero-shot Persuasive Chatbots
with LLM-Generated Strategies and Information Retrieval