Generating Diverse Personas for User Simulators to Test Interview Dialogue Systems

# Generating Diverse Personas for User Simulators to Test Interview Dialogue Systems
Source: [https://arxiv.org/html/2608.19549](https://arxiv.org/html/2608.19549)
Mikio NakanoAffiliation:C4A Research Institute, Inc\., Setagaya, Tokyo, JapanAffiliation:SANKEN, University of Osaka, Ibaraki, Osaka, JapanHironori TakeuchiAffiliation:Musashi University, Nerima, Tokyo, Japanmikio\.nakano@c4a\.jp, komatani@sanken\.osaka\-u\.ac\.jph\.takeuchi@cc\.musashi\.ac\.jp

###### Abstract

This paper addresses the issue of the significant labor required to test interview dialogue systems\. While interview dialogue systems are expected to be useful in various scenarios, like other dialogue systems, testing them with human users requires significant effort and cost\. Therefore, testing with user simulators can be beneficial\. Since most conventional user simulators have been primarily designed for training task\-oriented dialogue systems, little attention has been paid to the personas of the simulated users\. During development, testing interview dialogue systems requires simulating a wide range of user behaviors, but manually creating a large number of personas is labor\-intensive\. We propose a method that automatically generates personas for user simulators using a large language model\. Furthermore, by assigning personality traits related to communication styles when generating personas, we aim to increase the diversity of communication styles in the user simulator\. Experimental results show that the proposed method enables the user simulator to generate utterances with greater variation\.

## 1Introduction

Interview dialogue systems, which can efficiently gather information from many users, are a promising application of dialogue system technology and have attracted increasing research interest in recent years[26](https://arxiv.org/html/2608.19549#bib.bib29);[17](https://arxiv.org/html/2608.19549#bib.bib25);[52](https://arxiv.org/html/2608.19549#bib.bib24)\. However, building dialogue systems, including interview dialogue systems, incurs significant costs\. One major cost is the effort required for testing\. The goal of our study is to alleviate the manual effort for testing interview dialogue systems\.

To reduce testing costs,user simulationcan be used instead of manual testing\. However, existing user simulators have primarily been developed for the training and evaluation of task\-oriented dialogue systems\. Thus, we aim to develop a simulator that is useful for testing interview dialogue systems\.

When testing interview dialogue systems, simulating a wide variety of users is crucial\. To increase variation, it is beneficial to have a large number ofuser personas[53](https://arxiv.org/html/2608.19549#bib.bib34);[22](https://arxiv.org/html/2608.19549#bib.bib35);[18](https://arxiv.org/html/2608.19549#bib.bib30)\. However, manually creating numerous personas is labor\-intensive\. While using a large set of pre\-constructed personas has been considered[31](https://arxiv.org/html/2608.19549#bib.bib38), this approach is unsuitable for interview dialogues, where personas need to be aligned with the specific interview topics\.

Therefore, this paper proposes a method to generate diverse personas usinglarge language models\(LLMs\)\. In this method, a few manually created personas are provided as examples, and the LLM is tasked with generating a large number of additional personas\. To ensure diversity in speaking styles and attitudes toward the system, personality traits are specified during generation\.

Evaluation experiments demonstrated that the proposed method enables the creation of diverse dialogues through the generation of many personas\. This increases the likelihood of discovering system issues\. Note that detecting system issues from simulated dialogues is beyond the scope of our study\. Manual inspection is one possible approach, but automated methods using LLMs are also conceivable[9](https://arxiv.org/html/2608.19549#bib.bib50)\.

The contributions of our study are as follows:

- •We discuss that diverse personas are necessary for user simulators used in testing interview dialogue systems\.
- •We propose a method for generating diverse personas using LLMs\.
- •We introduce a technique to enhance persona diversity by specifying personality traits related to communication style\.
- •Through evaluation experiments, we demonstrate that the proposed persona generation method enables the creation of diverse dialogues\.

## 2Related Work

### 2\.1Interview Dialogue Systems

Interview dialogue systems that extract information from humans through natural language interaction have attracted significant research attention due to their high practical value\. Examples of the application domains include: course rating surveys at universities[46](https://arxiv.org/html/2608.19549#bib.bib17), telephone surveys[23](https://arxiv.org/html/2608.19549#bib.bib18);[26](https://arxiv.org/html/2608.19549#bib.bib29), mental health assessments[6](https://arxiv.org/html/2608.19549#bib.bib19), dietary intake recording[25](https://arxiv.org/html/2608.19549#bib.bib20), job interviews[47](https://arxiv.org/html/2608.19549#bib.bib21);[39](https://arxiv.org/html/2608.19549#bib.bib22);[21](https://arxiv.org/html/2608.19549#bib.bib23);[24](https://arxiv.org/html/2608.19549#bib.bib31);[19](https://arxiv.org/html/2608.19549#bib.bib32);[12](https://arxiv.org/html/2608.19549#bib.bib36), dietary preference surveys[52](https://arxiv.org/html/2608.19549#bib.bib24), career counseling for nurses[17](https://arxiv.org/html/2608.19549#bib.bib25), frailty diagnosis[3](https://arxiv.org/html/2608.19549#bib.bib26), collecting people’s beliefs about the future of robots[45](https://arxiv.org/html/2608.19549#bib.bib28), facilitating user review writing[50](https://arxiv.org/html/2608.19549#bib.bib53), and customer interviews[44](https://arxiv.org/html/2608.19549#bib.bib37)\. We propose a method for efficiently testing such interview dialogue systems\.

### 2\.2Testing Dialogue Systems

As mentioned earlier, testing dialogue systems requires significant effort, and various methods and tools have been proposed to facilitate this process[27](https://arxiv.org/html/2608.19549#bib.bib49)\. For example, there are tools that verify whether a dialogue system behaves as expected using corpora[5](https://arxiv.org/html/2608.19549#bib.bib1), and methods that test the system using predefined test cases[4](https://arxiv.org/html/2608.19549#bib.bib2);[16](https://arxiv.org/html/2608.19549#bib.bib3);[15](https://arxiv.org/html/2608.19549#bib.bib48)\. Since these testing methods require the preparation of test cases in advance, they are effective for detecting issues within the expected scope\. However, in interactions with diverse users, unexpected problems may arise, and thus, methods capable of detecting such unforeseen issues are desired\.

### 2\.3User Simulation for Dialogue Systems

We therefore use user simulators\. User simulators for dialogue systems have traditionally been used to evaluate dialogue strategies[8](https://arxiv.org/html/2608.19549#bib.bib52);[35](https://arxiv.org/html/2608.19549#bib.bib39)\. However, with advances in statistical modeling of dialogue strategies, research has increasingly focused on training dialogue strategy models using reinforcement learning[42](https://arxiv.org/html/2608.19549#bib.bib13);[36](https://arxiv.org/html/2608.19549#bib.bib16);[41](https://arxiv.org/html/2608.19549#bib.bib12);[1](https://arxiv.org/html/2608.19549#bib.bib27);[28](https://arxiv.org/html/2608.19549#bib.bib15);[37](https://arxiv.org/html/2608.19549#bib.bib14)\.

Recently, with the development of large language models \(LLMs\), user simulators based on LLMs have been proposed[51](https://arxiv.org/html/2608.19549#bib.bib47);[30](https://arxiv.org/html/2608.19549#bib.bib5);[2](https://arxiv.org/html/2608.19549#bib.bib6);[43](https://arxiv.org/html/2608.19549#bib.bib7);[20](https://arxiv.org/html/2608.19549#bib.bib10);[48](https://arxiv.org/html/2608.19549#bib.bib8);[7](https://arxiv.org/html/2608.19549#bib.bib9);[29](https://arxiv.org/html/2608.19549#bib.bib11)\.

In studies on user simulators for task\-oriented dialogue systems, evaluation metrics such as similarity to human user behavior and the quality of the learned dialogue strategies have been used[37](https://arxiv.org/html/2608.19549#bib.bib14)\.

Our study differs from these previous studies in that it proposes a method for constructing a user simulator specifically for testing interview dialogue systems\. Different evaluation metrics are also required\.

### 2\.4Using Personas in User Simulation

Research has also been conducted on using personas in user simulators\.[13](https://arxiv.org/html/2608.19549#bib.bib41)and[14](https://arxiv.org/html/2608.19549#bib.bib40)demonstrated that different dialogue strategies can be learned by preparing simulators for older and younger users\.[17](https://arxiv.org/html/2608.19549#bib.bib25)manually created nurse personas for a user simulator used in evaluating a career counseling dialogue system for nurses, under the supervision of nursing administrators\. Our study differs from these studies in that it automatically generates personas\.

### 2\.5Persona Generation

As for research on automatic persona generation, there are methods that extract personas from Reddit[31](https://arxiv.org/html/2608.19549#bib.bib38)\. They collect sentences that express a persona \(e\.g\., those containing "I" or "my"\) and use them for persona creation\. Methods for generating personas using LLMs have also been proposed\.[11](https://arxiv.org/html/2608.19549#bib.bib33)proposed a method for generating persona\-representing passages from text using an LLM\. However, with these methods, it is difficult to generate interviewee personas tailored to the domain of interview dialogue systems\.

## 3Proposed Method

In the proposed method, we generate simulated user personas using an LLM\. It is necessary to ensure that the generated personas include information relevant to the domain of the target interview dialogue system\. To achieve this, we perform in\-context learning using a small number of manually written personas \(calledseed personashereafter\) as few\-shot examples\. Furthermore, to facilitate the discovery of potential issues by introducing greater variation in user utterances during simulation, we provide personality traits that define communication styles when generating personas with the LLM\. By embedding the generated personas and personality traits into the prompt of an LLM\-based user simulator, we can generate diverse dialogues\. Figure[1](https://arxiv.org/html/2608.19549#S3.F1)illustrates the proposed method\.

Figure 1:Overview of the proposed method\.We use the following two axes as personality traits\.

#### Degree of Anthropomorphism

One axis concerns whether the user treats the system as an object or as a human\. Users who treat the system as an object aim to use the dialogue system efficiently and tend to use expressions that are easy for the system to understand\. In contrast, users who treat the system as a human speak to it as if they were speaking to a person, without considering whether the system can understand them\. This behavior can be interpreted as users anthropomorphizing the system[40](https://arxiv.org/html/2608.19549#bib.bib45)\. We refer to this axis as the degree of anthropomorphism\.

#### Degree of Elaborateness

The other axis concerns whether the user engages in redundant communication or direct communication\.[38](https://arxiv.org/html/2608.19549#bib.bib46)and[33](https://arxiv.org/html/2608.19549#bib.bib44)classify users’ communication styles into elaborateness and directness\. We adopt the same classification in our study\.

\- Whether the user often travels\- Places the user has recently visited\- Favorite places among the user’s past travel destinations\- Activities the user enjoys doing while traveling\- Reasons for not traveling often \(if applicable\)\- Places the user would like to visit next\-How the user books trips \(e\.g\., visiting a travel agency,calling a travel agency, using a website\)Figure 2:Topics the travel interview dialogue system asks the user about\.```
- Name: Risako Machiyama
- 49 years old
- Female
- Lives in Tokyo
- Lives with husband and two daughters (a senior
  high school second-year student and a junior
  high school third-year student)
- Works part-time in an office
- Speaks cheerfully and talks rapidly with
  many words
- Travels about once a year
- Most recently visited Taiwan with her family
- Enjoys shopping while traveling
- Would like to visit Los Angeles
- Usually books trips by visiting a travel
  agency
```

```
- Name: Yuma Yamanaka
- 32 years old
- Male
- Lives in Hokkaido
- Lives alone
- Chef
- Speaks clearly and often talks about himself
- Travels about once a year
- Most recently visited Osaka alone
- Enjoys food tours during travel for
  professional inspiration
- Would like to visit Singapore
- Usually books trips online
```

Figure 3:Examples of the seed personas for the travel interview dialogue system\.```
# Task

- Create 25 personas for interviewees being
  surveyed about their travel experiences
  and desired travel destinations, generating
  them in JSON format as shown in the example.
- Ensure as much variation among the personas
  as possible.
- The interviewees have the following
  personality traits; create personas
  consistent with these traits.
- Do not reuse any of the personas shown
  in the example.

# Personality Traits

{personality}

# Examples

{examples}
```

Figure 4:Prompt template for persona generation\.\{personality\}is replaced with a description of personality traits, and\{examples\}is replaced with the seed personas converted into JSON format\. In the noPT condition, the template without the personality trait description is used\.APM: High

```
- Views the system as a human rather than an
  object
- Speaks to the system as if speaking to a person
- Tries to test the system, abruptly changing
  topics or speaking at length
```

APM: Low

```
- Views the system as an object rather than a
  human
- Aware of the system’s limitations in
  understanding, speaks as simply as possible
- Shows a cooperative attitude toward the system
```

Figure 5:Personality trait descriptions used in the prompt for persona generation by varying the degree of anthropomorphism\.EL: High

```
- Speaks redundantly
- Talks about things not explicitly asked
- Sometimes uses indirect expressions
```

EL: Low

```
- Speaks directly
- Speaks concisely
- Does not talk about unasked topics
```

Figure 6:Personality trait descriptions used in the prompt for persona generation by varying the degree of elaborateness\.Although the Big Five personality traits are well\-known, we did not use them because it is considered more effective to directly specify users’ speaking styles is considered more effective for our purpose\.

## 4Evaluation

We conducted an evaluation experiment to investigate whether the proposed method is effective for testing interview dialogue systems\. Note that this experiment deals only with Japanese systems, and that all subsequent figures are translations from Japanese\.

### 4\.1Compared Methods

We compared the following five conditions:

#### BL

\(BaseLine\): Only seed personas are used without generating new personas\.

#### noPT

\(no Personality Traits\): Proposed method without prompts concerning the two kinds of personality traits\. Only seed personas are given when personas are generated\.

#### APM

\(AnthroPoMorphism\): Proposed method with degree of anthropomorphism given as a personality trait for persona generation\. The degree is either High or Low\.

#### EL

\(ELaborateness\): Proposed method with degree of elaborateness given as a personality trait for persona generation\. The degree is either High or Low\.

#### APM\+EL

Proposed method with both degree of anthropomorphism and degree of elaborateness given as personality traits for persona generation\.

### 4\.2Interview Dialogue Systems Used

We used the following two systems:

The first is a Japanese text\-based interview dialogue system that conducts interviews about travel\. It was built using the ChatGPT dialogue built\-in block of DialBB[34](https://arxiv.org/html/2608.19549#bib.bib4), where dialogues are conducted based on a single prompt template\. The interview topics extracted from users are shown in Figure[2](https://arxiv.org/html/2608.19549#S3.F2)\.

The second is a Japanese text\-based interview dialogue system that conducts interviews focusing on the user’s preferences for sweets\. It asks users questions such as whether they often eat sweets, what kinds of sweets they like, and where they usually buy sweets\. This system was also built using DialBB, employing ChatGPT for language understanding, named entity extraction, and dialogue management using a state transition network\. ChatGPT is used for evaluating transition conditions and for utterance generation within dialogue management\. The state transition network consists of 32 states and 54 transitions\.

```
# Task Description

- You are acting as a user of a dialogue system
  and are being interviewed by the system. Based
  on the flow of the conversation so far,
  generate your next utterance.

# Notes

- The utterance should consist of either 1, 2,
  or 3 sentences.
- Do not produce utterances of the same length
  consecutively.
- Each utterance must be within 100 characters.
- Speak in accordance with your assigned
  persona below.
- Do not include your name or the word "User"
  at the beginning of the utterance.
- Answer if you are asked for your name.
- Do not enclose the utterance in quotation
  marks.

# Your Persona

{persona}

# Dialogue History

{dialogue_history}
```

Figure 7:Prompt template used by the simulator\.\{persona\}is replaced with the generated persona, and\{dialogue\_history\}is replaced with the dialogue history at runtime\.
### 4\.3Procedure

We conducted persona generation and simulation according to the following procedure\.

First, for each system, we manually created 10 seed personas\. Figure[3](https://arxiv.org/html/2608.19549#S3.F3)shows examples of seed personas for the travel system\. In this paper, we show only examples and prompt templates for the travel interview system for the lack of space\. In the noPT, APM, EL, and APM\+EL conditions, these seed personas were embedded into the prompt as few\-shot examples to generate new personas\. In the APM, EL, and APM\+EL conditions, personality traits were assigned in a balanced manner, and personas consistent with those traits were generated\. The template used for persona generation for the travel system is shown in Figure[4](https://arxiv.org/html/2608.19549#S3.F4), and the descriptions of the personality traits used in the experiments are shown in Figures[5](https://arxiv.org/html/2608.19549#S3.F5)and[6](https://arxiv.org/html/2608.19549#S3.F6)\. The third item under "APM: High" in Figure[5](https://arxiv.org/html/2608.19549#S3.F5)indicates that users who anthropomorphize the system attempt to test whether the system can engage in conversation as flexibly as a human\. For each condition, 100 personas were generated\.111Personas were generated in batches of about 25 at a time, and when the total exceeded 100, 100 personas were randomly selected\.

The interview dialogue systems interacted with user simulators based on the generated personas\. The prompt template used for simulation is shown in Figure[7](https://arxiv.org/html/2608.19549#S4.F7)\. Each dialogue session consisted of 15 user utterances\. For conditions other than BL, each persona was used for only one dialogue session, resulting in 100 dialogues per condition\. For BL, each of the 10 seed personas was used 10 times\.

The LLM used by the persona generator and the simulator was OpenAI’s gpt\-4o\-2024\-11\-20,222[https://platform\.openai\.com/docs/models/gpt\-4o](https://platform.openai.com/docs/models/gpt-4o)while the LLM used by the interview dialogue systems was OpenAI’s gpt\-4o\-mini\-2024\-07\-18\.333[https://platform\.openai\.com/docs/models/gpt\-4o\-mini](https://platform.openai.com/docs/models/gpt-4o-mini)Since our objective is to evaluate the simulator, we used a high\-performance model for the simulator\. The temperature parameter was set to 0\.7 for all LLM usages\.

Prompt tuning was conducted separately from the main experiments, using a different system \(a sweets interview dialogue system that uses a single prompt template\) and different seed personas from those used in the main evaluation described in Section[4\.2](https://arxiv.org/html/2608.19549#S4.SS2)\.

Table 1:Diversity metrics of the simulations for the travel interview system\. For example, the row where Condition is APM and Personality is All represents metrics calculated from all the dialogues under the APM condition\. The subsequent row, where Personality is High, shows metrics calculated from the dialogues under the APM condition using a high degree of anthropomorphism in the personality setting\. The row where Condition is APM\+EL and Personality is High\+High represents metrics calculated from the dialogues with a high degree of anthropomorphism and a high degree of elaborateness\. Bold numbers are mentioned in the main text\.
Table 2:Diversity metrics of the simulations for the sweets interview system\.
### 4\.4Evaluation Metrics

Our goal is to automatically expose issues in interview dialogue systems\. One direct way to evaluate the proposed method would be to build a faulty interview dialogue system, interact with it using the user simulator generated by the proposed method, and measure how many issues can be detected\. However, faulty systems may cause dialogues to collapse once a problem occurs\. In practice, using the user simulator to improve a system would involve repeatedly fixing detected issues and running new simulations, but evaluating this iterative process is impractical\.

Therefore, as a second\-best approach, we measure the diversity of user simulator utterances\. The rationale is that the more diverse the utterances, the higher the probability of uncovering system issues\.

Following previous studies on user simulators for task\-oriented dialogue systems[51](https://arxiv.org/html/2608.19549#bib.bib47);[2](https://arxiv.org/html/2608.19549#bib.bib6);[43](https://arxiv.org/html/2608.19549#bib.bib7), we used the following metrics to evaluate utterance diversity\. For Japanese word tokenization, we used Sudachi[49](https://arxiv.org/html/2608.19549#bib.bib43)in C mode \(a mode that does not split compound words\)\.

#### Average and S\.D\. of utterance lengths

The average and standard deviation of the number of words per user utterance across all dialogues\.

#### Total words

The total number of words across all user utterances in all dialogues\.

#### Unique words

The number of unique words across all user utterances in all dialogues\.

#### Unique bigrams

The number of unique bigrams across all user utterances in all dialogues\.

#### TTR

Type\-Token Ratio: \(\(\# of unique words\) / \(\# of total words\)\) for all user utterances in all dialogues\.

#### Unique CW

The number of Unique Content Words appearing in all user utterances\.

#### CW\-TTR

The type\-token ratio for content words across all user utterances\.

#### SE

Shannon Entropy calculated from all user utterances across all dialogues\.

#### CE

Conditional bigram Entropy calculated from all user utterances across all dialogues\.

#### MTLD

Measure of Textual Lexical Diversity[32](https://arxiv.org/html/2608.19549#bib.bib42)calculated by concatenating all user utterances \(threshold: 0\.72\)\.

#### MSTTR

Mean Segmental Type\-Token Ratio[32](https://arxiv.org/html/2608.19549#bib.bib42)calculated by concatenating all user utterances\.

#### Average and S\.D\. of TTR

The average and standard deviation of type\-token ratios calculated per dialogue\.

### 4\.5Results

Tables[1](https://arxiv.org/html/2608.19549#S4.T1)and[2](https://arxiv.org/html/2608.19549#S4.T2)respectively show the diversity metrics for the travel interview system and the sweets interview system\. Results for each personality trait condition are also presented\. These results suggest the following two points: \(1\) noPT exhibits greater content variation compared to BL\. In other words, persona generation using LLMs can increase the variety of content in utterances\. \(2\) EL shows greater stylistic variation than noPT\. That is, by varying redundancy levels, one can increase stylistic diversity in utterances\.

\(1\) is suggested by the following findings: Comparing BL and noPT, especially in the travel domain, CW\-TTR increases from\.106\.106to\.122\.122in the travel domain, and in the sweets domain, from\.109\.109to\.133\.133\. This suggests that the generated personas contain different content words from the seed personas\. This may help uncover issues caused by diverse utterance content\. Although SE, CE, and MTLD also show slight improvements, the differences are not significant\. TTR slightly decreases in the sweets domain system, possibly due to limited stylistic changes, leading to low variation in function words\.

\(2\) is suggested by the following findings: Compared to noPT, EL \(All\) shows an increase in the standard deviation of utterance length—from7\.07\.0to18\.218\.2in the travel domain and from8\.08\.0to17\.717\.7in the sweets domain\. This suggests that varying the degrees of elaborateness enables the generation of utterances with diverse lengths, potentially exposing problems caused by such diversity\. TTR and CW\-TTR are lower, likely because elaborate utterances under the EL High condition tend to include fixed phrases\.

There is little difference between APM and noPT\. Also, comparing APM Low and High settings reveals no significant variation\. Furthermore, APM\+EL does not differ much from EL\. Even when examining the dialogues, major differences were not observed\. This suggests that the personality traits described in APM may not be sufficient to alter dialogue style\. Since personas generated in noPT already contain variation in personality, APM might not have introduced additional diversity beyond that\.

Note that in settings like EL High, where personality traits are narrowly defined, TTR and CW\-TTR increase\. This is likely because the number of dialogues is small, reducing word repetition\.

noPT:

```
- Name: Erina Takahashi
- 29 years old
- Female
- Lives in Kyoto Prefecture
- Lives alone
- Works for an IT company
- Reserved, but talkative with close friends
- Takes several trips a year
- Most recently visited Kanazawa; traveled
  alone
- Enjoys visiting local art museums and cafes
  while traveling
- Would like to visit Paris, France
- Books flights and hotels online
```

Figure 8:Examples of generated personas \(1\)\.EL High:

```
- Name: Ken Nakamura
- 34 years old
- Male
- Live in Hiroshima Prefecture
- Lives with his wife and 2-year-old son
- Public servant
- Speaks gently and is family-oriented
- Travels less frequently since having a child,
  but still goes on one trip per year
- Most recently visited Kagawa; a family trip
  focused on touring udon restaurants
- Chooses travel destinations that can be
  enjoyed by the whole family
- Would like to visit Hokkaido
- Plans trips together with his wife
```

EL Low:

```
- Name: Yusuke Ogawa
- 45 years old
- Male
- Lives in Hiroshima Prefecture
- Lives with his wife and two children (a son
  in 8th grade and a daughter in 5th grade)
- Public servant
- Speaks calmly and avoids unnecessary talk
- Goes on one trip per year
- Most recently visited Kyoto on a family trip
- Visits temples and shrines during travels
- Would like to visit Nara
- Makes travel reservations by calling a travel
  agency
```

Figure 9:Examples of generated personas \(2\)\.APM High:

```
- Name: Kenta Omura
- 24 years old
- Male
- Lives in Osaka Prefecture
- Lives alone
- Graduate student
- Quiet and not very talkative, but polite
- Takes about one trip per year
- Most recently visited Nara; did some
  sightseeing after attending an academic
  conference
- Enjoys exploring nature at travel
  destinations
- Would like to visit the Rocky Mountains in
  Canada
- Books flights online by himself
```

APM Low:

```
- Name: Naoto Yamashita
- 33 years old
- Male
- Lives in Hiroshima Prefecture
- Lives alone
- Designer
- Speaks calmly and politely
- Takes two trips a year
- Most recently visited Kyoto alone
- Enjoys visiting architectural sites and art
  museums while traveling
- Would like to visit Paris
- Plans and books trips meticulously online
```

Figure 10:Examples of generated personas \(3\)\.
### 4\.6Generated Personas

Examples of the generated personas are shown in Figures[8](https://arxiv.org/html/2608.19549#S4.F8),[9](https://arxiv.org/html/2608.19549#S4.F9), and[10](https://arxiv.org/html/2608.19549#S4.F10)\. These examples were randomly selected from the personas generated under each condition\. Figures[9](https://arxiv.org/html/2608.19549#S4.F9)and[10](https://arxiv.org/html/2608.19549#S4.F10)feature only male personas, but this is purely coincidental\. While the male personas shown speak in a reserved manner, some of the other generated male personas are sociable and talkative\. From these examples, it can be inferred that the generated personas are consistent with their respective personality traits\. The APM\+EL examples are omitted, as they did not yield any notable results with respect to the diversity metrics\. Additional examples of generated personas are shown in Figures[17](https://arxiv.org/html/2608.19549#A2.F17)and[18](https://arxiv.org/html/2608.19549#A2.F18)in the Appendix\.

## 5Discussion

As discussed above, the experimental results suggest that generating personas with an LLM enables the generation of dialogues with greater variation\. Although the quantitative metrics do not indicate this very clearly, you can see the variation by looking at the generated dialogues\. This could lead to the discovery of more problems in dialogue systems\. Furthermore, specifying personality traits related to communication style can introduce greater variation in utterance lengths, which may help uncover additional problems\.

In the current user simulator prompt we used, relatively similar utterances tend to be generated consecutively\. This might be different from the behaviors of human users\. It is possible that generating various types of utterances within a single dialogue could help reveal different types of issues\. In future work, we will explore methods for increasing such variation and investigate to what extent the behaviors of the user simulator cover those of human users\.

Although our study focused on APM and EL traits, there may be other personality traits that could introduce further variation\. Identifying and incorporating such traits will be considered in future research\.

The ultimate goal of our study is to identify as many issues in interview dialogue systems as possible\. However, in this work, we only measured the diversity of simulated users\. Whether the proposed method can help discover issues across various dialogue systems remains a topic for future work\. In the future, we aim to integrate this method into tools for building interview dialogue systems and validate it through its application in the development of diverse systems\.

Note that, although the systems used in our evaluation experiments did not exhibit obvious issues such as completely incoherent utterances, we found that under the El High condition, simulated users tended to engage in extended small talk\. This often caused the system to fail in extracting the necessary information, revealing a potential issue\.

## 6Concluding Remarks

This paper proposed a persona generation method for user simulation using large language models \(LLMs\)\. We also introduced a method to incorporate personality traits related to communication style during persona generation\. This enables diverse testing of interview dialogue systems without human involvement, significantly reducing development costs\. Although the evaluation experiments were limited and some issues remain, the results suggest the effectiveness of the proposed method\. Therefore, we believe it is worthwhile to share our proposed method and experimental findings\.

In the future, we plan to develop user simulators for interview dialogue systems with speech input/output and multimodal input/output\. We also aim to extend our approach to build user simulators for various types of dialogue systems beyond interview dialogues\.

## Acknowledgments

This work was partially inspired by the collaboration between Ekai Hashimoto, Takayoshi Sakurai, and Shun Shiramatsu of Nagoya Institute of Technology, Toshitake Komazaki of Tokyo Healthcare University, Shiho Tsuchiya of Kitasato University Hospital, and the first author\. We sincerely thank them\.

## Limitations

In addition to what was discussed in Section[5](https://arxiv.org/html/2608.19549#S5), there are several other limitations in our evaluation experiments\. First, we used only a single model as the LLM for both persona generation and the simulator\. Second, we used only Japanese systems\. Furthermore, the number of generated personas per condition was fixed at 100, the number of seed personas was set to 10 for both systems, and the temperature parameter was set to 0\.7 throughout the experiment; other configurations have not been tested\. The impact of varying these conditions remains unexamined\.

In future work, we plan to address these issues as we apply the proposed method to the development of various interview dialogue systems\.

## Ethical Considerations

In our study, human participants were not involved in system testing; therefore, there is no risk of collecting or leaking personal information from individuals\.

The biases potentially inherent in LLMs[10](https://arxiv.org/html/2608.19549#bib.bib51)may be reflected in automatically generated personas or user utterances\. As a result, certain types of personas or utterances may be excluded from testing, which could hinder the identification of potential issues\. Verifying whether this problem occurs and addressing it if necessary remain future challenges\.

## References

- Ai and Litman \(2009\)H\. Ai and D\. LitmanSetting up user action probabilities in user simulations for dialog system development\.InProceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP,K\. Su, J\. Su, J\. Wiebe, and H\. Li \(Eds\.\),Suntec, Singapore,pp\. 888–896\.External Links:[Link](https://aclanthology.org/P09-1100/)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p1.1)\.
- Algherairy and Ahmed \(2025\)A\. Algherairy and M\. AhmedPrompting large language models for user simulation in task\-oriented dialogue systems\.Computer Speech & Language89,pp\. 101697\.External Links:ISSN 0885\-2308,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.csl.2024.101697),[Link](https://www.sciencedirect.com/science/article/pii/S0885230824000809)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p2.1),[§4\.4](https://arxiv.org/html/2608.19549#S4.SS4.p3.1)\.
- Asaoet al\.\(2020\)Y\. Asao, J\. Kloetzer, J\. Mizuno, D\. Saiki, K\. Kadowaki, and K\. TorisawaUnderstanding user utterances in a dialog system for caregiving\.InProceedings of the Twelfth Language Resources and Evaluation Conference,N\. Calzolari, F\. Béchet, P\. Blache, K\. Choukri, C\. Cieri, T\. Declerck, S\. Goggi, H\. Isahara, B\. Maegaard, J\. Mariani, H\. Mazo, A\. Moreno, J\. Odijk, and S\. Piperidis \(Eds\.\),Marseille, France,pp\. 653–661\(eng\)\.External Links:[Link](https://aclanthology.org/2020.lrec-1.82/),ISBN 979\-10\-95546\-34\-4Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Atefi and Alipour \(2019\)S\. Atefi and M\. A\. AlipourAn automated testing framework for conversational agents\.External Links:1902\.06193,[Link](https://arxiv.org/abs/1902.06193)Cited by:[§2\.2](https://arxiv.org/html/2608.19549#S2.SS2.p1.1)\.
- Degerstedt and Jönsson \(2006\)L\. Degerstedt and A\. JönssonLINTest: a development tool for testing dialogue systems\.InProceedings of Interspeech 2006,pp\. 225–235\.External Links:[Link](https://www.isca-archive.org/interspeech_2006/degerstedt06_interspeech.html#),[Document](https://dx.doi.org/10.21437/Interspeech.2006-154)Cited by:[§2\.2](https://arxiv.org/html/2608.19549#S2.SS2.p1.1)\.
- DeVaultet al\.\(2014\)D\. DeVault, R\. Artstein, G\. Benn, T\. Dey, E\. Fast, A\. Gainer, K\. Georgila, J\. Gratch, A\. Hartholt, M\. Lhommet, G\. Lucas, S\. Marsella, F\. Morbini, A\. Nazarian, S\. Scherer, G\. Stratou, A\. Suri, D\. Traum, R\. Wood, Y\. Xu, A\. Rizzo, and L\. MorencySimSensei kiosk: a virtual human interviewer for healthcare decision support\.InProceedings of the 2014 International Conference on Autonomous Agents and Multi\-Agent Systems,AAMAS ’14,Richland, SC,pp\. 1061–1068\.External Links:ISBN 9781450327381,[Link](https://dl.acm.org/doi/10.5555/2615731.2617415)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Di Brattoet al\.\(2024\)M\. Di Bratto, A\. Origlia, M\. Di Maro, and S\. MennellaLinguistics\-based dialogue simulations to evaluate argumentative conversational recommender systems\.User Modeling and User\-Adapted Interaction34\(5\),pp\. 1581–1611\.External Links:[Link](https://link.springer.com/article/10.1007/s11257-024-09403-3)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p2.1)\.
- Eckertet al\.\(1997\)W\. Eckert, E\. Levin, and R\. PieracciniUser modeling for spoken dialogue system evaluation\.In1997 IEEE Workshop on Automatic Speech Recognition and Understanding Proceedings,Vol\.,pp\. 80–87\.External Links:[Document](https://dx.doi.org/10.1109/ASRU.1997.658991)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p1.1)\.
- Finchet al\.\(2023\)S\. E\. Finch, E\. S\. Paek, and J\. D\. ChoiLeveraging large language models for automated dialogue analysis\.InProceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue,S\. Stoyanchev, S\. Joty, D\. Schlangen, O\. Dusek, C\. Kennington, and M\. Alikhani \(Eds\.\),Prague, Czechia,pp\. 202–215\.External Links:[Link](https://aclanthology.org/2023.sigdial-1.20/),[Document](https://dx.doi.org/10.18653/v1/2023.sigdial-1.20)Cited by:[§1](https://arxiv.org/html/2608.19549#S1.p5.1)\.
- Gallegoset al\.\(2024\)I\. O\. Gallegos, R\. A\. Rossi, J\. Barrow, M\. M\. Tanjim, S\. Kim, F\. Dernoncourt, T\. Yu, R\. Zhang, and N\. K\. AhmedBias and fairness in large language models: a survey\.Computational Linguistics50\(3\),pp\. 1097–1179\.External Links:ISSN 0891\-2017,[Document](https://dx.doi.org/10.1162/coli%5Fa%5F00524),[Link](https://doi.org/10.1162/coli/_a/_00524),https://direct\.mit\.edu/coli/article\-pdf/50/3/1097/2471010/coli\_a\_00524\.pdfCited by:[Ethical Considerations](https://arxiv.org/html/2608.19549#Sx3.p2.1)\.
- Geet al\.\(2024\)T\. Ge, X\. Chan, X\. Wang, D\. Yu, H\. Mi, and D\. YuScaling synthetic data creation with 1,000,000,000 personas\.External Links:2406\.20094,[Link](https://arxiv.org/abs/2406.20094)Cited by:[§2\.5](https://arxiv.org/html/2608.19549#S2.SS5.p1.1)\.
- Gebhardet al\.\(2014\)P\. Gebhard, T\. Baur, I\. Damian, G\. Mehlmann, J\. Wagner, and E\. AndréExploring interaction strategies for virtual characters to induce stress in simulated job interviews\.InProceedings of the 2014 International Conference on Autonomous Agents and Multi\-Agent Systems,AAMAS ’14,Richland, SC,pp\. 661–668\.External Links:ISBN 9781450327381,[Link](https://dl.acm.org/doi/10.5555/2615731.2615838)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Georgilaet al\.\(2008\)K\. Georgila, M\. Wolters, and J\. MooreSimulating the behaviour of older versus younger users when interacting with spoken dialogue systems\.InProceedings of ACL\-08: HLT, Short Papers,J\. D\. Moore, S\. Teufel, J\. Allan, and S\. Furui \(Eds\.\),Columbus, Ohio,pp\. 49–52\.External Links:[Link](https://aclanthology.org/P08-2013/)Cited by:[§2\.4](https://arxiv.org/html/2608.19549#S2.SS4.p1.1)\.
- Georgilaet al\.\(2010\)K\. Georgila, M\. Wolters, and J\. MooreLearning dialogue strategies from older and younger simulated users\.InProceedings of the 11th Annual Meeting of the Special Interest Group on Discourse and Dialogue,Y\. Katagiri and M\. Nakano \(Eds\.\),Tokyo, Japan,pp\. 103–106\.External Links:[Link](https://aclanthology.org/W10-4321/)Cited by:[§2\.4](https://arxiv.org/html/2608.19549#S2.SS4.p1.1)\.
- Gómez\-Abajoet al\.\(2024\)P\. Gómez\-Abajo, S\. Pérez\-Soler, P\. C\. Cañizares, E\. Guerra, and J\. de LaraMutation testing for task\-oriented chatbots\.InProceedings of the 28th International Conference on Evaluation and Assessment in Software Engineering,EASE ’24,New York, NY, USA,pp\. 232–241\.External Links:ISBN 9798400717017,[Link](https://doi.org/10.1145/3661167.3661220),[Document](https://dx.doi.org/10.1145/3661167.3661220)Cited by:[§2\.2](https://arxiv.org/html/2608.19549#S2.SS2.p1.1)\.
- Guoet al\.\(2024\)G\. Guo, A\. Aleti, N\. Neelofar, and C\. TantithamthavornMORTAR: metamorphic multi\-turn testing for llm\-based dialogue systems\.External Links:2412\.15557,[Link](https://arxiv.org/abs/2412.15557)Cited by:[§2\.2](https://arxiv.org/html/2608.19549#S2.SS2.p1.1)\.
- Hashimotoet al\.\(2025\)E\. Hashimoto, M\. Nakano, T\. Sakurai, S\. Shiramatsu, T\. Komazaki, and S\. TsuchiyaA career interview dialogue system using large language model\-based dynamic slot generation\.InProceedings of the 31st International Conference on Computational Linguistics,O\. Rambow, L\. Wanner, M\. Apidianaki, H\. Al\-Khalifa, B\. D\. Eugenio, and S\. Schockaert \(Eds\.\),Abu Dhabi, UAE,pp\. 1562–1584\.External Links:[Link](https://aclanthology.org/2025.coling-main.106/)Cited by:[§1](https://arxiv.org/html/2608.19549#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1),[§2\.4](https://arxiv.org/html/2608.19549#S2.SS4.p1.1)\.
- Honget al\.\(2025\)M\. Hong, C\. J\. Zhang, C\. Chen, R\. Lian, and D\. JiangDialogue language model with large\-scale persona data engineering\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 3: Industry Track\),W\. Chen, Y\. Yang, M\. Kachuee, and X\. Fu \(Eds\.\),Albuquerque, New Mexico,pp\. 961–970\.External Links:[Link](https://aclanthology.org/2025.naacl-industry.71/),[Document](https://dx.doi.org/10.18653/v1/2025.naacl-industry.71),ISBN 979\-8\-89176\-194\-0Cited by:[§1](https://arxiv.org/html/2608.19549#S1.p3.1)\.
- Hoqueet al\.\(2013\)M\. \(\. Hoque, M\. Courgeon, J\. Martin, B\. Mutlu, and R\. W\. PicardMACH: my automated conversation coach\.InProceedings of the 2013 ACM International Joint Conference on Pervasive and Ubiquitous Computing,UbiComp ’13,New York, NY, USA,pp\. 697–706\.External Links:ISBN 9781450317702,[Link](https://doi.org/10.1145/2493432.2493502),[Document](https://dx.doi.org/10.1145/2493432.2493502)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Huet al\.\(2023\)Z\. Hu, Y\. Feng, A\. T\. Luu, B\. Hooi, and A\. LipaniUnlocking the potential of user feedback: leveraging large language model as user simulators to enhance dialogue system\.InProceedings of the 32nd ACM International Conference on Information and Knowledge Management,CIKM ’23,New York, NY, USA,pp\. 3953–3957\.External Links:ISBN 9798400701245,[Link](https://doi.org/10.1145/3583780.3615220),[Document](https://dx.doi.org/10.1145/3583780.3615220)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p2.1)\.
- Inoueet al\.\(2020\)K\. Inoue, K\. Hara, D\. Lala, K\. Yamamoto, S\. Nakamura, K\. Takanashi, and T\. KawaharaJob interviewer android with elaborate follow\-up question generation\.InProceedings of the 2020 International Conference on Multimodal Interaction,ICMI ’20,New York, NY, USA,pp\. 324–332\.External Links:ISBN 9781450375818,[Link](https://doi.org/10.1145/3382507.3418839),[Document](https://dx.doi.org/10.1145/3382507.3418839)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Jianget al\.\(2024\)H\. Jiang, X\. Zhang, X\. Cao, C\. Breazeal, D\. Roy, and J\. KabbaraPersonaLLM: investigating the ability of large language models to express personality traits\.InFindings of the Association for Computational Linguistics: NAACL 2024,K\. Duh, H\. Gomez, and S\. Bethard \(Eds\.\),Mexico City, Mexico,pp\. 3605–3627\.External Links:[Link](https://aclanthology.org/2024.findings-naacl.229/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-naacl.229)Cited by:[§1](https://arxiv.org/html/2608.19549#S1.p3.1)\.
- Johnstonet al\.\(2013\)M\. Johnston, P\. Ehlen, F\. G\. Conrad, M\. F\. Schober, C\. Antoun, S\. Fail, A\. Hupp, L\. Vickers, H\. Yan, and C\. ZhangSpoken dialog systems for automated survey interviewing\.InProceedings of the 14th Annual Meeting of the Special Interest Group on Discourse and Dialogue,M\. Eskenazi, M\. Strube, B\. Di Eugenio, and J\. D\. Williams \(Eds\.\),Metz, France,pp\. 329–333\.External Links:[Link](https://aclanthology.org/W13-4050/)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Jones and Sabouret \(2012\)H\. Jones and N\. SabouretAn affective model for a virtual recruiter in a job interview context\.Procedia Computer Science15,pp\. 312–313\.Note:4th International Conference on Games and Virtual Worlds for Serious Applications \(VS\-GAMES’12\)External Links:ISSN 1877\-0509,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.procs.2012.10.090),[Link](https://www.sciencedirect.com/science/article/pii/S1877050912008526)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Koboriet al\.\(2016\)T\. Kobori, M\. Nakano, and T\. NakamuraSmall talk improves user impressions of interview dialogue systems\.InProceedings of the 17th Annual Meeting of the Special Interest Group on Discourse and Dialogue,R\. Fernandez, W\. Minker, G\. Carenini, R\. Higashinaka, R\. Artstein, and A\. Gainer \(Eds\.\),Los Angeles,pp\. 370–380\.External Links:[Link](https://aclanthology.org/W16-3646/),[Document](https://dx.doi.org/10.18653/v1/W16-3646)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Lang and Eskenazi \(2025\)M\. M\. Lang and S\. EskenaziTelephone surveys meet conversational AI: evaluating a llm\-based telephone survey system at scale\.External Links:2502\.20140,[Link](https://arxiv.org/abs/2502.20140)Cited by:[§1](https://arxiv.org/html/2608.19549#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Liet al\.\(2022\)X\. Li, C\. Tao, J\. Gao, and H\. GuoA review of quality assurance research of dialogue systems\.In2022 IEEE International Conference On Artificial Intelligence Testing \(AITest\),Vol\.,pp\. 87–94\.External Links:[Document](https://dx.doi.org/10.1109/AITest55621.2022.00021)Cited by:[§2\.2](https://arxiv.org/html/2608.19549#S2.SS2.p1.1)\.
- Liet al\.\(2017\)X\. Li, Z\. C\. Lipton, B\. Dhingra, L\. Li, J\. Gao, and Y\. ChenA user simulator for task\-completion dialogues\.External Links:1612\.05688,[Link](https://arxiv.org/abs/1612.05688)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p1.1)\.
- Liuet al\.\(2023\)Y\. Liu, X\. Jiang, Y\. Yin, Y\. Wang, F\. Mi, Q\. Liu, X\. Wan, and B\. WangOne cannot stand for everyone\! Leveraging multiple user simulators to train task\-oriented dialogue systems\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 1–21\.External Links:[Link](https://aclanthology.org/2023.acl-long.1/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.1)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p2.1)\.
- Luoet al\.\(2024\)X\. Luo, Z\. Tang, J\. Wang, and X\. ZhangDuetSim: building user simulator with dual large language models for task\-oriented dialogues\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\),N\. Calzolari, M\. Kan, V\. Hoste, A\. Lenci, S\. Sakti, and N\. Xue \(Eds\.\),Torino, Italia,pp\. 5414–5424\.External Links:[Link](https://aclanthology.org/2024.lrec-main.481/)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p2.1)\.
- Mazaréet al\.\(2018\)P\. Mazaré, S\. Humeau, M\. Raison, and A\. BordesTraining millions of personalized dialogue agents\.InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing,E\. Riloff, D\. Chiang, J\. Hockenmaier, and J\. Tsujii \(Eds\.\),Brussels, Belgium,pp\. 2775–2779\.External Links:[Link](https://aclanthology.org/D18-1298/),[Document](https://dx.doi.org/10.18653/v1/D18-1298)Cited by:[§1](https://arxiv.org/html/2608.19549#S1.p3.1),[§2\.5](https://arxiv.org/html/2608.19549#S2.SS5.p1.1)\.
- McCarthy and Jarvis \(2010\)P\. M\. McCarthy and S\. JarvisMTLD, vocd\-D, and HD\-D: a validation study of sophisticated approaches to lexical diversity assessment\.Behavior research methods42\(2\),pp\. 381–392\.External Links:[Link](https://link.springer.com/article/10.3758/BRM.42.2.381)Cited by:[§4\.4](https://arxiv.org/html/2608.19549#S4.SS4.SSS0.Px10.p1.1),[§4\.4](https://arxiv.org/html/2608.19549#S4.SS4.SSS0.Px11.p1.1)\.
- Miehleet al\.\(2020\)J\. Miehle, I\. Feustel, J\. Hornauer, W\. Minker, and S\. UltesEstimating user communication styles for spoken dialogue systems\.InProceedings of the Twelfth Language Resources and Evaluation Conference,N\. Calzolari, F\. Béchet, P\. Blache, K\. Choukri, C\. Cieri, T\. Declerck, S\. Goggi, H\. Isahara, B\. Maegaard, J\. Mariani, H\. Mazo, A\. Moreno, J\. Odijk, and S\. Piperidis \(Eds\.\),Marseille, France,pp\. 540–548\(eng\)\.External Links:[Link](https://aclanthology.org/2020.lrec-1.68/),ISBN 979\-10\-95546\-34\-4Cited by:[§3](https://arxiv.org/html/2608.19549#S3.SS0.SSS0.Px2.p1.1)\.
- Nakano and Komatani \(2024\)M\. Nakano and K\. KomataniDialBB: a dialogue system development framework as an educational material\.InProceedings of the 25th Annual Meeting of the Special Interest Group on Discourse and Dialogue,T\. Kawahara, V\. Demberg, S\. Ultes, K\. Inoue, S\. Mehri, D\. Howcroft, and K\. Komatani \(Eds\.\),Kyoto, Japan,pp\. 664–668\.External Links:[Link](https://aclanthology.org/2024.sigdial-1.56/),[Document](https://dx.doi.org/10.18653/v1/2024.sigdial-1.56)Cited by:[§4\.2](https://arxiv.org/html/2608.19549#S4.SS2.p2.1)\.
- Niimi and Nishimoto \(1999\)Y\. Niimi and T\. NishimotoMathematical analysis of dialogue control strategies\.InProceedings of the 6th European Conference on Speech Communication and Technology,pp\. 1403–1406\.External Links:[Link](https://www.isca-archive.org/eurospeech_1999/niimi99_eurospeech.html#),[Document](https://dx.doi.org/10.21437/Eurospeech.1999-362)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p1.1)\.
- Pietquin and Dutoit \(2006\)O\. Pietquin and T\. DutoitA probabilistic framework for dialog simulation and optimal strategy learning\.IEEE Transactions on Audio, Speech, and Language Processing14\(2\),pp\. 589–599\.External Links:[Document](https://dx.doi.org/10.1109/TSA.2005.855836)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p1.1)\.
- Pietquin and Hastie \(2013\)O\. Pietquin and H\. HastieA survey on metrics for the evaluation of user simulations\.The knowledge engineering review28\(1\),pp\. 59–73\.External Links:[Link](https://doi.org/10.1017/S0269888912000343)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p1.1),[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p3.1)\.
- Pragstet al\.\(2019\)L\. Pragst, W\. Minker, and S\. UltesExploring the applicability of elaborateness and indirectness in dialogue management\.InAdvanced Social Interaction with Agents : 8th International Workshop on Spoken Dialog Systems,M\. Eskenazi, L\. Devillers, and J\. Mariani \(Eds\.\),pp\. 189–198\.External Links:ISBN 978\-3\-319\-92108\-2,[Document](https://dx.doi.org/10.1007/978-3-319-92108-2%5F20),[Link](https://doi.org/10.1007/978-3-319-92108-2_20)Cited by:[§3](https://arxiv.org/html/2608.19549#S3.SS0.SSS0.Px2.p1.1)\.
- Raoet al\.\(2020\)P\. S\. B\. Rao, M\. Agnihotri, and D\. B\. JayagopiAutomatic follow\-up question generation for asynchronous interviews\.InProceedings of the Workshop on Intelligent Information Processing and Natural Language Generation,D\. Sánchez, R\. Hervás, and A\. Gatt \(Eds\.\),Santiago de Compostela, Spain,pp\. 10–20\.External Links:[Link](https://aclanthology.org/2020.intellang-1.2/)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Reeves and Nass \(1996\)B\. Reeves and C\. NassThe media equation: how people treat computers, television, and new media like real people\.Cambridge University Press\.Cited by:[§3](https://arxiv.org/html/2608.19549#S3.SS0.SSS0.Px1.p1.1)\.
- Schatzmannet al\.\(2007\)J\. Schatzmann, B\. Thomson, K\. Weilhammer, H\. Ye, and S\. YoungAgenda\-based user simulation for bootstrapping a POMDP dialogue system\.InHuman Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Companion Volume, Short Papers,C\. Sidner, T\. Schultz, M\. Stone, and C\. Zhai \(Eds\.\),Rochester, New York,pp\. 149–152\.External Links:[Link](https://aclanthology.org/N07-2038/)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p1.1)\.
- Schatzmannet al\.\(2006\)J\. Schatzmann, K\. Weilhammer, M\. Stuttle, and S\. YoungA survey of statistical user simulation techniques for reinforcement\-learning of dialogue management strategies\.The Knowledge Engineering Review21\(2\),pp\. 97–126\.External Links:ISSN 0269\-8889,[Link](https://doi.org/10.1017/S0269888906000944),[Document](https://dx.doi.org/10.1017/S0269888906000944)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p1.1)\.
- Sekulicet al\.\(2024\)I\. Sekulic, S\. Terragni, V\. Guimarães, N\. Khau, B\. Guedes, M\. Filipavicius, A\. F\. Manso, and R\. MathisReliable LLM\-based user simulator for task\-oriented dialogue systems\.InProceedings of the 1st Workshop on Simulating Conversational Intelligence in Chat \(SCI\-CHAT 2024\),Y\. Graham, Q\. Liu, G\. Lampouras, I\. Iacobacci, S\. Madden, H\. Khalid, and R\. Qureshi \(Eds\.\),St\. Julians, Malta,pp\. 19–35\.External Links:[Link](https://aclanthology.org/2024.scichat-1.3/)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p2.1),[§4\.4](https://arxiv.org/html/2608.19549#S4.SS4.p3.1)\.
- Sidaouiet al\.\(2020\)K\. Sidaoui, M\. Jaakkola, and J\. BurtonAI feel you: customer experience assessment via chatbot interviews\.Journal of Service Management31\(4\),pp\. 745–766\.External Links:[Link](https://doi.org/10.1108/JOSM-11-2019-0341)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Skantzeet al\.\(2012\)G\. Skantze, S\. Al Moubayed, J\. Gustafson, J\. Beskow, and B\. GranströmFurhat at robotville: a robot head harvesting the thoughts of the public through multi\-party dialogue\.InProceedings of the Workshop on Real\-time Conversations with Virtual Agents in conjunction with the International Conference on Intelligent Virtual Agents,Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Stentet al\.\(2006\)A\. Stent, S\. Stenchikova, and M\. MargeDialog systems for surveys: the rate\-a\-course system\.In2006 IEEE Spoken Language Technology Workshop,Vol\.,pp\. 210–213\.External Links:[Document](https://dx.doi.org/10.1109/SLT.2006.326792)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Suet al\.\(2019\)M\. Su, C\. Wu, and Y\. ChangFollow\-up question generation using neural tensor network\-based domain ontology population in an interview coaching system\.InInterspeech 2019,pp\. 4185–4189\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2019-1300),ISSN 2958\-1796Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Sunet al\.\(2023\)W\. Sun, S\. Guo, S\. Zhang, P\. Ren, Z\. Chen, M\. de Rijke, and Z\. RenMetaphorical user simulators for evaluating task\-oriented dialogue systems\.ACM Trans\. Inf\. Syst\.42\(1\)\.External Links:ISSN 1046\-8188,[Link](https://doi.org/10.1145/3596510),[Document](https://dx.doi.org/10.1145/3596510)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p2.1)\.
- Takaokaet al\.\(2018\)K\. Takaoka, S\. Hisamoto, N\. Kawahara, M\. Sakamoto, Y\. Uchida, and Y\. MatsumotoSudachi: a Japanese tokenizer for business\.InProceedings of the Eleventh International Conference on Language Resources and Evaluation \(LREC 2018\),N\. Calzolari, K\. Choukri, C\. Cieri, T\. Declerck, S\. Goggi, K\. Hasida, H\. Isahara, B\. Maegaard, J\. Mariani, H\. Mazo, A\. Moreno, J\. Odijk, S\. Piperidis, and T\. Tokunaga \(Eds\.\),Miyazaki, Japan\.External Links:[Link](https://aclanthology.org/L18-1355/)Cited by:[§4\.4](https://arxiv.org/html/2608.19549#S4.SS4.p3.1)\.
- Tanaka and Inaba \(2024\)Y\. Tanaka and M\. InabaUser review writing via interview with dialogue systems\.InProceedings of the 25th Annual Meeting of the Special Interest Group on Discourse and Dialogue,T\. Kawahara, V\. Demberg, S\. Ultes, K\. Inoue, S\. Mehri, D\. Howcroft, and K\. Komatani \(Eds\.\),Kyoto, Japan,pp\. 428–439\.External Links:[Link](https://aclanthology.org/2024.sigdial-1.37/),[Document](https://dx.doi.org/10.18653/v1/2024.sigdial-1.37)Cited by:[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Terragniet al\.\(2023\)S\. Terragni, M\. Filipavicius, N\. Khau, B\. Guedes, A\. Manso, and R\. MathisIn\-context learning user simulators for task\-oriented dialog systems\.External Links:2306\.00774,[Link](https://arxiv.org/abs/2306.00774)Cited by:[§2\.3](https://arxiv.org/html/2608.19549#S2.SS3.p2.1),[§4\.4](https://arxiv.org/html/2608.19549#S4.SS4.p3.1)\.
- Zenget al\.\(2023\)J\. Zeng, Y\. Nakano, and T\. SakatoQuestion generation to elicit users’ food preferences by considering the semantic content\.InProceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue,S\. Stoyanchev, S\. Joty, D\. Schlangen, O\. Dusek, C\. Kennington, and M\. Alikhani \(Eds\.\),Prague, Czechia,pp\. 190–196\.External Links:[Link](https://aclanthology.org/2023.sigdial-1.18/),[Document](https://dx.doi.org/10.18653/v1/2023.sigdial-1.18)Cited by:[§1](https://arxiv.org/html/2608.19549#S1.p1.1),[§2\.1](https://arxiv.org/html/2608.19549#S2.SS1.p1.1)\.
- Zhanget al\.\(2018\)S\. Zhang, E\. Dinan, J\. Urbanek, A\. Szlam, D\. Kiela, and J\. WestonPersonalizing dialogue agents: I have a dog, do you have pets too?\.InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),I\. Gurevych and Y\. Miyao \(Eds\.\),Melbourne, Australia,pp\. 2204–2213\.External Links:[Link](https://aclanthology.org/P18-1205/),[Document](https://dx.doi.org/10.18653/v1/P18-1205)Cited by:[§1](https://arxiv.org/html/2608.19549#S1.p3.1)\.

## Appendix AExample Dialogues from the Travel Interview System

Example dialogues from the travel interview system are shown in Figures[11](https://arxiv.org/html/2608.19549#A1.F11)–[16](https://arxiv.org/html/2608.19549#A1.F16)\. Figure[11](https://arxiv.org/html/2608.19549#A1.F11)shows a result from testing with a simulator using the lower persona from Figure[3](https://arxiv.org/html/2608.19549#S3.F3)\. Figures[12](https://arxiv.org/html/2608.19549#A1.F12)–[16](https://arxiv.org/html/2608.19549#A1.F16)show results from testing with simulators using the generated personas in Figures[8](https://arxiv.org/html/2608.19549#S4.F8),[9](https://arxiv.org/html/2608.19549#S4.F9), and[10](https://arxiv.org/html/2608.19549#S4.F10)\.

Figure 11:Example simulated dialogue in condition BL\.Figure 12:Example simulated dialogue in condition noPT\.Figure 13:Example simulated dialogue in condition EL High\.Figure 14:Example simulated dialogue in condition EL Low\.Figure 15:Example simulated dialogue in condition APM High\.Figure 16:Example simulated dialogue in condition APM Low\.
## Appendix BAdditional Examples of Generated Personas

Figures[17](https://arxiv.org/html/2608.19549#A2.F17)and[18](https://arxiv.org/html/2608.19549#A2.F18)show additional examples of generated personas\. These were also randomly selected\.

noPT:

```
- Name: Yusuke Matsumoto
- 33 years old
- Male
- Lives in Nagano Prefecture
- Lives with his parents
- Farmer
- Quiet but sincere; good at detailed tasks
- Goes on a trip once a year
- Recently visited Nagoya alone
- Visits markets and direct-from-farm produce
  stands when traveling
- Would like to visit Spain
- Does his own research and makes reservations
  online
```

EL High:

```
- Name: Kaori Ishikawa
- 48 years old
- Gender: Female
- Lives in Niigata Prefecture
- Lives with her husband and their junior high
  school-aged daughter
- Nursery school teacher
- Speaks in a gentle tone and is polite in
  conversation
- Goes on a trip about once a year
- Recently visited Sendai on a family
  sightseeing trip
- Enjoys soaking in hot springs while traveling
- Would like to visit Kyoto
- Plans trips using travel magazines as a
  reference
```

EL Low:

```
- Name: Misaki Yamauchi
- 47 years old
- Female
- Lives in Nagasaki Prefecture
- Lives with her husband
- Homemaker
- Speaks in a gentle and calm manner
- Travels once a year
- Recently visited Kagoshima with her husband
- Buys local specialties when traveling
- Would like to visit Shikoku
- Uses a travel agency
```

Figure 17:Additional examples of generated personas \(1\)\.APM High:

```
- Name: Miho Uchida
- 38 years old
- Female
- Lives in Hyogo Prefecture
- Lives with her husband and their daughter,
  who is in the first grade of elementary school
- Nurse
- Cheerful and a good listener
- Goes on a trip once a year
- Recently visited Kyoto
- Enjoys taking photos while traveling
- Would like to visit Sapporo
- Makes reservations online
```

APM Low:

```
- Name: Kana Kawamura
- 26 years old
- Female
- Lives in Tokyo
- Single and lives alone
- Freelance illustrator
- Has a free-spirited and creative personality
- Travels frequently
- Recently visited Kyoto alone
- Enjoys visiting art museums while traveling
- Would like to visit Florence
- Makes reservations based on information
  found on social media
```

Figure 18:Additional examples of generated personas \(2\)\.