Keywords

1 Introduction

In this research, we propose a driving support agent which speak based on politeness theory [1]. Driving support agent is drawing attention as a new form of driving assistance for driving, but knowledge on effective utterance for driving support has not been established. The proposed agent selects an utterance based on the politeness theory in consideration of the age of driver, gender and driving characteristics. In the previous research, it has been pointed out that the driving support agent needs to be designed according to the driver’s age and ability [2].

Politeness theory is proposed by Brown and Levinson is employed in this research. This research focuses on positive politeness strategies (PPS) and negative politeness strategies (NPS). PPS and NPS are representative utterance strategies in the five strategies defined in politeness theory. PPS is a strategy for actively reducing psychological distance to the opponent. They are categorized into fifteen sub-strategies. NPS is strategy for keeping psychological distance to the opponent and keeping it, there are ten sub-strategies.

In the research field of Human-Robot Interaction, politeness theory has been used for robot’s utterance design in the situation of cooking, but it does not correspond to languages and cultures other than American English [3]. So this research we are uses Japanese language and Japanese native speakers.

2 Politeness Theory

Of the two individuals interacting with one another, we define the speaker as S and the listener as H [1]. According to Brown and Levinson, S and H both desire to form an interpersonal relationship with one another. This desire is called face [1]. In general, in dialog, S wishes to hold H’s face; however, depending on the action, the result may be to threaten H’s face. Such an action is called a face-threatening act (FTA).

When S needs to perform an FTA to H, S estimates the weight possessed by the FTA. Here, the weight of the FTA is calculated as

$$ {\text{Wx }} = {\text{ D }}\left( {{\text{S}},{\text{ H}}} \right) \, + {\text{ P }}\left( {{\text{H}},{\text{ S}}} \right) \, + {\text{ Rx,}} $$
(1)

where D is a value indicating the social distance of S and H, P is the amount of force H exerts on S, and Rx is a value indicating how much the FTA is considered a burden in the given culture and society. More specifically, weight Wx of the FTA is the sum of D, P, and R. Since P and R fluctuate in different cultures and societies, the resulting weight of the FTA varies depending on the given culture and society, even in the same speech act. Here, if the degree of the FTA is relatively high, NPS is selected; however, if the degree of the FTA is relatively low, PPS is selected. Furthermore, PPS comprises 15 strategies, whereas NP comprises 10 [1].

As noted above, PPS consists of 15 strategies, including praising the other individual, sympathizing with the other individual, giving a gift to the other individual, and so on.

3 Proposed System

An overview of the proposed system is shown in Fig. 1. The proposed system supports the driver with the following flow. The system gains information concerning the driver. Each driver inputs their gender, age, driving characteristics and so on before driving. In this research, we use RoBoHoN [4] as an agent. Also, driver’s individual characteristics is used to evaluate driving characteristics [5].

Fig. 1.
figure 1

Outline of the proposed system.

Full size image
  1. 1.

    During driving, in the situation when a driver needs assistance, the system sends the acquired driving situation to stochastic model and estimate politeness strategy according to the driver and the driving. The construction of the stochastic model is based on sample data collected from the result of the questionnaire on the driver’s information, and the questionnaire asking impressions of agent which uses PPS and/or NPS.

  2. 2.

    An instruction sends to extract the utterance to the utterance DB based on the estimation result by the stochastic model that can estimate the politeness strategy according to the information of the user. The utterances, utterances according to the driving situation are categorized into each politeness strategy and stored in DB. The utterance DB is constructed by converting instructor’s utterances to NPS and PPS. They are accumulated them in the DB, utilizing the utterance data from the driver’s instructor. The utterances extracted from the utterance DB is sent to the utterance timing control mechanism.

  3. 3.

    The agent utters sentences sent from utterance timing control mechanism to the driver by speech synthesis. It should be assumed that the utterance timing control can be performed sufficiently.

As described above, the proposed system is thought to select a utterance according to the attributes and circumstances of the driver, and to make it possible to support driving with high acceptance for a wide range of users. Moreover, in this paper, situations requiring a high degree of urgency and momentary judgment, such as when a pedestrian has popped out, are excluded from the scope of the system that the proposed system supports.

4 Experiment

4.1 Experiment Set Up

The purpose of this experiment is to investigate the impression on the utterance strategy used by the agent in the driving support scene as a preliminary experiment for developing the proposed system. In this experiment, we also focus on the difference between the end-of-sentence style (honorific words - non-honorific words), which is a representation method of distinctive psychological distances in Japanese conversation. Sample data collection for constructing a probabilistic model of the proposed system will be conducted for participants with a wide range of attributes based on this experiment.

The experiment participants watch movies in which the authors manually converted RoBoHoN’s utterance in a part of the reference movie [6] to NPS and PPS in the participant plan. Figure 2 is the image of the experimental video. In this experiment, NPS is “to show respect” (use of honorifics), “to question/to obscurate”, “to apologize”, PPS, in addition to non-honorific words, Incorporate, “exaggerate sympathy” are used. Take counter balance on the order of the videos you watch. The length of the video is about 2 min for both NPS and PPS. Table 1 examples of utterance contents of agents. In the utterance example of the first NPS in Table 1, we show respect by using honorific expressions and it is said that “Do you mind if I enter behind a white car?” We are doing support. Language behaviors that do not compel actions against opponents correspond to NPS’s “Question and ambiguity” [7]. In the first utterance example of PPS, the proximity of psychological distance is expressed by using non-polite words. In the second example of NPS, for the utterance that the driver “avoids pedestrians,” it is dangerous because you do not know when to jump out “While using the ambiguous utterance” I do not think so “while agreeing, I use NPS’s” “Question and ambiguity”. In the example of PPS, we use “to emphasize empathy” by PPS by not only empathizing but also emphasizing expressions such as “it is true!” Participants will answer questionnaires by question paper for agents of NPS and PPS after watching movies.

Fig. 2.
figure 2

Image of the experimental video.

Full size image
Table 1. Example of agent’s utterance.
Full size table

The experiment participants were 46 university students/graduate students (20 men, 26 women, average age 21.7 years old, driving history: 9 years or less).

As the evaluation criterion, we use the characteristic adjective scale [8] (Table 2) and the impression evaluation items created by the authors (Table 3). The characteristic adjective scale is a seven-step SD method, the impression evaluation item is a seven-step Likert scale method (1: absolutely not applicable, 2: hardly applicable, 3: it does not apply, 4: neither, 5: Somewhat true, 6: fairly applicable, 7: perfectly applicable). The following 15 items are used for the evaluation items.

Table 2. Characteristic adjective scale [8].
Full size table
Table 3. Evaluation items.
Full size table

Of the 15 items above, Q5, Q8, and Q10 are reversal items. For example, in “Q5: Irritated against a robot”, as the evaluation value approaches seven points, the evaluation is “not irritated”.

4.2 Experimental Result

Figures 3 and 4 shows the results of impression evaluation. In the figure, the numerals attached to the graph represent the average point of the evaluation with one digit after the decimal point. This also applies to subsequent graphs. Also, “#” is added to the item number as a mark in the reversal item. In order to compare the evaluation values of NPS and PPS in each item, significance judgment was made by using Wil-coxon’s signed rank order test with significance levels of 1% and 5%. A significant difference in the result of the test, “Q 7: I felt that I could get along with the robot” (p < 0.01), “Q 8: I think that the robot gets tired quickly”, “Q 9: I want the robot to be like a family member or best friend”, “Q12: If there is a robot, it seems to be more fun than driving by alone”, “Q14: I want to use a robot” (p < 0.05). In addition, a marginally significant (p < 0.1) was observed in “Q 6: I can feel good with robot”. In the other items, no statistical difference was found.

Fig. 3.
figure 3

Result of the evaluation (a).

Full size image
Fig. 4.
figure 4

Result of the evaluation (b).

Full size image

Figures 5 and 6 shows the results of the characteristic adjective scale. In this figure, for example, in the item of “1 Aggressive-Passive”, one point is “Aggressive” and seven points are “Passive”. As a result of Wilcoxon’s signed ranking test, significant difference in “1 aggressive - Passive”, “3 Sassy-Cheeky”, “4 Friendly-Severe”, “8 Responsible-No sense of responsibility”, “9 Impolite-Prudent”, “10 Shameless-Shy”, “11 Heavy-Frivolous”, “12 Sunk-Excited”, “15 Down to earth-Indiscriminate”, “16 Approachable-Friendless”, “18 Not confident-Confident” (p < 0.01), “2 Of bad character-Of good character”, “13 Dignified-Subservient”, “17 Lethargic-Ambitious”, “20 Unkind-kind” (p < 0.05). In addition, a significant trend (p < 0.1) was observed in “19 long-short temper”. In the other items, no statistical difference was found.

Fig. 5.
figure 5

Result of the Characteristic adjective scale (a).

Full size image
Fig. 6.
figure 6

Result of the Characteristic adjective scale (b).

Full size image

5 Discussion

In impression evaluation, PPS overall got high evaluation. Especially in “Q7: I felt that I could get along with the robot”, the average PPS was 5.2 points compared to the average NPS score of 4.2, and a score difference of about 1.2 times was seen. From this, it is suggested that driving support agent using PPS support is more effective for improving familiarity than NPS. On the other hand, NPS did not have any items that received a significantly high evaluation. However, since the experiment participants are all students in their twenties in this experiment, different results may be seen when widening the age range of participants. For example, when conducting experiments with relatively high age group participants, it is considered that PPS may be evaluated poorly by agents. In previous research, it has been reported that acceptance of driving support agents differs for each age of participants [2].

According to the results of the characteristic adjective measure, PPS has an average score of 4.0 (neither) on the average of 5.7 (cautious) of NPS in “9 Impolite-Prudent”, about 1.4 times the score difference Was observed. From this, it is suggested that by using NPS, careful support is provided and the user may feel that it is conveying accurate information. On the other hand, it also gives the impression that it is “inaccessible” and “less friendly” compared with PPS, so suggesting that continuing support by NPS for too long may not be expected to improve familiarity.

Experimental results suggest that agent using PPS is effective for improving familiarity between agent and participants. Since agent using NPS gave impression that it was carefully supporting, it was suggested that driver feel that they convey accurate information. Our study approach will help design the driving support agent. The experiment participants are limited to students in this research. Future research requires to conduct experiments on a wider range of users.

6 Conclusion

In this paper, we selected agents that support driving by selecting a politeness strategy according to driver’s attributes and operating conditions. In addition, we focused on the difference in style of end of sentence, and conducted an impression evaluation experiment of agents. Experimental results suggested that PPS could be more effective in improving familiarity than NPS to university students and graduate students. While NPS gave the impression of careful support, it gave an impenetrable impression compared to PPS. By verifying the utterance effect of the driving support agent for each utterance strategy and targeting participants with a wide range of age and driving characteristics as in this paper, we established a support method with a language corresponding to a wider attribute It is thought that it will be possible to go.

In the future, we plan to develop a proposal system and examine interaction experiments considering utterance strategy effective other than the end-of-sentence style.