Exploring the Effects of Japanese Font Designs on Impression Formation and Decision-Making in Text-Based Communication

There has been some prior research on the potential of typography to convey emotions, particularly in the context of European scripts. For instance, Choi, Yamazaki and Aizawa [5] demonstrated the capability of typography to convey emotions based on core affect [12] and basic emotions [13]. Furthermore, previous research has also proposed text-based communication systems that leverage different typographical forms [4, 6], and it has been reported that altering the typographical form in text-based communication can be an effective means of expressing emotions [5].

In contrast, the majority of the empirical studies that have examined the relationship between font shape and emotion in CMC have primarily focused on fonts for the Latin alphabet, with limited examination of Japanese fonts. Many of the previous studies that have established a relationship between Japanese fonts and emotion have employed the semantic differential (SD) method [14] to evaluate font impressions; however, the adjectives used in the evaluation were rather arbitrary, making it difficult to generalize the findings [15]. To address these limitations, Toyosawa [16] focused on core affect [12], which comprises two dimensions of emotional valence and arousal that have been widely employed in traditional CMC research. However, one limitation of this examination was the inclusion of only five fonts.

In regards to the relationship between fonts and confidence, previous research has established a correlation in handwritten texts [17, 18]. However, studies that have investigated the relationship between the design of fonts presented on computers and confidence are scarce.

As such, the primary objective of this study is to examine the relationship between Japanese fonts and impression formation by utilizing core affect and basic emotions as measures of emotion, and confidence as a measure of their influence on decision-making. The following research questions (RQs) were established:

RQ1: How consistent are the core affect and confidence impression ratings for Japanese fonts among raters?

RQ2: What type of Japanese fonts are associated with high and low impressions of emotional valence (positive and negative), arousal (aroused and calm), and confidence (confident and unconfident).

RQ3: What type of Japanese fonts are associated with each of the basic emotions?

We conducted an examination of prior research on the utilization of paralinguistic emotional cues, such as emojis and typography, within text-based chat applications. Despite that numerous previous studies have substantiated that the incorporation of additional emotional cues enhances the conveyance of emotions, there remains a dearth of literature that goes into detail about the specific behavioral changes, especially decision-making, that result from the transmission of these emotions in text chats.

Much of the previous literature on the efficacy of online advice has been conducted within text-based communication that does not incorporate paralinguistic emotional cues (e.g., [19, 20, 21]) In comparison to other forms of text-based communication, such as e-mail, text-based chats possess the advantage of facilitating the conveyance of emotions through paralinguistic emotional cues such as emoticons and stamps, and there exists some evidence that impressions conveyed with these cues may influence decision-making. Prior research on the correlation between emotional cues and decision-making in text-based chat has primarily focused on the utilization of emoticons [7]. However, it is important to note that emoticons, which often depict emotional facial expressions, possess limitations as emotional cues. That is, previous research has demonstrated that emoticons are commonly employed not only to express the sender’s emotions but also to emphasize the content of the message. [22, 3, 23]. Therefore, the relationship between the emotions conveyed in paralinguistic cues and the degree of acceptance of advice in text-based interactions remains uncertain.

On the other hand, there exists a large body of literature that has highlighted the impact of impressions regarding the sender of a message on decision-making. One measure that has been posited as a particularly influential factor is confidence. For instance, when the sender of a message conveys a numerical value of confidence in the advice in conjunction with the advice itself, message recipients are more likely to make a decision in accordance with the advice if they receive a higher confidence score [8, 9, 10, 11]. This is thought to be due to the propensity of humans to make decisions through the confidence heuristic [24], whereby individuals rely on the perceived confidence level of a source when making decisions. It is plausible that when confidence is communicated through typography instead of a numerical value in text-based chats, the degree of advice acceptance may be affected in a similar manner as in previous studies.

Hence, in this study, we focused on confidence as an impression about the other person that influences decision-making, and our second objective of this study was to investigate how decision-making is affected when confidence is communicated through typefaces within text-based chats. We hypothesized (H1):

H1: Advice presented in a typeface that evokes a sense of confidence in the recipient is more likely to be accepted in comparison to advice presented in a typeface that evokes lack of confidence in the recipient.

For the first experiment, stimuli were generated from the complete set of 93 fonts belonging to the Japanese font series ”FONT1000 ³³3http://www.font1000.com/,” included in the ”Adobe Fonts ⁴⁴4https://fonts.adobe.com/” library. The complexity of the Japanese writing system and the abundance of characters present challenges in creating a diverse range of Japanese fonts. To circumvent these challenges, the FONT1000 series offers a wide variety of fonts by limiting the number of characters to 1,000 commonly used kanji, in addition to including all hiragana and katakana characters as well as punctuation marks. Additionally, the design of each typeface in the FONT1000 series ensures a consistent appearance of the characters, irrespective of their variations (kanji, hiragana or katakana).

During each trial of the experimental task, the message “Please tell me more about Plan A” was presented in a Japanese font (Figure 1A). The participants were instructed to consider this message as a reply to the question “Which do you think is better, Plan A or Plan B?” They were then asked to rate their impressions of the sender of the message (Figure 1A) in terms of core affect (Figure 1B and Figure 1C), confidence (Figure 1D), and basic emotions (Figure 1E). For the core affect evaluation, they were prompted to respond to the emotional valence and arousal level of the sender’s emotions using the Self-Assessment Manikin (SAM) test [25], which was used in a previous study on European fonts  [5], on a 9-point scale (Figure 1B and Figure 1C), respectively. For the confidence assessment, the participants were prompted to respond to the level of confidence they perceived the sender felt using a 9-point scale ranging from “not at all confident” to “very confident” (Figure 1D). In the basic emotion assessment, they were asked to select which category of the following: anger, disgust, fear, happiness, sadness, surprise, and neutral the sender’s emotions fit the best (Figure 1E).

The experimental tasks were developed as web applications utilizing Next.js and Firebase Realtime Database, and the participants completed the tasks on a web browser via their own personal computers or other devices.

We hired 177 Japanese participants for this experiment (139 males and 36 females, $M_{age}=46.4$ years, $SD=11.3$, excluding non-response) via a Japanese crowdsourcing platform Yahoo! Crowdsourcing ⁵⁵5https://crowdsourcing.yahoo.co.jp/. None of the participants had any background information about the experiment before joining the study. They were compensated for their participation in accordance with the rules of our university.

To assess the fonts, all 93 fonts were sorted into four sets, and the participants were tasked to evaluate each font once per set, while also responding to a concentration check test three times during the assessment of each set. The concentration check test was administered by displaying the message (Figure 1A) as “Please select the most positive emotion for the following question” in lieu of “Please tell me more about Plan A”, and the participants were instructed to respond using the 9-point Self-Assessment Manikin (SAM) scale.

Participants had the option to participate in the experiment for a minimum of one set and a maximum of five sets. Data from 167 participants were analyzed, as ten participants provided inaccurate responses in the concentration check test. 50 raters were recruited for each of the 93 fonts for a total of 4650 responses. Data from 10 participants who provided incorrect responses in the concentration check test were excluded. As a result, an average of 47.2 (45-49) participants per font, for a total of 4393 responses, were utilized in the analyses.

Initially, to examine the consistency of impression ratings of emotion and confidence in the Japanese fonts across raters (RQ1), the correlation of each participant’s impression ratings and the average of those of the other participants for each dimension of the core affect (emotional valence and arousal) and confidence were calculated following the process in previous work [26]. High inter-rater correlations were found for emotional valence and confidence, and moderate inter-rater correlation was found for arousal (emotional valence: $M=.568$, 95% CI [.536, .600]; arousal: $M=.382$, 95% CI [.339, .425]; confidence: $M=.553$, 95% CI [.520, .588]).

Subsequently, in order to investigate which Japanese fonts were characteristic of each of the evaluation dimensions of emotional valence, arousal, and confidence (RQ2), the mean of each evaluation value for each font was calculated. Figure 2 illustrates two fonts with the highest and two fonts with the lowest mean assessment values for each dimension (valence, arousal and confidence).

In order to investigate which Japanese fonts were associated with each of the basic emotions (RQ3), we determined the selection rate of each basic emotion for each font and identified the fonts that had a significantly higher selection rate than chance level by a binomial test. Firstly, of the 93 fonts, we excluded those with significantly higher selection rates for ”Neutral” compared to chance level, leaving 13 fonts. Among these, three fonts were associated with fear (Figure 3A), four with happiness (Figure 3B), and one each for sadness and surprise (Figure 3C and Figure 3D). Additionally, two fonts were associated with both fear and sadness (Figure 3E) and one font with both happiness and surprise (Figure 3F). We identified no fonts that were associated with anger or disgust at higher than chance level.

First, we examined the degree of consistency (RQ1) among raters in their impression ratings of emotional valence, arousal and confidence towards Japanese fonts by utilizing inter-rater correlation coefficients. The results revealed high inter-rater correlations for emotional valence and confidence. Notably, these correlation coefficients were comparable to those of facial impression evaluations [26], which have been considered as having high inter-rater correlation. This finding is congruent with prior research on European fonts that demonstrated fonts to be effective in communicating emotional valence [4], implying that fonts can serve as reliable emotional cues with consistent evalutions among raters.

Second, in order to investigate what type of Japanese fonts were associated with high and low impressions of emotional valence, arousal and confidence (RQ2), we analyzed the mean of each evaluation value for each font. Figure 2 illustrates that fonts that were rated positively in terms of emotional valence and high in confidence had thicker lines compared to fonts that were rated negatively and low in confidence. This finding is consistent with previous research [27] which found that fonts with thin lines were associated with “ambiguity”. In future work, we are interested in investigating in detail the parameters (e.g., line thickness) that contribute to the emotional valence and confidence perceptions in Japanese fonts. Additionally, there was a tendency for fonts rated negatively in terms of emotional valence and low in confidence to have inconsistent line widths, as well as low readability. This is in contrast to previous research on European fonts [5] which found that complex and difficult-to-read fonts are not suitable for expressing emotion. Future research should investigate the relationship between readability and the emotional valence and confidence impressions of fonts.

Third, in order to investigate what type of Japanese fonts were associated with each basic emotion (RQ3), we determined the selection rate of each basic emotion for each font, and identified the fonts that had a significantly higher selection rate of basic emotions compared to chance level. The results revealed that fonts that only conveyed fear were characterized by non-uniformly thick lines and trembling lines (Figure 3A). This finding is consistent with previous research [27], which found that fonts with wavy line thickness were associated with feelings of fear and worry. Conversely, fonts that conveyed both fear and sadness tended to have extremely thin lines, which is distinct from fonts that only conveyed fear (Figure 3E). This suggests that fonts may be capable of conveying complex emotions that are composed of Ekman’s (1992) basic emotions. Additionally, all of the fonts that conveyed happiness were characterized by thicker lines, however, some of these fonts were difficult to read. Further investigation is required to explore this contradiction with previous research on European fonts [5], which found that typefaces that are difficult to read are not suitable for expressing emotion.

Lastly, it is important to highlight that only a small subset of the Japanese fonts in our dataset were associated with specific basic emotions. As such, our results indicate the ambiguity of standard font designs in terms of conveying basic emotions, specifically, but at the same time call to attention the opportunities for design of Japanese fonts to support emotional expression in text-chats, particularly in terms of negative emotions such as anger and disgust.

We conducted Experiment 2 to investigate the impact of font-communicated confidence on decision making. The experimental paradigm employed was the Judge-Advisor System [8], a widely-utilized method in decision making research that allows for assessment of the degree of advice acceptance by comparing decisions made before and after receiving advice.

The task used for decision making was the dot counting task [28, 29]. Participants were presented with a square shape on a computer screen containing an arbitrary number of dots on each side (randomly determined on each trial) of the gaze point. The number of dots on one side was controlled by the program, with $200+d$ dots on either side and $200-d$ dots on the other side. The dots were randomly distributed on a 20 x 20 invisible grid inside a square. The value of $d$ was manipulated by the 2-down-1-up method [30] to ensure equal correct response rates for all participants.

The flow of a single iteration of the dot counting task is depicted in Figure 4. To initiate the task, participants were instructed to press the start button, after which the dot patterns were presented for a duration of 500 milliseconds on both sides of the screen (Figure 4A). The participants were subsequently prompted to indicate which side they believed contained the greater number of dots with ten different options, ranging from “I am 100% sure it is left” to “I am 100% sure it is right”, with increments of 20% (Figure 4B). A middle choice to indicate uncertain was not allowed. The participants were required to make their selection and indicate their level of confidence by clicking on the corresponding radio button with their mouse, and then clicking the “Send” button.

Upon activation of the “Send” button, the options were instantly obscured, and three seconds later, only the text message corresponding to the selected choice reappeared on the screen. Six seconds after the “Send” button was activated, a statement in the form of “I believe there are more dots on the right than on the left.” or “I believe there are more dots on the left than on the right.” was displayed in Japanese, and in black letters on a white background, as purported advice from the computer (Figure 4C).

The advice was displayed in one of eight typefaces. The typefaces employed comprised of the four fonts that had the highest mean confidence scores (high confidence fonts) and the four fonts that had the lowest mean confidence scores (low confidence fonts) in Experiment 1.

Finally, after reviewing the purported advice, the participants responded again using the same 10-point response box as before the advice (as depicted in Figure 4D). At this time, the previously selected options were outlined in red to indicate the pre-advice response. Upon completion of the response, the next trial was immediately initiated.

The dot-counting task consisted of 48 trials, which were divided into six blocks. In each block, each of the eight fonts was used once in a randomized order as the font for the advice message. Thus, all fonts were presented six times during the experiment. Additionally, five practice trials were conducted prior to the initiation of the 48 primary trials. Prior to commencing the main trials and upon completion of the main trials, the participants were evaluated to ensure their concentration on the task. Specifically, during the dot pattern display, 400 dots were presented on one side and one dot on the other, and the participants were prompted to indicate which side they perceived to contain the greater number of dots. Similar assessments were conducted post the completion of the 48 main trials.

The experimental tasks were developed as web applications utilizing Next.js and Firebase Realtime Database, and the participants completed the tasks on a web browser via their own personal computers or other devices.

We hired 100 Japanese paricipants for this experiment (80 males and 20 females, $M_{age}=46.3$ years, $SD=11.0$) via the crowdsourcing platform Yahoo! Crowdsourcing ⁶⁶6https://crowdsourcing.yahoo.co.jp/. None of the participants had any background information about the experiment before joining the study. They were compensated for their participation in accordance with the rules of our university.

Data analysis was performed utilizing data from 95 participants out of the total 100 after excluding two participants who did not fully complete the experiment and three participants whose responses remained unchanged prior and subsequent to receiving advice in all trials of the dot-counting task. Notably, no participants provided inaccurate responses during the concentration check task.

The acceptance of advice, denoted as $\delta_{C}$, was calculated in accordance with the methodology outlined in a previous study [29] by utilizing the following formula for determining the difference in confidence between pre-advice and post-advice responses.

When the pre-advice answer and the advice direction were consistent:

When the answer before the advice and the direction of the advice were inconsistent:

$C_{pre}$ represents the confidence level of the response prior to receiving advice, ranging from 20% to 100%; i.e., $C_{pre}$ does not take into account whether participants thought the left or right side contained more dots. $C_{post}$ represents the confidence level of the response following advice, ranging from -100% to 100%. If the sides chosen by the participants changed between before and after the advice, $\delta_{C}$ was coded as negative, and conversely, as positive if the sides were unchanged. As a result, $\delta_{C}$ has a maximum value of 200% (when a participant chooses either the left or right answer with 100% confidence prior to receiving advice, receives advice that differs from their pre-advice answer and chooses a different orientation after receiving advice with 100% confidence) and a minimum value of -200% (when a participant chooses either the left or right answer with 100% confidence, receives the same advice as their pre-advice answer and chooses a different orientation than before receiving advice with 100% confidence).

Our hypothesis (H1) posited that the advice written in high confidence fonts would be more accepted than the advice written in low confidence fonts. To test H1, $\delta_{c}$ was subjected to a two-way ANOVA with font (high or low confidence) and direction of advice (whether the advice was consistent or inconsistent with the pre-advice response) as within-participant factors (Figure 5). Results revealed a significant main effect of font ($F[1,94]=7.90$, $p=.006$, $\eta^{2}_{p}=.078$), with the acceptance of advice response being greater for high confidence fonts than for low confidence fonts. H1 was supported. The main effect of direction of advice was also significant ($F[1,94]=53.72$, $p<.001$, $\eta^{2}_{p}=.364$), with the degree of advice acceptance being greater for advice that differed from the pre-advice response than for the same pre-advice response. The interaction effect between font type and degree of advice was not significant ($F[1,94]=0.63$, $p=.429$, $\eta^{2}_{p}=.007$).

We examined the effect on the degree of advice acceptance when advice was given in text using Japanese fonts that conveyed the impression of high or low confidence (H1). We found that advice presented in a font that conveyed a high level of confidence was more likely to be accepted in comparison to advice presented in a font that conveyed a low level of confidence. In other words, this result suggested that the font used in text-based communication can influence the degree of advice acceptance by the recipient. This finding is congruent with previous research [8, 9, 10, 11], which posited that when confidence is communicated numerically, advice is more likely to be accepted when a high level of confidence is conveyed. Our results imply that a “confidence heuristic” [24] may be employed in decision-making even when fonts are employed to convey the level of confidence.

It is noteworthy that a significant main effect of advice direction was also observed in the experiment. This result may can attributed to the experimental design. In this study, the degree of belief was set at five levels, ranging from 20% to 100% on the left and right sides, respectively, thus when the pre-advice response and the direction of the advice were the same, there was less scope for modifying the response in accordance with the advice than when the pre-advice response and the direction of the advice were dissimilar. Prior research on advice acceptance has established that decision makers tend to disregard advice due to a high level of confidence in their own judgment [31], and that the more divergent their own judgment and advice are, the less likely they are to accept the advice [32]. Therefore, it is possible that advice that is congruent with the pre-advice response is more likely to be accepted than advice that is incongruent from the pre-advice response if an experimental design is employed that does not bias the maximum value of the degree of advice acceptance depending on the direction of the advice. Further research is required to determine the effect of advice direction on the degree of advice acceptance when the font is altered in text-based communication.