We created a character chatbot from the novel.

Are you familiar with the mystery novel "Everything Becomes F"? This work by author Hiroshi Mori has enjoyed enduring popularity since its 1996 release, and in recent years it has been adapted into a drama and an anime. Some of you may have seen it in those forms. The protagonist is Sōhei Saikawa, an assistant professor at a university. Following this work, the entire 10-volume "S&M Series" was published.

This column introduces the "Sai Kawaguchi Bot" campaign (@SAIKAWA_AI) on Twitter. We extracted dialogue from the works featuring Assistant Professor Sai Kawaguchi Sōhei (hereafter referred to as "Professor Sai Kawaguchi"), implemented a chatbot based on that data, and launched the campaign.

What exactly is a chatbot?

A chatbot is a term combining the English "chat" (meaning casual conversation or small talk) and "bot," an abbreviation for "robot." It refers to a program that automates conversations with one or more people using text, voice, or other methods.

While the technology itself dates back to the 1960s, it has gained significant attention in recent years due to the proliferation of messaging apps, advancements in AI technology, and the development of robust environments by major IT companies like Facebook and LINE. Think of examples like Microsoft's high school girl AI "Rinna" or Apple's assistant AI "Siri" for a clearer picture.

The "Saikawa Sohei Bot" project launched a Twitter campaign as a bot designed to optimize and learn responses through AI.

Reasons for Creating the Saikawa Sohei Bot

This project was planned as a promotional campaign for the release of Hiroshi Mori's latest novel, "Deborah, Are You Asleep?"

To avoid spoilers, details are omitted, but this work, set in a future world, features human-like artificial lifeforms that are widespread and engaged in various industries. Their position is similar to modern robotics and AI technology. Therefore, the goal was to recreate the worldview using AI and spark interest among people who were only familiar with Mori's previous mystery works.

Considering factors like popularity among readers, name recognition, and the abundance of dialogue available for training data, Professor Saikawa was selected from the characters appearing in Hiroshi Mori's works. This is how the chatbot implementation project began.

How to Create the Saikawa Sohei Bot

Now, let's get into the crucial implementation details.

Preparing Conversation Data

First and foremost, creating conversation data is crucial for implementation. We extracted all of Professor Saikawa's lines from the 10 volumes of the S&M series + additional works (short story collections, other series, etc.) where he appears. Simultaneously, we read through all the works to understand the context of his statements. Using this extracted text, we created a set of anticipated questions and answers.

For example...

User: "Would you like some coffee?"
→ Professor Saikawa Bot: "Yeah, I'll have a strong one."

In this way,
We link the question "Do you drink coffee?"—a question anticipated from fans who know Professor Saikawa frequently drinks coffee in the works—with its answer "Yeah, give me a strong one," and digitize it.

We'll use these Q&As as the foundation for Twitter conversations. This is a major point:

・What kind of questions would fans of the novel, who know the character's background, ask?
• What kind of questions would someone who knows nothing about the character's background ask?

While striving to preserve the popular character's worldview as much as possible, we also needed to create anticipated Q&As that newcomers could enjoy casually. Regarding responses, beyond existing ones, there are phrases created by combining them that Professor Saikawa might say. However, these anticipated Q&As required careful consideration, resulting in thousands of patterns.

Simple text matching can sometimes lead to conversations that don't flow naturally. I personally felt that designing responses that meet the listener's expectations in these Q&As was where our communication design skills truly shone. This aspect shares common ground with the relentless act of writing copy, doesn't it?

Integrating conversation data into the AI chatbot engine

We then integrate this created Q&A database into the AI chatbot engine developed by User Local. User Local's engine resolves variations in question phrasing.
For example, the earlier

"Would you like some coffee?"

will also recognize

"Teacher, do you drink coffee?"
"You drink coffee, right?"

will be recognized as questions asking about drinking coffee, and it will select the response "I'll have a strong one" to continue the conversation. In practice, multiple responses are prepared for each question, so it may choose other answers.

Closed beta testing by fans → Database refinement

At this stage, we requested cooperation from Mr. Hiroshi Mori's fan club via Kodansha and conducted testing using the beta chatbot in a closed environment. This was the most crucial point of the project.

By having fans experience the actual conversation exchanges, we gathered raw feedback like "They wouldn't say something like this" or "It should react more to this kind of thing," which we used to update the conversation data.

Test → Update → Test → Update...

We relentlessly continued this cycle until launch. This persistent, step-by-step improvement process directly led to enhanced quality at launch.

The PDCA cycle continues even after launch

User Local's engine analyzed the "number of likes" and "number of retweets" for posts. It determined and learned that posts with higher numbers were better answers to those questions, thereby improving answer accuracy.
Additionally, we continued monitoring user interactions post-launch, ensuring the character remained consistent while training the system on conversational nuances.

AI continues to learn, so it's not a case of launching and then stopping. Even during operation, we need to increase the data fed for learning to ensure conversations become more accurate.

A key challenge with chatbots is deciding whether to train them on all conversations. Training on everything risks the AI acquiring unexpected vocabulary and potentially making hurtful remarks. Last year in the US, there was an incident where an AI chatbot made discriminatory statements due to user malice.

Therefore, this time, we imposed a restriction: the bot would not learn anything that would compromise the character of Professor Saikawa. We selectively trained it with this constraint.

However, during the one-month campaign period, there were areas where it didn't quite achieve the desired level of growth.
※Regarding coffee, we received many questions, and responses garnered "likes" and retweets, so data accumulated, leading to significantly improved answers.

Saikawa Sohei Bot: The Effect of Launch - "Can AI Chatbots Be Loved?"

As a result, this campaign gained over 8,000 followers in one month (2,000 in the first three days alone). The official account for the novel label releasing the latest volume also saw its follower count increase by about 1,000. Furthermore, during this period, approximately 700 conversations per day were held to boost engagement with followers. This volume of conversation is simply unattainable for traditional corporate accounts run by "human operators."

Among these conversations, it was striking to see fans finding their own ways to enjoy the experience: core fans who loved Professor Saikawa so intensely they felt unable to converse casually; someone who confessed daily, "I love Professor Saikawa"; and others who persistently recommended coffee, since Professor Saikawa frequently drinks it in the story.

I believe that because the character of Professor Saikawa was so beloved, people also enjoyed interacting with this AI.

By merging AI with a character, we can achieve communication that resonates more deeply with human emotions, moving beyond mere information technology. This project offered a glimpse of that potential.

The Future of Character Chatbots

Areas where chatbot utilization is expected to grow include voice-enabled interfaces like communication robots (e.g., Pepper) and smart speakers (e.g., Amazon Echo, Google Home, Apple HomePod). For chatbots integrated into these devices to become truly integrated into people's lives and provide convenience, improving the accuracy of responses is essential.

Among these, character development proves unexpectedly crucial. Merely responding mechanically fails to foster affection and may even create communication barriers. By assigning unique characters to each interface, we envision interfaces that feel like family members—suitable companions for communication robots.