A Robot that Recommends Products with Few Samples
Interactive-SmartClerk
When building a machine learning model, a large number of samples is typically required. Even when fine-tuning a model to match an individual’s preferences, a considerable amount of preference data is necessary. On the other hand, a store clerk can recommend products by simply asking about 2 or 3 preferences. This research utilizes a mathematical model revealed through cognitive science experiments that explain how humans infer others’ preferences. By asking about preferences for just four dresses, it can recommend a preferred dress.
Missed Conversation Support System
In online meetings such as Zoom, it’s easy to miss what others are saying when you’re distracted by something else. Many people have experienced the frustration of being asked for their opinion when they weren’t paying attention. We developed an algorithm that identifies important missed statements (which we call SCAINs) from the conversation. Additionally, we created the SCAINs Presenter, which presents these important missed statements to help people navigate through situations where they missed parts of the conversation. With the practical application of this research, individuals will be able to participate in multiple meetings or handle multiple tasks more efficiently.
In the video above, you and Person A are having a conversation about a picture next to you, while Person B is not paying attention (in this video, Person A and Person B are using GPT-4, and you are providing human input). When important parts of the conversation between you and Person A arise that would cause a loss of meaning if missed, the corresponding sections are highlighted in yellow and red (the colors are used to make the algorithm’s judgments easier to understand). In the video, we verify what happens if Person B misses these key sections by directing the conversation to them. You’ll notice that Person B, who missed the colored sections, starts saying things that are off-topic.
Automatic English Cloze Test Generation System
CLOZER
Q. Fill in the blank with the correct word: “If you want to go to Keio University, you should ____ hard.”
This type of fill-in-the-blank question is called a cloze test. In this research, we define a good cloze test as one where no other word fits the context besides the correct answer. We developed an algorithm called CLOZER that automatically generates such questions. CLOZER uses BERT, a large-scale natural language model, to simulate human answer tendencies and calculates a Gap Score based on the Gini coefficient to generate questions. This research aims to determine whether machine learning models can simulate human thought processes when solving problems and whether they can create good questions.
A Robot That Engages in Conversation While Judging Relationships
People engage in conversations while considering whether they agree or disagree with others’ opinions. Sometimes, this leads to conflicts during discussions. This research utilizes a large-scale language model to infer the relationships between participants and applies Heider’s balance theory, which describes the dynamics of likes and dislikes among three people, to simulate conversations where participants either become allies or experience discord. Let’s take a look at a conversation between Keio-fan Pepper-kun, Waseda-fan Pepper-chan, and a human.
Conversation Robot System for Travel Memoir Generation
We aim to create a system that generates travel diaries through dialogue with a robot. While image captioning technology can now provide descriptive text about photos, it cannot capture the stories or memories behind the moment when the photo was taken or the experiences associated with it. To create a travel diary, it is necessary to gather information directly from the person who experienced the trip. In this research, we developed a robot that engages in conversation about photos taken during a trip. The system then uses the information gathered through this conversation to generate a travel diary filled with personal memories.