<Presentation Content>
Presenter:Shinichi Nakagawa
Title: Analysis of Layer Configuration in Composite Service Recommendation using Hierarchical Multi-label Classification
Abstract:A composite service is a new service created by combining multiple web services on the Internet. Composite services enable the construction of services that meet user needs. However, selecting the necessary web services for constructing composite services is difficult. In this research, we propose a method for selecting combinations of services that meet requirements from natural language user requirement definitions using pre-trained models of BERT. Specifically, we formulate the problem as a multi-label classification problem and solve it using BERT's pre-trained model. Furthermore, to improve selection accuracy, we categorize services hierarchically and perform hierarchical multi-label classification. Additionally, to realize hierarchical multi-label classification, it is necessary to set the number of stages of functional decomposition and the number of functional candidates, and we analyze the layer configuration.

Slide:

<Presentation Content>
Presenter:Yuria Mori
Title: Impact of Voice Conversion on Interaction Distance in Metaverse Environments
Abstract:With the development of the metaverse, interest in the behavior and psychology of users in virtual environments has been increasing. The Proteus effect, where avatar characteristics in virtual space influence behavior, has been reported, but most previous research has focused on visual features, and the impact of voice has not been sufficiently examined. In this research, we analyze the impact of voice conversion on interaction distance and aim to verify the Proteus effect mediated by voice.

Slide:

<Presentation Content>
Presenter:Sakie Kaoku
Title: Experimental Evaluation of Structured Prompts in Translation Agents
Abstract:With recent globalization, there is a growing demand for high-accuracy machine translation systems supporting multilingual communication. Neural Machine Translation (NMT) still faces challenges in resolving ambiguity when dealing with polysemous words and culturally dependent expressions, often leading to mistranslations when contextual information is insufficient. To address this, we propose a translation agent that resembles human interpreters by clarifying ambiguous expressions through dialogue with users. Large Language Models (LLMs), with their high dialogue and context understanding capabilities, enable interactive disambiguation and translations that better reflect user intentions, supporting the realization of translation agents. However, natural language prompts tend to become ambiguous and bloated as dialogue progresses, making it difficult to consistently control agent behavior. Therefore, we design a YAML-based structured prompt notation with explicitly defined syntax and semantics to achieve stable and versatile agent control. This notation structures agent behavior through state transitions, events, and actions, achieving consistent behavior across different LLMs. In our evaluation experiments using gpt-4.1-mini, we compared YAML prompts with naturally written prompts. Additionally, we included Google Translate and DeepL as comparison targets representing existing machine translation methods without dialogue functionality. Results showed that YAML-based prompts prevented scenario deviations, whereas natural language prompts exhibited deviations. Regarding translation accuracy, while some results were worse than existing methods, the system generated translations requiring minimal manual correction. These findings suggest that structured prompts are an effective approach for improving the stability and consistency of translation agent dialogue.

Slide:

<Presentation Content>
Presenter:Yuki Tobisawa
Title: Web Service Composition using LLM Multi-Agent
Abstract:API selection for constructing composite services from natural language requirements is a critical challenge in automatic service composition. Existing single LLM agent approaches face scalability problems: as the number of APIs increases, prompt length grows linearly, leading to context length limitations and reduced inference accuracy. In this paper, we propose a multi-agent approach where each API specification is individually maintained by contractor agents, and recommendation is performed through bidding based on the Contract Net Protocol (CNP). Furthermore, we externalize the reasoning process into distinct tasks: core function decomposition, category mapping, API matching, and API selection, comparing different task allocation schemes (manager-led, contractor-led, and cooperative). Evaluation using the ProgrammableWeb dataset shows that the manager-led scheme achieves the most stable and high accuracy, revealing that the design of task allocation determines the balance between accuracy and cost.

Slide:

<Presentation Content>
Presenter:Kai Honda
Title: LLM Agent Architecture for Context-Aware IoT Service Composition
Abstract:Recently, numerous IoT devices have been deployed in homes and offices, with sensors that acquire environmental information and actuators such as air conditioning and lighting widely distributed in everyday spaces. Accordingly, the demand for services that enable optimal device control in response to natural language instructions has increased. However, in the real world, time and external environment fluctuate, and physical conflicts between devices can occur, making it difficult to generate flexible and safe control plans tailored to the situation. Therefore, in this research, we propose an LLM agent that uses large-scale language models to collect real-world context information from natural language user instructions and dynamically compose IoT services according to the situation.

Slide:

<Presentation Content>
Presenter:Yuzuki Masuo
Title: A Dialectic Discussion for LLM Multiagent Question Answering
Abstract:In recent years, question-answering using large-scale language models (LLMs) has been utilized in various services. In a multicultural coexistence society with residents from different cultures, not only accurate answers to questions are required, but also answers that lead to mutually acceptable consensus between conflicting diverse values. However, in existing LLM multi-agent discussions, the debate format competing between conflicting positions is central, and discussions tend to converge toward winning or losing. As a result, opposing value structures are not logically integrated, and consensus acceptable to both parties is often not reached. Therefore, in this research, we propose a dialogue protocol based on the framework of dialectical reasoning and abstract argumentation, in which multiple LLM agents sequentially execute argument construction, counter-argumentation, mutual rebuttal, and integration.

Slide:

<Presentation Content>
Presenter:Ryo Yamauchi
Title: LLM Multi-Agent Cross-Domain QA Based on Shared Plans
Abstract:In scenarios described in natural language, ambiguity, increased interactions, and elongated contexts often lead to deviations from the intended scenario. To address this issue, this research proposes a structured prompt notation for controlling agents. Specifically, the prompt is defined using two elements: "syntax" and "scenario," and is written in YAML format to reduce ambiguity. This notation eliminates ambiguities caused by polysemy and colloquial expressions, enabling flexible translations that adapt to the given context. An evaluation of the constructed agent's performance in handling polysemous words and culturally dependent expressions showed that, through user interactions, the agent was able to provide more accurate and contextually appropriate translations.

Slide: