井手 佑翼 | D, 中間発表 | 自然言語処理学 | 渡辺 太郎, | 荒牧 英治, | 大内 啓樹 | ||
title: CoAM: Corpus of All-Type Multiword Expressions
abstract: Multiword expressions (MWEs) refer to idiomatic sequences of multiple words. MWE identification, i.e., detecting MWEs in text, can play a key role in downstream tasks such as machine translation, but existing datasets for the task are inconsistently annotated, limited to a single type of MWE, or limited in size. To enable reliable and comprehensive evaluation, we created CoAM: Corpus of All-Type Multiword Expressions, a dataset of 1.3K sentences constructed through a multi-step process to enhance data quality consisting of human annotation, human review, and automated consistency checking. Additionally, for the first time in a dataset for MWE identification, CoAM's MWEs are tagged with MWE types, such as NOUN and VERB, enabling fine-grained error analysis. Annotations for CoAM were collected using a new interface created with our interface generator, which allows easy and flexible annotation of MWEs in any form. Through experiments using CoAM, we find that a fine-tuned large language model outperforms MWEasWSD, which achieved the state-of-theart performance on the DiMSUM dataset. Furthermore, analysis using our MWE type tagged data reveals that VERB MWEs are easier than NOUN MWEs to identify across approaches. For more details, see our paper at https://arxiv.org/pdf/2412.18151. language of the presentation: *** Japanese *** | |||||||
SUPRANES MICHAEL VAN BALLESCAS | D, 中間発表 | ソーシャル・コンピューティング | 荒牧 英治, | 渡辺 太郎, | 若宮 翔子, | PENG SHAOWEN | |
title: Plug-and-play Counterfactual Text Generation for Data Augmentation in Classification Tasks
abstract: Online texts are a valuable source of data for studying language, social constructs, and human behavior. However, such curated datasets often suffer from sampling bias and contain spurious data artifacts, which may reduce model generalizability. A promising solution is counterfactual data augmentation (CDA), which involves creating synthetic examples by minimally editing original texts to express target attribute/s. While CDA has been shown to improve model robustness (Sen et al., 2022; Madaan et al., 2023), generating valid counterfactuals remains a challenge. Manual generation is effective but labor-intensive. Fine-tuning large language models (LLMs) requires significant resources, and prompt-based approach may struggle with sensitive or subjective attributes (Li et al., 2023). Also, In-context learning is often dependent on large but closed-source models. This study investigates plug-and-play controlled text generation as a scalable, low-resource alternative for CDA. These methods enable conditional text generation without extensive retraining and are compatible with open-source models. Experiments in the hate speech classification domain demonstrate that the proposed approach, using Flan-T5-L and Roberta-L, achieves performance comparable to systems powered by more advanced LLMs such as GPT-4o-mini. language of the presentation: English | |||||||
THUWIBA QAIS ABDALLA IBRAHIM | D, 中間発表 | 数理情報学 | 池田 和司, | 荒牧 英治, | 久保 孝富, | 日永田 智絵, | LI YUZHE |
title:Exploring the Relationship Between the Stochastic Maximum Principle and the Free Energy Principle in Decision-Making under Uncertainty in OCD Patient
abstract: Decision-making under uncertainty is a hallmark of obsessive-compulsive disorder (OCD). In this study, we present a novel computational model that integrates the Free Energy Principle (FEP) and the Stochastic Maximum Principle (SMP) to explain distorted recognition–action coupling in OCD patients. Our model simulates how uncertainty evolves through the interaction of stochastic noise and misaligned recognition/action dynamics, formalized by distinct temperature parameters. We validate this model using two datasets: a publicly available dataset of 617 healthy participants performing the Iowa Gambling Task (IGT), and a synthetic dataset simulating OCD patient behavior over 95, 100, and 150 trials. Simulation results capture recency and response consistency effects and reproduce hallmark behavioral patterns in OCD language of the presentation: English | |||||||
倉井 龍太郎 | D, 中間発表 | サイバネティクス・リアリティ工学 | 清川 清, | 加藤 博一, | 内山 英昭, | Perusquia Hernandez Monica, | 平尾 悠太朗 |
title: Research on Using LLMs to Support 3D Content Creation in the Metaverse
abstract: The proliferation of metaverse platforms has revealed a significant challenge: the number of content creators is extremely low relative to the number of users. This disparity stems from three technical barriers: (1) the complexity of programming object behavior, (2) the specialized knowledge required for 3D modeling, and (3) platform-specific integration tasks. This study proposes a comprehensive system that uses large language models (LLMs) to allow anyone to create 3D content on commercial metaverse platforms. The proposed system consists of three main components. The first component is “MagicItem,” which automatically generates ClusterScript from natural language to enable motion programming. Second is “MagicCraft,” which generates dynamic 3D objects from text descriptions. Third is the "Affordance Estimation System" (currently under development), which automatically determines where objects can sit or be held. In an evaluation experiment of MagicItem involving 63 participants, 41% of those with no programming experience successfully created functional object motions. MagicCraft was evaluated by 51 participants and scored 71.8 on the System Usability Scale, demonstrating excellent usability. The affordance estimation system, which is currently under development, has confirmed the basic effectiveness of position estimation using LLMs. language of the presentation: Japanese 発表題目: メタバースにおけるLLMを用いた3Dコンテンツ作成支援に関する研究 発表概要: メタバースプラットフォームの普及により、ユーザー数に対してコンテンツ作成者が極めて少ないという課題が明らかになった。この格差は、(1)オブジェクト動作のプログラミングの複雑さ、 (2)3Dモデリングに必要な専門知識、(3)プラットフォーム固有の統合作業という3つの技術的障壁に起因する。 本研究では、大規模言語モデル(LLM)を活用し、商用メタバースプラットフォームでの3Dコンテンツ作成を誰もが行えるようにする包括的システムを提案する。 提案システムは3つの主要コンポーネントで構成される。第一に、自然言語からClusterScriptを自動生成し、動作プログラミングを可能にする「MagicItem」。 第二に、テキスト記述から動的3Dオブジェクトを生成する「MagicCraft」。第三に、オブジェクトの座る・握る位置を自動推定する「アフォーダンス推定システム」(開発中)である。 63名を対象としたMagicItemの評価実験では、プログラミング未経験者の41%が機能的なオブジェクト動作の作成に成功した。 51名によるMagicCraftの評価では、System Usability Scaleで71.8点を記録し、優れた使いやすさを実証した。 現在開発中のアフォーダンス推定では、LLMを用いた位置推定の基礎的な有効性を確認している。 | |||||||
XIA WEI | D, 中間発表 | サイバネティクス・リアリティ工学 | 清川 清, | 加藤 博一, | 内山 英昭, | Perusquia Hernandez Monica, | 平尾 悠太朗 |
title: A Visual Guidance Method Using
Head Redirection for Re-Experience in VR Environments language of the presentation: Japanese | |||||||
MUHAMMAD YEZA BAIHAQI | D, 中間発表 | 自然言語処理学(ロボット対話知能) | 渡辺 太郎☆, | 吉野 幸一郎, | Angel Garcia Contreras | ||
title: Rapport Building Strategies: Dialogue, Motion, and Evaluation in Human-Agent Interaction.
abstract: Rapport is a conversational dynamic centered on relationship building, which significantly influences outcomes in collaborative tasks. This study investigates comprehensive strategies for fostering rapport in human-agent interaction by developing dialogue strategies, generating nonverbal behaviors, and conducting multi-layered evaluations. We first introduced a rapport-building dialogue strategy by integrating utterances—derived from psychological theories of human-human rapport—into small talk with a virtual agent. This strategy was implemented through a dialogue system employing two dialogue management approaches—free-form and predefined scenarios—by prompting a large language model (LLM). Building on this, we proposed an enhanced strategy incorporating proactive behaviors, personalization, and aizuchi (backchannels) to deepen user engagement. To evaluate rapport more holistically, we conducted multiple experiments within a multi-dimensional framework that combines subjective metrics—such as Likert scales, pairwise comparisons, and thematic analysis—with objective analyses of behavioral patterns and emotional responses. Furthermore, we address the challenge of eliciting rapport not only through verbal but also nonverbal communication by generating contextually appropriate co-speech motion using LLMs, enabling more natural and socially expressive interactions in both humanoid robots and virtual agents. language of the presentation: English | |||||||