Dialogue contains a wealth of information about speakers’ preferences and experiences, which can be leveraged to personalize and suggest advanced information in various systems. The task of a conversational recommender system is to make recommendations through dialogue. However, existing approaches do not adequately consider the preference information obtained from the dialogue. Moreover, dialogue-based recommendations face unique challenges, such as noise during dialogue and a lack of detailed item information. We introduce the SumRec framework for dialogue-based recommendations, which utilizes information from speaker summaries and recommendation sentences. In this framework, a large language model (LLM) generates summaries focusing on the speaker and sentences recommending items, thereby extracting features of both the speaker and the item. A speaker summary condenses the dialogue to highlight the speaker’s interests, preferences, and experiences. Recommendation sentences describe the type of users who would prefer the item, facilitating an appropriate link between the speaker and the item information. The score estimator then uses this information to predict how likely the speaker is to appreciate the item. To train and evaluate SumRec, we developed ChatRec, a dataset for recommending tourist attractions based on chat dialogues between two individuals. This dataset includes information on tourist destinations, their rating scores by speakers, and predicted scores by third parties. Experimental results using ChatRec showed that SumRec outperformed the baseline method, which relied solely on dialogue and item information. Further experiments with REDIAL, an existing recommendation dialogue dataset, demonstrated similar performance improvements with SumRec.

One of the main issues in the development of an adaptive dialogue system is to estimate a user’s sentiment state since the user’s self-reported sentiment does not necessarily appear in the user utterances. To mitigate the issue that the true sentiment state is not expressed as observable signals, psychophysiology and affective computing studies have focused on physiological signals that capture involuntary changes related to emotions. We address the issue by proposing a new attention mechanism based on the time-series physiological signals and word sequences. Our proposed method, called Time-series Physiological Transformer which captures sentiment changes based on both linguistic and physiological information, significantly outperformed the previous best result (p < 0.05).

Recent studies revealed that Deep Neural Network (DNN) models misrecognize adversarial examples, crafted with malicious modifications to the input. Modifying inputs to induce errors in DNN models is known as adversarial attack, and this vulnerability is a concern even in DNNs that handle natural language. Generally, adversarial attack research on DNNs that manage natural language has either been language-agnostic or predominantly focused on English. Meanwhile, the need to examine vulnerabilities in specific languages has also arisen. Therefore, this paper proposes ANJeL (Adversarial attack to Neural networks for JapanesE Language) to detect vulnerabilities particular to DNNs specialized in Japanese. The proposed method creates adversarial examples based on character types and grammatical features of the Japanese language under the black-box condition. Experimental results have shown that the proposed method successfully detected vulnerabilities within open-source language models and a commercial cloud service by revealing the presence of adversarial examples incorporating perturbations based on Japanese linguistic features. In particular, the detected adversarial examples by ANJeL exhibited greater naturalness and similarity to the input than those detected through previous approaches.

While medical devices based on Artificial Intelligence (AI) are beginning to be approved in various countries/regions, there is concern over how to address biases inherent in big data and machine learning algorithms that could disadvantage specific patient groups. This study analyzes which sources of bias regulatory authorities in each countries/regions are focusing on in the governance of three types of unwanted bias in post-market updates of AI-based medical devices, and what requirements are assumed for risk management. Among the three types of unwanted bias, “Data bias” had many sources of bias that were commonly addressed by all three regions, Japan, the U.S., and Europe. On the other hand, many “human cognitive bias” sources were unaware of the bias, and it was found that there were different responses to the “bias introduced by engineering decisions”. By analyzing the considerations for unwanted bias by stage of the AI system life cycle, all countries confirmed the considerations for the deployment stage to be described. In the other stages, there were differences in the countries where the considerations were described, revealing that there are differences in the considerations at each life cycle stage among countries. Based on these results, we propose three measures to promote the development and approval of AI-based medical devices more fairly, with reduced unwanted bias, and globally. The first is to adapt to the differences in requirements for unwanted bias that exist between countries in order to expand internationally. The second is to utilize interdisciplinary experts when designing rule-based systems in order to ensure transparency about the possible existence of cognitive biases. The third is to disclose the results of research into cognitive biases toward groups. We believe that this measure, which addresses areas that current governance does not clearly set as requirements, will contribute to the early adoption of AI systems, which are currently evolving.