This paper reports on the development of a web-based learning system called “Information Theory Hands-on Center” for combined learning of information theory. While general textbooks and materials allow students to understand individual concepts, it has been difficult to grasp the overall communication flow by combining these concepts. Our system provides an interactive learning environment covering fundamental concepts such as information content and entropy, source coding with Huffman codes, and channel coding with Hamming codes. Students can experience not only individual theories and algorithms, but also the complete process where a sender transmits messages through noisy channels using Huffman and Hamming codes, and a receiver decodes messages while correcting errors. In an evaluation experiment conducted with the cooperation of undergraduate students, no significant difference was found in test scores before and after system use. However, the five-point scale questionnaire showed an average rating of 4.48 for “promotion of understanding” and 4.28 for “content richness.”

In information recommendation, a session refers to a sequence of user actions within a specific time frame. Session-based recommender systems aim to capture short-term preferences and generate relevant recommendations. However, user interests may shift even within a session, making appropriate segmentation essential for modeling dynamic behaviors. In this study, we propose a supervised session segmentation method based on similarity features derived from action embeddings and attributes. We compute the similarity scores between items within a fixed-size window around each candidate segmentation point, using item co-occurrence embeddings, text embeddings of titles and brands, and price information as sources for these similarity features. These features are used to train supervised classification models to predict the session boundaries. We construct a manually annotated dataset from real browsing histories and evaluate the segmentation performance using F1-score, PR-AUC, and ROC-AUC. The LightGBM model achieves the best performance, with an F1-score of 0.806 and a PR-AUC of 0.831. These results demonstrate the effectiveness of the proposed method for session segmentation and its potential to capture dynamic user behaviors.

This study aims to automatically generate patent maps by estimating technical problems and solutions from a set of patent documents in a specific field. Specifically, our method extracts sentences relevant to technical problems and solutions from patent documents by using RoBERTa. Next, using the extracted sentences as training data, our method estimates terms related to technical problems and solutions by few-shot learning with a large language model (LLM). The estimated terms are clustered based on semantic similarity. Then, a patent map is automatically generated using these clustered terms as axis labels. In the evaluation, the appropriateness of the estimated technical problems and solutions for the patent map is evaluated by calculating the similarity between the axis labels of a manually created patent map and those estimated by the proposed method.

In decision-making within small groups, individual members are afforded the opportunity to freely articulate their opinions, which can foster both compromise and the emergence of novel ideas. However, previous studies have not sufficiently examined how such emergent interactions influence consensus formation and individual satisfaction. In the present study, we develop a reinforcement learning model that integrates individual satisfaction as a reward metric, and we employ simulation experiments to investigate the impact of agents generating new opinions during discussions on both consensus formation and satisfaction. The simulation results suggest that the introduction of new opinions not only enhances individual satisfaction but also reduces the cost—measured in the number of steps required—to achieve consensus. Moreover, we conduct a linguistic simulation using a large-scale language model (LLM) to assess the applicability of this approach to real-world contexts and to compare decision-making processes across different group sizes. Our findings indicate that, in small groups, agents tend to conclude discussions with uniformly high satisfaction levels and rapid convergence, whereas in larger groups significant divergence in satisfaction and opinion is observed, complicating the consensus process. Additionally, differences in the vocabulary and structural characteristics of decisions emerge, highlighting the importance of carefully designing group size and composition for complex decision-making tasks.

In this paper, we propose the Judging Manner Interaction System (Jmis), an interaction system that employs large language models (LLMs) to evaluate manners based on environmental information, with the aim of enabling robots to behave with manners and demonstrate flexible, human-considerate actions. In shared spaces between humans and robots, it is essential for robots to actively account for the presence of humans, act with flexibility, and be capable of judging given instructions from an ethical perspective, including refusing inappropriate ones. By integrating Jmis into robots, the system’s Manner Judgment Module and Manner Improvement Module allow robots to embody manners and adopt considerate, flexible behaviors. The Manner Judgment Module uses an LLM to assess whether the robot’s planned actions (goals and the actions taken to achieve them) comply with social manners. Furthermore, in the Manner Improvement Module, when actions are judged as violating manners, the system employs an LLM to determine whether manners can be improved through communication with surrounding humans, and it generates appropriate utterances. To validate the effectiveness of the Manner Judgment and Manner Improvement Modules of Jmis, we conducted evaluation experiments examining impressions of robot behaviors with Jmis implemented. The results showed that, in both simulated and real environments, robots using Jmis gave participants stronger impressions of likability and perceived intelligence. Moreover, it was demonstrated that robots equipped with manners conveyed the intentions behind their actions more clearly.

History-dependent responses (HDRs) are phenomena in which prior stimulus history influences subsequent neuronal responses; they are considered fundamental to higher-order information processing such as memory and decision-making. In this study, we propose a method utilizing cultured neuronal networks to extract history-dependent neuronal activity from response activities to two sequential stimuli. The neuronal activity was binned into 1-ms time windows, and the spatial distribution of simultaneously firing electrodes was defined as an instantaneous spatial pattern (ISP). This sequential ISP data was then converted into images (visualized) and input into a Deep Convolutional Neural Network (CNN). These converted images of neural response activity, obtained under various inter-stimulus interval (ISI) conditions, were trained and classified by a deep CNN. The estimated discrimination accuracy for each class (i.e., the stimulating electrode) was evaluated as an index for the emergence (manifestation) of history-dependent neural activity. As a result, it was confirmed that the discrimination accuracy decreased depending on the interval of the two sequential stimuli. Specifically, under long ISI conditions, the evoked response patterns became similar, leading to this reduced discrimination accuracy. It was shown that the two evoked response patterns were distinct, and discrimination accuracy was high, approximately 1–2 s after the evoked stimulation, indicating that the history (memory) persisted during this period. Furthermore, an analysis of the contribution ratio for discrimination suggested that under short ISI conditions, the firing times tended to align across sweeps, suggesting that the internal state of the neuronal network may be forming a temporary stable state.

This study addresses the challenge that the eCO₂ output of MOX-type CO₂ sensors, estimated from TVOC, is strongly affected by environmental factors such as atmospheric pressure, temperature, and humidity, making it difficult to use directly for occupancy estimation. We propose a calibration method based on regression models that explicitly incorporate atmospheric pressure as an explanatory variable. A MOX sensor and an NDIR sensor, used as a reference, were colocated in the same environment, and continuous multi-day data of CO₂, TVOC, temperature, humidity, and pressure were collected to construct the calibration models. Incorporating pressure improved accuracy, and comparison of linear regression, multiple linear regression, second-order polynomial regression, and random forest (RF) showed that RF performed best, with RMSE reduced by approximately 42% compared with the uncalibrated values. For occupancy estimation, we examined a regression model using RF with explanatory variables including CO₂, illuminance, temperature, humidity, and temporal features. Although the NDIR-only approach achieved the highest accuracy, for eCO₂, calibration improved performance, and combining it with illuminance and temporal features enabled accuracy close to that of the NDIR sensor.

This study analyzes the actual usage of onomatopoeia in product descriptions and its impact on sales performance, focusing on items in the “Women’s” category on Mercari, one of Japan’s largest online flea market platforms. The analysis reveals that terms such as pittari (perfect fit), kirakira (sparkling), and saratto (smooth/light touch) appear frequently. By extracting co-occurring words with these high-frequency onomatopoeic expressions and performing cluster analysis, this study examines the specific contexts in which they are used. The results suggest that the likelihood of an item being sold is influenced by the use of onomatopoeia within appropriate contexts, specifically through effective combinations with relevant co-occurring words.