This paper proposes a system that supports watching Mahjong games for novices. Two factors make watching Mahjong games difficult for novices: it is difficult to know what combination of tiles the players are aiming for, and it is difficult to instantly grasp what tiles each player has and what tiles they discard. The system predicts and visualizes the combination of tiles at the goal that each player is aiming for and also visualizes the status of each player’s holding and discarded tiles side by side in the horizontal line. In the evaluation experiment, participants watched Mahjong games through the proposed system. Utterances during the experiment were analyzed based on Think-aloud protocols. For comparison, we prepared a comparison system that reproduced a typical Mahjong game, had the participants watch the game similarly, and analyzed the content of their speech. Participants watched two games, and the proposed system increased utterances indicating an understanding of the game’s situation in the second time more than that in the first time. This indicates that the proposed system can help the spectator understand the game’s contents and watch the Mahjong game.
In recent years, the development of large-scale language models (LLM) has dramatically improved text generation performance. However, general-purpose LLMs have the problem that they do not always perform optimally in tasks in specific fields. The approach of fine-tuning with data from specific fields is commonly used to address this problem, but collecting training data from limited fields is difficult. In particular, data related to human values have problems such as difficulty in annotation, insufficient amount of data, and large variability. In addition, the training and inference of LLMs is expensive. This study focused on the task of human values estimation and tested the effectiveness of an approach that uses an LLM to generate pseudo-data and trains a small-scale model on that data. In the experiment, we augmented a 3,870-items human values dataset with 10 categories to four times its original size with the generated pseudo-data. The training dataset with pseudo-data increased human values estimation accuracy by 17
Important bug report comments can help developers efficiently fix bugs. This paper proposes a bug report comment ranking model called BRCR based on deep learning models. This paper also proposes a pseudo-labeling approach to address the few-shot problem. In the experiments, bug reports of three GitHub software projects are used to evaluate the proposed BRCR ranking model. The experimental results show that BRCR with LSTM and pseudo-labeled data can effectively outperform other ranking models.
This paper proposes a new evaluation index for evaluating human fallibility in shogi. By analyzing the differences between the decision-making processes of AI and humans, we found that while humans rely on intuition and decisions made within a limited timeframe, AI often makes different decisions because it carries out exhaustive searches. In this study, we developed a new evaluation index that extracts features prone to human error using a policy network trained based on professional game records. This evaluation model uses logistic regression to predict the probability of making a mistake in the endgame. We have shown that the proposed indicator is effective through testing with test data. In the future, we plan to construct similar evaluation indicators for the opening and middle game phases and verify their effectiveness.
Given the surge in digital communication, the spread of fraud messages across various communication platforms poses an increasingly significant threat to the public due to their accessibility and ubiquity. Existing fraud detection methodologies predominantly focus on recognized fraud tactics, thus diminishing their effectiveness against unfamiliar or evolving deceptive strategies, a challenge known as domain adaptation. To address this issue, we introduce a novel approach, Graph-Based Syntactic Structure Generation, capable of autonomously learning common syntactic structures within fraud messages without the need for predefined fraud patterns. This method leverages co-occurrence graphs and graph auto-encoder techniques to extract connection state features of each token from extensive texts, thereby acquiring global information essential for identifying complex syntactic structures indicative of fraud messages. Further, we employ an attention-based model for syntactic structure transformation of texts, converting each token into word, part of speech (POS), or # (fixed valued) to facilitate effective syntactic pattern recognition. This approach not only enhances the model’s understanding of textual syntactic structures but also enriches semantic information for subsequent large language model analysis. Experiments conducted on datasets sourced from Cofacts and the 165 anti-fraud platform demonstrate our method’s robust recognition performance on cross-domain datasets, even with limited training data, surpassing traditional methods’ limitations when confronting newly emerged fraud tactics.
