Kyoji Umemura

Shintaro Uchiyama, Kyoji Umemura, Yusuke Morita

This paper proposes a system that uses large language models to provide immediate feedback to students in flipped classroom preparation learning. This study aimed to solve challenges in the flipped classroom model, such as ensuring that students are emotionally engaged and motivated to learn. Students often have questions about the content of lecture videos in the preparation of flipped classrooms, but it is difficult for teachers to answer them immediately. The proposed system was developed using the ChatGPT API on a video-watching support system for preparation learning that is being used in real practice. Answers from ChatGPT often do not align with the context of the student's question. Therefore, this paper also proposes a method to align the answer with the context. This paper also proposes a method to collect the teacher's answers to the students' questions and use them as additional guides for the students. This paper discusses the design and implementation of the proposed system.

Masato Kikuchi, Mitsuo Yoshida, Kyoji Umemura et al.

In natural language processing (NLP), the likelihood ratios (LRs) of N-grams are often estimated from the frequency information. However, a corpus contains only a fraction of the possible N-grams, and most of them occur infrequently. Hence, we desire an LR estimator for low- and zero-frequency N-grams. One way to achieve this is to decompose the N-grams into discrete values, such as letters and words, and take the product of the LRs for the values. However, because this method deals with a large number of discrete values, the running time and memory usage for estimation are problematic. Moreover, use of unnecessary discrete values causes deterioration of the estimation accuracy. Therefore, this paper proposes combining the aforementioned method with the feature selection method used in document classification, and shows that our estimator provides effective and efficient estimation results for low- and zero-frequency N-grams.

Masato Kikuchi, Kento Kawakami, Kazuho Watanabe et al.

Likelihood ratios (LRs), which are commonly used for probabilistic data processing, are often estimated based on the frequency counts of individual elements obtained from samples. In natural language processing, an element can be a continuous sequence of $N$ items, called an $N$-gram, in which each item is a word, letter, etc. In this paper, we attempt to estimate LRs based on $N$-gram frequency information. A naive estimation approach that uses only $N$-gram frequencies is sensitive to low-frequency (rare) $N$-grams and not applicable to zero-frequency (unobserved) $N$-grams; these are known as the low- and zero-frequency problems, respectively. To address these problems, we propose a method for decomposing $N$-grams into item units and then applying their frequencies along with the original $N$-gram frequencies. Our method can obtain the estimates of unobserved $N$-grams by using the unit frequencies. Although using only unit frequencies ignores dependencies between items, our method takes advantage of the fact that certain items often co-occur in practice and therefore maintains their dependencies by using the relevant $N$-gram frequencies. We also introduce a regularization to achieve robust estimation for rare $N$-grams. Our experimental results demonstrate that our method is effective at solving both problems and can effectively control dependencies.

Ryosuke Homma, Yoshifumi Seki, Mitsuo Yoshida et al.

Dwell time has been widely used in various fields to evaluate content quality and user engagement. Although many studies shown that content with long dwell time is good quality, contents with short dwell time have not been discussed in detail. We hypothesize that content with short dwell time is not always low quality and does not always have low user engagement, but is instead related to user interest. The purpose of this study is to clarify the meanings of short dwell time browsing in mobile news application. First, we analyze the relation of short dwell time to user interest using large scale user behavior logs from a mobile news application. This analysis was conducted on a vector space based on users click histories and then users and articles were mapped in the same space. The users with short dwell time are concentrated on a specific position in this space; thus, the length of dwell time is related to their interest. Moreover, we also analyze the characteristics of short dwell time browsing by excluding these browses from their click histories. Surprisingly, excluding short dwell time click history, it was found that short dwell time click history included some aspect of user interest in 30.87% of instances where the cluster of users changed. These findings demonstrate that short dwell time does not always indicate a low level of user engagement, but also level of user interest.

Kenshin Sekimoto, Yoshifumi Seki, Mitsuo Yoshida et al.

The purpose of this study is to clarify what kind of news is easily retweeted and what kind of news is easily Liked. We believe these actions, retweeting and Liking, have different meanings for users. Understanding this difference is important for understanding people's interest in Twitter. To analyze the difference between retweets (RT) and Likes on Twitter in detail, we focus on word appearances in news titles. First, we calculate basic statistics and confirm that tweets containing news URLs have different RT and Like tendencies compared to other tweets. Next, we compared RTs and Likes for each category and confirmed that the tendency of categories is different. Therefore, we propose metrics for clarifying the differences in each action for each category used in the $χ$-square test in order to perform an analysis focusing on the topic. The proposed metrics are more useful than simple counts and TF-IDF for extracting meaningful words to understand the difference between RTs and Likes. We analyzed each category using the proposed metrics and quantitatively confirmed that the difference in the role of retweeting and Liking appeared in the content depending on the category. Moreover, by aggregating tweets chronologically, the results showed the trend of RT and Like as a list of words and clarified how the characteristic words of each week were related to current events for retweeting and Liking.

Masato Kikuchi, Mitsuo Yoshida, Kyoji Umemura

In Japanese scientific news articles, although the research results are described clearly, the article's sources tend to be uncited. This makes it difficult for readers to know the details of the research. In this paper, we address the task of extracting journal names from Japanese scientific news articles. We hypothesize that a journal name is likely to occur in a specific context. To support the hypothesis, we construct a character-based method and extract journal names using this method. This method only uses the left and right context features of journal names. The results of the journal name extractions suggest that the distribution hypothesis plays an important role in identifying the journal names.

Ryosuke Homma, Keiichi Soejima, Mitsuo Yoshida et al.

There is dwell time as one of the indicators of user's behavior, and this indicates how long a user looked at a page. Dwell time is especially useful in fields where user ratings are important, such as search engines, recommender systems, and advertisements are important. Despite the importance of this index, however, its characteristics are not well known. In this paper, we analyze the dwell times of various websites by desktop and mobile devices using data of one year. Our aim is to clarify the characteristics of dwell time on non-news websites in order to discover which features are effective for predicting the dwell time. In this analysis, we focus on device types, access times, behavior on the website, and scroll depth. The results indicated that the number of sessions decreased as the dwell time increased, for both desktop and mobile devices. We also found that hour and month greatly affected the dwell time, but day of the week had little effect. Moreover, we discovered that inside and click users tended to have longer dwell times than outside and non-click users. However, we can not find a relationship between dwell time and scroll depth. This is because even if a user browsed the bottom of the page, the user might not necessarily have read the entire page.

Ayaka Takamoto, Mitsuo Yoshida, Kyoji Umemura et al.

This paper proposes a novel method that can replace compression-based dissimilarity measure (CDM) in composer estimation task. The main features of the proposed method are clarity and scalability. First, since the proposed method is formalized by the information quantity, reproduction of the result is easier compared with the CDM method, where the result depends on a particular compression program. Second, the proposed method has a lower computational complexity in terms of the number of learning data compared with the CDM method. The number of correct results was compared with that of the CDM for the composer estimation task of five composers of 75 piano musical scores. The proposed method performed better than the CDM method that uses the file size compressed by a particular program.

Ayaka Takamoto, Mayu Umemura, Mitsuo Yoshida et al.

In this paper, we propose a way to improve the compression based dissimilarity measure, CDM. We propose to use a modified value of the file size, where the original CDM uses an unmodified file size. Our application is a music score analysis. We have chosen piano pieces from five different composers. We have selected 75 famous pieces (15 pieces for each composer). We computed the distances among all pieces by using the modified CDM. We use the K-nearest neighbor method when we estimate the composer of each piece of music. The modified CDM shows improved accuracy. The difference is statistically significant.

Kana Oomoto, Haruka Oikawa, Eiko Yamamoto et al.

In this paper, we propose a statistical test to determine whether a given word is used as a polysemic word or not. The statistic of the word in this test roughly corresponds to the fluctuation in the senses of the neighboring words a nd the word itself. Even though the sense of a word corresponds to a single vector, we discuss how polysemy of the words affects the position of vectors. Finally, we also explain the method to detect this effect.

Hayato Okumoto, Mitsuo Yoshida, Kyoji Umemura

When we watch a video, in which human hands manipulate objects, these hands may obscure some parts of those objects. We are willing to make clear how the objects are manipulated by making the image of hands semi-transparent, and showing the complete images of the hands and the object. By carefully choosing a Half-Diminished Reality method, this paper proposes a method that can process the video in real time and verifies that the proposed method works well.

Yuki Kondo, Masatsugu Hangyo, Mitsuo Yoshida et al.

We can extract useful information from social media data by adding the user's home location. However, since the user's home location is generally not publicly available, many researchers have been attempting to develop a more accurate home location estimation. In this study, we propose a method to estimate a Twitter user's home location by using weather observation data from AMeDAS. In our method, we first estimate the weather of the area posted by an estimation target user by using the tweet, Next, we check out the estimated weather against weather observation data, and narrow down the area posted by the user. Finally, the user's home location is estimated as which areas the user frequently posts from. In our experiments, the results indicate that our method functions effectively and also demonstrate that accuracy improves under certain conditions.