Summarization of the Methodology of Applying N-gram to Obtain Factor Scores of Q&A Statements

Yuya Yokoyama — 2025-01-19

With a view to solving the troubles of mismatches between the questioners and respondents of Question and Answer (Q&A) sites, an impression evaluation experiment resulted in obtaining nine factors of impressions for Q&A statements. Factor scores were then estimated through multiple regression analysis utilizing feature values of statements. The factor scores obtained and estimated were subsequently employed for finding appropriate respondents who would be likely to answer a posted question. However, this methodology so far has substantially depended on the syntactic information extracted through morphological analysis. In addition, this method has a significant drawback of demanding manifold variables and complex multiple regression equations to estimate factor scores. Thus, another course has been taken by applying N-gram instead of morphological analysis. So far, the analyses of 2-gram through 5-gram have shown good estimation accuracy. In order to strengthen these tendencies, in this paper, 6-gram is applied to the feature values. Further analysis has shown that 6-gram would also be applicable to the method. In terms of estimation accuracy, N-grams also outscore morphological analysis; above all 2-gram and 3-gram show the best accuracy. Hence, it could be suggested that N-gram should play a more important role in estimating factor scores than mere morphological analysis.

Contrastive Learning for Fine-Grained Reading Detection

Md. Rabiul Islam — 2025-03-28

Reading is a cognitive activity that we perform aiming at various purposes, such as gaining knowledge and entertaining ourselves, with different scripts and layouts. Therefore, automatic reading detection gives useful information about users’ reading activities. Deep learning enables automatic feature extraction and model creation but needs large-sized labeled data. The self-supervised learning devised to overcome this limitation work as noncontrastive self-supervised learning (SSL) and contrastive self-supervised learning (contrastive learning). Although SSL is well explored for reading analysis, contrastive learning is not still well explored. This paper explores contrastive learning that works in several ways. A Simple Framework for Contrastive Learning of Visual Representations (SimCLR) is one way that has attracted much attention in many research domains because of its superior performance. We explore SimCLR for the cognitive activity recognition task of finegrained reading detection employing electrooculography datasets. These datasets describe eye movements that have been recorded for in-the-wild condition. The obtained results are compared against SSL and supervised baselines. The results show that, for an equal number of training samples, the SimCLR method obtains a maximum performance gain of 3.02 and 3.96 percentage points compared to the two baselines, respectively. Besides, SimCLR shows the best performance for large-sized data with a data efficiency of about 80%, whereas SSL shows the best performance for small-sized data. The analysis conducted in this paper shows a direction for researchers and system designers to employ self-supervised learning for automatic reading detection.

Information Engineering Express

Summarization of the Methodology of Applying N-gram to Obtain Factor Scores of Q&A Statements

Contrastive Learning for Fine-Grained Reading Detection